> How functional does the code have to be before it's no longer a hallucination?
At all, preferably. Hallucinating `the_hard_bit()` from a library that doesn't exist isn't particularly useful. (That said, I do use GitHub Copilot, because when it's looking at actually related stuff, that's pretty good. Should we just hand an unbounded search and ingest to ChatGPT? Probably not!)
> How functional does the code have to be before it's no longer a hallucination?
From LLMs, at least, I expect it will always be a hallucination. Code is never the point. Code is the working medium by which people solve problems for other people.
One way to see this is to realize that code bases on their own are generally worthless in the sense that people rarely pay much money for them. They pay for users. They pay for teams. But they don't generally pay for raw code.
Another way to look at it: imagine that a manager looks at ChatGPT and says, "At last, I can fire all the programmers. Fuck those guys." They set out to build an app. How long do you think it will take before they are forced to admit defeat and hire somebody who can read and edit the code?
Even if you think they make it all the way to a revenue-generating product, the manager will have become a prompt engineer, creating a large mass of interrelated prompts that are used to make code that makes the app. We have not eliminated the programmer; we have turned a manager into a programmer who has been forced to discover a new programming language. One that is clearly more English-like, but it also lacking in precision and, on current evidence, is much harder to use.
But I think the more likely outcome is that they will need actual programmers pretty quickly. That at best they will have speeded the creation of some more or less generic code. Which is exactly what we saw with the code-generation wizards of earlier eras: you got a fast initial result as long as it was pretty standard. But then you were generally worse off, because you had a bunch of only semi-coherent code that somebody had to understand before they could do novel or difficult things.
> Another way to look at it: imagine that a manager looks at ChatGPT and says, "At last, I can fire all the programmers. Fuck those guys." They set out to build an app. How long do you think it will take before they are forced to admit defeat and hire somebody who can read and edit the code?
How is this different from "At last, I can fire all [those senior programmers and hire junior ones, and keep a team lead around, to keep them on track]. Fuck those guys."?
Maybe I'm not getting the point of your question. It's different in that junior programmers are still people, capable of much more than LLMs are. But yes, it's the same shortsighted executive intent.
My point is that the direction and management that you outline is as relevant to GPT as it is to a team of (sufficiently) junior developers.
And, given that people find utility in Github copilot, I'm not convinced that the difference between GPT and a junior developer is qualitative rather than quantitative. Regardless of whether LLMs are demonstrating intelligence or creativity, they are already writing productive code: they aren't a hallucination.
> One way to see this is to realize that code bases on their own are generally worthless in the sense that people rarely pay much money for them. They pay for users. They pay for teams. But they don't generally pay for raw code
This is something I've thought for a while. If Google's source code was leaked tomorrow, would it even matter? Almost certainly not, and most people probably couldn't do much with it either.
I’ve worked at another FAANG, and I used to joke all the time that if our code leaked it would give us a competitive advantage because the folks copying would have to spend so much time implementing our bloated code and god-awful build tools that they would probably go bankrupt.
>How functional does the code have to be before it's no longer a hallucination?
IDK, I haven't tried copilot, but does the code it generates work on the first try with no human intervention?
>What if we are all not much more than only drawing "on the (admittedly vast) proportion of [our experiences] ingested at training time?
ugh, this argument again. Despite the machine learning community using words such as "training" and "learning" to describe the way they tune parameters, it has never been proven that any existing AI resembles human cognition. This is something that needs to be demonstrated empirically.
Imagine yourself, for a minute, 50,000 years ago in the past. By many estimates this was just before humans had developed the ability for normal speech. The bleeding edge of technology is "poke him with the pointy end" and the epitome of human speech and literature is *angry grunting sound*. Even if somehow one was able to insert the entire quantifiable knowledge of humanity into an e.g. ChatGPT style system you're not really going to get anything more out of it than what you put into it. Ask the computer for how to develop nuclear energy and you're going to get back *confused grunting sound*.
