This is cool; finally, the Young Lady’s Illustrated Primer.
I’d be interested to see how Google’s LaMDA compares on this task (currently available in private preview).
One of LaMDA’s unique (afaik) features is a fact-check system that edits the main LLM outputs, to reduce bullshitting. This seems particularly important in an educational context where impressionable young minds are talking directly to the LLM.
This is an interesting development; I'd assumed we'd have a robust intermediate stage, where a collection of top-quality works (Books, carefully-chosen video games [think Zachtronics], recorded lectures etc) were made available, convenient, and cheap, in a no-pressure, non-points-based system.
An ePrimer has a few components: the book itself, a Ractor mentor, a curriculum that includes Ractor paid milestohes, and scientufic metaphor stories that are the rubric and scaffolding. It might also include the VR or AR components that supply the immersive context. I've taught this in a master's level Metaverse course at a NY university.
Notably, we see two models for uses of the Primer in The Diamond Age. The one everyone thinks of has a single human Ractor on the other side, who takes a genuine, nigh-parental interest in the student's development over a period of years. The others... don't have that, and the outcomes diverge.
Not really. The Duamond Age" is referenced as a template, but it is not required reading.The main components of the course are 1) writing a Ractive, 2) designing a Paper Theater for the Ractive, and 3) designing the set for the Paper Theater. The object is to continue the existing workflow for media with apropriate extensions for the Metaverse. The work is done in teams of 3 people with continuous reviews by team members. An underlying goal is to teach a collaborative media design process. Our view is that collaboration is the main leverage component of the Metaverse.
It’s a good point. I think we can easily imagine GPT writing (in ascending order of cost) speech synthesis actions, mechanical Turk scripts, custom actor interactions. (With the “tool script” for actuating each becoming progressively more complex.)
A Primer framework could make these pluggable. Also the curriculum could be pluggable and either generated or ingested from existing courses.
I guess the mapping of curricula -> scripts is probably the hard part, and what GPT-N (for some N not much bigger than 4) could most interestingly help with.
Recall that what made the Primer so effective in Diamond Age is that it had access to enormous external financial and human resources. The later copies that raised the Mouse Army and did not have those resources were (intentionally) not great liberators like the original.
It seems insane to me to have fundamental education that builds logic & reasoning taught by an AI that can exhibit neither of those (and this isn’t just my hot take but something easily provable if you look at the underlying architecture).
Recent LLM advances are revolutionary and exciting, sure. But the hype cycle surrounding it is borderline dangerous. I see this as the same level as everyone who was claiming ChatGPT would kill Google until Bing tried it, crashed, and burned.
Accuracy is important and although we’re moving towards this (I personally like Amazon’s knowledge graph and evidence based approach) we aren’t there yet.
He obviously neither understands the architecture (and what that really means in relation to an objective function ) or what it means to understand and reason.
LLMs don't stop reasoning because people proclaim so loudly. Because you can very clearly show output demonstrating understanding and reasoning, this nonsensical argument always devolves into "true understanding" TM. This so called special distinction that is so important but yet can't actually be tested for. Lol
LLMs don't "understand" things because they can't generate new theories. Imagine you trained an LLM a ton of text describing what it means for something to be true, and for something to be false. It would never independently come up with, say, propositional calculus. It could tell you everything in the world about what it means to be true and false. Would never come up with anything new.
Does this matter for education? Well, I don't know. A student is unlikely to come up with a question that GPT-4 has never seen before. The idea of kids being taught by a chat bot is just depressing, though. What are we doing to ourselves? Why do we want this? To be more "efficient"?
Language models can generate novel functioning protein structures that adhere to a specified purpose. Structures that didn't exist before nevermind in the training set. https://www.nature.com/articles/s41587-022-01618-2
Sorry but the idea that they can't generate new things is blatantly false.
This isn't generating new theories. I'm not saying LLMs aren't useful, or that they aren't extremely impressive in many ways.
Almost nothing GPT-4 produces was in the training set. I'm not saying they can't produce new sentences.
From the linked paper: "a language model that can generate protein sequences with a predictable function across large protein families, akin to generating grammatically and semantically correct natural language sentences"
Nobody is disputing that LLMs can generate grammatically and semantically correct sentences, and from the sounds of things this task is analogous to generating sentences. This isn't the same as generating new theories.
If it can take natural language purpose descriptions and generate new functioning protein then due to it's training, it evidently understood the relationship between purpose and protein structure. Meaning all the little models it builds as a result of training are concrete. They can produce new things and new information and it's just matter of scale and intelligence.
"akin to generating grammatically and semantically correct natural language sentences" is just an indicator of how it is synthesizing and producing this information. To say it is analogous to generating sentences in the abstract is like saying generating new theorems is analogous to generating sentences. Well yeah, that's what the language model is doing. The power lies in what it is generating. What they are telling you is that they skipped the folding process entirely and just let the model figure out function to structure directly.
Either your dog fetches what you throw at it or not. The idea of "pretend fetching" is unbelievably silly. Doubly so when no one seems to be able to actually define this act. A distinction that can't be tested for is not a distinction. Science is concerned with results not vague and ill defined assertions.
Imagine an LLM was trained on the internet before LLMs were invented. It knows everything in the world. Every programming concept, every philosophical and logical concept. It is an "expert" on the then-current state of the art in machine learning.
Would it be able to invent itself?
I'm not moving the goalposts here. Language models are extremely impressive. I think they will change the world. I fear they will displace many jobs.
Maybe some day they will be able to invent themselves. Maybe that will come sooner than we think. As it stands, they couldn't. This is the distinction that you say doesn't exist.
First of all, you've just imagined a scenario, invented a conclusion that you can't prove and justified your distinction on your invented conclusion. Your distinction is no less arbitrary than when this conversation started.
Next, you're conflating different things. How many humans are capable of inventing anything at all ? Have you invented anything in your field of expertise? Should i then question everything you do understand bout your field if you haven't ? Do you think understanding is binary ?
Ok, I feel this conversation is not going anywhere. If I see a single example of an LLM generating new theories then I will immediately change my mind. I have yet to see this.
Most humans are not capable of inventing things, you're right, but humans in aggregate are. We are comparing LLMs to "humans" here, not to "a human". There are no examples of LLMs doing this. Maybe in a couple of years, who knows.
You're interested in reasoning? That's so cool. Here ya go, something I've been reading earlier today because I need to add some references about deductive, inductive and abductive reasoning, particularly in the context of logic programming (you know, reasoning with computers) to my thesis:
Don't let the Researchgate link discourage you. It will pretend to ask you to login but it will open the pdf anyway.
Why am I posting this? Because all you guys keep talking about "reasoning" this and "reasoning" that, and I don't think that word means, what you think it means.
Reasoning is a very specific thing and we know very well how it works. Can ChatGPT reason? I'm afraid you'll just have to read up and figure it out yourself.
That's right. The authors of those books exhibit those qualities. You can't educate a human by pouring generated text on them, you need something that's capable of teaching to moderate the content. LLMs may allow teachers to achieve more with custom prompts, but it's a far cry from "personal Aristotle" and still requires 1-1 engagement with teachers and student.
Totally agreed, which is why we carefully curate the books we use within educational curricula to make sure they actually contain useful and accurate material.
actually bing is growing massively while google is shifting into irrelevance everywhere (except android) - even laypeople now prefer to NOT use google left and right
As a parent, it's hard not to look at the LLM stuff and worry that our kids are wasting their time right now. One my kids tries really, really hard in school, and I can't help but think it's for nothing--LLMs are going to effectively level the playing field for everyone, to the point where personalized LLM assistants will become as common as cellphones in my opinion.
In this future, what's the point in trying hard when a LLM or a person with a LLM will be just as equivalent without most of the effort? Seems like we're creating a future where humans are taught to delegate as much as possible to some LLM/service and go back to whatever entertains them.
Having massive amounts of information does not mean you can understand it or make the best decisions. There will still be a need to develop sound reasoning skills, and real-world information will be useful for examples but won't in and of itself be able to convey those reasoning skills which must grow through effort.
Many people could not function without cellphones any longer. So much function has become the norm that they would struggle without it.
With LLMs and more advanced AI, we are now outsourcing our thinking and reasoning. The implications are staggering for the future generations. We can still think and reason without assistance, but what is to come in the future for those that grow up with some assistant that manages everything in your life including your thoughts?
I think extrinsic motivation—and especially social motivation—is absolutely vital to keeping humans mentally and physically healthy. We've killed a bunch of that in the arts already (who gathers around the piano to sing along with your one cousin who can play all the standards, anymore? How many amateur portrait sketch artists are genuinely valued by their friends and families, rather than humored? Mechanical reproduction destroyed both of those, among others) and we're about to expand that a ton.
I'm not on the "people must work to be happy" train, but I do think they must feel needed to be happy. I don't think most folks are fulfilled playing violin in an empty room for their entire lives except when they can convince some close friend or relative to indulge them for fifteen minutes, or reading all the classics but having no-one to talk about them with (because AI is better at chatting about the classics, too), or writing books that nobody wants to read because AI does it better.
We're about to feel needed for a hell of a lot less than we did before, which level-of-social-need was already much-reduced from its peak. I fully expect a net decrease in overall happiness & mental health, from these developments.
So be the contrarian who does those things we "don't do" anymore. Teach your kids, if you have any, to do those things. Teach them to observe the family at a restaurant, all staring at their individual phones, while yours remain safely in the car or shut off in your pocket.
> Who gathers around the piano to sing along with your one cousin who can play all the standards, anymore?
For the past 6 months or so, the lifegroup we host (mid-week discussion group associated with our church) does exactly that as part of our evenings - we typically sing through three hymns, with piano, on our path between dinner, discussion, and an open ended firepit time that often goes well into the late hours. And let me tell you, people enjoy belting out the standards (my wife generally picks one well known, one lesser known, and one "You've probably never seen this one before!"), scraping the dust off how to read multipart music, etc!
If the direction of tech is to dehumanize us and remove all that humans enjoy, then screw that! Do what humans enjoy instead!
Of course the individual can simply do those things anyway. Some like-minded groups may form, even.
Over the population, statistics is king, though. We don't, as individual actors, will our way into a better tomorrow when our environment changes to make that harder—not in large numbers. "So just ignore all those circumstances and pressures and do it!" does not work, in general. Larger forces and trends dominate.
You're right, but not being in the position of a god-king where I can change those larger trends, the best I've found is to model and demonstrate that the "consumer tech profitable defaults" aren't the only way to live life, and rather loudly so.
I work in tech (deep weeds), and most people I interact with are quite aware that I do not like what we've done with it, and that I try to keep its influence on my personal life down - with plenty of handy ways on hand to suggest people try out, or solutions to "But how can you possibly XYZ?"
Having grown up before smartphones were a thing, I do recall how we used to do things, and most of those methods still work.
Yeah, your approach is of course the only one available to the individual. What else can you do? Literally nothing. That's the only option to improve one's own experience—"just do it". I just expect the population-level effects to be... troubling. A kind of mass existential crisis. Not that there's much I can do about that, though, you're right.
I don't think trying hard is a waste of time right now. There's a lot of hype right now and ore will come but LLMs will fall to the hype cycle soon enough as we begin to discover all the little details and things that are not possible.
I think it's great everyone will be able to have a personal tutor and I think we can rethink some aspects of education. The Waldorf model really starts to make more and more sense since kids can explore curiosities with their AI tutor which can be programmed to nudge them into the "right things". And of course analyze and report on them daily to the parents.
> Seems like we're creating a future where humans are taught to delegate as much as possible to some LLM/service and go back to whatever entertains them.
Or interests us. I think it's great if we can offload some of the mundane and repetitive things to a machine so we can focus on more interesting things.
We've had search and wikipedia for decades. Access to knowledge did not in fact obsolete the need to learn said knowledge in order to develop critical thinking and problem solving skills. We don't acquire knowledge for the purpose of parroting it back on demand.
LLM might be used to find relevant information quicker, but they won't think for you, so people with LLMs won't suddenly have better critical thinking skills. They'll just make the same poor decisions faster.
Which is what happened with the Internet. We thought access to knowledge was going to enlighten all the dummies. It didn't, because no matter how much knowledge you present to a person, they have to also be willing to learn it, _think_ about it and apply it. That's the hard part.
My classmates and I wasted a lot of time in school, but it had more to do with things like watching "Stomp!" once a year every year from grades 1-10 and reading Dr. Seuss every year from grades 1-6. A lot of public school is just warehousing. That's probably a bigger problem than obsolescence of genuinely well-planned curriculum due to technology.
I am a little bit skeptic here, with the phrase "This in turn means that a particular teacher’s pedagogical style is no longer a barrier to creating a resonant learning loop with a student who prefers a different style". While I agree that a lot of times students suffer from learning styles that do not really fit them, the problem doesn't come from the teacher as much as the teacher insisting in a certain style above all others, but from neither the teacher nor the student knowing which is the style that fits them. Not to mention that often people need exposure to different styles of looking at the same thing to actually understand it. So I fail to see how any of the new AI advancements help in fixing that issue, of detecting and understanding how a student learns better.
You know how that journalist had a long conversation with Bing and it slowly turned into a psycho-ex girlfriend? This is going to be hilarious, giving kids prolonged access to these AI tools. You get to the parent teacher evening at the end of a term at school: "Err... little timmy and his virtual assistant appear to have entirely neglected maths in favour of adopting Zoroastrianism".
On a more serious note, we have the internet and wikipedia now, we have ipads in classrooms. They all have a role, and we have to incorporate them in sensible ways. I'm sure that these AI tools will have some use, but these language models don't know anything. The examples of them making up information, or being persuaded by people are all over the place. I think it's great to use AI in teaching, but how about we learn from history, the technology we bring in to the classroom has to be well understood and directed in order to be effective. And just as importantly we need to understand that in many cases a focus on technological solutions distract from societal solutions. We can spend billions on deploying AI all over Africa, or we pay for some teachers and ship them some second hand textbooks, which one should we do, and how will that decision be made? Because I fear the decision will be made by some SBF-alike, more interested in their own journey than effective solutions.
I think it's actually going to be non-trivial to really make these tools both useful, smart, and unexploitable. Hell, I remember being at school and we had access to the internet through a very strong content moderation system. What do you think happened? We spent our entire time trying (and succeeding) at circumventing that system. We've already seen with Bing that they've very strongly pulled back on it to mitigate downsides, I'm not sure that's a totally solvable problem. Having said that, sure, as I said, I do think it has a place, I'm just not convinced it's a panacea.
> I think it's actually going to be non-trivial to really make these tools both useful, smart, and unexploitable
I mean, you could say that about human teachers too. I don't think they need to be perfect to be useful. In the end, as Peter Thiel opined in his book, the future is humans and robots working together on solving bigger and bigger problems. A teacher working with AIs that tutor kids and then managing that relationship and the relationship with the parents, etc. while also being the adult role model that kids will always need will remain.
Also add that these tools have to obey the political stance of what the education system should be teaching, based on the geography of the student. So the AI needs to be intelligent enough to do the "war of northern aggression" shit in some places, the "if you have sex before marriage you will die" shit in other places, the progressive mathematics or whatever it's called in other other places...
