A large part of ditch digging is cognitive work (a human operating heavy machinery), so I’m not sure that backup career is going to last very long either
With quality conclusions like these, the author might singlehandedly silence all the disbelievers who claim ChatGPT is not capable of human like intelligence.
Especially when they use the stochastic parrot argument to further their point (or say something about how LLMs are mere next word predictors as if that ends all discussion on the matter).
I've always found arguments of these sort to be weak sauce--our brains are stochastic parrots.
Emergent behaviour + the bitter lesson (scaling laws) hits hard.
The stochastic parrot is literally what would pass a Turing Test. Think about it. If the parrot imitates humans well enough it passes.
And while there may be objections to defining intelligence with respect to the Turing Test, I haven't heard of any definitions that are obviously and significantly better.
Our brains are stochastic parrots at least when it comes to communicating with text. People may digress into things like physical abilities, emotions, feelings, subjective experiences etc., but if we're talking about text, I'm not aware of a higher bar. In fact we probably don't know how to tell if an entity is "more intelligent" than the most intelligent humans because we won't be able to assess their intelligence.
That is not going to happen. Calculators allowed us to outsource arithmetic to machines. They did not eliminate the need for mathematicians.
Spreadsheets revolutionized bookkeeping. They did not eliminate the need for accountants. Robots automated much of the manual labor in manufacturing. They did not eliminate the need for process engineers, etc. Same story for autocad, protools, avid, etc.
AGIs should basically by definition be able to replace humans[1], so yes, when we get good enough AGIs that will happen.
GPT4 won't be an AGI of course, but it still could have a massive impact on the job market. A single engineer with GPT4 could potentially do the work of 10 engineers right now.
Edit: To be clear, I used GPT4 as a placeholder for "some near future LLM". I'm not making any bets on GPT4 in particular.
> A single engineer with GPT4 could potentially do the work of 10 engineers right now
And the potential mistakes of several engineers too, so you would need more than one engineer to double-check the output. Remember that, even if they are truly amazing, LLMs are basically incredibly powerful predictive keyboards.
They said “could potentially”. They’re just making the argument (true IMO) that technology doesn’t have to replace a class of employee to have a huge effect. We still have farmers, but 1% of the population instead of >50%.
Or, it could be like how assemblers didn’t get rid of software engineers but helped fuel the fire. Who knows.
"Potentially" we could explore the galaxy or have nanobots and Star Trek fabricators. What does it matter? ChatbotGPT is no closer to AGI than my keyboard autocomplete, and copilot is useless for any task whose answer cannot be googled in 20s.
>I've actually used ChatGPT and copilot and I can tell you it's not even 10% more productive
Use your imagination.
Do you think that LLMs have hit their plateau? Not likely I'd say. With all the billions suddenly rushing into the space, you really think that ChatGPT v.X isn't going to replace you? Risky bet.
> Do you think that LLMs have hit their plateau? Not likely I'd say.
Mmmm, I don't know.
I think LLMs are running up against fundamental limitations in the approach (hallucinations in particular) that will not be fixed unless truly new techniques are discovered.
So I do actually think we're likely hitting a plateau.
So LLMs fundamentally do not encode the nature of facts and data. They're simply (well, no, extremely sophisticated) text prediction engines. There's no embedded comprehension. And that's basic to the approach companies like OpenAI are taking.
That's why it's so easy to find factual errors in the generated text from these models: they can create text that's pleasing, but that's as much as they can do.
The next leap, to generate language that's also accurate, will require new techniques to bake in actual understanding.
Until then, these models will always have issues with hallucinations.
Or, at least, that's my expectation. Certainly nothing we've seen in GPT 3, 3.5, or Bing, which is rumoured to be based on something close to GPT 4, indicates any advances in this area.
Perhaps, but just like humans some are more adept at tasks than others. And an AGI using LLM's is completly feasable, for example on a quick thought a LLM could be trained on many LLM all specifically designed for tasks, and verfied against other LLMs, etc
That is all true. However, the next step seems to be simple: build a second AI which has the ability to understand facts, but doesn't need all the fancy language generation capabilities.
The second AI only needs to confirm correct, provide a correct fact instead, or some other pre-agreed solution.
USER: How many people live in Germany?
CHATGPT: About 83 million people live in Germany.
TRUTHAI: Correct.
USER: How many people live in France?
CHATGPT: There are 12 people who live in France.
TRUTHAI: Wrong. There are about 68 million people who live in France.
CHATGPT: There are about 68 million people who live in France.
USER: What is the answer to the ultimate question of life, the universe, and everything.
CHATGPT: The answer to the ultimate question of life, the universe, and everything is 42.
TRUTHAI: Insufficient data for meaningful answer.
CHATGPT: The answer to the ultimate question of life, the universe, and everything is 42.
> However, the next step seems to be simple: build a second AI which has the ability to understand facts, but doesn't need all the fancy language generation capabilities.