You're not really going to get from there (as a machine) to man on the moon, Shakespeare, and nuclear energy through anything like a normal recombination of what's already known. Yet, somehow, humanity did. And extremely fast. The time frame from then to now is but 1000-2000 human generations. And that with endless war, fallible memory, dark ages, knowledge being lost (or burnt), and so on endlessly. A "intelligent" computer system, without such flaws, ought be able to replicate our progress in a negligible amount of time. But whatever technology this may be that I'm appealing to, I don't really see it on our current path with natural language search/recombination.
I burned an hour with chat gpt insisting an AddOrUpdate function existed in microsoft entity framework. When I called bullshit it hallucinated that another library contained it. Then it hallucinated versions... Then I gave up...
A week later I noticed that Update in the new version does an Upsert... by reading the f*cking docs... google also didn't know this answer nor did SO.
Thius has been my experience with ChatGPT and code,it hallucinates a lot of stuff.
I don't get the point of this article. ChatGPT and it's ilk are not going away. They are here, now, and our friends and families will start using them very soon. I have a contrarian view from this article that ChatGPT is actually good for the mainstream audience because it will teach them to think about how trust worthy the information is coming from a machine. Most people will understand they have to take what is being said by a machine with a grain of salt (or they will soon). Maybe they'll start applying some of those discernment skills to popular mainstream media. The cat's out of the bag. You can no longer bury it.
Does TikTok, Android/Google,Instagram use make the same friends and family think about privacy and data? From a cursory look at...nearly everyone...i'd say the answer is a resounding no
When Facebook came around we all knew it would be a disaster, and it was.
When TikTok/Instagram replaced it, we knew it would be worse, and it was.
Now we have LLMs that vanishingly few people in the greater public actually understand and we know they'll be used for every possible evil imaginable at scale. I'm not optimistic.
Actually, Instagram was pretty sane for a long time. Until recently it just showed me who I followed and nothing else.
Now they're trying to complete with TikTok and I get all these crappy suggested reels and it's turning me off the platform. But for a long time it was pretty good unlike Facebook that ruined itself with its algorithmic feed.
> When Facebook came around we all knew it would be a disaster, and it was.
Not really. When Facebook started as college exclusive social site it seem quite innocent. Later when it created news-feed, opened up to everyone etc it became a cesspool. Once people started getting their news from it, that's when it became dangerous.
Privacy concerns and ChatGPT have different failure modes.
If ChatGPT fabricates a link or makes something up, there's a potential feedback mechanism. E.g. I ask ChatGPT to explain a science term, and then I get told that my understanding is incorrect in class when I use the ChatGPT definition.
It doesn't exist for every use case, but I'm hopeful everyone will be "bitten" once by ChatGPT etc., and then folks will verify its output appropriately.
This doesn't seem much different to me than how people should use wikipedia. It looks like it will be a useful tool to use, you just can't be careless.
Wikipedia has moderation and standards for sourcing, etc. that has made it generally very reliable. ChatGPT is a fancy parrot that sounds reliable. To actually verify that its output is correct requires more work and cross-referencing sources, etc.
That's a good point. For more complex topics, the act of verifying correctness is also more complex. I think it's a fair tradeoff though. A more powerful tool also requires more work to use CORRECTLY. Now, whether people will actually put in due diligence is another matter.
You might be more confident than me in people's ability to discern BS. I think I am reasonably intelligent, but I get fooled all the time. Compounding the problem is that it is easier to fool someone than to tell him he has been fooled.
There's no cat and no bag - just statistical sleight of hand. The people playing the trick know this, and are just collecting investor dollars while they can (and astroturfing every HN article that explains what it is). Relying on the output of chatgpt is no different than doing a rain dance and expecting it to rain.
I'd say it's more like relying on a minimum wage employee. It gets right (what I'm asking it to do) most of the time, but not always, and you gotta check up on it.
I don't understand why people act like it has to be 100% reliable to be useful.
Because if you need to check up on it, just do what you would do to check up on it in the first place. Coding is a very niche area where often you can tell if it is giving good answers. I think that's why HN people think it is so dope. In almost any other domain finding information from reliable sources is key, and chatgpt can't help with that at all.