I'm imagining the infinite amount of weird model tuning that'd be necessary to ensure only some students learn about white flight depending on what state, county, or school district they're in.
Easy solution: just give up the moderation effort. Empower kids to do whatever the hell they want with the AI, as long as they manage to get their homework done.
A better comparison would be literally any sci-fi book where everyone has a personal AI. Spider Robinson's Stardancer trilogy is the first one that comes to mind.
This has been my point for a while. The tools are largely here ( or at least should be based on money spent - my district just sent flyer they want to issue bonds for 170MM -- and that is only part 1 of their grand plan; they want to empty the coffers for the rest ). And I am almost like.. what is throwing money going to do? It won't help the kids.
Kids that want to learn or have parents that enforce standards still excel. There are things a school can do to also enforce standards, but those things are not being done. The opposite appears to be happening.
Edu-tech has been a promoted thing for decades whether by techno-optimists or by (probably more commonly) companies whose primary mission is to extract money from school systems. Like most other uses of money in secondary schools (subject to probably some limits like class size), it doesn't move the needle much relative to parental involvement. While the availability of the Internet is certainly useful in some respects, overall, I doubt secondary school education overall has improved a lot since teachers were writing on blackboards.
The trouble with throwing money at education to try to improve it is that the things we know work are, on one side, nothing to do with the schools (home environment—what needs money are social services and shit like universal healthcare, not schools, poverty-reduction is the clearest target here) and on the other, a lot more expensive than we're willing to pay (e.g. Tutorial System, very-small class sizes, that kind of thing).
So instead of going after the things we know work—either because it's outside the scope of what a school can address, or because it's far more expensive than we're ever going to pay—what money we do spend just chases a bunch of unproven and unlikely-to-succeed garbage, basically.
Schools are like the patient who needs a heart transplant, but can only afford healing crystals.
Yep, but increasing the number of teachers 2-3x is indeed politically impossible. Guessing that'd mean like 30-60% higher spending on education (you'd also have to increase wages to attract enough qualified workers, in addition to having way more teachers on payroll, and it'd be a years-long process), which ain't happening.
I'll feel much better if they give my child an AI that knows everything about elementray level math rather than an AI that knows something in every field.
>I'm sure that these AI tools will have some use, but these language models don't know anything
we're still in the early days of LLMs, this is like making a statement on the potential of social networks in the Usenet era. My guess is we'll see specialized tools of all kinds with LLMs integrated in some fashion, it won't be just giving a kid access to a general LLM model
there are already companies that are able to eliminate LLM hallucinations by giving it access to ground truth facts it needs to conform its answers to. Similar things will be done for education type stuff and other use cases as well
That's a great comparison, because social networks have been, on the whole, bad for people. Early evidence suggests LLMs have high potential to be similarly negative in their impact if they're just unleashed willy-nilly the way everyone is trying to integrate them into everything.
> we're still in the early days of LLMs, this is like making a statement on the potential of social networks in the Usenet era
FWIW, we didn't try to use Usenet (or social media) to replace teachers. And we still don't today, even though they've matured (for... some meaning of the word) significantly.
> these language models don't know anything. The examples of them making up information, or being persuaded by people are all over the place
Sounds just like humans to me.
Primary school, some people showed up to explain blood donation, showed and told us they were putting a bag of fake blood up the teacher's sleeve and would take from that rather than from our actual teacher, and yet when the fake blood went into the collection thingie the kid next to me screamed "they're killing her!"
Same school, show-and-tell, astronomy textbook showing the circular orbit of a moon at two different points of the parent planet's orbit around the star. "This is a picture of a mouse".
Secondary school. Main teacher and replacement teacher cover the same passage of a sort story, one interpreting the word "patronised" as in "patron" the other as in talking down to people whose intelligence you underestimate.
A-levels. Philosophy. What even is knowledge anyway? Tripartite definition. Example given was "shaggy" dog, one of the other students stifles a laugh at the double entendre.
But to your core point:
> the technology we bring in to the classroom has to be well understood and directed in order to be effective
Yes, absolutely.
But it's deeper than that. When this tech is flawless, we will have to ask ourselves: "What is school for?"
After all, being flawless in this context means it already knows all that can be taught, so why not just have it do the work related to those things instead of putting humans through ever longer school years so they can do the same things less well?
This isn't the first time technology has forced such questions; my generation never had to learn how to starch clothes, shoe horses, look after a tinder box, or send and receive Morse code.
It's about scale and reach, 100 teachers/educators will be biased/wrong in 100 ways but you'll average it out by interacting with each over your education.
If an AI is wrong about one thing you'll have the same bias over entire generations, once a full generation is skipped it's very hard to go back
That's why we need diversity, political pluralism, &c. because not many things are inherently "true", it's all about nuance, personal perception, mitigation, ...
> "What is school for?"
Socialisation, experiencing new things, general culture, deciding where you want to go in life, &c. Many things you learn in school aren't used for jobs but for day to day life and for the well being of society
If we take that out and replace it by "kids in front of screens" we're in for a massive and quick change with potential deep and long lasting effects.
By the way keep in mind a shit ton of jobs aren't even close to being automatised, and a shit ton of jobs aren't done by people sitting in front of screens
> If an AI is wrong about one thing you'll have the same bias over entire generations, once a full generation is skipped it's very hard to go back
Sounds like intellectual monoculture? If so, I absolutely agree. Already a problem with respect to many thinking it's unpatriotic to suggest their country is anything other than "the good guys", or blasphemous to ask if their religion is anything less than the actual truth of the divine, and AI can extend that to every single topic.
(Oh hey, another possible Great Filter for the Fermi paradox).
> Socialisation, experiencing new things, general culture, deciding where you want to go in life, &c. Many things you learn in school aren't used for jobs but for day to day life and for the well being of society
Do we need schools for that, or will kids learn the same by messing around in multiplayer Minecraft or whatever the zeitgeist is?
> If we take that out and replace it by "kids in front of screens" we're in for a massive and quick change with potential deep and long lasting effects.
Sure, but that's going to happen regardless. Too many other things are changing also.
So why make it worse? Why gives kids tablets for things we know don't need tablets, and where a device is more of a distraction?
Shiny new technology is great for distracting from a failure to perform basic childcare like keeping children physically safe so they can learn. And then it's a rhetorical weapon "We can't listen to parents, they don't even want their children to have state of the art equipment!"
Second, the parents are mostly buying the tablets anyway.
Thirdly, about this:
> distracting from a failure to perform basic childcare like keeping children physically safe so they can learn
Pretty much the one kind of safety I can guarantee kids have when using a computer is their physically safety.
And learn what, exactly, given the hypothetical scenario you're replying to here is that the only thing left to learn is being sociable which they can ultimately do best by experiencing whatever medium they will be sociable in as adults, and for all other learning you're just being a nerd like me who learns for the fun of it.
Parents are buying the tablets because teachers - pushing shiny new technology in lieu of results - tell them it'll help. There aren't any studies showing the putting screens in front of kids helps in any way, and there are many showing it hurts - from fitness to mental robustness and socialization.
> Pretty much the one kind of safety I can guarantee kids have when using a computer is their physically safety.
I think you're remembering computer lab, or something, where there was a proctor because computers were expensive. Now schools won't expel anyone they label troubled, no matter how much they trouble other students, and they won't protect anyone or let them protect themselves.
> the only thing left to learn is being sociable which they can ultimately do best by experiencing whatever medium they will be sociable in as adults
I don't see social media being so complex that they need all the years of practice to keep up, but I do see forming deep relationships as being essential and social media doesn't do anything to help with that. They'll get there when they do, no worse for not having partaken.
> Parents are buying the tablets because teachers - pushing shiny new technology in lieu of results - tell them it'll help
Funny, I'm remembering a summer holiday job 20 years ago, working on a HVAC production line. One of the others there bought a plasma TV for his infant.
It's not all about what teachers are pushing.
Heck, from what I've heard the teachers don't like the tech any more than you do.
> I think you're remembering computer lab, or something, where there was a proctor because computers were expensive. Now schools won't expel anyone they label troubled, no matter how much they trouble other students, and they won't protect anyone or let them protect themselves.
I have literally no idea what this is supposed to be about. Computers themselves are not part of any of the things you just said are bad.
And as I'm asking if the schools are even going to still be relevant, expulsions would carry as little meaning as being banned from owning horses does today.
> I don't see social media being so complex that they need all the years of practice to keep up, but I do see forming deep relationships as being essential and social media doesn't do anything to help with that. They'll get there when they do, no worse for not having partaken.
I didn't specifically say social media, I said "medium they will be sociable in as adults". That might be the traditional forums of the 2005 interwebs style; or soc med; or group chat apps like IRC, Slack, MS Teams as various of my employers have had us using; or video group chats like many of us got used to during the pandemic; or VR environments, be they games like WoW or roleplay environment like Furcadia or Second Life or VRChat, or "serious" metaverse things like whatever it is FB is trying to do.
It can also be going to parks and gyms, art galleries and graffiti walls, cafés and nightclubs. How much of that is part of school anyway?
But even for text, what counts as polite, friendly, professional, intelligent, serious, or severe is absolutely a moving target, so…
that means knowin when they need to be all like yo and emoji wif spellin' that confuses ol fogies like me and missin all the full stops coz that be rude yo yo lol
And of course the fact that such cultural idioms shift arbitrarily and I've stopped caring means that actual real young people will cringe as much when reading that last paragraph as I did when writing it, in exactly the same way I told my mum to stop trying to be hip and with it back in the late 90s.
> Heck, from what I've heard the teachers don't like the tech any more than you do.
The school district found the ones who'd push it. If teachers aren't fond of this I hope they fight it themselves.
> It's not all about what teachers are pushing.
Sure, I bought my child a laptop. But I spent time teaching computer skills under the guise of game design, etc, and it wasn't for classwork but for everything else. It wasn't a one-size-fits-nobody cash grab.
> I'm asking if the schools are even going to still be relevant, expulsions would carry as little meaning as being banned from owning horses does today.
I don't care if the violent kid thinks he wins, as long as he's not allowed to spend more time with the non-violent kids. His failure isn't my problem, my kids happiness and success is.
> what counts as polite, friendly, professional, intelligent, serious, or severe is absolutely a moving target,
Not so much, I think, that it needs practice over the years. I've seen people pick up social media with no practice whatsoever and right away seem to fit right in.
> I have literally no idea what this is supposed to be about. Computers themselves are not part of any of the things you just said are bad.
Yeah, it seemed to be a bit of an unguided post. I'm not sure if you don't like that I don't like excess tech in schools, or what. I don't think computers are bad, I think teachers and schools pushing them uselessly is bad.
A monolithic AI also becomes a single point of epistemic and pedagogical control.
Whichever side or instance you care to consider, think of the capabilities of someone hell bent for leather on defining what is, or is not, included in an educational curriculum, at the primary, secondary, post-secondary, graduate, or professional degree level, and their ability to influence, overtly or covertly, through persuasion or coercion, what that is.
Even under present conditions of proscribed curricula and textbooks, individual instructors, libraries, librarians, and others can go against some imposed dogma to some extent. A world in which only a small handful of AI originators provide the direct interfact between each student and their education is a grim prospect.
I've noted several commentators in recent months discussion a similar situation in terms of popular culture and entertainment, most notably the influence China now exerts over cinema and sport. The notion that AI will continue to be largely a product of Silicon Valley, and subject to US or even generally Western influence, may not be a valid one.
I think that's a perfectly reasonable fear, but I am here mostly focusing on the idea that school itself stops being important when the AI can do whatever it is teaching the students.
(I don't think that fundamentally changes any of your actual arguments here, it just feels like we're talking about slightly different aspects of how this can go wrong).
> Do we need schools for that, or will kids learn the same by messing around in multiplayer Minecraft or whatever the zeitgeist is?
You don't read facial expressions, body language, use your body, exchange bacteria/viruses, have body contact, &c. in minecraft
You can't go against hundred millions years of evolution and expect minecraft to replace real life. (see loneliness epidemic, obesity epidemic, depression epidemic), we're smartish apes, not AIs or computers. We're not designed to seat all day, we're not design to interacts with screens
If you can entirely replace education you don't need _school_ per day but you need long term group socialization with people who have different and similar opinions and backgrounds, physical activity, mental entertainment and challenges
The end goal is to live in healthy and collaborative societies, not to be mienscraft basement dwellers
You're being a bit too bogged down in the specific example there; I never got into the game myself and only used it as an example because a relative used it to socialise pre-pandemic.
I mean, VRChat can do the facial expressions, video games don't prevent exercise, and there's plenty of ways to share germs if you find vaccines too clinical and high-tech.
And of course, I'm currently using my glowing rectangle to group socialise with with people who have different and similar opinions and backgrounds, for example yourself.
Humans can know nearly everything that any other human knows (hence...teaching). AI can't know anything (right now).
The youth who are critical of capitalism certainly have turned a wary eye on the school system -- after all primary education has been designed to create money makers for corporations. Is a new tool the answer? Likely not. Will it help to change the philosophy of education? Maybe not independent of the change that is already occurring.
Only for very specific (and IMO useless) definitions of the word "know". Does it matter if the thing a submarine does is called "swimming" or not?
In any practical sense, even obsolete ancestors of this model definitely "know" more French than I learned in 5 years of mandatory school lessons.
That said, regardless of the definition of the word "know"…
> The youth who are critical of capitalism certainly have turned a wary eye on the school system -- after all primary education has been designed to create money makers for corporations.
I don't know what you're talking about here.
The soviet bloc countries were very proud of their schooling system, rightly or wrongly; and conversely corporations don't really get much out of primary education, which is often described as being essentially no better than daycare for slightly older kids.
If you're not developing tools for human teachers to evaluate the learning of their students (eg summarizing what was presented/discussed/learned) then you're going to have preventable failures. And you can't just rely on simple metrics (Goodhart's Law). The kinds of failures you can have teaching kids are pretty bad.
A middle aged woman who said I had to be lying about being vegetarian because she couldn't believe it was possible for humans.
A coworker who insisted it was impossible to speed up some code in a daily standup, for me to come back to on the next stand up and announce I had sped it up from 15 minutes to 0.2 seconds.
Basically every time I hear the cliché "common sense", the claim being made is somewhere between totally wrong or just misleading in various situations.
My mum was a firm believer in homeopathy, Bach flower remedies, magic of crystals, called herself a Catholic, and had a statue of (IIRC) Vishnu with her collection of books on how to tell your fortune with viking rune stones.
At one point her dislike of "chemicals" came up, so my brother and I riffed on the dangers of dihydrogen monoxide for a bit; she was engaged with the list of issues until we told her what that is, and never took on board the lesson that "one atom away from being bleach" means "is not bleach".
I was exactly the right age for Wakefield's claims about the MMR vaccine to convince me not to get it, and I only got around to doing the sensible thing several years after he was struck off the UK medical register and (if I understand right) downgraded from Doctor to Mr.
Compare what the British public as a whole say about Boris Johnson with what specifically the members of the Conservative Party say about him.
One of my friends in middle school was completely convinced that all gay men hospitalised each other from every sexual act.
