The research from Purdue University, first spotted by news outlet Futurism, was presented earlier this month at the Computer-Human Interaction Conference in Hawaii and looked at 517 programming questions on Stack Overflow that were then fed to ChatGPT.
“Our analysis shows that 52% of ChatGPT answers contain incorrect information and 77% are verbose,” the new study explained. “Nonetheless, our user study participants still preferred ChatGPT answers 35% of the time due to their comprehensiveness and well-articulated language style.”
Disturbingly, programmers in the study didn’t always catch the mistakes being produced by the AI chatbot.
“However, they also overlooked the misinformation in the ChatGPT answers 39% of the time,” according to the study. “This implies the need to counter misinformation in ChatGPT answers to programming questions and raise awareness of the risks associated with seemingly correct answers.”
So this issue for me is this:
If these technologies still require large amounts of human intervention to make them usable then why are we expending so much energy on solutions that still require human intervention to make them usable?
Why not skip the burning the planet to a crisp for half-formed technology that can’t give consistent results and instead just pay people a living fucking wage to do the job in the first place?
Seriously, one of the biggest jokes in computer science is that debugging other people’s code gives you worse headaches than migraines.
So now we’re supposed to dump insane amounts of money and energy (as in burning fossil fuels and needing so much energy they’re pushing for a nuclear resurgence) into a tool that results in… having to debug other people’s code?
They’ve literally turned all of programming into the worst aspect of programming for barely any fucking improvement over just letting humans do it.
Why do we think it’s important to burn the planet to a crisp in pursuit of this when humans can already fucking make art and code? Especially when we still need humans to fix the fucking AIs work to make it functionally usable. That’s still a lot of fucking work expected of humans for a “tool” that’s demanding more energy sources than currently exists.
They don’t require as much human intervention to make the results usable as would be required if the tool didn’t exist at all.
Compilers produce machine code, but require human intervention to write the programs that they compile to machine code. Are compilers useless wastes of energy?
Compilers are deterministic and you can reason about how they came to their results, and because of that they are useful.
No, they’re useful because they produce useful machine code.
That’s a distinction without a difference. The code is useful because we can reason how it was made and we can then make deterministic changes. Try using a compiler that gives you a qualitatively different result each time it runs even though the inputs are the same.
It’s useful because it does the stuff we want it to do.
You’re focusing on a very high level philosophical meaning of “usefulness.” I’m focusing on what actually does what I need it to do.
I’m providing explicit examples of compilers doing “the stuff we want it to do”. LLMs do what the want 50% of the time and it still needs modifications afterwards. Imagine having to correct a compiler output and calling that compiler “useful”.
So if something isn’t perfect it’s not “useful?”
I use LLMs when programming. Despite their imperfection they save me an enormous amount of time. I can confidently confirm that LLMs are useful from personal direct experience.
I honestly don’t know how well AI is going to scale when it comes to power consumption vs performance. If it’s like most of the progress we’ve seen in hardware and software over the years, it could be very promising. On the other hand, past performance is no guarantee for future performance. And your concerns are quite valid. It uses an absurd amount of resources.
The usual AI squad may jump in here with their usual unbridled enthusiasm and copium that other jobs are under threat, but my job is safe, because I’m special.
Eye roll.
Meanwhile, thousands have been laid off already, and executives and shareholders are drooling at the possibility of thinning the workforce even more. Those who think AI will create as many jobs as it destroys are thinking wishfully. Assuming it scales well, it could spell massive layoffs. Some experts predict tens of millions of jobs lost to AI by 2030.
To try and answer the other part of your question…at my job (which is very technical and related to healthcare) we have found AI to be extremely useful. Using Google to search for answers to problems pales by comparison. AI has saved us a lot of time and effort. I can easily imagine us cutting staff eventually, and we’re a small shop.
The future will be a fascinating mix of good and bad when it comes to AI. Some things are quite predictable. Like the loss of creative jobs in art, music, animation, etc. And canned response type jobs like help desk chat, etc. The future of other things like software development, healthcare, accounting, and so on are a lot murkier. But no job (that isn’t very hands-on-physical) is 100% safe. Especially in sectors with high demand and low supply of workers. Some of these models understand incredibly complex things like drug interactions. It’s going to be a wild ride.
There is a good chance that it is instrumental in discoveries that lead to efficient clean energy. It’s not as if we were at some super clean, unabused planet before language models came along. We have needed help for quite some time. Almost nobody wants to change their own habits(meat, cars, planes, constant AC and heat…), so we need something. Maybe AI will help in this endevour like it has at so many other things.
There is exactly zero chance… LLMs don’t discover anything, they just remix already existing information. That is how it works.
This is a common misunderstanding of what it means to discover new things. New things are just remixing old things. For example, AI has discovered new matrix multiplications, protein foldings, drugs, chess/go/poker strategies, and much more that are all far superior to anything humans have ever come up with in these fields. In all these cases, the AI was just combining old things in new ways. Even Einstein was just combining old things into new ways. There is exactly zero chance that AI will all of a sudden quit making new discoveries all of a sudden.
no, people have discovered. llms were just a tool used to manipulate large sets of data (instructed and trained by people for the specific task) which is something in which computers are obviously better than people. but same as we don’t say “keyboard made a discovery”, the llm didn’t make a discovery either.
that is just intentionally misleading, as is calling the technology “artificial intelligence”, because there is absolutely no intelligence whatsoever.
and comparing that to einstein is just laughable. einstein understood the broad context and principles and applied them creatively. llm doesn’t understand anything. it is more like a toddler watching its father shave and then moving a lego piece accross its face pretending to shave as well, without really understaning what is shaving.