I can't even teach my baby boomer aged inlaws to understand that obvious bullshit social media chainposts (along the lines of "copy this into your timeline to stop facebook from selling your data") are fake. There's no way they're going to see past ML-generated content, esp ones where feedback from success/fail ratios will cause it to 'improve' over time.
We might not be able to bury the cat, but that's hardly a good thing. Shits going to get really awkward for the next few decades.
Clickbait article. Same was said about the Internet back then!
Also reminds of people doing symbolic AI (decision trees and stuff) that kept criticizing NNs telling they don't work, will never be able to tell why they have this output (it's even wrong), etc.
They had some good points though. The article says that this technology centralizes power in the hands of rich tech companies who exploit low paid workers and unpaid creators who do the value creating work. The tech companies organize the data then endlessly skim off the top. At the same time they maintain undemocratic control of who can see what information, guided by profit and sometimes authoritarian governments.
These are all the same criticisms of the walled garden internet and theyre just as relevant for LLM’s
It's nice to see a little realism starting to poke through the hype. The rush for automatic bullshit generation for me is very much in the bucket of, "Your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should." As with every new technology, we should be looking at LLMs through a lens of maximizing the benefits and minimizing the harms.
There is one thing that ChatGPT has done, which is highlight the amount of bullshit we are exposed to. If ChatGPT does nothing else but replace the bullshit generators, then we know on the whole we are safe to ignore all bullshit (a pretty safe course of action).
Over the past few months I've found myself second guessing whether the article I am reading was generated bullshit or "real" bullshit. It's certainly made me re-think what I waste my time reading. If I can't get through the first paragraph before asking myself "is this AI bullshit?" it's a really good indication that the entire article is going to be bullshit, at which time it doesn't matter whether a human wrote it or not.
For sure. And it's worth asking what we were getting out of those articles before.
As an example, there's an Android video game called Splash Wars. It has algorithmically generated levels with no sense of progression. In some sense it's not a good game, in that there's no narrative, no progression, no increasing cleverness. But I have played an embarrassing number of levels of it because for me it's soothing to have something modestly challenging but ultimately familiar. I realized that there are some video games I use to think. With the ADHD lets-go-ride-bikes portion of my brain pacified, I can ruminate on problems that I might otherwise get distracted from.
Similarly, I suspect that bullshit articles fill other unspoken needs for people. That in the morning over cereal, maybe what I want from my newspaper is not actual news so much as news-shaped or news-flavored textual product. Maybe I want comfortable familiarity or confirmation of my biases.
So I hope you're write, that others start asking those same questions. That the coming vast bullshit surplus puts many of us off it for good.
Hah! I am honestly not even sure about that. There's clearly investment money to be raised. But sustained ways to generate value in the world? That's not as clear to me. I wonder if this more in the category of Groupon, which everybody thought was worldchanging for about 20 minutes. Or Shazam, which was an absolutely fucking mindblowing technology at first, and then it quickly became banal. Or perhaps the closest analogy was Qwiki, which was sort of inspired by the Star Trek voice interface where the computer would tell you about things on request. It was a hot startup, something people were watching closely. And then when people actually got to use it, turned out nobody cared.
The key thing is that most people are obsessed with get rich quick, as long as you're first and you're fast you can make your money, damn the consequences.
Bringing up that the AI model is racist or some other kind of -ist to bolster your argument is a bullshit tactic. People are -ists, too. I agree with the rest of their points, though.
Why do “smart” people like trying to connect all the different current news items into one grand narrative. Is he trying to use the political capital generated by opposition to chatgpt for other issues? Chatgpt has less than nothing to do with underpaid home health aids. They don’t even know what it is!
> ChatGPT is, in technical terms, a 'bullshit generator'.
This bs generator speeded up my coding by 10-20x. It's like a superpower... not talking about ChatGPT specifically but its stochastic parrot cousin (Copilot).