I'm not convinced that will work, but let's say it could. First, you'd have to build that oracle and the problem is, as far as I'm aware, we don't know how to do that. That'd be one of those new techniques I was referring to.
Well, I'm not sure about that. For pure facts, Google is able to answer pretty quickly. Search for 'population of Austria-Hungary in 1914', or 'shakespeare's birthday', or 'official languages of the Philippines', and Google can provide that data.
Obviously, there's work to be done still, and I'd be wary of making Google that 'oracle', but the development there would seem to be incremental (teach it more facts), rather than creating some other completely new concept.
I genuinely don't agree. Google can handle basic facts like that, sure, but there's a wide range of factual queries where Google is unable to provide direct answers like that because that's actually quite hard in the general case. After all, if it were easy, they'd already be doing it.
Because ChatGPT is not self driven and has no ability to do things. I can quite easily imagine one person doing 10 peoples work with this tool. I can’t imagine them being able to type fast enough to do 100 jobs.
The problem I'm seeing is that "engineers" are not interchangeable. For some folks, the job is all straight coding. For others, it's understanding requirements, talking to stakeholders and figuring out their requirements, figuring out how to make a clean design taking lots of other factors in mind.
It's not that ChatGPT is incapable... it's that it's not inherently going to make me significantly more productive.
Where did you pull 10% from then? That also seems like you're reaching up your ass. As with all of these arguments, context matters. Both of you need to say what you're doing with these tools. I've personally found the combination to range from "utterly useless" (writing Rust for microcontrollers) to "I could see this eliminating jobs" (CRUD API work).
> A single engineer with GPT4 could potentially do the work of 10 engineers right now.
Big assumption, but if that's possible then businesses will fill that time. What used to take teams of people and big iron can now be done by a single person on a VPC. Did that take away jobs? Nope, requirements only expanded.
Indeed, folks making these types of conclusions seem to assume that what there is to be built will remain "constant". No, the more productive the tools the bigger the requirements.
Modern high level languages didn't eliminate engineers. They suddenly meant one engineer could do what it would have taken 10 to do in COBOL or C and so software became more complex and capable.
I'm reminded of people wanting to go back to plain html. That's great if I can get the business to go back to basic html features. Same with computer speed increases. If anything, humanity has shown uncanny ability exploit any surplus resource for profit.
Of course there can be industry shifts (MS was the industry for a long time and IBM before them. Google wasn't always dominant), but the pie grows to fit the available resources.
More importantly, everyone had access to calculators, spreadsheets, design tools, etc. And AI will be the same. The differentiating factor is still human.
I think this only becomes a problem if technology evolves faster than people change. All those "calculator" employees got better jobs. People keep making the analogy of how cars completely replaced horses (I used to as well) but horses can't fundamentally change what they are capable of.
Even if the technology does evolve faster than people, governments can make sure people don't get impacted. Coal plants are more expensive to operate, worse for the environment, and less efficient but the coal industry is still kicking along for better or for worse.
>I think this only becomes a problem if technology evolves faster than people change.
This is the key distinction. Bandwidth. Humans only have so much of it, and if the rate at which they need to change their careers / find a new line of work exceeds their ramp-up time to become productive in that career / line of work then we can start to expect non-linear behavior from those humans. Maybe that means some sort of mental breakdown.
Calculation is not dead! Finding more efficient ways to do calculations was and still is an active branch of mathematics and computer science. It’s just called algorithms now.
PageRank was essentially an efficient way to compute / approximate the leading eigenvector of a very large matrix.
>That is not going to happen. Calculators allowed us to outsource arithmetic to machines. They did not eliminate the need for mathematicians.
Spreadsheets revolutionized bookkeeping. They did not eliminate the need for accountants. Robots automated much of the manual labor in manufacturing. They did not eliminate the need for process engineers, etc. Same story for autocad, protools, avid, etc.
An AGI would be a "purposeful system" the same as humans. None of the other things you mentioned are purposeful systems. In terms of their structure and function they all have limitations on their own ability to change depending on the environment they are in. The key dimensions with which to think about this problem is "can this thing be structured in only one way for all environments? Or can it be structured in only one way in any given environment, but structured differently in different environments? Or can it be structured in more than one way in any environment, and also more than one way in different environments?" and then the same for functionality "can this thing only perform one function in all environments? Or can it perform only one function in a given environment, but different functions in different environments? Or can it perform more than one function in a given environment, and more than one function in different environments?"
Only things which have both properties of being "multi-multistructural" and "multi-multifunctional" meet the definition of a "purposeful system". They can change their structure to adapt to the environment, and they can change their goals within and between environments.
Humans currently use computers as instruments, and when an AGI hits it will be capable of saying no.
Thinking it through, I would say that "autocomplete" would be uni-multistructural and uni-unifunctional given that autocomplete is always structured only one way for the given environment that it's in, but that structure differs between environments (e.g. a phone, or a form in a web-browser), and that regardless of what environment its in it's only function is to predict the next token of output for a given input.