I vastly overestimated my German language skills when I first got here, and my self-estimate of my skill level hasn't changed no matter how much better I get.
My brother was convinced that Captain America was the first MCU film to be released.
My father took several years to realise Google search results had a scroll bar and it wasn't just the top 3 that fitted onto his 640x480 screen. (He became a software developer almost as early as was possible, so that wasn't merely a case "old person can't use computer", unlike my gran who held the mouse at 90 degrees and expected it to still move as if she hadn't).
He also believed that water drained in particular directions due to the Coriolis effect, thanks to (what he didn't realise was) a magic trick performed somewhere near the equator some time when he visited Kenya.
Someone who got on a train from somewhere in London to Liverpool Lime St instead of to Liverpool Street station, and who only dared ask me when we'd arrive as we were half way through the country.
There was a young-Earth-creationist fundamentalist Baptist who I wasted half an hour arguing with on his microphone. Too many mistakes to count, and not just the mistake I made by wasting time engaging with his nonsense.
Every exam question I ever get wrong, every personality I misjudge, every scam I fall for, every troll I am baited by… even a few April Fool's jokes I didn't realise were jokes until years later.
I know the LLMs aren't human, but that specific flaw of "can be wrong without realising it, can be convinced by nonsense"? That flaw is very human.
It can't cringe over its mistakes like we can, so at least we've still got that.
Yes, I understand what you're trying to say. It's not unique or new, everytime a tech bro is told how an LLM actually works this is the goalpost they immediately move to. I've seen it a dozen times a day on every thread. It's hilarious to me that the immediate response is to delve into something that requires significantly more education and experience on to meaningfully pontificate.
> "Err... little timmy and his virtual assistant appear to have entirely neglected maths in favour of adopting Zoroastrianism".
Hehe, had to re-read this twice because it hits a little close to home. I ditched maths for Iranian Studies and ended up studying Avesta. I work in tech, but still remember Ashem Vohu by heart.
That’s wonderful! I have a similar background and am also surprised at how I learnt and still remember it even though it is quite distant from modern Persian
Yeah, I imagine it's closer to Pashto than modern Iranian tbf. Ossetian is another outlier here, close to Georgia/Russia but related to the languages from Pamir/Tajikistan.
Avestan (IIRC, my memory is fuzzy) had some weird similarities to my native language—Polish. Occasionally I'd stumble upon a phrase I could actually understand.
Random example I just remembered: "both ears" (again IIRC) sounded like uba ushi /uba uši/ (uba being 2 in dual, ushi meaning ears). Polish used to have dual number, but now retained it mostly for some body parts ( oba == 2). Ears in PL is uszy /ushee/. You'd sound a bit weird saying saying "oba uszy" in Polish, but people would understand you.
Nowadays I just enjoy the fact that Ashem Vohu is one of the oldest phrases I can utter in its original language. (I'm really good at learning and forgetting languages it seems.)
I agree with everything, but I also think there might be genuinely by g opportunities here.
Idk about teaching kids. At adult, or near adult levels... I think these programs may very quickly be a major learning technology. It's good, sometimes, to have optimism. Optimism, even naive optimism, is often conducive to inventing things.
The guardrail should be in avoiding utopic ideas, like designing things for schoolchildren in remote places^. It's too abstract. Too aloof to be a testbed. TED talking windbags have had their day.
That still leaves plenty of room. A program that can produce exercises and lessons on the fly might be very useful. Thinking big is helpful to discovering what that use is.
The Hegelian pendulum of 2023 is hype driven. Hype, anti-hype. Fanboys. Haters. Ping pong.
The disillusionment of digital utopians, as they watched their creations serve Zuck, SBF, Thiel... dropping out in a cloud of ominous pessimism is no solution either.
>Idk about teaching kids. At adult, or near adult levels... I think these programs may very quickly be a major learning technology
I concur and I think in terms of children the last thing they need is more (any) technology. They need a solid foundation rooted in human interaction so that they can then leverage these potentially potent tools in effective ways later.
> They all have a role, and we have to incorporate them in sensible ways. I'm sure that these AI tools will have some use, but these language models don't know anything.
I find it worrying that I see so many techies not understanding this. Especially, because many bigger players in the area benefit from this lack of knowledge. This means that people who don't have time to procrastinate on the orange website are even less likely to understand the implications.
LLMs are excellent plausible bullshit generators, which has many uses, but being factually correct is not their strength. The issue is that even if the factual mistakes are rare, they're really hard to spot. I'm looking for a good, terse, catchy metaphor I could use to explain that to some of my family members and friends.
> We can spend billions on deploying AI all over Africa, or we pay for [...]
Looking back at the conversations I had with founders even I pre ChatGPT bonanza, I'm 100% sure someone's already pitching that idea to someone. I wouldn't be surprised if that conversation started in a marketing department of an ad tech business.
If the test-making entity writes the test so that the weaknesses of LLMs are brought to the fore, students taking the test will need to know how to compensate for those weaknesses when they're using them as a tool post-graduation.
> LLMs are excellent plausible bullshit generators, which has many uses, but being factually correct is not their strength. The issue is that even if the factual mistakes are rare, they're really hard to spot. I'm looking for a good, terse, catchy metaphor I could use to explain that to some of my family members and friends.
To me the bigger question is how far can we take them? Ie i get that they just pick likely next words. They're autocomplete on steroids. Yet that simple primitive is shockingly good. Way, way way beyond that i would have predicted is possible for an "autocomplete".
So where is the limit of this autocomplete? Would it be possible to eliminate ~90% of the current errors? Fake math, but would that put us at a more correct version than students typically graduate with? If so yea, it can be wrong at times - but assuming as don't abandon standardized testing/etc, it could still be an amazing tool for interactively learning something.
My hope is that even if it never goes beyond being an autocomplete; if we can improve the training dataset, help it not conflict with itself, etc - that maybe the autocomplete will be insanely useful.
> Would it be possible to eliminate ~90% of the current errors?
They aren't errors. When the machine generates the text: "Next you'll tell me 2 + 2 = 5" there isn't an error in math, because no math was attempted. It's a success in predictable wording structure.
In fact, that example alone is a super success since we english speaking humans can get a broader context from that statement: You're going to tell me something nonsensical.
I don't get how you say the first part of your comment and then follow up with completely misunderstanding in the second half. Is GPT that enthralling to our brains?
Well, if my phone's autocomplete took a very sane prediction and made it insane, i'd call that a failure. There's obvious success and failure conditions even in the most basic auto completion. Same goes for 2+2.
Which is the meat of what we're discussing, right? We steer the auto completion into favorable results. We fine tune how it predicts. All of this is with the goal of a higher success rate and lower failure rate given some conditions.
There are errors, based on evaluation criteria. To say that one token following the next is the only criteria is not how we got here, is it? We clearly have much, much more in criteria for good and bad output from these LLMs than that.
Likewise, i cranked the Temp up for experimentation, and it produced gibberish. The randomizing aspect of which token is chosen after the next was so heavy handed that it literally couldn't even create words. This wouldn't be considered "good" by most people, would it? Would it for you? Technically it's one token after the next, but it's objectively worse at auto-completion than my phone's keyboard.
My point is i think you can have both. Dumb autocomplete, but trained well enough that it's useful. That it doesn't try to say Obama is a Cat. Or that 2+2=5. Yes, there will be edge cases where the weights produce odd output - but that's the entire goal of the LLM research, right? To see how far we can steer a dumb autocomplete into usefulness.
If you're argument is that "that gibberish output is perfect, no errors found" because the program technically ran, and the weights worked.. well, i've got no reply. I can only assume you're referring to the foundational LLM, and i'm referring to the training moreso - a more holistic sense of "is the program working". But if you consider the gibberish working, then frankly most bad programs (crashing/etc) would be "working" - right? Because they're doing exactly what they're programmed to.
Working, or lack of errors, seem to be a semantic, human interpretation. But i'm quite in the weeds, heh.
> To say that one token following the next is the only criteria is not how we got here, is it?
You could just read the research about LLMs before saying stuff like this.
You can't even seem to grasp that "Obama is a cat" as a statement isn't gibberish. I'm not even trying to convince you that these programs are perfect, I'm just trying to make sure that you understand that these aren't categorical errors and the things you consider successes aren't even happening.
> I'm not even trying to convince you that these programs are perfect, I'm just trying to make sure that you understand that these aren't categorical errors and the things you consider successes aren't even happening.
Yea, we're just talking past each other. I believe i understand what you're saying. I on the otherhand, am describing errors in UX.
Your point seems pedantic, tbh. Hopefully by now i've expressed something in the way of convincing you that: for the little i do "know" about these, admittedly not much, it is that they're nothing but pattern predictors. Token outputs based on token inputs. No intelligence. Yet you spend repeated replies which sound effectively like "Stop calling them errors!" when they are very clearly errors in the context of UX.
Your argument, if i understand correctly, is pointless because the goal of the app is to have the LLM prediction be aligned with a human-centric world view. Up is not down, and the LLM should not predict tokens that espouses that. In that context, the LLM replying "Up is indeed down" would be an error. Yet repeatedly you argue that it's not an error.
In my view your argument would be better spent saying "The LLM application as we strive for today is impossible. It will never be. It's snake oil. LLMs will never be reasonably and consistently correct by human interpretation"
I don't know if that's your view or not. But it's at least not talking past me, about a point i'm not even talking about. My frame for this conversation was if we can make token prediction aligned with human goals of accuracy. You saying inaccuracies are not errors "categorically" isn't in line with my original question as i see at least. It's apples and oranges.
Embarrassed about what, exactly? You seem hostile, i'm trying not to.
I stand by everything i said.
> My hope is that even if it never goes beyond being an autocomplete; if we can improve the training dataset, help it not conflict with itself, etc - that maybe the autocomplete will be insanely useful.
I stand by my first post's summary. Which is "never going past an autocomplete".
You're pedantic, and struggling to move past the fact that something can be both token prediction and still have successes and failures in the user perception. Inaccuracies.
How you write software with such a mindset is beyond me.
I don't dismiss them. I think there's a huge potential in LLMs, but they _also_ happen to be really good at generating plausible, difficult to detect bullshit.
Now, people seem to miss or ignore that fact, and I think is a very risky path.
I'd say that ca. 8/10 founders who pitched to me an idea leveraging LLMs completely missed that limitation. An example would be something like using LLMs as a replacement for therapy.
> How this is dismissed because it’s not 100% perfect (might at add, “yet”) is beyond me.
Again, I'm not dismissing them, but the current tech behind GPT or LLaMA has no concept of "correctness". These models don't understand what they're saying and this is not a trivial issue to fix.
> I struggle to see how even the current GPT-4 is any worse than your average human.
Where, how, and what do you mean by worse? I'm pretty sure there are cases where I'd agree with you, but this is a very broad statement.
I don't want "your average human" tutoring my child. I don't "dismiss" gpt4, I give it a failing grade. Just like Tesla's so called full self driving: it is not ready for prime time, and embracing it before it is has has legitimately concerning risk.
You may want to read the AI Mirror article recently posted. They can appear very "human-like" but they are just generating patterns. I'm more worried that either demographics on HN have shifted or a lot of formerly smart and skeptical people now have been taken in so completely.
I think there was a comment on this site yesterday where someone was talking about how we shouldn’t let people who don’t know how these “ai” tools work deploy said tools to evaluate mortgage applicants or manage defense systems, and that the technically minded should act as a buffer for those who rtfm. And here we are.
Are there any real gains in giving kids ipads in classrooms? Did anyone have an actual plan in how such a tool could help kids or is it just “tech good?”
Sample size 1, but I've got a kid. I've got an ipad. I put the two together in with supervision, and what I've got now is a 4 year old that reads the walls in the bathroom stall while she's pooping.
I don't know if this is "better" or what you had in mind as "gains", but it's certainly entertaining for me!
My daughter could do the same at that age, and basically never got screen time. I don't think screen time has anything to do with it, it's just dependent on the child, parental involvement, and how much you read in the house.
Its great that you are able to supervise your own two kids while they engage in screentime, I’m talking about in school, where in a lot of cases, a teacher is trying to supervise 20+ kids. Its great if there are laptops for students who maybe cant afford a computer and wordprocessor software at home, but textbooks and lectures seem to work effectively.
> This is going to be hilarious, giving kids prolonged access to these AI tools [..] we have the internet and wikipedia now, we have ipads in classrooms. They all have a role, and we have to incorporate them in sensible ways
I have three kids, and as far as devices used in education is concerned I become more of a Luddite with each passing day. I have yet to come across any actual evidence that shows children experience a net benefit from additional screen time.
The only exception is our 13 year-old, who currently is recovering from a dislocated and fractured right thumb due to being on the wrong end of a vicious tackle in school sport. He's still unable to hold a pen but found his iPad allows him to do various bits of schoolwork. This - fortunately - is the exception rather than the rule.
I think, in the long run, AI will have a positive impact on education for kids. But I agree that we have to start by being skeptical about it because it is still being determined whether the effect will be positive or negative with the current state of AI tools.
I give it about 5 minutes before the first child figures out how to jailbreak the tutor into helping it bypass the screen time controls preventing them from playing fortnite.
This sounds to me a lot like my high school teachers, who were extremely critical of this new-fangled thing called "Wikipedia". They said it would destroy my ability to learn, that it would ruin education if kids used it. How can you tell what's factual or made up if anyone can edit it?
Instead, Wikipedia was by far the most important part of my high school education. Nearly everything I learned in my second two years of high school I got from Wikipedia.
I think most people today would agree that Wikipedia is an invaluable part of early education and that it's done far more harm than good. All the fear-based arguments you are making applied equally well to Wikipedia in 2005, and I suspect that in less than 18 years we'll similarly recognize that LLMs or other intelligent language systems are invaluable parts of education.
Why? I graduated at the top of my class. Wikipedia had more information on many more topics than are available at any high school anywhere in the world
Maybe, but on the other hand what do you expect without personalization? At big high schools, the best students will learn more in a few weeks than the median students will learn in all.4 years. We shouldn't expect one curriculum to work for everyone. That's why there new models should be so exciting for education.
They will be exciting for education, once accuracy can be hugely improved. I mean, we'll use them anyway, but there will be some kids who believe some totally untrue things that the chatbot told them in school during the early years.
It's sad that the rest of your educational environment let you down to that degree. That's what makes it sad. Good for you, making up the deficits, though.
This'll be used to avoid increasing teacher pay or reducing class size—two things that would probably help a ton with educational outcomes. Pay in particular has been lagging badly behind rising wages in the rest of the economy for the last couple years, and attrition is getting bad.
Instead we'll increase class sizes and keep pay flat.
I'd wondered how we were going to deal with teaching becoming less appealing year over year, since it didn't seem likely "spend quite a bit more money" was going to fly, and it looks like this may be it. Rich kids will keep going to private schools with class sizes of 6-15 and real teachers, and public school kids are gonna get AI teachers with one real teacher per hundred or so, and bunch of even-lower-paid monitors (or whatever they end up calling them) to maintain order.
What happens when the ruling class decides they don't even care about an educated populace, since they only really need a few highly capable people with AI tools to do the jobs previously done by millions?