GPT has been great for exploring new frameworks too. Copilot is amazing when I know what roughly what I'm doing, but GPT is like personalized framework / API documentation at times
I could leave my webcam on, it'll maybe regularly notify me its on, and hopefully block taking pictures, instead it'll basically just watch my eye movements and know which screen to focus on, and take voice prompts to find the window and subwindow i.e. chrome+hackernews tab, or vscode+filename. It'd basically be like vim but for outside the terminal and without memorizing shortcuts and it'd have a listening mode so I can just ask it questions whenever and it'll maybe even interface with copilot and give extra 'context' to things.
This is what i want, basically a mini-version of "HER" to give me more super powers until it takes my job, lol.
It's not integrated into an IDE like GH co-pilot yet is it?
I only find I reach for it for stuff like navigating some curly regex or forgetting date time format syntax for the millionth time. But I would be very keen to understand if I'm missing out.
Posted elsewhere, but this is what I suspect is happening. Folks that are primarily slapping together a lot of code are benefitting from ChatGPT. It's pretty useless for the other stuff except maybe boilerplate stuff for a design doc.
I find this interesting. There's a few things ChatGPT has been useful for with coding but in general I've found it to produce outright incorrect answers and it can't really reason about higher level problems. Even for spark questions in scala it generates code w/ totally made up methods or code that doesn't work properly.
Its writing style is also pretty tedious for many prompts.
I've noticed "full stack" folks really enjoying ChatGPT but dedicated infra and ML folks (not working on toy problems) finding it less valuable.
A good chunk of my work involves talking to folks, understanding requirements, putting together design docs, etc. vs. actual coding. I've found it not particularly valuable even for doc writing.
How does Copilot help you code faster? I've watched some coworkers use it but it appears to be more distraction than help, as they accept CoPilot suggestions and then have to rewrite it all anyway.
It helps me a ton, but I suspect that's because I disabled auto-suggest, so it only gives me a suggestion when I specifically ask. I know by now what kinds of problems it's good at, so I can selectively prompt it when I know I'm working on something that it will able to do faster.
One example is when working on a front end form in Vue or React: if I have a bunch of variables in the component state and create a form tag, Copilot is great at creating all of the inputs with the correct bindings and input component types, so I'll usually prompt it after creating the first one by hand as an example.
Since I'm selectively prompting it instead of just letting it suggest left and right, I also find it helpful to sometimes write out a quick comment explaining what it is that I want in more detail than the function name can provide, then prompt it on the next line after the comment. That often helps to get better results, though I'm still careful to read through everything.
While i find the downplaying of chatgpt to be a bit much, lets not go too far in the other direction.
ChatGPT is not suitable in applications where accuracy is more important than plausability, is i think the neutral way of putting the limitation.
So it can't write your PhD thesis for you and hasn't literally replaced himans in every field of endeavour. Big whoop. If that's the bar, it is a pretty high bar.
it can help you write that PhD thesis. It can summarize all current research that might be considered relevant to your current thesis, vastly improving your productivity!
Using chatGPT requires expertise, i think, rather than the other way around! If you have expertise in your field, you can easily check correctness (or at least plausibility) of the output, and fix up the bits it gets wrong. This will make any work you need to undertake faster.
It has lots of value. But it's not the AI it pretends to be. Mostly because we and it can't always tell what's fact or fiction and it has 0 cognition. It's up to us to figure out where it can add value. It's way too early to announce its death.
Little known fact, Wheels, most of these "AIs" are really just hallucination machines with a lot of cross-referencing. Most of their inputs were Philip K. Dick, the artwork of J.K. Potter, and Duran Duran lyrics.
This seems to be the number one problem with lay understanding of ML masquerading as AGI. People immediately jump to the conclusion that the computer can think. Not yet, not even close.
I’m not sure if I can think, despite having a PhD in EECS. We humans prize our ability to think very highly but it’s a nebulous concept. I believe these models will get much better at simulating “thinking” in the next few years, and we soon won’t be able to say the model isn’t thinking, we’ll move on to criticizing the quality of its responses the way we criticize other bad ideas.