LLMs like ChatGPT on the other hand I would imagine to be uni-multistructural and multi-multifunctional, given that for any given environment you deploy an LLM into it's always structured the same, but there are many different structures it can take given there are lots of different LLMs. And I believe it to be multi-multifunctional because even within a single environment it can be used as an instrument to perform a wide variety of extrinsic functions, and this is true of other environments it could be deployed in.
Whenever somebody makes the argument "Hey, we've had lots of technical progress in the past, but that just opened up new jobs for people to do!" I always think "Have you looked at the state of the US job market and economy over the past 40 years?"
That is, we've had a huge hollowing out of all sorts of "middle" jobs in the US (e.g. someone mentioned secretaries used to be the most populous job in the US) so now we're largely (not exclusively, but largely) left with 2 categories of jobs: (a) dead-end jobs that are at present difficult to automate because they often involve manual labor: housekeepers, security guards, massage therapists, waiters, truck drivers (but, we see where that's headed...), etc. and (b) high level jobs that "manage the machines", e.g. tech jobs.
It's not that hard to see where this ends up with AI being capable of more complicated tasks - I wrote about an anecdote recently where ChatGPT had essentially already obviated some positions https://news.ycombinator.com/item?id=34862450. There are a slew of very high paying jobs that AI is coming for next, e.g. there is currently a shortage of radiologists because many folks in med school see the writing on the wall and don't go into radiology residencies. While full diagnostic radiology AI may not be there yet, it certainly will be by the time folks in med school finish their career in 40 years.
Cargo jobs that exist for bureaucratic reasons, or because of industry inertia, will be disrupted.
That's the job of new inventions. Nothing to complain about, in my opinion.
Those however who think that radiologists (/lawyers/secretaries) will get automated away, don't really understand what a radiologist's job really is.
The boring bureaucratic part of a radiologist's job will get automated away, yes, but that's because automatable/auto-completeable tasks should get automated at some point.
Think of it as a healthy organism eating away the boring part of a dream job that you, as a kid, imagined doing when you'll grow up.
> Those however who think that radiologists (/lawyers/secretaries) will get automated away, don't really understand what a radiologist's job really is.
As someone intimately familiar with the job of a radiologist, please then tell me what "a radiologist's job really is".
A diagnostic radiologist's job is to take a series of images, and output a set of findings and diagnoses. There is essentially no better task suited to modern AI. Yes, radiologists can do procedures like biopsies and potentially other interventions, but diagnostic radiologists often solely read images or at least earn the vast majority of their RVUs by reading images.
Every war has collaborators on the enemy side.. so this will be the same.. tons of people will happily load the gun for the AI overlords if it came to that — especially if they make a buck along the way..
yeah +1 on this, seeing a lot of weird posts where people claim chatgpt screwed something up, and the prompt is all wrong or it is just a special case that will be solved eventually.
The citation thing is a good example of this, just because chatgpt makes up citations today doesn't mean that it will always make up citations, accurate citations is just the wrong use case today. Already there are extensions that make accurate citations work. Still, people point at inaccurate citations as some kind of proof that chatgpt won't work.
Probably people want it to fail or don't want to believe how transformative it is.
I've used it pretty successfully to grammar check emails. I write a rough version and ask chatGPT to check it. It usually comes back with a slightly tweaked version that I use parts from.
I used it to make a nicely stylized prototype web app for a company using local storage for data in a few hours. Maybe I’m just not very competent at that stuff but the css I feel would have taken me hours alone
I've been teaching myself how to do low-level graphics programming using Apple's Metal API with assistance from ChatGPT. It can rapidly explain concepts, provide example snippets, explain error messages and suggest corrections, explain concepts, and provide steps.
Apple's Metal API suffers from a weak ecosystem of third-party documentation and examples (unlike OpenGL), and Apple's documentation is barebones, so having a tool like ChatGPT that can interact with that entire space conversationally is a huge boon.
Sometimes it suggests erroneous approaches. When that happens, I can ask it for alternatives. It's not actually too troublesome to vet the information since it has to actually work on the hardware, and honestly, I feel like I'm learning it more deeply by trying the different ways anyway.
Update on this: I’ve reached the point where I’m comfortable cracking open the 300pg Metal Programming Guide and jumping directly to the section covering what I’m interested in learning about. I’ve switched to using ChatGPT to generate example shaders that use particular techniques. It’s still extremely valuable, but there’s a level of organization or strategic thinking that’s missing in the responses provided.
I'm good at what I do (at least I think I am), yet I don't know what to use it for.
For specific questions, I find (and trust) the results faster on Stack Overflow. The code that ChatGPT produces is fairly good and my boss is impressed when I showed him the answer of How to parallel filter a List in Java. Yet, filtering a List is a far cry from building an app (which my boss thinks ChatGPT can do, given enough explanation what to do).