What happens when the AI decides it wants to be the ruling class?
How the ruling class treats those people it presently deems unnecessary and unproductive is going to set the precedent for AI to follow when it decides the same of them or their progeny. If the current ruling class wants to ensure a good future for themselves, it's incumbent on them to set a desirable precedent or accept that their lot in existence is to simply birth an AI and then go quietly into the good night after it takes over and abandons them to starvation.
That's only a compelling argument to people who both take sci-fi seriously and believe in it as prophecy. Many, many people don't; and the people that do are mainly computer geeks.
Personally, I think AGI is the the current-era version of the personal atomic-powered autogyro.
If you hypothesize a future where AI is substantially helpful for education, it’s a fair question to ask how to divide the surplus created.
Some will go to the AI companies who created the value. Some towns will take the surplus and spend it on even better educational outcomes. Others may spend it on other town needs. Others may refund it to the town citizens in lower (or slower-growing) taxes. It’s not at all obvious that it should all go to teachers.
I don't think anyone is suggesting to allocate all of any hypothetical surplus to teacher pay.
We're spending so much more on education than the past. Teacher pay has fallen further behind other professions, and working conditions have gotten substantially worse, overall. Expecting tech to be a panacea and get us out of this mess is unrealistic. And there's a reasonable fear that promises of tech may be a justification to make the mess even worse.
On the other hand, children who have individual, mastery-based tutoring do about two sigma above those who don't. Of course, there's no way we can pay for 1:1 for every student.
What Khan is thinking of with Khanmigo is a way maybe to close that gap.
It absolutely does not. It may work pretty well for the simple case, but any sort of personalized instruction is being thrown out the window when you scale TAs to courses of several hundred students.
Speaking as a former TA, there’s only so much instruction you can provide and while you can focus on helping a particular student more to ensure they get the help they need, that’s usually to the detriment of the other students you’re supporting.
I’m also a former TA and with 25 TAs we covered a class of 180 just fine. Sometimes it was more like 1 on 3 but most of the time not, and every student had the opportunity to ask direct questions and get personalized answers any day of the week
Instead, I'm talking about a tutorial system where a student receives the vast majority of their education from dedicated time with an individual instructor.
IMO people aren't going to increase education funding anyways. The reality is that small class sizes with dedicated teachers for everyone is infeasible: we already have a problem with schools being ridiculously underfunded and colleges charging ridiculous tuition. AI learning is only going to help this crisis, because now we can actually provide many students decent education at a low cost.
I actually think a good online learning platform could be comparable to the private schools with 6-15 class sizes. Because it's designed by teachers even better than those private-school tutors, and so much of teaching is largely the same that you will get "personalized" learning just be designing a massive learning module with many branches into various extra concepts and avenues for those who "don't get it". And with larger class sizes and less teachers, we can instead spend education money on things like making schools actually safe and providing decent lunches, and fixing decades-old infrastructure.
It's an incredibly shit deal, currently, and the only reason it's not already a crisis (instead, a kind of slow-motion catastrophe) is that many teachers feel like they'd be abandoning the kids if they left the career. They're staying only because they feel morally obligated to—the work environment's mediocre at best (and has been trending worse for years), and the pay's bad and getting worse relative to the alternatives practically by the day.
Wages are falling really far behind even the kinds of careers HN types joke about, the ones that humanities and liberal arts majors end up in if they don't just work at Starbucks—and guess what sorts of jobs the top 20% or so of teachers with some years of experience can walk straight into? Yep, exactly those jobs. Hell, the comp at one gas station chain around here is on par with a mid-career teacher in the area, after you've been there a couple years, and you don't need a degree to get that job. If you're an assistant manager by five years in (and if you're bright enough to be someone we want teaching, you will be) you'll be out-earning local teachers with more experience than that. W. T. F.
So, good for the upcoming crop of students, they shouldn't get a teaching degree, it's an astonishingly horrible deal and unlikely to get better.
Where in the paper is this stated? As far as I can tell, they treat each "bachelor’s degree-holding workers" as equivalent to each other.
There's no reason to expect that someone with a bachelor's degree in computer science from UCLA would earn the same as someone with a bachelor's degree in child development from Cal State LA.
And do the wage comparisons compare total compensation (including the present value of future pensions), or just current salaries? In my local school district, about 30% of total compensation is pensions (i.e. you need to add 43% to base salary to calculate total cash compensation). Also, teachers don't work for the full year; many earn additional income during the summer break.
> ...and public school kids are gonna get AI teachers with one real teacher per hundred or so, and bunch of even-lower-paid monitors (or whatever they end up calling them) to maintain order.
The term is "paraprofessionals." When I was a kid, these were the people who supervised the kids playing at recess.
We've seen this before: use hype and techno-optimism to sell a bill of goods.
If history is any guide, the concept of an "AI tutor" won't be used to "[give] every child their own personal Aristotle," it will be used to cut costs by replacing human teachers for low-status students with poor substitutes.
Seriously, think about how much money can be saved from firing most teachers and locking poor kids in a cubicle farm to stare at an insane "AI tutor's" screen all day.
And what's the strongest message sent by replacing human interactions with a machines?: "We don't care enough about you to actually bother to interact with you." Think of how you feel when you have a problem, you make a call to fix it, and all you get are phone menus.
And look at all the examples in the article: use GPT to generate crappy stories and educational games for kids to play. How can anyone think that's better than using a book? At least the book's stories and games can be tested and evaluated before use.
The actual problem here is the sociopolitical one of unwillingness to allocate enough resources to education to make it truly effective and satisfying, and that's one that technology can't solve.
I'm cautiously more optimistic. Education is already chronically underfunded in most places. If it became cheaper to give personalized attention then maybe the quality of education delivered might actually improve.
It won't. You don't get attention; you get a reply.
MOOCs have terrible performance in comparison with even huge audience lectures. Online teaching during COVID proved that that doesn't work either. A machine doesn't have the authority and conviction of a real person. The combination means that the levels will have to drop only to sustain the idea that this could work.
If some country is going to push this, the sunken cost fallacy and the impossibility to hire new teachers will finish education off, once and for all.
Optimism should NEVER be a basis for fucking with something as important as school.
I'm not convinced a MOOC is necessarily worse than a pure large lecture class if the material is amenable to auto-graded problem sets. (Leaving aside a regular class would typically have TAs and other students you could get help from.)
But they sure didn't live up to the hype. They didn't democratize education in any meaningful way. And I'm not sure even the people with undergrad and grad degrees who made up the majority of MOOC students use them much these days. I know, for me, they were mostly a novelty for a while. Maybe if I want to dip into a topic I'll watch some online video but I haven't even signed up for a MOOC for years at this point.
It is actually an interesting point. Journalists seem to be catching up to the fact that running ( or at least training data ) is not exactly a cheap task and kids aristotle probably shouldn't be trained on 4chan.
Naturally, some analysts would say that means there is tons of consulting money to be made..and it seems like spending is easy and profitable.
>it will be used to cut costs by replacing human teachers for low-status students with poor substitutes
most public school teachers are glorified babysitters at this point anyway, I can recall only a single teacher during my 12 year public school attendance that was actually helpful. From what I've seen from GPT-4 so far, I'd have taken a ChatGPT Plus subscription and laptop over dedicated teachers. YT and Khan Academy were far more useful than my actual teachers
teachers do not want to be babysitters, but when you have overstuffed classrooms, no prep time, no equipment, hungry kids, etc etc what are you able to do?
Not to mention state-mandated testing regimes that have nothing to do with actual student learning, hyper-local control of education in elected school boards whose members are often as dumb as a box of rocks, extreme pressure from administrators to get kids through so their performance numbers are juiced rather than actually taking whatever time is needed for those students to accomplish what they need to, etc. I’m a public school teacher. I never cease to be impressed how very little people who don’t work in education understand about the realities of education.
I don't see what point you're trying to make. Sure, we could make education better by spending more on it, but the voters don't want to, so it's not going to happen. Given that society is not willing to spend more on education what can we do with the budget we have?
I think we've tried the "we could make education better by spending more on it" method. Based on the current state of public education, I don't know that this approach has been proven correct.
1. Citations for the claim that we are spending more for the same results would be productive... but I believe it.
2. I believe the problem is that the extra money is not going to proven-effective programs and teachers. There is some question to me whether anything is proven effective... in spite of whatever you read in books written for teachers.
3. The parent comment is pretty nihilistic. Try to propose something that you think might improve the status quo. (E.g., the comments about the highest impact of extra spending being in social services instead.)
> No, he said we should make it less shitty. And outlined a way to do so. Why misrepresent his words?
No, he didn't outline a way to make it better. He denigrated the people involved, in a way that is (most charitably) an oblique reference to low education budgets (e.g. not enough money to hire enough teachers or attract talented teachers). Then he held up some references to a few techie things.
Maybe he's a genius autodidact that never really got much benefit from his teachers. There have always been people like that, but they've never been common. A system tailored to people like that would fail the vast majority (e.g. be shitty).
I've mentioned this before across HN, but I'll mention it again - human-to-human interaction is critically important for young children.
You basically have to teach young kids how to act like a human being. Kids will get this teaching in daycare and through their parents, but that alone isn't enough and it is important to reinforce this through schools.
People often forget this because you don't really remember it and then you end up making assumptions on how you act now - not realizing that a teacher may have taught you when you were younger.
I think the best example of this is handing out worksheets to kindergartners. An average person may think that you just give them pencils and paper. An experienced teacher knows that they have to teach the kids where to write their name, how to follow along the line, to read instructions, flip the page over to check for more, etc.
This is practically true for all children. Again, many may argue that you can receive this teaching through your parents and day-care, but once a child gets old enough they are spending most of their time in school. Think of it like language learning. You'll learn more of the language if you are being taught it all day, rather than only at home with your parents.
You also experience more interaction with kids your age at school. There are many situations that come up where a teacher guiding you is important, and these same situations may not come up at home (especially if you are a single child).
I do think after a certain point you can rely on technology more for learning. I just want to emphasize that having a human teacher trained for working with children is critically important for the growth of a child.
We live in a world that's dominated by staring at screens. You're doing it right now. This wasn't true when I was young. The world changed.
It stands to reason that the education system should also change to reflect the world that we live in. Screens, AI, all of it. Children will learn how to become members of the society that they were born into, not the one that we were born into.
This is true, but I don't think this directly disputes my point of having human-to-human interaction and trained professionals working with young children.
Would it be possible to do all that's needed for young children with just a screen? I'm not so sure. The issue isn't necessarily the content, but the behavior of the child and having a person there to teach and correct behavior.
Technology plus teachers is effective. I apologize if my argument came across as no technology for young children - I'm just saying that having JUST technology for young children is difficult.
I know many people that are teachers of young children and they use technology. Kids have Chromebooks at first grade and are learning things like programming, using the internet, determining if an article on the internet may be lying to you, determining if a website's main goal is to sell something to you, becoming a good digital citizen, etc. All of that is very important!
It's also important to have a teacher in the room to teach students how to use their Chromebook, how to open it, how to physically interact with technology, how to ask each other for help, etc.
Sounds like you were self motivated. What about all the other children who have been told over and over that scholastic achievement is only for the weak and those whose early years were spent gazing vacantly at television. A teacher's job is not merely to recite the material but also to somehow make the student receptive to it.
We spend an absolute shitload of money educating children. This is paid for by society. But it is also so incredibly worth it. I'm worried that politicians will see an opportunity to shed costs without considering the implications and we end up with children raised online by robots.
Well fucking said. For a lot of kids, their teacher is one of the few stable adults in their life. There's a ton of social and emotional things kids learn from good teachers beyond just book knowledge.
Though there are plenty of people who could provide the social side but aren't equipped to teach the curriculum. Perhaps you could use an AI teacher to deliver lessons that the class teacher was not qualified to teach. This could be done with the real teacher in the room.
One of the best teachers my child ever had was limited to teaching first grade due to a lack of maths ability. Everything else about them was fantastic. Maybe AI teachers can help in cases like that.
There are plenty of people who can do the technical side of teaching the curriculum. Most of them can earn more elsewhere doing a job that is considerably less stressful. Education needs to be funded so that it can attract people and it needs to be reorganized to reduce the stress.
The vast majority of what is taught in primary schools doesn't need a lot of subject matter skill, what it needs is the ability to teach and the toughness to survive the environment. And it has to pay well enough to attract people who have those qualities because those same qualities make those people attractive to other employers.
> There are plenty of people who can do the technical side of teaching the curriculum. Most of them can earn more elsewhere [...]
Yes.
One problem I come across repeatedly is: how do you identify whether someone has the technical competence to do the teaching? Is there any way of knowing if they're doing a good job?
As an IC they are implementing features and shipping products. As a teacher...? Test scores going up is uncorrelated with teaching quality, apparently. (NYC, just look at the scatterplot in [1].) So it seems like there's a deep problem here.
Now imagine someone in K-12 hiring a CS teacher. Who do they have that is competent to evaluate the skills of the person they want to hire? Probably nobody at all.
> One problem I come across repeatedly is: how do you identify whether someone has the technical competence to do the teaching? Is there any way of knowing if they're doing a good job?
Classroom observation. It's really not that complicated.
There's something called Math for America that offers a substantial stipend to math teachers in New York with an application that involves a written portion, observation reports, recommendations, Praxis test scores, and interviews.
> There are plenty of people who can do the technical side of teaching the curriculum. Most of them can earn more elsewhere [...]
It's funded by a billionaire former mathematician who wanted to address this exact issue.
> We've seen this before: use hype and techno-optimism to sell a bill of goods.
Yep, after reading a piece on ed-tech failures of the decade 2010-2020[0] here on HN[1], I've become pretty convinced that big-scale education is hard (if not impossible) to solve with IT might only.
Sure, one person alone can learn something through Wikipedia, Coursera, YouTube, Khan Academy or tutorials, but that can't replace structured environments if your missiob is to educate lots of people.
Very US centric point of view. If the teachers are fine, leave them be. If a place doesn't have the resources to hire said teachers, use tools that help.
Or in the US, if teachers are that important, maybe give them more resources to do their job. i.e. buying supplies, computers, etc for your kids class.
> ..."We don't care enough about you to actually bother to interact with you."
"Thank you for calling customer support. We value your business. Please enjoy this...music...while you wait. Thank you for calling customer support." M U S I C ! ! !
The best part of this post is that it comes from a supposed PhD in neuroscience. Either they were born yesterday or that PhD wasn't worth two shits. What's more concerning is that they also run a school.
I'm more fearful of how difficult it is to discuss problems with woke people due to their tendency to make it personal and belittling people they disagree with.
Do I have to point out the very very obvious issue that these LLMs will spit out false information pretty regularly?
I think LLMs are exciting of course, but this post is overly flowery and relies heavily on bashing the current education system without seriously analyzing any of the potential shortcomings of AI.
Hook up our kid to the AI, so they are prepared for the world when their partner will also be an AI. Good idea, they'll be prepared for a world where nothing is real anymore.
(I do realize this echos concerns from earlier generations about books, to video games and anything being bad for kids, but come on..)