The point of the Turing test, to me, was that if a process responds in a logical way, you can’t really tell whether it is thinking or computing - and it doesn’t matter.
Yep. The same argument was put forward against chess engines - that they didn't really play chess! All they did was search all possibilities and spat out the "best" outcome based on some scoring system.
And yet, these engines have beaten the best humans.
The forms underneath intelligence, or thinking, doesn't matter. The outcome is what matters.
With 100% certainty Donald Trump has said more lies/bullshit than ChatGPT (porcentually speaking for all his public statements) and he became President of the US, some said due malice, others from stupidity, and some from both; where does people like him fall under the implied fantasy by this article that humans do better than to make "good guess(es) to pass your sense-making filter"? The social beneffits of people like him are less than speculative and the damages are empirically demostrated.
The nerve of talking about "ghost work" for $2 while publishing a page on the internet, where a giant chunk of the hardware used to mantain and use such network is made with raw materials mined and refined with slave labor, and even the parts where is debatable if it constitutes slave labor are jobs that people would never chose over the content filtering for $2.
the most irritating thing about this piece is the headline, a headline only a GPT could write.
Shakespeare's famous opening, "Friends, Romans, countrymen, lend me your ears; I come to bury Caesar, not to praise him" is the lead-in to a speech by Antony in which he very much praises Caesar and buries his enemies.
Hopefully people treat AIs like they treat other people on the internet, with disrespect and zero sympathy. We have been trained for this. Don't go soft on polite robots, that will be your undoing.
I'm starting to think the people who write these articles are terrible at harnessing search engines, so also terrible at getting good answers from an AI.
I've had pretty decent luck, it's really good at providing a bit of context to glue two things together with the missing pieces I couldn't figure out, most recent with JSON and JMESpath filters.
> However, as I spell out in my book, the concept of AGI is inseparable from the kind of hierarchy of intelligence that has underpinned ideas of innate supremacy since the days of empire and colonialism.
I think this line discredits the entire article, which was already a few obvious points blown at gale force into the reader’s ear.
I disliked the article, but this particular quote is a relatively strong one. Colonialism, racism, and aristocratic structures were often justified by the idea that certain classes of people are more intelligent, enlightened, or were selected by dieties. Here the author is arguing that a similar process is happening with AI, where atrocious policies by people in power are justified by claiming the AI (which was trained to encode the powerful groups values) is more intelligent.
> Saying, as the OpenAI CEO does[1], that we are all 'stochastic parrots' like large language models, statistical generators of learned patterns that express nothing deeper, is a form of nihilism. Of course, the elites don't apply that to themselves, just to the rest of us. The structural injustices and supremacist perspectives layered into AI put it firmly on the path of eugenicist solutions to social problems.
[1] Sam Altman: i am a stochastic parrot, and so r u
That this isn't even sentence-to-sentence consistent is somehow one of the less egregious aspects of it.
As an aside, I wrote my previous comment in “stochastic parrot” mode.
That is, I scattered commas where it felt natural to breathe.
Then, I edited it. Maybe, there was a non-essential clause in there. Also, possibly, there was an Oxford comma. Or maybe not! I’m not really clear about English grammar rules. But what this stochastic parrot says is usually intelligible to other English speakers.
Likewise that author seems to be disappointed that models have found a connection between Islam and violence.
The author may find this discovery distasteful, but that is irrelevant: Islam is an idea, many ideas are associated with violence. Islam was famously “spread through the sword” and to this day violent retribution is often visited upon ex-Muslims by their families in the name of Islam. This isn’t hate speech: Islam is an idea that people adopt and leave, and not an innate quality. To treat any idea like an innate quality denies people the right to not accept the ideas of those around them.
I think ChatGTP is pretty amazing and worthy of all of its recent praise. However, is it a bit of a bullshit generator. It doesn’t know if it’s right or wrong or anything, it’s just word soup trained enough to where it’s usual correct in general knowledge areas. Here is a good example:
424 comments
[ 2.7 ms ] story [ 318 ms ] threadHow functional does the code have to be before it's no longer a hallucination?