And any other general questions, while again, impressive...are too general for someone with a specific knowledge and useful only for somebody not familiar with the subject. Maybe the trick is how the answers are presented with confidence and giving the feeling that the converser actually knows something about the subject, yet I know it just machine generated babbling without any real understanding...
StackOverflow seems to be the ultimate litmus test for ChatGPT at this point for coding work. Pros: it customizes the code more than StackOverflow will. Cons: Subtle errors that can be difficult to spot at first glance, but could be easily caught when run/tested. Like for example I will see it trying to import functions from modules, when the function doesn't exist in that module but another one, and the code is mostly written like its okay otherwise.
We need models that can read long texts like a bunch of scientific papers, a book, or the whole git repo of an app. And we need models that can do iterative changes - diff models, they can be trained of course on git commits.
the models we have today will never be able to read .. they are just able to produce something that we cannot distinguish from human output.
we should be more careful on how we use them. in my opinion, ChatGPT and similar will be horrible for search on the web, as we are flooded with text that looks like a human wrote it but it does not add any knowledge or new insight.
That is because it has seen literature or code that is syntactically similar enough to what you give it, so that it can autocomplete its way to a semi-intelligent response.
Then we have a different definition of understanding. A good half of these debates come down to definitions at this point.
To me, ChatGPT clearly displays degrees of understanding. Often when I ask it to do things like make a certain modification to a piece of code, I'll ask it "why did you choose to do it like that?" and it gives a coherent explanation. That it correctly did what I asked and that it's explanation of why it did that way comport with one another implies understanding in my view. How could it not? What else would that imply? And if I ask a human to do the same thing and get the same result you'd have no problem with saying "well it's able to do that because it understands, and if it didn't understand then it would have zero capability of doing that". So, what's with the double standard?
For me, it's matrix multiplications on data made to output things that are approximations to what a human has written before on that question or text you're prompting it.
I know how training a deep neural network works and there's no reading involved for me :)
It's just estimating what a person would write (being trained on a really large data set of what people wrote). If it's trained on gibberish, it will just output gibberish. If I ask you to read a book with gibberish words would you do it? ChatGPT would (according to you) "read it" and recite the gibberish I gave it without a problem.
The matrix multiplications used to build these models don't show any comprehension, agency, or consciousness. Their point is to estimate the next data point given the previous data point, they will perform the same task independent from which training data we give them. That's not reading.
I don't say this might not change in the future, yet treating ChatGPT like a human using words like "reading," "understanding" etc. or intelligence for that matter is just anthropomorphizing for me.
I look at it from a general system theory point of view first and foremost as that's a level of abstraction higher than both humans and machines, yet you can find clear sets of definitions that describe all of these things. They go so far as to include knowledge, understanding and degrees thereof. I've started working my way through "On Purposeful Systems" as I've been finding most people's definitions far too hand wavy to be useful in thinking about the topic. That's been quite interesting.
One thing that stands in my mind from having read that is that the structural properties of the system matter in terms of its possible function, and I'm referring to the exact structure that includes the model trained on the non-gibberish text and treating that as the specific system I'm referring to. It seems when you say ChatGPT you're referring to the entire functional class of objects that implement the transformer architecture, which I'm not.
After thinking about it for a while the second thing I would note is that you're focused only on it's intrinsic function of predicting the next token, which as you say it always performs that function regardless of the training data, and I agree with that. Where we differ, I think, is I'm considering the specific structure which includes the model trained on non-gibberish data as an instrument (not a human!) that we use for it's extrinsic functions. Normally an individual or system's extrinsic functions are so narrow they're a lot easier to define, and we've never interacted with a system until now that the extrinsic function is like "understand what I say and reply accordingly". Think about it from the perspective of would you use or have any utility for GibberishGPT? No, same way I wouldn't read a book of pure gibberish.
to make my point easier: for my definition of to read, to understand and comprehend you need a body, an embodied mind. you need agency and consciousness. ChatGPT has none of those, it's an implementation of several matrix functions.
the specific model you are referring to does nothing, it's not acting (it's not reading anything). you have to press enter for it to work. so it's not reading, you (and OpenAi engineers) are doing its actions for it.
Making the context 10x longer would 'only' cost about 100x as much compute. Presumably you could take GPT-3 (trained with the normal context) and then finetune it on a comparitively small amount of data with the new context length, so it shouldn't be enormously expensive to train either.
Would be interesting to see how well it could 'understand' a paper, or if more layers etc. would be needed.
It would be expensive but probably worth it. Most companies would pay hundreds to have an AI model scan their whole codebase and be able to answer questions about it.