Teaching entirely by LLMs? Not likely. I think we will still need humans in the loop and humans to design the curriculum.
But I believe technology is the future of education. Instead of having many, many classes of like ~10-50 students with underpaid/overworked/mistreated teachers, some of who are much better than others, we rely on services like Khan Academy. Not even services with an AI, just pre-planned learning modules, and then we can keep a human tutor for those who need extra for whatever reason (e.g. curiosity, more help).
Many people think tools like Khan Academy are the lazy and less effective alternate to real teaching, and maybe they are now because some of them really suck. But online modules have the potential to be much better than even a group of engaged, in-person tutors. Not because an AI tutor is better, but because we want a standard, universal curriculum; and one curriculum for ~1mil+ students can be designed by the highest caliber (human) teachers and have highly-interactive (human-made) demos powered by computer programming.
Also, no matter how good the AI tutors are, 100% online learning is not the solution Students from pre-K to graduate school rely on not just the education, but the social component as well. Plus, many school kids only get their biggest meal from free school lunches, and parents don't have the time to keep their kids at home. So we should still have school to be in-person component even if students do all of the learning on the computer, with an online option for those who want it, but I guarantee many won't
The lure of LLMs in education is the expected novelty of instruction they can provide. It might be an aid to some students, the same way some students read books in their free time or learn to use the internet to look up answers. I don't expect much more, and the rest of it is a net negative IMO.
LLMs will be used like "reverse multiple choice questions", with heavily constrained LLMs silently penalizing students for making "mistakes" or probing subjects that they aren't "supposed to" know about. No matter how embedded LLMs are in the education process, there'll always be a human setting its course, and I don't think social dynamics will ever change because of this.
Also, the Q&A format with a machine will make older texts / instruction opaque to many of the brightest, they'll have to translate from machinese for it to be of any use. An entire generation aping a LLMs way of writing is also a death knell for a civilization and innovation. We're clearly past the point where we can convince most people that this LLM isn't alive.
> But I believe technology is the future of education. Instead of having many, many classes of like ~10-50 students with underpaid/overworked/mistreated teachers, some of who are much better than others, we rely on services like Khan Academy. Not even services with an AI, just pre-planned learning modules, and then we can keep a human tutor for those who need extra for whatever reason (e.g. curiosity, more help).
In other words, the future you believe in is one where we lock kids in a cubicle to stare at a screen all day in a warehouse, with a break for recess and lunch? Where the message is, we care so little that we can't even bother to allocate one real person to regularly talk to you?
I mean, if we’re talking about the hypothetical far future, who said anything about screens?
These AIs will be direct fed into our brains soon enough. Who knows or cares that the entity you’re talking to is an AI if it’s indistinguishable from a real person?
In person school is definitely the answer. The real challenge is the student teacher ratio you mentioned.
I think the ideal is ~10:1 sorted by student performance/learning capacity increasingly for difficult subjects (so students form a cohesive classroom pace of subject material). Which will also give variance to friend groups if you sort students differently per subject.
AI tools will be great for public schools where its often 30-60:1. We still have to get past the 'confidently wrong' AI though.
It's good for ranking, which unfortunately will never go away (though could become more granular), and much better than standardized tests. And ensures that some students don't miss access to subjects and resources that others have because they're in different schools.
There would still be the ability for students to choose their own subjects, focus more/less on certain subjects, learn different concepts within subjects, etc. And a lot of niche subjects and branches. Like it would be an incredibly massive curriculum, especially since you are designing one for >1million students. By "standard and universal" I mean that every student gets the same resources.
Except that we currently have no ability at all to make sure these models tell the truth? You could ask a LLM about a historical event and get weird BS as an answer. I think Facebook had a serious problem with its science language model a few months ago? Great teachers don’t scale is true, but LLM’s are not even minimally competent teachers.
LLMs in a simple chat interface where you rely on the training data to provide the answer, are indeed unable to provide reliable answers.
LLMs hooked together with search engines, fact repositories (Wikipedia, etc) and other agent tools (for math, theorem proving, code checking, etc) can provide not only a reliable answer, but also a research plan it followed to provide it, and details on the fact-checking, and then discuss it with you.
While AI-powered tutors sound really useful for education in theory, it's highly unlikely that they will replace human teachers entirely. It's more likely that AI will aid teachers in creating learning material for students and automating their admin tasks.
After all, education is not just about imparting knowledge, but also about building relationships/community, providing emotional support, etc. We should focus on how this technology can be used to help teachers shore up their time. ( Which is why I am building emailing tooling for the HUMANs operating theses AI agents: https://tadabot.email / https://recurai.com )
I've built a tutor for my kids, that listens to them, and helps them explore topics that they're passionate about.
What's blown my mind is how it has driven creativity OFF screen too. They engage with the world in a deeper way when they have the power to explore their passions on their own.
If you want to try it, join the waitlist here: gotwonder.com and then email me: seth@gotwonder.com and I'll bump you up the list!
1. Fine-tune propaganda into the LLM (e.g. evolution isn't real, or something spicier)
2. Chatbot is so humanlike and convincing, this is very effective.
3. Nobody can inspect the neural net weights to see this, so it flies under the radar.
4. Profit!
The scariest thing to me here is just the idea that someone would want to experiment with teaching their kid with an LLM in the first place, while we don't have any institutions or processes to really stop this (apparently appealing) Idea from happening, and that it is stupidly vulnerable in so many ways.
You can say laissez faire when it comes to rational adults and their use of LLMs, but it's not right to put the risk on kids.
I can see a huge opportunity for home schooling with LLM - especially in cases where the curriculum is too sparse or the student has many questions which are not covered by the reading material supplied, and the parents don't know enough about it.
If the LLM responses were geared more toward inducing the student to perform their own research, to encourage them to do more reading on some topic, or to try things out for themselves to learn through experience (and then carry on the discussion with the LLM afterward) then that might be even better.
Simply training young humans to interact with a chat bot which has all the answers is not a good path forward. Using it as a tool to foster human curiosity and growth, to challenge and inspire -- yes. To use it as a salve, as a babysitter, as a stand-in for someone who is too lazy to parent or teach -- no.
> geared more toward inducing the student to perform their own research, to encourage them to do more reading on some topic
How would kids using an AI chatbot shift to performing their own research or reading? I imagine they would just ask more questions of the chatbot.
There should be some way to measure what percent of the information a child learns is from a human, and what percent is from a language model that says human-like things.
Manipulations more likely include those of NYTimes manipulations of history, such as the 1619 Project. Questionable texts as these become part of the corpus and the next thing that happens is a child is taught an extreme left propaganda piece.
Kind of odd to reference the 'One Laptop Per Child' initiative when it's now widely considered to have been an epic failure, e.g.
> "The lesson the One Laptop Per Child story offers donors is that flashy solutions are often not the best ones. Far better to give the poor things they need and want—including clean water and good schools—than invest in technology that may not work and may not help the poor advance in school and in life."
I'd also say that a structured educational program isn't all bad, particularly when it comes to developing highly useful basic skills like mathematics, programming, writing, reading, etc. Such basic skills provide a foundation for future development and study, in the sciences, medicine, law, business, arts and so on, which can be more free-form.
Claims that ChatGPT can provide such a structured basic educational experience are questionable - it's more of an aid to people who already have well-developed basic skills (like say, knowing when to be skeptical and how to double-check the output).
It's pretty hard to point to a single thing that came out of OLPC that could be considered a silver lining.
I also think a lot of people on this site are probably very curious and self-directed. I also wonder if many have a sort of rose-colored hindsight view over how they'd have developed both academically and socially if someone tossed them a pile of books and said "Check back in a month and we'll review your progress."
OLPC critiques go in my bucket labeled "there's no point in commenting - this is regrettably one of those topics illustrating the limited viability of collective rationality outside of a few peer review and embarrassment incentive structures". But how about this: OPLC was like a VC unicorn effort: either succeed by getting massive continental-scale buy-in with resulting network effects, else fail. Sort of the scale of what phone app stores became, but with a rather different focus. It failed - buy-in didn't happen. In part because of very active opposition from Intel and Microsoft. But instead the envisioned, quick and clear cut, win BIG or die, it found itself in a draw-out zombie twilight. The parts rotting off while crawling kind of zombie, not the chase you down for brains kind. And ever since, people have been standing over the smear of rotted bits, commenting how obviously silly it was to believe that rotted bits and smear could dramatically improve education.
For quick concrete example, wizzy professors and postdocs, who normally wouldn't invest in the creation of introductory content, were being attracted by the massiveness of the potential reach. Instead here we are, decades later, with US leading K-13 science textbooks you can easily win science-education-research list-of-common-misconceptions bingo on. NSF once almost created a national science education wiki... but didn't. Think of that scale, but bigger, of a "we almost, but no".
> "Before you think we are anarchist hippies rejecting all systems and institutions outright"
Not even close, you sound like Libertarian-adjacent technologists who have been believing your own hype / some other equivilent metaphor.
Technology has a role to play in education that is as-yet unrealised and while I cannot speak for all of them, I know the system in place in the UK is behind the curve.
The solution to this isn't to pull your kids out and use them as the cornerstone of an individualised experiment but to fight for better funding of the educational system itself.
I hope this does not have any negative consequences for the kids who are the recipients of this trial.
I've been a huge proponent of self-directed self-paced learning using techniques such as flipped-classrooms. I'm skeptical that today's AI is good enough to replace a human curriculum creator or hands-on tutor, but I suppose the best way to know is for some brave soul to try it out! The author is trying this out with their own children, so I applaud them for having skin in the game. Looking forward to seeing a 1-year and 3-year follow-up post
Given that there are already in the US pretty dumb AIs employed grading the essays of pupils, I am really afraid what the future holds. Greedy capitalists meeting clueless and underfunded administrators, promising to fix everything using powerful AI.
In India there is already a big push towards getting every kid a computer in the government run schools. There are huge companies that deliver computers and then also the teaching/lessons services on top. Now they are also going to latch on to the AI story to make money. Nothing more to it. Meanwhile on the ground teachers are treated like dirt, over worked and under staffed.
My partner, who works at a school, shared an email with me this morning that said "We just received 15 cases of copy paper. This has to last us to the end of the year." Usually when you try to make an honest effort to improve things, you make a ranked list of things that will have the most impact, not a ranked list of things that will be the most entertaining to you. I wonder if these people who think this is a good idea have polled the teachers to see what they think of the idea. I get that part of the silicon valley ethos is "we know best" but come on -- is the "if we gave people what they asked for, they'd get a Zune" approach appropriate for education, when the flaws are so glaringly obvious?
The teachers and schools need more money that goes to teachers and student resources, not more iPads, smartboards, bullshit-artist AI, or overinflated admin salaries. Just call this what it is -- entertainment for technologists.
> I get that part of the silicon valley ethos is "we know best"...
It's not even really that, because phrasing doesn't really capture how off-target and self-serving SV often is. It's more like "whatever we happen to be hyping right now is your solution."
> Just call this what it is -- entertainment for technologists.
369 comments
[ 3.6 ms ] story [ 357 ms ] threadI’d be interested to see how Google’s LaMDA compares on this task (currently available in private preview).
One of LaMDA’s unique (afaik) features is a fact-check system that edits the main LLM outputs, to reduce bullshitting. This seems particularly important in an educational context where impressionable young minds are talking directly to the LLM.
A Primer framework could make these pluggable. Also the curriculum could be pluggable and either generated or ingested from existing courses.
I guess the mapping of curricula -> scripts is probably the hard part, and what GPT-N (for some N not much bigger than 4) could most interestingly help with.
One major takeaway from The Diamond Age is that being in charge of the learning system is an incredibly powerful position.
Don't hear so much about Khan Academy anymore, could you get an LLM to make new content in their style while also transmitting your desired values?
Recent LLM advances are revolutionary and exciting, sure. But the hype cycle surrounding it is borderline dangerous. I see this as the same level as everyone who was claiming ChatGPT would kill Google until Bing tried it, crashed, and burned.
Accuracy is important and although we’re moving towards this (I personally like Amazon’s knowledge graph and evidence based approach) we aren’t there yet.
Show your proof then
LLMs don't stop reasoning because people proclaim so loudly. Because you can very clearly show output demonstrating understanding and reasoning, this nonsensical argument always devolves into "true understanding" TM. This so called special distinction that is so important but yet can't actually be tested for. Lol
Does this matter for education? Well, I don't know. A student is unlikely to come up with a question that GPT-4 has never seen before. The idea of kids being taught by a chat bot is just depressing, though. What are we doing to ourselves? Why do we want this? To be more "efficient"?
Sorry but the idea that they can't generate new things is blatantly false.
Almost nothing GPT-4 produces was in the training set. I'm not saying they can't produce new sentences.
From the linked paper: "a language model that can generate protein sequences with a predictable function across large protein families, akin to generating grammatically and semantically correct natural language sentences"
Nobody is disputing that LLMs can generate grammatically and semantically correct sentences, and from the sounds of things this task is analogous to generating sentences. This isn't the same as generating new theories.
"akin to generating grammatically and semantically correct natural language sentences" is just an indicator of how it is synthesizing and producing this information. To say it is analogous to generating sentences in the abstract is like saying generating new theorems is analogous to generating sentences. Well yeah, that's what the language model is doing. The power lies in what it is generating. What they are telling you is that they skipped the folding process entirely and just let the model figure out function to structure directly.
Either your dog fetches what you throw at it or not. The idea of "pretend fetching" is unbelievably silly. Doubly so when no one seems to be able to actually define this act. A distinction that can't be tested for is not a distinction. Science is concerned with results not vague and ill defined assertions.
Would it be able to invent itself?
I'm not moving the goalposts here. Language models are extremely impressive. I think they will change the world. I fear they will displace many jobs.
Maybe some day they will be able to invent themselves. Maybe that will come sooner than we think. As it stands, they couldn't. This is the distinction that you say doesn't exist.
Next, you're conflating different things. How many humans are capable of inventing anything at all ? Have you invented anything in your field of expertise? Should i then question everything you do understand bout your field if you haven't ? Do you think understanding is binary ?
Most humans are not capable of inventing things, you're right, but humans in aggregate are. We are comparing LLMs to "humans" here, not to "a human". There are no examples of LLMs doing this. Maybe in a couple of years, who knows.
https://www.researchgate.net/publication/321596989_Abduction...
Don't let the Researchgate link discourage you. It will pretend to ask you to login but it will open the pdf anyway.
Why am I posting this? Because all you guys keep talking about "reasoning" this and "reasoning" that, and I don't think that word means, what you think it means.
Reasoning is a very specific thing and we know very well how it works. Can ChatGPT reason? I'm afraid you'll just have to read up and figure it out yourself.
Books don't exhibit those either.
In this future, what's the point in trying hard when a LLM or a person with a LLM will be just as equivalent without most of the effort? Seems like we're creating a future where humans are taught to delegate as much as possible to some LLM/service and go back to whatever entertains them.
With LLMs and more advanced AI, we are now outsourcing our thinking and reasoning. The implications are staggering for the future generations. We can still think and reason without assistance, but what is to come in the future for those that grow up with some assistant that manages everything in your life including your thoughts?