What if we are all not much more than only drawing "on the (admittedly vast) proportion of [our experiences] ingested at training time?
At all, preferably. Hallucinating `the_hard_bit()` from a library that doesn't exist isn't particularly useful. (That said, I do use GitHub Copilot, because when it's looking at actually related stuff, that's pretty good. Should we just hand an unbounded search and ingest to ChatGPT? Probably not!)
Well, at the point when it works but I can't understand it well enough to verify that it is correct, we may have some issues.
I know someone who has been using it to write complex regexs. To mutate the joke, to me it just sounds like now they have three problems.
From LLMs, at least, I expect it will always be a hallucination. Code is never the point. Code is the working medium by which people solve problems for other people.
One way to see this is to realize that code bases on their own are generally worthless in the sense that people rarely pay much money for them. They pay for users. They pay for teams. But they don't generally pay for raw code.
Another way to look at it: imagine that a manager looks at ChatGPT and says, "At last, I can fire all the programmers. Fuck those guys." They set out to build an app. How long do you think it will take before they are forced to admit defeat and hire somebody who can read and edit the code?
Even if you think they make it all the way to a revenue-generating product, the manager will have become a prompt engineer, creating a large mass of interrelated prompts that are used to make code that makes the app. We have not eliminated the programmer; we have turned a manager into a programmer who has been forced to discover a new programming language. One that is clearly more English-like, but it also lacking in precision and, on current evidence, is much harder to use.
But I think the more likely outcome is that they will need actual programmers pretty quickly. That at best they will have speeded the creation of some more or less generic code. Which is exactly what we saw with the code-generation wizards of earlier eras: you got a fast initial result as long as it was pretty standard. But then you were generally worse off, because you had a bunch of only semi-coherent code that somebody had to understand before they could do novel or difficult things.
How is this different from "At last, I can fire all [those senior programmers and hire junior ones, and keep a team lead around, to keep them on track]. Fuck those guys."?
And, given that people find utility in Github copilot, I'm not convinced that the difference between GPT and a junior developer is qualitative rather than quantitative. Regardless of whether LLMs are demonstrating intelligence or creativity, they are already writing productive code: they aren't a hallucination.
but even if it's true, I think it's a case I've covered in the paragraph after the one you quote.
This is something I've thought for a while. If Google's source code was leaked tomorrow, would it even matter? Almost certainly not, and most people probably couldn't do much with it either.
IDK, I haven't tried copilot, but does the code it generates work on the first try with no human intervention?
>What if we are all not much more than only drawing "on the (admittedly vast) proportion of [our experiences] ingested at training time?
ugh, this argument again. Despite the machine learning community using words such as "training" and "learning" to describe the way they tune parameters, it has never been proven that any existing AI resembles human cognition. This is something that needs to be demonstrated empirically.
Sometimes, it certainly does. Does your code always work on the first try with no intervention?
You're not really going to get from there (as a machine) to man on the moon, Shakespeare, and nuclear energy through anything like a normal recombination of what's already known. Yet, somehow, humanity did. And extremely fast. The time frame from then to now is but 1000-2000 human generations. And that with endless war, fallible memory, dark ages, knowledge being lost (or burnt), and so on endlessly. A "intelligent" computer system, without such flaws, ought be able to replicate our progress in a negligible amount of time. But whatever technology this may be that I'm appealing to, I don't really see it on our current path with natural language search/recombination.
A week later I noticed that Update in the new version does an Upsert... by reading the f*cking docs... google also didn't know this answer nor did SO.
Thius has been my experience with ChatGPT and code,it hallucinates a lot of stuff.
When TikTok/Instagram replaced it, we knew it would be worse, and it was.
Now we have LLMs that vanishingly few people in the greater public actually understand and we know they'll be used for every possible evil imaginable at scale. I'm not optimistic.
See you all in the hellish afterscape.
Now they're trying to complete with TikTok and I get all these crappy suggested reels and it's turning me off the platform. But for a long time it was pretty good unlike Facebook that ruined itself with its algorithmic feed.