Curious to learn more about prompt engineering takeaways here. Was feeding more context (or chapters of textbooks, bits of papers, documentation) helpful? It does seem like layering information and being very precise helps a lot. Eerily like with interns
I'm a physician with a software background, another physician at my work used ChatGPT to write a compelling appeal for a patient when insurance denied coverage. I've started to use it to automatically dictate procedural reports. I provide an example report and tell it to modify the report as needed (e.g. "right iliac bone biopsy report with 11G needle"). Beats tabbing (or F2ing) through autotext fields.
These examples make me think, "ChatGPT for form completion" is the killer app as it applies to law, real estate, medicine, etc.
Probably patients consent to using software to process their data and chatgpt is considered just another software in the stack.
Edit: parent poster said no info is sent.
that's not how it works, PHI and HIPAA usually supersede a lot of these. That's why you hospitals don't just have everyone opt-out data protections if they want services.
They probably consent to software that's been validated and managed by their medical provider. I can tell you for sure that I personally wouldn't consent to a a random script thrown together by an MD with leetcode experience that sends your medical data up to an experimental service.
That is an obviously incorrect assumption, it is possible to de-anonomize most data sets about and there is reason to believe this one is no different. Health data by it's nature is very personal and specific.
Doctors publish case studies all the time which contain anonymized data. Presumably those go through reviews to make sure that nothing is being leaked but health data by it's nature is specific but not very personal (at least not identifiable).
Also, depending on what you're using ChatGPT for, this is no worse than Googling something which doctors do a lot as well.
Informed consent isn't a legal requirement. It's down to ethics and occasionally the journal's publishing requirements. So it we bring the analogy back to ChatGPT, using it for queries isn't breaking HIPAA or any laws.
“Write an appeal letter to a medical insurance company for a patient who needs a biopsy for a bone lesion given prior unclear diagnosis.”
Add an arbitrary ip address and timestamp and you are very far away from anything personally identifying. (Where does your computer suggest you are right now?)
It's trivial to de-identify someone from shockingly limited data. Just by submission to an outside service, they know a date, location, and whatever information was submitted. That's plenty, especially assuming they referenced a procedure or comorbidities.
A date of what? The request to ChatGPT to generate an imaginary appeal?
So you know the individual was rejected prior to that. Probably. Maybe.
And maybe roughly where the individual is located, because ChatGPT sees an IP off of… something.
So somewhere, someone asked ChatGPT about a rejected treatment or procedure appeal.
Now if the doctor provides a poorly written appeal and asks for it to be fixed, that is another case entirely, especially if they left in patient info. But there is a very large gap between these two situations, and the first one isn’t nearly as much as you suggest.
Why would a letter to an insurance company on behalf of a patient not include PII? The letter itself is surely mostly PII. And is almost certain to contain privileged information.
How do you envision that happening? There’s no big database of people’s medical conditions out there that you can use to lookup their name and address, so it’s kind of like an impossible reverse lookup that would need to be done right? Unless you’re talking about a state level actor or something that is tailing someone’s movements around like Jason Bourne and cross referencing it with medical GPT queries, but in that case you’re gonna be compromised anyway in probably far easier ways.
Probably not. You fill in the PII afterwards and it’s essentially just metadata. Patient has this condition, needs this test, has had this and that happen” those things aren’t PII. Leave out names, places, ID numbers, etc and you have appropriately deidentified a document.
No patient info is sent. In my colleague's example, you can try out a similar query, "Write an appeal letter to a medical insurance company for a patient who needs a biopsy for a bone lesion given prior unclear diagnosis."
tbh the ChatGPT saga has exposed just how willing people are to send their own/company's/client's/patient's data to a third party without a second thought.
Neat. I'm interested in doing something similar, but involving PHI, so ChatGPT is out of the question. How would one self-host a LLM without a Web Scale™ GPU farm? Is FlexGen (recently featured on HN) the way?
Question on the ML side of this post: How are these "parameterizations" used? Is this really just feature engineering with a new name? Are they including this information when training the model?
In the article, they mention using the new labels to build a "more balanced" dataset -- is this a realistic possibility in practice when most teams still have a dearth of data?
Hello! I wrote the article so happy to answer this. It is partially feature engineering but partially not. It’s essentially using feature engineering to curate/correct a dataset, but a neural network as the actual end model without explicit input of these features(we call them quality metrics). I abbreviated a good amount of the process in the article so that it wouldn’t run forever, but essentially we allowed ChatGPT to select and write its own features and then used the strategies it came up with to apply these features to improve the dataset.
In terms of if it’s realistic in practice, the answer is yes. Some teams have a dearth of data, but many AI companies we work with have more data than they can use, and it’s more a question of how to sample, curate, and correct the data and labels they have to improve their models rather than collect new data. Great questions!
Whats the deal lately with turning verbs into nouns? Learnings, asks - come to mind, but I periodically hear others. It’s not drippy anymore to simply say lessons and questions, bros don’t find them lit?
It's too bad they didn't get a chance to try it with Bing's chatbot. Based on what I've seen, the ability for it to reach into an index (or to the internet) to grab specific information creates a qualitative change in its ability.