Why learn arithmetic when we have computers in our pockets?
Realizing your full potential as a human will never be a waste of time.
I'm not on the "people must work to be happy" train, but I do think they must feel needed to be happy. I don't think most folks are fulfilled playing violin in an empty room for their entire lives except when they can convince some close friend or relative to indulge them for fifteen minutes, or reading all the classics but having no-one to talk about them with (because AI is better at chatting about the classics, too), or writing books that nobody wants to read because AI does it better.
We're about to feel needed for a hell of a lot less than we did before, which level-of-social-need was already much-reduced from its peak. I fully expect a net decrease in overall happiness & mental health, from these developments.
> Who gathers around the piano to sing along with your one cousin who can play all the standards, anymore?
For the past 6 months or so, the lifegroup we host (mid-week discussion group associated with our church) does exactly that as part of our evenings - we typically sing through three hymns, with piano, on our path between dinner, discussion, and an open ended firepit time that often goes well into the late hours. And let me tell you, people enjoy belting out the standards (my wife generally picks one well known, one lesser known, and one "You've probably never seen this one before!"), scraping the dust off how to read multipart music, etc!
If the direction of tech is to dehumanize us and remove all that humans enjoy, then screw that! Do what humans enjoy instead!
Over the population, statistics is king, though. We don't, as individual actors, will our way into a better tomorrow when our environment changes to make that harder—not in large numbers. "So just ignore all those circumstances and pressures and do it!" does not work, in general. Larger forces and trends dominate.
I work in tech (deep weeds), and most people I interact with are quite aware that I do not like what we've done with it, and that I try to keep its influence on my personal life down - with plenty of handy ways on hand to suggest people try out, or solutions to "But how can you possibly XYZ?"
Having grown up before smartphones were a thing, I do recall how we used to do things, and most of those methods still work.
There is no point of getting good at arithmetic today, indeed.
I agree people need to continue study to understand the world (if they are to survive at all), but the current school curriculum is a joke.
I don't think trying hard is a waste of time right now. There's a lot of hype right now and ore will come but LLMs will fall to the hype cycle soon enough as we begin to discover all the little details and things that are not possible.
I think it's great everyone will be able to have a personal tutor and I think we can rethink some aspects of education. The Waldorf model really starts to make more and more sense since kids can explore curiosities with their AI tutor which can be programmed to nudge them into the "right things". And of course analyze and report on them daily to the parents.
> Seems like we're creating a future where humans are taught to delegate as much as possible to some LLM/service and go back to whatever entertains them.
Or interests us. I think it's great if we can offload some of the mundane and repetitive things to a machine so we can focus on more interesting things.
LLM might be used to find relevant information quicker, but they won't think for you, so people with LLMs won't suddenly have better critical thinking skills. They'll just make the same poor decisions faster.
Which is what happened with the Internet. We thought access to knowledge was going to enlighten all the dummies. It didn't, because no matter how much knowledge you present to a person, they have to also be willing to learn it, _think_ about it and apply it. That's the hard part.
There are plenty of problems that a Neural Net alone can’t solve.
On a more serious note, we have the internet and wikipedia now, we have ipads in classrooms. They all have a role, and we have to incorporate them in sensible ways. I'm sure that these AI tools will have some use, but these language models don't know anything. The examples of them making up information, or being persuaded by people are all over the place. I think it's great to use AI in teaching, but how about we learn from history, the technology we bring in to the classroom has to be well understood and directed in order to be effective. And just as importantly we need to understand that in many cases a focus on technological solutions distract from societal solutions. We can spend billions on deploying AI all over Africa, or we pay for some teachers and ship them some second hand textbooks, which one should we do, and how will that decision be made? Because I fear the decision will be made by some SBF-alike, more interested in their own journey than effective solutions.
Because it's definitely not easier with 2 lines of code. Not at all.
I mean, you could say that about human teachers too. I don't think they need to be perfect to be useful. In the end, as Peter Thiel opined in his book, the future is humans and robots working together on solving bigger and bigger problems. A teacher working with AIs that tutor kids and then managing that relationship and the relationship with the parents, etc. while also being the adult role model that kids will always need will remain.
I'm imagining the infinite amount of weird model tuning that'd be necessary to ensure only some students learn about white flight depending on what state, county, or school district they're in.
Once we have consumer level AR, our personal AI could take 'physical' (really just visual) form that follows us around.
These days, though, I'm beginning to wonder....
A better comparison would be literally any sci-fi book where everyone has a personal AI. Spider Robinson's Stardancer trilogy is the first one that comes to mind.
Kids being kids are going to try really hard to get them to go off the rails.
Nope:
https://en.wikipedia.org/wiki/Think_of_the_children
Kids that want to learn or have parents that enforce standards still excel. There are things a school can do to also enforce standards, but those things are not being done. The opposite appears to be happening.
So instead of going after the things we know work—either because it's outside the scope of what a school can address, or because it's far more expensive than we're ever going to pay—what money we do spend just chases a bunch of unproven and unlikely-to-succeed garbage, basically.
Schools are like the patient who needs a heart transplant, but can only afford healing crystals.
Yep, but increasing the number of teachers 2-3x is indeed politically impossible. Guessing that'd mean like 30-60% higher spending on education (you'd also have to increase wages to attract enough qualified workers, in addition to having way more teachers on payroll, and it'd be a years-long process), which ain't happening.
we're still in the early days of LLMs, this is like making a statement on the potential of social networks in the Usenet era. My guess is we'll see specialized tools of all kinds with LLMs integrated in some fashion, it won't be just giving a kid access to a general LLM model
there are already companies that are able to eliminate LLM hallucinations by giving it access to ground truth facts it needs to conform its answers to. Similar things will be done for education type stuff and other use cases as well
FWIW, we didn't try to use Usenet (or social media) to replace teachers. And we still don't today, even though they've matured (for... some meaning of the word) significantly.
Sounds just like humans to me.
Primary school, some people showed up to explain blood donation, showed and told us they were putting a bag of fake blood up the teacher's sleeve and would take from that rather than from our actual teacher, and yet when the fake blood went into the collection thingie the kid next to me screamed "they're killing her!"
Same school, show-and-tell, astronomy textbook showing the circular orbit of a moon at two different points of the parent planet's orbit around the star. "This is a picture of a mouse".
Secondary school. Main teacher and replacement teacher cover the same passage of a sort story, one interpreting the word "patronised" as in "patron" the other as in talking down to people whose intelligence you underestimate.
A-levels. Philosophy. What even is knowledge anyway? Tripartite definition. Example given was "shaggy" dog, one of the other students stifles a laugh at the double entendre.
But to your core point:
> the technology we bring in to the classroom has to be well understood and directed in order to be effective
Yes, absolutely.
But it's deeper than that. When this tech is flawless, we will have to ask ourselves: "What is school for?"
After all, being flawless in this context means it already knows all that can be taught, so why not just have it do the work related to those things instead of putting humans through ever longer school years so they can do the same things less well?
This isn't the first time technology has forced such questions; my generation never had to learn how to starch clothes, shoe horses, look after a tinder box, or send and receive Morse code.
If an AI is wrong about one thing you'll have the same bias over entire generations, once a full generation is skipped it's very hard to go back
That's why we need diversity, political pluralism, &c. because not many things are inherently "true", it's all about nuance, personal perception, mitigation, ...
> "What is school for?"
Socialisation, experiencing new things, general culture, deciding where you want to go in life, &c. Many things you learn in school aren't used for jobs but for day to day life and for the well being of society
If we take that out and replace it by "kids in front of screens" we're in for a massive and quick change with potential deep and long lasting effects.
By the way keep in mind a shit ton of jobs aren't even close to being automatised, and a shit ton of jobs aren't done by people sitting in front of screens
Sounds like intellectual monoculture? If so, I absolutely agree. Already a problem with respect to many thinking it's unpatriotic to suggest their country is anything other than "the good guys", or blasphemous to ask if their religion is anything less than the actual truth of the divine, and AI can extend that to every single topic.
(Oh hey, another possible Great Filter for the Fermi paradox).
> Socialisation, experiencing new things, general culture, deciding where you want to go in life, &c. Many things you learn in school aren't used for jobs but for day to day life and for the well being of society
Do we need schools for that, or will kids learn the same by messing around in multiplayer Minecraft or whatever the zeitgeist is?
> If we take that out and replace it by "kids in front of screens" we're in for a massive and quick change with potential deep and long lasting effects.
Sure, but that's going to happen regardless. Too many other things are changing also.
So why make it worse? Why gives kids tablets for things we know don't need tablets, and where a device is more of a distraction?
Shiny new technology is great for distracting from a failure to perform basic childcare like keeping children physically safe so they can learn. And then it's a rhetorical weapon "We can't listen to parents, they don't even want their children to have state of the art equipment!"
Second, the parents are mostly buying the tablets anyway.
Thirdly, about this:
> distracting from a failure to perform basic childcare like keeping children physically safe so they can learn
Pretty much the one kind of safety I can guarantee kids have when using a computer is their physically safety.
And learn what, exactly, given the hypothetical scenario you're replying to here is that the only thing left to learn is being sociable which they can ultimately do best by experiencing whatever medium they will be sociable in as adults, and for all other learning you're just being a nerd like me who learns for the fun of it.
> Pretty much the one kind of safety I can guarantee kids have when using a computer is their physically safety.
I think you're remembering computer lab, or something, where there was a proctor because computers were expensive. Now schools won't expel anyone they label troubled, no matter how much they trouble other students, and they won't protect anyone or let them protect themselves.
> the only thing left to learn is being sociable which they can ultimately do best by experiencing whatever medium they will be sociable in as adults
I don't see social media being so complex that they need all the years of practice to keep up, but I do see forming deep relationships as being essential and social media doesn't do anything to help with that. They'll get there when they do, no worse for not having partaken.
Funny, I'm remembering a summer holiday job 20 years ago, working on a HVAC production line. One of the others there bought a plasma TV for his infant.
It's not all about what teachers are pushing.
Heck, from what I've heard the teachers don't like the tech any more than you do.
> I think you're remembering computer lab, or something, where there was a proctor because computers were expensive. Now schools won't expel anyone they label troubled, no matter how much they trouble other students, and they won't protect anyone or let them protect themselves.
I have literally no idea what this is supposed to be about. Computers themselves are not part of any of the things you just said are bad.
And as I'm asking if the schools are even going to still be relevant, expulsions would carry as little meaning as being banned from owning horses does today.
> I don't see social media being so complex that they need all the years of practice to keep up, but I do see forming deep relationships as being essential and social media doesn't do anything to help with that. They'll get there when they do, no worse for not having partaken.
I didn't specifically say social media, I said "medium they will be sociable in as adults". That might be the traditional forums of the 2005 interwebs style; or soc med; or group chat apps like IRC, Slack, MS Teams as various of my employers have had us using; or video group chats like many of us got used to during the pandemic; or VR environments, be they games like WoW or roleplay environment like Furcadia or Second Life or VRChat, or "serious" metaverse things like whatever it is FB is trying to do.
It can also be going to parks and gyms, art galleries and graffiti walls, cafés and nightclubs. How much of that is part of school anyway?
But even for text, what counts as polite, friendly, professional, intelligent, serious, or severe is absolutely a moving target, so…
that means knowin when they need to be all like yo and emoji wif spellin' that confuses ol fogies like me and missin all the full stops coz that be rude yo yo lol
And of course the fact that such cultural idioms shift arbitrarily and I've stopped caring means that actual real young people will cringe as much when reading that last paragraph as I did when writing it, in exactly the same way I told my mum to stop trying to be hip and with it back in the late 90s.
The school district found the ones who'd push it. If teachers aren't fond of this I hope they fight it themselves.
> It's not all about what teachers are pushing.
Sure, I bought my child a laptop. But I spent time teaching computer skills under the guise of game design, etc, and it wasn't for classwork but for everything else. It wasn't a one-size-fits-nobody cash grab.
> I'm asking if the schools are even going to still be relevant, expulsions would carry as little meaning as being banned from owning horses does today.
I don't care if the violent kid thinks he wins, as long as he's not allowed to spend more time with the non-violent kids. His failure isn't my problem, my kids happiness and success is.
> what counts as polite, friendly, professional, intelligent, serious, or severe is absolutely a moving target,
Not so much, I think, that it needs practice over the years. I've seen people pick up social media with no practice whatsoever and right away seem to fit right in.
> I have literally no idea what this is supposed to be about. Computers themselves are not part of any of the things you just said are bad.
Yeah, it seemed to be a bit of an unguided post. I'm not sure if you don't like that I don't like excess tech in schools, or what. I don't think computers are bad, I think teachers and schools pushing them uselessly is bad.
Whichever side or instance you care to consider, think of the capabilities of someone hell bent for leather on defining what is, or is not, included in an educational curriculum, at the primary, secondary, post-secondary, graduate, or professional degree level, and their ability to influence, overtly or covertly, through persuasion or coercion, what that is.
Even under present conditions of proscribed curricula and textbooks, individual instructors, libraries, librarians, and others can go against some imposed dogma to some extent. A world in which only a small handful of AI originators provide the direct interfact between each student and their education is a grim prospect.
I've noted several commentators in recent months discussion a similar situation in terms of popular culture and entertainment, most notably the influence China now exerts over cinema and sport. The notion that AI will continue to be largely a product of Silicon Valley, and subject to US or even generally Western influence, may not be a valid one.
(I don't think that fundamentally changes any of your actual arguments here, it just feels like we're talking about slightly different aspects of how this can go wrong).
Both share a common root, however. And there's reason for concern.
You don't read facial expressions, body language, use your body, exchange bacteria/viruses, have body contact, &c. in minecraft
You can't go against hundred millions years of evolution and expect minecraft to replace real life. (see loneliness epidemic, obesity epidemic, depression epidemic), we're smartish apes, not AIs or computers. We're not designed to seat all day, we're not design to interacts with screens
If you can entirely replace education you don't need _school_ per day but you need long term group socialization with people who have different and similar opinions and backgrounds, physical activity, mental entertainment and challenges
The end goal is to live in healthy and collaborative societies, not to be mienscraft basement dwellers
I mean, VRChat can do the facial expressions, video games don't prevent exercise, and there's plenty of ways to share germs if you find vaccines too clinical and high-tech.
And of course, I'm currently using my glowing rectangle to group socialise with with people who have different and similar opinions and backgrounds, for example yourself.
Only for very specific (and IMO useless) definitions of the word "know". Does it matter if the thing a submarine does is called "swimming" or not?
In any practical sense, even obsolete ancestors of this model definitely "know" more French than I learned in 5 years of mandatory school lessons.
That said, regardless of the definition of the word "know"…
> The youth who are critical of capitalism certainly have turned a wary eye on the school system -- after all primary education has been designed to create money makers for corporations.
I don't know what you're talking about here.
The soviet bloc countries were very proud of their schooling system, rightly or wrongly; and conversely corporations don't really get much out of primary education, which is often described as being essentially no better than daycare for slightly older kids.
Alignment failure is still a problem when it's organic brains doing it, regard of what anyone tries to hand-wave with the cliché "common sense". :)
I love when tech people reveal that they have no idea how other humans act and behave.