Not really. When Facebook started as college exclusive social site it seem quite innocent. Later when it created news-feed, opened up to everyone etc it became a cesspool. Once people started getting their news from it, that's when it became dangerous.
https://news.ycombinator.com/item?id=1289997
If ChatGPT fabricates a link or makes something up, there's a potential feedback mechanism. E.g. I ask ChatGPT to explain a science term, and then I get told that my understanding is incorrect in class when I use the ChatGPT definition.
It doesn't exist for every use case, but I'm hopeful everyone will be "bitten" once by ChatGPT etc., and then folks will verify its output appropriately.
People care very much about making good decisions so they will be careful about blindly trusting the output of a machine.
I don't understand why people act like it has to be 100% reliable to be useful.
We might not be able to bury the cat, but that's hardly a good thing. Shits going to get really awkward for the next few decades.
https://langchain.readthedocs.io/en/latest/modules/agents/ge...
Also reminds of people doing symbolic AI (decision trees and stuff) that kept criticizing NNs telling they don't work, will never be able to tell why they have this output (it's even wrong), etc.
They had some good points though. The article says that this technology centralizes power in the hands of rich tech companies who exploit low paid workers and unpaid creators who do the value creating work. The tech companies organize the data then endlessly skim off the top. At the same time they maintain undemocratic control of who can see what information, guided by profit and sometimes authoritarian governments.
These are all the same criticisms of the walled garden internet and theyre just as relevant for LLM’s
Over the past few months I've found myself second guessing whether the article I am reading was generated bullshit or "real" bullshit. It's certainly made me re-think what I waste my time reading. If I can't get through the first paragraph before asking myself "is this AI bullshit?" it's a really good indication that the entire article is going to be bullshit, at which time it doesn't matter whether a human wrote it or not.
As an example, there's an Android video game called Splash Wars. It has algorithmically generated levels with no sense of progression. In some sense it's not a good game, in that there's no narrative, no progression, no increasing cleverness. But I have played an embarrassing number of levels of it because for me it's soothing to have something modestly challenging but ultimately familiar. I realized that there are some video games I use to think. With the ADHD lets-go-ride-bikes portion of my brain pacified, I can ruminate on problems that I might otherwise get distracted from.
Similarly, I suspect that bullshit articles fill other unspoken needs for people. That in the morning over cereal, maybe what I want from my newspaper is not actual news so much as news-shaped or news-flavored textual product. Maybe I want comfortable familiarity or confirmation of my biases.
So I hope you're write, that others start asking those same questions. That the coming vast bullshit surplus puts many of us off it for good.
This bs generator speeded up my coding by 10-20x. It's like a superpower... not talking about ChatGPT specifically but its stochastic parrot cousin (Copilot).
I could leave my webcam on, it'll maybe regularly notify me its on, and hopefully block taking pictures, instead it'll basically just watch my eye movements and know which screen to focus on, and take voice prompts to find the window and subwindow i.e. chrome+hackernews tab, or vscode+filename. It'd basically be like vim but for outside the terminal and without memorizing shortcuts and it'd have a listening mode so I can just ask it questions whenever and it'll maybe even interface with copilot and give extra 'context' to things.
This is what i want, basically a mini-version of "HER" to give me more super powers until it takes my job, lol.
It's not integrated into an IDE like GH co-pilot yet is it?
I only find I reach for it for stuff like navigating some curly regex or forgetting date time format syntax for the millionth time. But I would be very keen to understand if I'm missing out.
I find it very hard to believe you can build software 10-20x faster with GPT, since the majority of building software isn't writing code.
I'm also curious what you're programming that can be completely offloaded to GPT.
Its writing style is also pretty tedious for many prompts.
I've noticed "full stack" folks really enjoying ChatGPT but dedicated infra and ML folks (not working on toy problems) finding it less valuable.
A good chunk of my work involves talking to folks, understanding requirements, putting together design docs, etc. vs. actual coding. I've found it not particularly valuable even for doc writing.