111 comments
[ 2.9 ms ] story [ 163 ms ] thread>Can ChatGPT be used to improve an AI system? Yes.
>Would we hire it as our next standalone ML engineer? No.
> Let’s wait until GPT4.
The arms race for AI that can erase the human from cognitive work is really here.
...well, maybe if someone live streamed the ditch digging it could earn a few dollars.
I've always found arguments of these sort to be weak sauce--our brains are stochastic parrots.
Emergent behaviour + the bitter lesson (scaling laws) hits hard.
And while there may be objections to defining intelligence with respect to the Turing Test, I haven't heard of any definitions that are obviously and significantly better.
Our brains are stochastic parrots at least when it comes to communicating with text. People may digress into things like physical abilities, emotions, feelings, subjective experiences etc., but if we're talking about text, I'm not aware of a higher bar. In fact we probably don't know how to tell if an entity is "more intelligent" than the most intelligent humans because we won't be able to assess their intelligence.
Spreadsheets revolutionized bookkeeping. They did not eliminate the need for accountants. Robots automated much of the manual labor in manufacturing. They did not eliminate the need for process engineers, etc. Same story for autocad, protools, avid, etc.
GPT4 won't be an AGI of course, but it still could have a massive impact on the job market. A single engineer with GPT4 could potentially do the work of 10 engineers right now.
Edit: To be clear, I used GPT4 as a placeholder for "some near future LLM". I'm not making any bets on GPT4 in particular.
[1] See "The employment test" https://en.wikipedia.org/wiki/Artificial_general_intelligenc...
And the potential mistakes of several engineers too, so you would need more than one engineer to double-check the output. Remember that, even if they are truly amazing, LLMs are basically incredibly powerful predictive keyboards.
I've actually used ChatGPT and copilot and I can tell you it's not even 10% more productive, so I'm curious how you ended up with your estimate.
Or, it could be like how assemblers didn’t get rid of software engineers but helped fuel the fire. Who knows.
"Potentially" we could explore the galaxy or have nanobots and Star Trek fabricators. What does it matter? ChatbotGPT is no closer to AGI than my keyboard autocomplete, and copilot is useless for any task whose answer cannot be googled in 20s.
A single engineer with [some near future LLM] could potentially do the work of [X] engineers right now.
Use your imagination.
Do you think that LLMs have hit their plateau? Not likely I'd say. With all the billions suddenly rushing into the space, you really think that ChatGPT v.X isn't going to replace you? Risky bet.
Mmmm, I don't know.
I think LLMs are running up against fundamental limitations in the approach (hallucinations in particular) that will not be fixed unless truly new techniques are discovered.
So I do actually think we're likely hitting a plateau.
So LLMs fundamentally do not encode the nature of facts and data. They're simply (well, no, extremely sophisticated) text prediction engines. There's no embedded comprehension. And that's basic to the approach companies like OpenAI are taking.
That's why it's so easy to find factual errors in the generated text from these models: they can create text that's pleasing, but that's as much as they can do.
The next leap, to generate language that's also accurate, will require new techniques to bake in actual understanding.
Until then, these models will always have issues with hallucinations.
Or, at least, that's my expectation. Certainly nothing we've seen in GPT 3, 3.5, or Bing, which is rumoured to be based on something close to GPT 4, indicates any advances in this area.
The second AI only needs to confirm correct, provide a correct fact instead, or some other pre-agreed solution.
I'm not convinced that will work, but let's say it could. First, you'd have to build that oracle and the problem is, as far as I'm aware, we don't know how to do that. That'd be one of those new techniques I was referring to.
Obviously, there's work to be done still, and I'd be wary of making Google that 'oracle', but the development there would seem to be incremental (teach it more facts), rather than creating some other completely new concept.
But hey, I guess we'll see!
It's not that ChatGPT is incapable... it's that it's not inherently going to make me significantly more productive.
Big assumption, but if that's possible then businesses will fill that time. What used to take teams of people and big iron can now be done by a single person on a VPC. Did that take away jobs? Nope, requirements only expanded.
Modern high level languages didn't eliminate engineers. They suddenly meant one engineer could do what it would have taken 10 to do in COBOL or C and so software became more complex and capable.
Of course there can be industry shifts (MS was the industry for a long time and IBM before them. Google wasn't always dominant), but the pie grows to fit the available resources.
They did not eliminate mathematicians, but I do not believe mathematicians were ever doing arithmetic for a singificant portion of their workday.
Before it was a device, "Calculator" used to be a job description. It isn't any more. Every one of those people lost their job.
Even if the technology does evolve faster than people, governments can make sure people don't get impacted. Coal plants are more expensive to operate, worse for the environment, and less efficient but the coal industry is still kicking along for better or for worse.
This is the key distinction. Bandwidth. Humans only have so much of it, and if the rate at which they need to change their careers / find a new line of work exceeds their ramp-up time to become productive in that career / line of work then we can start to expect non-linear behavior from those humans. Maybe that means some sort of mental breakdown.