I didn’t say LLMs made these discoveries. They didn’t. AI made those discoveries. Yes, it is true that humans made AI, so in a way, humans made the discoveries, but if that is your take, then it is impossible for AI to ever make any discovery. Really, if we take this way of thinking to its natural conclusion, then even humans can never make discoveries, only the universe can make discoveries, since humans are a result of the universe “universing”. It is arbitrary to try to credit humans with anything that happens further down their evolution.
Humans tried for a long time to get good at chess, and AI came along and made the absolute best chess players utterly irrelevant even if we give a team of the worlds best chessplayers an endless clock and thr AI a single minute for the entire game. That was 20 years ago. This is happening in more and more fields and showing no sign of stopping. We don’t know yet if discoveries will come from future LLMs like theybm have from other forms of AI, but we do know that with each generation more and more complex patterns are being identified and utilized by LLMs. 3 years ago the best LLMs would have scored single digits on IQ test, now they are triple digits, it is laughable to think that anyone knows where the current rapid trajectory will stop for this new technology, and much more laughable to think we are already at the end.
if this is your take, then lot of keyboard made a lot of discovery.
AI could make a discovery if there was one (ai). there is none at the moment, and there won’t be any for any foreseeable future.
tool that can generate statistically probable text without really understanding meaning of the words is not an intelligence in any sense of the word.
your other examples, like playing chess, is just applying the computers to brute-force through specific mundane task, which is obviously something computers are good at and being used since we have them, but again, does not constitute a thinking, or intelligence, in any way.
it is also laughable to assume it will just continue indefinitely, because “there is a trajectory”. lot of technology have some kind of limit.
and just to clarify, i am not some anti-computer get back to trees type. i am eager to see what machine learning models will bring in the field of evidence based medicine, for example, which is something where humans notoriously suck. but i will still not call it “intelligence” or “thinking”, or “making a discovery”. i will call it synthetizing so much data that would be humanly impossible and finding a pattern in it, and i will consider it cool result, no matter what we call it.
This is literally my point. It is arbitrary to choose that all the good ideas came from “humans”. If we are going to give all credit for anything AI produces to humans, then it only seems fair to give all credit for human things to our common ancestors with chimpanzees, because if it were not for their clever ideas, we would never have been here. But wait, we can’t stop there, because we have to give credit to the original single-celled life forms, and eventually, back to the universe itself(like I mentioned before).
Look, I totally get the desire to want to glorify humans and think that we have something special that machines don’t/can’t have. It kinda sucks to think that we are not so special, and potentially extememly inferior to what is right around the corner. We can’t let that primal ego desire cloud our judgement, though. Our brains are physical machines doing calculations. There is not some magical difference between our calculations that make it so we can make discoveries and machines cannot.
Imagine you teach your little brother how to play chess, and then your brother thinks about it a bunch and comes up with a bunch of new strategies and starts to kick your butt every time, and eventually atatts crushing tournaments. Sure, you can cling to the fact that you taught him how to play, and you can go around telling everyone how “you” are winning all these tournaments because your brother is actually winning them, but it doesn’t change the fact that your brother is the one with the secret sauce that you simply are unable to comprehend.
Your whole point is that if people do it, then it is some special discovery thing, but if computers do it, then it is just computational brute force. There is actually no difference between the two, it is just two different ways of wording the same process. We made programs that could understand the rules, and then it went further and in the same direction that we were trying to go.
So far as continuing indefinitely because we are on a trajectory goes, sure, we will eventually hit some intelligence plateaus, but we are nowhere near this point. Why can I say this with such certainty? Because we have things that we know will work that we haven’t gotten around to combining yet. Some of this gets a bit technical, but a nice way to think of it is this. Right now, we are mainly using hardware designed to generate general graphics that we have hijacked to use for machine learning. The usual speedup when we go from using generalized hardware to specialized is about 5 orders of magnitude(10,000x). That kind of a gain has huge implications in the AI/ML world. This is just one out of many known improvements on the horizon, but it is one of the simplest to wrap your head around. I don’t know how familiar you are with things like crewAI or autogen, but they are phenomenal, they absolutely crush all of the greatest base LLMs, but they are still a bit slow due to how many LLM calls they take. When we have a 10,000x speedup(which is pretty much guarenteed), then everyone will be able to instantly use enormous agent frameworks like this in an instant.
I understand wanting to see humans as having a monopoly on “intelligence”, but quite frankly that era is coming to an end. It may be a bumpy ride, but the sooner humans learn to adjust to this new world, the better. I don’t think it is something that someone can really make someone else see, but once you do see it, it is very obvious. I suggest you check out the cutting-edge agent stuff out there and then imagine that the most impressive stuff will be routinely done from a single prompt in an instant. Then, on top of that, consider that the base LLMs that we have now are the worst there will ever be. We are in for a very wild ride.
It’s also “discovered” multitudes more that are complete nonsense.
Yeah, that’s the nature of discovery. Humans also “discovery” tons of things like chess strategies that are complete nonsense. Over time, we discard the most nonsense ones and keep the good ones as best as we can. It just turns out that this process is done way faster and efficiently by machines. That’s why nobody thinks humans are going to surpass AI at chess, go, poker, protein folding, matrix multiplation algorithm creation, and a whole bunch of other things.
Just a slight correction. ML/AI has aided in all sorts of discoveries, GenAI is a “remixing of existing concepts”. I don’t believe I’ve read, nor does the underlying principles really enable, anything regarding GenAI and discovering new ways to do things.