One example is when working on a front end form in Vue or React: if I have a bunch of variables in the component state and create a form tag, Copilot is great at creating all of the inputs with the correct bindings and input component types, so I'll usually prompt it after creating the first one by hand as an example.
Since I'm selectively prompting it instead of just letting it suggest left and right, I also find it helpful to sometimes write out a quick comment explaining what it is that I want in more detail than the function name can provide, then prompt it on the next line after the comment. That often helps to get better results, though I'm still careful to read through everything.
ChatGPT is not suitable in applications where accuracy is more important than plausability, is i think the neutral way of putting the limitation.
So it can't write your PhD thesis for you and hasn't literally replaced himans in every field of endeavour. Big whoop. If that's the bar, it is a pretty high bar.
it can help you write that PhD thesis. It can summarize all current research that might be considered relevant to your current thesis, vastly improving your productivity!
Using chatGPT requires expertise, i think, rather than the other way around! If you have expertise in your field, you can easily check correctness (or at least plausibility) of the output, and fix up the bits it gets wrong. This will make any work you need to undertake faster.
It's a super charged Cliff Clavin.
https://en.wikipedia.org/wiki/Cliff_Clavin
This seems to be the number one problem with lay understanding of ML masquerading as AGI. People immediately jump to the conclusion that the computer can think. Not yet, not even close.
Edit: Switching AI to AGI to avoid any confusion.
The point of the Turing test, to me, was that if a process responds in a logical way, you can’t really tell whether it is thinking or computing - and it doesn’t matter.
And yet, these engines have beaten the best humans.
The forms underneath intelligence, or thinking, doesn't matter. The outcome is what matters.
AI is a field that covers all kinds of highly limited intelligences, including bad guys in a video game walking back and forth indefinitely.
How well would ChatGPT perform if it played the final round of Jeopardy?
The nerve of talking about "ghost work" for $2 while publishing a page on the internet, where a giant chunk of the hardware used to mantain and use such network is made with raw materials mined and refined with slave labor, and even the parts where is debatable if it constitutes slave labor are jobs that people would never chose over the content filtering for $2.
Shakespeare's famous opening, "Friends, Romans, countrymen, lend me your ears; I come to bury Caesar, not to praise him" is the lead-in to a speech by Antony in which he very much praises Caesar and buries his enemies.
Its really really important that you ask it the right questions or it does tend to feed you some pretty average stuff.
and often its not just the right question, but the right question asked the right way.
I've had pretty decent luck, it's really good at providing a bit of context to glue two things together with the missing pieces I couldn't figure out, most recent with JSON and JMESpath filters.
I think this line discredits the entire article, which was already a few obvious points blown at gale force into the reader’s ear.
> Saying, as the OpenAI CEO does[1], that we are all 'stochastic parrots' like large language models, statistical generators of learned patterns that express nothing deeper, is a form of nihilism. Of course, the elites don't apply that to themselves, just to the rest of us. The structural injustices and supremacist perspectives layered into AI put it firmly on the path of eugenicist solutions to social problems.
[1] Sam Altman: i am a stochastic parrot, and so r u
That this isn't even sentence-to-sentence consistent is somehow one of the less egregious aspects of it.
It seems that, like ChatGPT, the author is also a bullshit generator, and therefore, by his own definition, unintelligent.
That is, I scattered commas where it felt natural to breathe.
Then, I edited it. Maybe, there was a non-essential clause in there. Also, possibly, there was an Oxford comma. Or maybe not! I’m not really clear about English grammar rules. But what this stochastic parrot says is usually intelligible to other English speakers.
The author may find this discovery distasteful, but that is irrelevant: Islam is an idea, many ideas are associated with violence. Islam was famously “spread through the sword” and to this day violent retribution is often visited upon ex-Muslims by their families in the name of Islam. This isn’t hate speech: Islam is an idea that people adopt and leave, and not an innate quality. To treat any idea like an innate quality denies people the right to not accept the ideas of those around them.
https://drive.google.com/file/d/1wALYKw59TqExbLiQJvql_sM7OhE...