A middle aged woman who said I had to be lying about being vegetarian because she couldn't believe it was possible for humans.
A coworker who insisted it was impossible to speed up some code in a daily standup, for me to come back to on the next stand up and announce I had sped it up from 15 minutes to 0.2 seconds.
Basically every time I hear the cliché "common sense", the claim being made is somewhere between totally wrong or just misleading in various situations.
My mum was a firm believer in homeopathy, Bach flower remedies, magic of crystals, called herself a Catholic, and had a statue of (IIRC) Vishnu with her collection of books on how to tell your fortune with viking rune stones.
At one point her dislike of "chemicals" came up, so my brother and I riffed on the dangers of dihydrogen monoxide for a bit; she was engaged with the list of issues until we told her what that is, and never took on board the lesson that "one atom away from being bleach" means "is not bleach".
I was exactly the right age for Wakefield's claims about the MMR vaccine to convince me not to get it, and I only got around to doing the sensible thing several years after he was struck off the UK medical register and (if I understand right) downgraded from Doctor to Mr.
Compare what the British public as a whole say about Boris Johnson with what specifically the members of the Conservative Party say about him.
One of my friends in middle school was completely convinced that all gay men hospitalised each other from every sexual act.
I vastly overestimated my German language skills when I first got here, and my self-estimate of my skill level hasn't changed no matter how much better I get.
My brother was convinced that Captain America was the first MCU film to be released.
My father took several years to realise Google search results had a scroll bar and it wasn't just the top 3 that fitted onto his 640x480 screen. (He became a software developer almost as early as was possible, so that wasn't merely a case "old person can't use computer", unlike my gran who held the mouse at 90 degrees and expected it to still move as if she hadn't).
He also believed that water drained in particular directions due to the Coriolis effect, thanks to (what he didn't realise was) a magic trick performed somewhere near the equator some time when he visited Kenya.
Someone who got on a train from somewhere in London to Liverpool Lime St instead of to Liverpool Street station, and who only dared ask me when we'd arrive as we were half way through the country.
There was a young-Earth-creationist fundamentalist Baptist who I wasted half an hour arguing with on his microphone. Too many mistakes to count, and not just the mistake I made by wasting time engaging with his nonsense.
Every exam question I ever get wrong, every personality I misjudge, every scam I fall for, every troll I am baited by… even a few April Fool's jokes I didn't realise were jokes until years later.
I know the LLMs aren't human, but that specific flaw of "can be wrong without realising it, can be convinced by nonsense"? That flaw is very human.
It can't cringe over its mistakes like we can, so at least we've still got that.
Hehe, had to re-read this twice because it hits a little close to home. I ditched maths for Iranian Studies and ended up studying Avesta. I work in tech, but still remember Ashem Vohu by heart.
https://en.wikipedia.org/wiki/Ashem_Vohu
Avestan (IIRC, my memory is fuzzy) had some weird similarities to my native language—Polish. Occasionally I'd stumble upon a phrase I could actually understand.
Random example I just remembered: "both ears" (again IIRC) sounded like uba ushi /uba uši/ (uba being 2 in dual, ushi meaning ears). Polish used to have dual number, but now retained it mostly for some body parts ( oba == 2). Ears in PL is uszy /ushee/. You'd sound a bit weird saying saying "oba uszy" in Polish, but people would understand you.
Nowadays I just enjoy the fact that Ashem Vohu is one of the oldest phrases I can utter in its original language. (I'm really good at learning and forgetting languages it seems.)
Idk about teaching kids. At adult, or near adult levels... I think these programs may very quickly be a major learning technology. It's good, sometimes, to have optimism. Optimism, even naive optimism, is often conducive to inventing things.
The guardrail should be in avoiding utopic ideas, like designing things for schoolchildren in remote places^. It's too abstract. Too aloof to be a testbed. TED talking windbags have had their day.
That still leaves plenty of room. A program that can produce exercises and lessons on the fly might be very useful. Thinking big is helpful to discovering what that use is.
The Hegelian pendulum of 2023 is hype driven. Hype, anti-hype. Fanboys. Haters. Ping pong.
The disillusionment of digital utopians, as they watched their creations serve Zuck, SBF, Thiel... dropping out in a cloud of ominous pessimism is no solution either.
I concur and I think in terms of children the last thing they need is more (any) technology. They need a solid foundation rooted in human interaction so that they can then leverage these potentially potent tools in effective ways later.
I find it worrying that I see so many techies not understanding this. Especially, because many bigger players in the area benefit from this lack of knowledge. This means that people who don't have time to procrastinate on the orange website are even less likely to understand the implications.
LLMs are excellent plausible bullshit generators, which has many uses, but being factually correct is not their strength. The issue is that even if the factual mistakes are rare, they're really hard to spot. I'm looking for a good, terse, catchy metaphor I could use to explain that to some of my family members and friends.
> We can spend billions on deploying AI all over Africa, or we pay for [...]
Looking back at the conversations I had with founders even I pre ChatGPT bonanza, I'm 100% sure someone's already pitching that idea to someone. I wouldn't be surprised if that conversation started in a marketing department of an ad tech business.
To me the bigger question is how far can we take them? Ie i get that they just pick likely next words. They're autocomplete on steroids. Yet that simple primitive is shockingly good. Way, way way beyond that i would have predicted is possible for an "autocomplete".
So where is the limit of this autocomplete? Would it be possible to eliminate ~90% of the current errors? Fake math, but would that put us at a more correct version than students typically graduate with? If so yea, it can be wrong at times - but assuming as don't abandon standardized testing/etc, it could still be an amazing tool for interactively learning something.
My hope is that even if it never goes beyond being an autocomplete; if we can improve the training dataset, help it not conflict with itself, etc - that maybe the autocomplete will be insanely useful.
They aren't errors. When the machine generates the text: "Next you'll tell me 2 + 2 = 5" there isn't an error in math, because no math was attempted. It's a success in predictable wording structure.
In fact, that example alone is a super success since we english speaking humans can get a broader context from that statement: You're going to tell me something nonsensical.
I don't get how you say the first part of your comment and then follow up with completely misunderstanding in the second half. Is GPT that enthralling to our brains?
Which is the meat of what we're discussing, right? We steer the auto completion into favorable results. We fine tune how it predicts. All of this is with the goal of a higher success rate and lower failure rate given some conditions.
No?
Both "2 + 2 = 4" and "2 + 2 = 5" are correct statements following the words "I don't think you understand [token]". There are no errors to fix.
Likewise, i cranked the Temp up for experimentation, and it produced gibberish. The randomizing aspect of which token is chosen after the next was so heavy handed that it literally couldn't even create words. This wouldn't be considered "good" by most people, would it? Would it for you? Technically it's one token after the next, but it's objectively worse at auto-completion than my phone's keyboard.
My point is i think you can have both. Dumb autocomplete, but trained well enough that it's useful. That it doesn't try to say Obama is a Cat. Or that 2+2=5. Yes, there will be edge cases where the weights produce odd output - but that's the entire goal of the LLM research, right? To see how far we can steer a dumb autocomplete into usefulness.
If you're argument is that "that gibberish output is perfect, no errors found" because the program technically ran, and the weights worked.. well, i've got no reply. I can only assume you're referring to the foundational LLM, and i'm referring to the training moreso - a more holistic sense of "is the program working". But if you consider the gibberish working, then frankly most bad programs (crashing/etc) would be "working" - right? Because they're doing exactly what they're programmed to.
Working, or lack of errors, seem to be a semantic, human interpretation. But i'm quite in the weeds, heh.
You could just read the research about LLMs before saying stuff like this.
You can't even seem to grasp that "Obama is a cat" as a statement isn't gibberish. I'm not even trying to convince you that these programs are perfect, I'm just trying to make sure that you understand that these aren't categorical errors and the things you consider successes aren't even happening.
Yea, we're just talking past each other. I believe i understand what you're saying. I on the otherhand, am describing errors in UX.
Your point seems pedantic, tbh. Hopefully by now i've expressed something in the way of convincing you that: for the little i do "know" about these, admittedly not much, it is that they're nothing but pattern predictors. Token outputs based on token inputs. No intelligence. Yet you spend repeated replies which sound effectively like "Stop calling them errors!" when they are very clearly errors in the context of UX.
Your argument, if i understand correctly, is pointless because the goal of the app is to have the LLM prediction be aligned with a human-centric world view. Up is not down, and the LLM should not predict tokens that espouses that. In that context, the LLM replying "Up is indeed down" would be an error. Yet repeatedly you argue that it's not an error.
In my view your argument would be better spent saying "The LLM application as we strive for today is impossible. It will never be. It's snake oil. LLMs will never be reasonably and consistently correct by human interpretation"
I don't know if that's your view or not. But it's at least not talking past me, about a point i'm not even talking about. My frame for this conversation was if we can make token prediction aligned with human goals of accuracy. You saying inaccuracies are not errors "categorically" isn't in line with my original question as i see at least. It's apples and oranges.
I stand by everything i said.
> My hope is that even if it never goes beyond being an autocomplete; if we can improve the training dataset, help it not conflict with itself, etc - that maybe the autocomplete will be insanely useful.
I stand by my first post's summary. Which is "never going past an autocomplete".
You're pedantic, and struggling to move past the fact that something can be both token prediction and still have successes and failures in the user perception. Inaccuracies.
How you write software with such a mindset is beyond me.
I struggle to see how even the current GPT-4 is any worse than your average human.
How this is dismissed because it’s not 100% perfect (might at add, “yet”) is beyond me.
Now, people seem to miss or ignore that fact, and I think is a very risky path.
I'd say that ca. 8/10 founders who pitched to me an idea leveraging LLMs completely missed that limitation. An example would be something like using LLMs as a replacement for therapy.
> How this is dismissed because it’s not 100% perfect (might at add, “yet”) is beyond me.
Again, I'm not dismissing them, but the current tech behind GPT or LLaMA has no concept of "correctness". These models don't understand what they're saying and this is not a trivial issue to fix.
> I struggle to see how even the current GPT-4 is any worse than your average human.
Where, how, and what do you mean by worse? I'm pretty sure there are cases where I'd agree with you, but this is a very broad statement.
In addition to techies believing LLMs ar emore than bullshit generators they also appear to think humans are incredibly simple machines.
I'm so glad the tech industry is given so much money and power over the rest of us without actually being forced to understand us.
https://news.ycombinator.com/item?id=34850890
The logical errors in other subjects will be fixed over time. Education is an easy domain because it is artificial by nature.
I can see a lot of people switching to ai homeschooling in a few years
Define “knowing”, and then say why it even matters if an LLM gives a great illusion of knowing.
Are there any real gains in giving kids ipads in classrooms? Did anyone have an actual plan in how such a tool could help kids or is it just “tech good?”
I don't know if this is "better" or what you had in mind as "gains", but it's certainly entertaining for me!
I have three kids, and as far as devices used in education is concerned I become more of a Luddite with each passing day. I have yet to come across any actual evidence that shows children experience a net benefit from additional screen time.
The only exception is our 13 year-old, who currently is recovering from a dislocated and fractured right thumb due to being on the wrong end of a vicious tackle in school sport. He's still unable to hold a pen but found his iPad allows him to do various bits of schoolwork. This - fortunately - is the exception rather than the rule.
Instead, Wikipedia was by far the most important part of my high school education. Nearly everything I learned in my second two years of high school I got from Wikipedia.
I think most people today would agree that Wikipedia is an invaluable part of early education and that it's done far more harm than good. All the fear-based arguments you are making applied equally well to Wikipedia in 2005, and I suspect that in less than 18 years we'll similarly recognize that LLMs or other intelligent language systems are invaluable parts of education.
Wow, and please understand this is not personal, but it must be one of the saddest things I read this decade.
Instead we'll increase class sizes and keep pay flat.
I'd wondered how we were going to deal with teaching becoming less appealing year over year, since it didn't seem likely "spend quite a bit more money" was going to fly, and it looks like this may be it. Rich kids will keep going to private schools with class sizes of 6-15 and real teachers, and public school kids are gonna get AI teachers with one real teacher per hundred or so, and bunch of even-lower-paid monitors (or whatever they end up calling them) to maintain order.
How the ruling class treats those people it presently deems unnecessary and unproductive is going to set the precedent for AI to follow when it decides the same of them or their progeny. If the current ruling class wants to ensure a good future for themselves, it's incumbent on them to set a desirable precedent or accept that their lot in existence is to simply birth an AI and then go quietly into the good night after it takes over and abandons them to starvation.
Personally, I think AGI is the the current-era version of the personal atomic-powered autogyro.
Some will go to the AI companies who created the value. Some towns will take the surplus and spend it on even better educational outcomes. Others may spend it on other town needs. Others may refund it to the town citizens in lower (or slower-growing) taxes. It’s not at all obvious that it should all go to teachers.
We're spending so much more on education than the past. Teacher pay has fallen further behind other professions, and working conditions have gotten substantially worse, overall. Expecting tech to be a panacea and get us out of this mess is unrealistic. And there's a reasonable fear that promises of tech may be a justification to make the mess even worse.
On the other hand, children who have individual, mastery-based tutoring do about two sigma above those who don't. Of course, there's no way we can pay for 1:1 for every student.
What Khan is thinking of with Khanmigo is a way maybe to close that gap.
This doesn't seem like an "of course" thing to me. University office-hours and the TA system work pretty well for this kind of thing.
Speaking as a former TA, there’s only so much instruction you can provide and while you can focus on helping a particular student more to ensure they get the help they need, that’s usually to the detriment of the other students you’re supporting.
https://en.wikipedia.org/wiki/Bloom%27s_2_sigma_problem
Instead, I'm talking about a tutorial system where a student receives the vast majority of their education from dedicated time with an individual instructor.
I actually think a good online learning platform could be comparable to the private schools with 6-15 class sizes. Because it's designed by teachers even better than those private-school tutors, and so much of teaching is largely the same that you will get "personalized" learning just be designing a massive learning module with many branches into various extra concepts and avenues for those who "don't get it". And with larger class sizes and less teachers, we can instead spend education money on things like making schools actually safe and providing decent lunches, and fixing decades-old infrastructure.
Why? Do you mean that it is literally infeasible or that the current political climate makes it difficult?
Surely education is the single best investment a country can make, but it doesn't easily make a few people rich.
Current Los Angeles teachers' wage penalty is 25% vs neighbors with same college degrees. [1]
One only need look at the current starvation in teacher training programs' rosters to see the coming implosion of supply of new teachers.
Paper with methodology and citations: https://utla.net/app/uploads/2022/08/UTLA_ShortageReport.pdf
Wages are falling really far behind even the kinds of careers HN types joke about, the ones that humanities and liberal arts majors end up in if they don't just work at Starbucks—and guess what sorts of jobs the top 20% or so of teachers with some years of experience can walk straight into? Yep, exactly those jobs. Hell, the comp at one gas station chain around here is on par with a mid-career teacher in the area, after you've been there a couple years, and you don't need a degree to get that job. If you're an assistant manager by five years in (and if you're bright enough to be someone we want teaching, you will be) you'll be out-earning local teachers with more experience than that. W. T. F.
