Large language model AIs might seem smart on a surface level but they struggle to actually understand the real world and model it accurately, a new study finds.
A few weeks back I got a parking ticket because I believed a google search result. Parking is free on Sundays and holidays, but the city’s website doesn’t specify which holidays. Google insisted that Halloween is a holiday and thus parking is free, but it isn’t actually federally recognized, which I found out the hard way.
They’re just like me fr
The headline is misleading. By “real-world use” they mean using ChatGPT and Claude for street navigation in New York. Which is one very specific use-case.
there is nothing misleading about the headline. street navigation is quite primitive use case compared to what some others were suggesting (like firing stuff of suicide hotline and replacing them with chatbots).
while machine learning can no doubt be useful tool for many of narrowly specified specific tasks, where all you need to do is evaluate lot of data and find pattern in it, the business behind it acts as if it already had invented GAI and unfortunately will keep pretending that and probably cause lot of damage in hunt for money.
I agree there is a lot of marketing BS around LLMs right now. But I would argue that they are quite useful for e.g. basic language and coding tasks and at least for me these are real-world use cases too.
We can’t even teach humans a coherent world model.
I would argue humans often have a world model that is too coherent. If you ask a flat earther about their beliefs they will always argue that the earth is flat and evidence to the contrary is manufactured or interpreted wrongly. That is a completely absurd world model, but perfectly coherent.
Because the world is full of contradictions
An important characteristic of a model is “stability.” Stability means that small changes in input produce small changes in output.
Stability is important for predictability. For instance, suppose you want to make a customer support portal. You add a bot hoping that it will guide the user to the desired workflow. You test the bot by asking it a bunch of variations of questions, probably with some RLHF. But then when it goes to production, people will start asking it variations of questions that you didn’t test (guaranteed). What you want ideally, is that it will map the variants to the best workflow that matches what the customer wants. Second best would be to say “I don’t know.” But what we have are bots who will just generate some crazy off-the-wall crap, and no way to prevent it.
Cue tech obsessives trying to defend or deflect from LLMs etc and their problems in 5…
What exactly is there to defend or deflect, LLM’s are useful for some things and not for others, this is well known
Wow. How did you know?
😉
As such, it raises concerns that AI systems deployed in a real-world situation, say in a driverless car, could malfunction when presented with dynamic environments or tasks.
This is currently happening with driverless cars that use machine learning - so this goes beyond LLMs and is a general machine learning issue. Last time I checked, Waymo cars needed human intervention every six miles. These cars often times block each other, are confused by the simplest of obstacles, can’t reliably detect pedestrians, etc.
I think they mean WE struggle to understand these things have no understanding, probably because they are struggling with it also.
it guesses the next word, that is literally all it does, it’s not trying to build a model of reality to more accurately guess. It has no fidelity and anyone taking it seriously has themselves failed the turing test.
I am by no means an AI fanboy, and I extremely dislike the fact that it is in the hands of big tech, uses so much energy and is built on the work of people who are not being rewarded in any way. It is a new technology that is being forced and abused in the most capitalist way possible.
I do think however, that what you declare here as fact is not as certain as you make it out to be. Research indicates that machine learning models do in fact form some sort of model of understanding of their problem domain. For example this research. I am all for being critical of AI, but oversimplifying the issue might not work in our favour.
Wow, a video just came out that explains my position on this topic almost perfectly
https://youtu.be/AqwSZEQkknU?t=273
tl;dw: I tried to time stamp the exact point …ok, You generally can’t deduce the rules of an underlying reality from an emergent level. She calls it decoupling of scales, and it’s essentially the same problem I have with simulation theory. These programs might form a model of reality but that reality would be at best human produced descriptions of reality and most likely just a model of how best to guess the next word.
tl;dr: put glue on your pizza to stop the cheese sliding off
That’s a very interesting point of view, and indeed well formulated in the video!
I don’t necessarily agree with it though. I as a human being have grown up and learned from experience and the experiences of previous humans that were documented or directly communicated to me. I can see no inherent difference with an artificial intelligence learning on the same data.
I never did all the experiments, nor the research previous scientists did, but I trust their reproducibility and logical conclusions. I think on the same way, artificial intelligence could theoretically also learn these things based on previous documented findings. This would be an ideal “général intelligence” AI.
The main problem I think, is that AI needs to be even more computationally intensive and complex for it to be able to get to these advanced levels of understanding. And at this point, I see it as a fun theoretical exercise without actual practical benefit: the cost (both in money, time and energy) seems far too large to eventually create something that we can already do as humans ourselves.
The current state of LLMs is one of very basic “semblance” of understanding, and close to what you describe as probability based conversation.
I feel that AI is best at doing very specific tasks, were the problem space is small enough for it to actually learn the underlying model. In the same way I think that LLMs are best at language: rewriting text or generating stuff. What companies seem to think though is because a model is wel at producing realistic language, that it is also competent at the contents of what it is writing. And again, for that to be true, it needs a much more advanced method of calculation than is currently available.
Take this all with a grain of salt though, as I am no expert on the matter. I am an electrical engineer who no longer works in the sector due to mental issues, but with an interest in computer science.
I as a human being have grown up and learned from experience and the experiences of previous humans that were documented or directly communicated to me. I can see no inherent difference with an artificial intelligence learning on the same data.
It’s a massive difference in scale. For one, before you even leave the womb you have millions of years of evolution shaping the initial structure of your brain. Then your “training” begins, but it’s infinitely richer than anything we’re giving to these LLMs. Sights, sounds, smells, feelings, so many that part of what your brain is learning is what it must ignore. You’re also benefitting from the interactivity of your environment, you can experiment with things and get feedback for what happens. As you get older and develop more skills, you can start integrating them together to do even more complex things, and the people around you will use their own incredible intelligence to specifically tailor your training to what you need as you learn and grow.
Meanwhile, an LLM is getting fed words, and learning how to predict the next word. It’s a pale shadow of the complex lives humans live. Words are one of the more powerful things we have for thinking and reasoning, so if you’re going to go all in on one skill, it’s a rich environment for learning and in theory the contents of all of humanity’s writing probably contains all the information necessary to recreate human intelligence, but our current technology doesn’t even come close to wringing every ounce of knowledge from the training sets.
I agree with this take, well formulated!
In other news, water is wet
“when we try to use a tape measure to hammer in nails it doesn’t really work, so tape measures are useless”
Nah, it’s either a skill issue, or your tape measure isn’t nearly chonky enough.