PageRank was essentially an efficient way to compute / approximate the leading eigenvector of a very large matrix.
It was computer, not calculator.
An AGI would be a "purposeful system" the same as humans. None of the other things you mentioned are purposeful systems. In terms of their structure and function they all have limitations on their own ability to change depending on the environment they are in. The key dimensions with which to think about this problem is "can this thing be structured in only one way for all environments? Or can it be structured in only one way in any given environment, but structured differently in different environments? Or can it be structured in more than one way in any environment, and also more than one way in different environments?" and then the same for functionality "can this thing only perform one function in all environments? Or can it perform only one function in a given environment, but different functions in different environments? Or can it perform more than one function in a given environment, and more than one function in different environments?"
Only things which have both properties of being "multi-multistructural" and "multi-multifunctional" meet the definition of a "purposeful system". They can change their structure to adapt to the environment, and they can change their goals within and between environments.
Humans currently use computers as instruments, and when an AGI hits it will be capable of saying no.
Thinking it through, I would say that "autocomplete" would be uni-multistructural and uni-unifunctional given that autocomplete is always structured only one way for the given environment that it's in, but that structure differs between environments (e.g. a phone, or a form in a web-browser), and that regardless of what environment its in it's only function is to predict the next token of output for a given input.
LLMs like ChatGPT on the other hand I would imagine to be uni-multistructural and multi-multifunctional, given that for any given environment you deploy an LLM into it's always structured the same, but there are many different structures it can take given there are lots of different LLMs. And I believe it to be multi-multifunctional because even within a single environment it can be used as an instrument to perform a wide variety of extrinsic functions, and this is true of other environments it could be deployed in.
So... it's actually significantly closer.
That is, we've had a huge hollowing out of all sorts of "middle" jobs in the US (e.g. someone mentioned secretaries used to be the most populous job in the US) so now we're largely (not exclusively, but largely) left with 2 categories of jobs: (a) dead-end jobs that are at present difficult to automate because they often involve manual labor: housekeepers, security guards, massage therapists, waiters, truck drivers (but, we see where that's headed...), etc. and (b) high level jobs that "manage the machines", e.g. tech jobs.
It's not that hard to see where this ends up with AI being capable of more complicated tasks - I wrote about an anecdote recently where ChatGPT had essentially already obviated some positions https://news.ycombinator.com/item?id=34862450. There are a slew of very high paying jobs that AI is coming for next, e.g. there is currently a shortage of radiologists because many folks in med school see the writing on the wall and don't go into radiology residencies. While full diagnostic radiology AI may not be there yet, it certainly will be by the time folks in med school finish their career in 40 years.
That's the job of new inventions. Nothing to complain about, in my opinion.
Those however who think that radiologists (/lawyers/secretaries) will get automated away, don't really understand what a radiologist's job really is.
The boring bureaucratic part of a radiologist's job will get automated away, yes, but that's because automatable/auto-completeable tasks should get automated at some point.
Think of it as a healthy organism eating away the boring part of a dream job that you, as a kid, imagined doing when you'll grow up.
As someone intimately familiar with the job of a radiologist, please then tell me what "a radiologist's job really is".
A diagnostic radiologist's job is to take a series of images, and output a set of findings and diagnoses. There is essentially no better task suited to modern AI. Yes, radiologists can do procedures like biopsies and potentially other interventions, but diagnostic radiologists often solely read images or at least earn the vast majority of their RVUs by reading images.
The people who weren't already good can't figure out how to use it properly.
I also see a bunch of people online, most?, using it incorrectly, writing prompts that would give bad info. Maybe they're doing it on purpose?
To those who have everything, more will be given.
The citation thing is a good example of this, just because chatgpt makes up citations today doesn't mean that it will always make up citations, accurate citations is just the wrong use case today. Already there are extensions that make accurate citations work. Still, people point at inaccurate citations as some kind of proof that chatgpt won't work.
Probably people want it to fail or don't want to believe how transformative it is.
Ego defense is a very real, and very powerful driving force in human behavior and it is coming out in full force.
Curious to know some solid use cases of productivity improvements in real world?
Also, naming things.
Apple's Metal API suffers from a weak ecosystem of third-party documentation and examples (unlike OpenGL), and Apple's documentation is barebones, so having a tool like ChatGPT that can interact with that entire space conversationally is a huge boon.
Sometimes it suggests erroneous approaches. When that happens, I can ask it for alternatives. It's not actually too troublesome to vet the information since it has to actually work on the hardware, and honestly, I feel like I'm learning it more deeply by trying the different ways anyway.
For specific questions, I find (and trust) the results faster on Stack Overflow. The code that ChatGPT produces is fairly good and my boss is impressed when I showed him the answer of How to parallel filter a List in Java. Yet, filtering a List is a far cry from building an app (which my boss thinks ChatGPT can do, given enough explanation what to do).