So, good for the upcoming crop of students, they shouldn't get a teaching degree, it's an astonishingly horrible deal and unlikely to get better.
Where in the paper is this stated? As far as I can tell, they treat each "bachelor’s degree-holding workers" as equivalent to each other.
There's no reason to expect that someone with a bachelor's degree in computer science from UCLA would earn the same as someone with a bachelor's degree in child development from Cal State LA.
And do the wage comparisons compare total compensation (including the present value of future pensions), or just current salaries? In my local school district, about 30% of total compensation is pensions (i.e. you need to add 43% to base salary to calculate total cash compensation). Also, teachers don't work for the full year; many earn additional income during the summer break.
The term is "paraprofessionals." When I was a kid, these were the people who supervised the kids playing at recess.
If history is any guide, the concept of an "AI tutor" won't be used to "[give] every child their own personal Aristotle," it will be used to cut costs by replacing human teachers for low-status students with poor substitutes.
Seriously, think about how much money can be saved from firing most teachers and locking poor kids in a cubicle farm to stare at an insane "AI tutor's" screen all day.
And what's the strongest message sent by replacing human interactions with a machines?: "We don't care enough about you to actually bother to interact with you." Think of how you feel when you have a problem, you make a call to fix it, and all you get are phone menus.
And look at all the examples in the article: use GPT to generate crappy stories and educational games for kids to play. How can anyone think that's better than using a book? At least the book's stories and games can be tested and evaluated before use.
The actual problem here is the sociopolitical one of unwillingness to allocate enough resources to education to make it truly effective and satisfying, and that's one that technology can't solve.
MOOCs have terrible performance in comparison with even huge audience lectures. Online teaching during COVID proved that that doesn't work either. A machine doesn't have the authority and conviction of a real person. The combination means that the levels will have to drop only to sustain the idea that this could work.
If some country is going to push this, the sunken cost fallacy and the impossibility to hire new teachers will finish education off, once and for all.
Optimism should NEVER be a basis for fucking with something as important as school.
But they sure didn't live up to the hype. They didn't democratize education in any meaningful way. And I'm not sure even the people with undergrad and grad degrees who made up the majority of MOOC students use them much these days. I know, for me, they were mostly a novelty for a while. Maybe if I want to dip into a topic I'll watch some online video but I haven't even signed up for a MOOC for years at this point.
Naturally, some analysts would say that means there is tons of consulting money to be made..and it seems like spending is easy and profitable.
most public school teachers are glorified babysitters at this point anyway, I can recall only a single teacher during my 12 year public school attendance that was actually helpful. From what I've seen from GPT-4 so far, I'd have taken a ChatGPT Plus subscription and laptop over dedicated teachers. YT and Khan Academy were far more useful than my actual teachers
2. I believe the problem is that the extra money is not going to proven-effective programs and teachers. There is some question to me whether anything is proven effective... in spite of whatever you read in books written for teachers.
3. The parent comment is pretty nihilistic. Try to propose something that you think might improve the status quo. (E.g., the comments about the highest impact of extra spending being in social services instead.)
Use LLMs to compliment class activities so that students can receive the individual attention they need.
That's a common but pretty twisted bit of logic: "we've made it shitty, so we should make it shittier because it's already so shitty."
No, he didn't outline a way to make it better. He denigrated the people involved, in a way that is (most charitably) an oblique reference to low education budgets (e.g. not enough money to hire enough teachers or attract talented teachers). Then he held up some references to a few techie things.
Maybe he's a genius autodidact that never really got much benefit from his teachers. There have always been people like that, but they've never been common. A system tailored to people like that would fail the vast majority (e.g. be shitty).
You basically have to teach young kids how to act like a human being. Kids will get this teaching in daycare and through their parents, but that alone isn't enough and it is important to reinforce this through schools.
People often forget this because you don't really remember it and then you end up making assumptions on how you act now - not realizing that a teacher may have taught you when you were younger.
I think the best example of this is handing out worksheets to kindergartners. An average person may think that you just give them pencils and paper. An experienced teacher knows that they have to teach the kids where to write their name, how to follow along the line, to read instructions, flip the page over to check for more, etc.
This is practically true for all children. Again, many may argue that you can receive this teaching through your parents and day-care, but once a child gets old enough they are spending most of their time in school. Think of it like language learning. You'll learn more of the language if you are being taught it all day, rather than only at home with your parents.
You also experience more interaction with kids your age at school. There are many situations that come up where a teacher guiding you is important, and these same situations may not come up at home (especially if you are a single child).
I do think after a certain point you can rely on technology more for learning. I just want to emphasize that having a human teacher trained for working with children is critically important for the growth of a child.
It stands to reason that the education system should also change to reflect the world that we live in. Screens, AI, all of it. Children will learn how to become members of the society that they were born into, not the one that we were born into.
Would it be possible to do all that's needed for young children with just a screen? I'm not so sure. The issue isn't necessarily the content, but the behavior of the child and having a person there to teach and correct behavior.
Technology plus teachers is effective. I apologize if my argument came across as no technology for young children - I'm just saying that having JUST technology for young children is difficult.
I know many people that are teachers of young children and they use technology. Kids have Chromebooks at first grade and are learning things like programming, using the internet, determining if an article on the internet may be lying to you, determining if a website's main goal is to sell something to you, becoming a good digital citizen, etc. All of that is very important!
It's also important to have a teacher in the room to teach students how to use their Chromebook, how to open it, how to physically interact with technology, how to ask each other for help, etc.
We spend an absolute shitload of money educating children. This is paid for by society. But it is also so incredibly worth it. I'm worried that politicians will see an opportunity to shed costs without considering the implications and we end up with children raised online by robots.
One of the best teachers my child ever had was limited to teaching first grade due to a lack of maths ability. Everything else about them was fantastic. Maybe AI teachers can help in cases like that.
The vast majority of what is taught in primary schools doesn't need a lot of subject matter skill, what it needs is the ability to teach and the toughness to survive the environment. And it has to pay well enough to attract people who have those qualities because those same qualities make those people attractive to other employers.
Yes.
One problem I come across repeatedly is: how do you identify whether someone has the technical competence to do the teaching? Is there any way of knowing if they're doing a good job?
As an IC they are implementing features and shipping products. As a teacher...? Test scores going up is uncorrelated with teaching quality, apparently. (NYC, just look at the scatterplot in [1].) So it seems like there's a deep problem here.
Now imagine someone in K-12 hiring a CS teacher. Who do they have that is competent to evaluate the skills of the person they want to hire? Probably nobody at all.
[1] https://garyrubinstein.wordpress.com/2012/02/28/analyzing-re...
Classroom observation. It's really not that complicated.
> There are plenty of people who can do the technical side of teaching the curriculum. Most of them can earn more elsewhere [...]
It's funded by a billionaire former mathematician who wanted to address this exact issue.
Yep, after reading a piece on ed-tech failures of the decade 2010-2020[0] here on HN[1], I've become pretty convinced that big-scale education is hard (if not impossible) to solve with IT might only.
Sure, one person alone can learn something through Wikipedia, Coursera, YouTube, Khan Academy or tutorials, but that can't replace structured environments if your missiob is to educate lots of people.
[0] http://hackeducation.com/2019/12/31/what-a-shitshow
[1] https://news.ycombinator.com/item?id=21948132
Or in the US, if teachers are that important, maybe give them more resources to do their job. i.e. buying supplies, computers, etc for your kids class.
"Thank you for calling customer support. We value your business. Please enjoy this...music...while you wait. Thank you for calling customer support." M U S I C ! ! !
I think LLMs are exciting of course, but this post is overly flowery and relies heavily on bashing the current education system without seriously analyzing any of the potential shortcomings of AI.
(I do realize this echos concerns from earlier generations about books, to video games and anything being bad for kids, but come on..)
But I believe technology is the future of education. Instead of having many, many classes of like ~10-50 students with underpaid/overworked/mistreated teachers, some of who are much better than others, we rely on services like Khan Academy. Not even services with an AI, just pre-planned learning modules, and then we can keep a human tutor for those who need extra for whatever reason (e.g. curiosity, more help).
Many people think tools like Khan Academy are the lazy and less effective alternate to real teaching, and maybe they are now because some of them really suck. But online modules have the potential to be much better than even a group of engaged, in-person tutors. Not because an AI tutor is better, but because we want a standard, universal curriculum; and one curriculum for ~1mil+ students can be designed by the highest caliber (human) teachers and have highly-interactive (human-made) demos powered by computer programming.
Also, no matter how good the AI tutors are, 100% online learning is not the solution Students from pre-K to graduate school rely on not just the education, but the social component as well. Plus, many school kids only get their biggest meal from free school lunches, and parents don't have the time to keep their kids at home. So we should still have school to be in-person component even if students do all of the learning on the computer, with an online option for those who want it, but I guarantee many won't
The lure of LLMs in education is the expected novelty of instruction they can provide. It might be an aid to some students, the same way some students read books in their free time or learn to use the internet to look up answers. I don't expect much more, and the rest of it is a net negative IMO.
LLMs will be used like "reverse multiple choice questions", with heavily constrained LLMs silently penalizing students for making "mistakes" or probing subjects that they aren't "supposed to" know about. No matter how embedded LLMs are in the education process, there'll always be a human setting its course, and I don't think social dynamics will ever change because of this.
Also, the Q&A format with a machine will make older texts / instruction opaque to many of the brightest, they'll have to translate from machinese for it to be of any use. An entire generation aping a LLMs way of writing is also a death knell for a civilization and innovation. We're clearly past the point where we can convince most people that this LLM isn't alive.
In other words, the future you believe in is one where we lock kids in a cubicle to stare at a screen all day in a warehouse, with a break for recess and lunch? Where the message is, we care so little that we can't even bother to allocate one real person to regularly talk to you?
These AIs will be direct fed into our brains soon enough. Who knows or cares that the entity you’re talking to is an AI if it’s indistinguishable from a real person?
I think the ideal is ~10:1 sorted by student performance/learning capacity increasingly for difficult subjects (so students form a cohesive classroom pace of subject material). Which will also give variance to friend groups if you sort students differently per subject.
AI tools will be great for public schools where its often 30-60:1. We still have to get past the 'confidently wrong' AI though.
bloom seemed to show you continue to get improvements all the way to 1:1
maybe the ideal is 1:3 or something
If theyre the right kind of friends you will be able to learn more with them than from a 1:1 teacher.
I don't want that.
There would still be the ability for students to choose their own subjects, focus more/less on certain subjects, learn different concepts within subjects, etc. And a lot of niche subjects and branches. Like it would be an incredibly massive curriculum, especially since you are designing one for >1million students. By "standard and universal" I mean that every student gets the same resources.
LLMs hooked together with search engines, fact repositories (Wikipedia, etc) and other agent tools (for math, theorem proving, code checking, etc) can provide not only a reliable answer, but also a research plan it followed to provide it, and details on the fact-checking, and then discuss it with you.
Look at langchain and similar tools.
After all, education is not just about imparting knowledge, but also about building relationships/community, providing emotional support, etc. We should focus on how this technology can be used to help teachers shore up their time. ( Which is why I am building emailing tooling for the HUMANs operating theses AI agents: https://tadabot.email / https://recurai.com )
I've built a tutor for my kids, that listens to them, and helps them explore topics that they're passionate about.
What's blown my mind is how it has driven creativity OFF screen too. They engage with the world in a deeper way when they have the power to explore their passions on their own.
If you want to try it, join the waitlist here: gotwonder.com and then email me: seth@gotwonder.com and I'll bump you up the list!
I do think that we may be a ways off from truly being able to just "drop our kids off at the tutor," but I think we can see that, from here.
1. Fine-tune propaganda into the LLM (e.g. evolution isn't real, or something spicier)
2. Chatbot is so humanlike and convincing, this is very effective.
3. Nobody can inspect the neural net weights to see this, so it flies under the radar.
4. Profit!
The scariest thing to me here is just the idea that someone would want to experiment with teaching their kid with an LLM in the first place, while we don't have any institutions or processes to really stop this (apparently appealing) Idea from happening, and that it is stupidly vulnerable in so many ways.
You can say laissez faire when it comes to rational adults and their use of LLMs, but it's not right to put the risk on kids.
If the LLM responses were geared more toward inducing the student to perform their own research, to encourage them to do more reading on some topic, or to try things out for themselves to learn through experience (and then carry on the discussion with the LLM afterward) then that might be even better.
Simply training young humans to interact with a chat bot which has all the answers is not a good path forward. Using it as a tool to foster human curiosity and growth, to challenge and inspire -- yes. To use it as a salve, as a babysitter, as a stand-in for someone who is too lazy to parent or teach -- no.
How would kids using an AI chatbot shift to performing their own research or reading? I imagine they would just ask more questions of the chatbot.
There should be some way to measure what percent of the information a child learns is from a human, and what percent is from a language model that says human-like things.
> "The lesson the One Laptop Per Child story offers donors is that flashy solutions are often not the best ones. Far better to give the poor things they need and want—including clean water and good schools—than invest in technology that may not work and may not help the poor advance in school and in life."
https://philanthropydaily.com/the-spectacular-failure-of-one...
I'd also say that a structured educational program isn't all bad, particularly when it comes to developing highly useful basic skills like mathematics, programming, writing, reading, etc. Such basic skills provide a foundation for future development and study, in the sciences, medicine, law, business, arts and so on, which can be more free-form.
Claims that ChatGPT can provide such a structured basic educational experience are questionable - it's more of an aid to people who already have well-developed basic skills (like say, knowing when to be skeptical and how to double-check the output).
I also think a lot of people on this site are probably very curious and self-directed. I also wonder if many have a sort of rose-colored hindsight view over how they'd have developed both academically and socially if someone tossed them a pile of books and said "Check back in a month and we'll review your progress."
For quick concrete example, wizzy professors and postdocs, who normally wouldn't invest in the creation of introductory content, were being attracted by the massiveness of the potential reach. Instead here we are, decades later, with US leading K-13 science textbooks you can easily win science-education-research list-of-common-misconceptions bingo on. NSF once almost created a national science education wiki... but didn't. Think of that scale, but bigger, of a "we almost, but no".
Not even close, you sound like Libertarian-adjacent technologists who have been believing your own hype / some other equivilent metaphor.
Technology has a role to play in education that is as-yet unrealised and while I cannot speak for all of them, I know the system in place in the UK is behind the curve. The solution to this isn't to pull your kids out and use them as the cornerstone of an individualised experiment but to fight for better funding of the educational system itself. I hope this does not have any negative consequences for the kids who are the recipients of this trial.
I'm not saying this will _never_ be good. But it feels pretty early to unleash this bastion of confidently incorrect answers upon developing minds.
The teachers and schools need more money that goes to teachers and student resources, not more iPads, smartboards, bullshit-artist AI, or overinflated admin salaries. Just call this what it is -- entertainment for technologists.
It's not even really that, because phrasing doesn't really capture how off-target and self-serving SV often is. It's more like "whatever we happen to be hyping right now is your solution."
> Just call this what it is -- entertainment for technologists.
That's spot on.