And any other general questions, while again, impressive...are too general for someone with a specific knowledge and useful only for somebody not familiar with the subject. Maybe the trick is how the answers are presented with confidence and giving the feeling that the converser actually knows something about the subject, yet I know it just machine generated babbling without any real understanding...
we should be more careful on how we use them. in my opinion, ChatGPT and similar will be horrible for search on the web, as we are flooded with text that looks like a human wrote it but it does not add any knowledge or new insight.
This would make me conclude that the models we have today can in fact read, and typically rather well.
Never say never
When we are talking about THE MODELS WE HAVE TODAY, then yes, we can say never.
They already do read.
To me, ChatGPT clearly displays degrees of understanding. Often when I ask it to do things like make a certain modification to a piece of code, I'll ask it "why did you choose to do it like that?" and it gives a coherent explanation. That it correctly did what I asked and that it's explanation of why it did that way comport with one another implies understanding in my view. How could it not? What else would that imply? And if I ask a human to do the same thing and get the same result you'd have no problem with saying "well it's able to do that because it understands, and if it didn't understand then it would have zero capability of doing that". So, what's with the double standard?
For me, it's matrix multiplications on data made to output things that are approximations to what a human has written before on that question or text you're prompting it.
I know how training a deep neural network works and there's no reading involved for me :)
It's just estimating what a person would write (being trained on a really large data set of what people wrote). If it's trained on gibberish, it will just output gibberish. If I ask you to read a book with gibberish words would you do it? ChatGPT would (according to you) "read it" and recite the gibberish I gave it without a problem.
The matrix multiplications used to build these models don't show any comprehension, agency, or consciousness. Their point is to estimate the next data point given the previous data point, they will perform the same task independent from which training data we give them. That's not reading.
I don't say this might not change in the future, yet treating ChatGPT like a human using words like "reading," "understanding" etc. or intelligence for that matter is just anthropomorphizing for me.
One thing that stands in my mind from having read that is that the structural properties of the system matter in terms of its possible function, and I'm referring to the exact structure that includes the model trained on the non-gibberish text and treating that as the specific system I'm referring to. It seems when you say ChatGPT you're referring to the entire functional class of objects that implement the transformer architecture, which I'm not.
After thinking about it for a while the second thing I would note is that you're focused only on it's intrinsic function of predicting the next token, which as you say it always performs that function regardless of the training data, and I agree with that. Where we differ, I think, is I'm considering the specific structure which includes the model trained on non-gibberish data as an instrument (not a human!) that we use for it's extrinsic functions. Normally an individual or system's extrinsic functions are so narrow they're a lot easier to define, and we've never interacted with a system until now that the extrinsic function is like "understand what I say and reply accordingly". Think about it from the perspective of would you use or have any utility for GibberishGPT? No, same way I wouldn't read a book of pure gibberish.
the specific model you are referring to does nothing, it's not acting (it's not reading anything). you have to press enter for it to work. so it's not reading, you (and OpenAi engineers) are doing its actions for it.
Reading and understanding are obviously different things. I can read a paper on quantum physics, but I can't understand it.
Would be interesting to see how well it could 'understand' a paper, or if more layers etc. would be needed.
https://arxiv.org/abs/2103.03206
These examples make me think, "ChatGPT for form completion" is the killer app as it applies to law, real estate, medicine, etc.
Also, depending on what you're using ChatGPT for, this is no worse than Googling something which doctors do a lot as well.
“Write an appeal letter to a medical insurance company for a patient who needs a biopsy for a bone lesion given prior unclear diagnosis.”
Add an arbitrary ip address and timestamp and you are very far away from anything personally identifying. (Where does your computer suggest you are right now?)
So you know the individual was rejected prior to that. Probably. Maybe.
And maybe roughly where the individual is located, because ChatGPT sees an IP off of… something.
So somewhere, someone asked ChatGPT about a rejected treatment or procedure appeal.
Now if the doctor provides a poorly written appeal and asks for it to be fixed, that is another case entirely, especially if they left in patient info. But there is a very large gap between these two situations, and the first one isn’t nearly as much as you suggest.
feed to ChatGPT
find replace ABCEDFG HIJKLMNOP with name?
Then we'd make sure that our org/administration understood the risk, through whatever channels are appropriate.
In the article, they mention using the new labels to build a "more balanced" dataset -- is this a realistic possibility in practice when most teams still have a dearth of data?
In terms of if it’s realistic in practice, the answer is yes. Some teams have a dearth of data, but many AI companies we work with have more data than they can use, and it’s more a question of how to sample, curate, and correct the data and labels they have to improve their models rather than collect new data. Great questions!
Btw we don't say "drip" and "lit" anymore, get with the times!!!!
---
[0] https://en.wiktionary.org/wiki/learnings
https://en.wiktionary.org/wiki/drippy