The self driving car is a great example, you're right, it was so good and yet never lived up to the hype.

Perhaps one difference is that a human could potentially get extremely good at textual tasks with nothing but text to learn from. You can read how to solve cryptic crosswords, see examples and extrapolate. In that sense language models have a somewhat complete training dataset. Yes this requires an understanding of the material, rather than just parroting, but the signal is there if you can separate it from noise.

Driving a car requires an understanding of a much wider context which is perhaps hard to acquire with just driving data. Understanding of rain, birds on the road, shaky drivers, balls rolling out from between cars, lane restrictions... You can't just throw petabytes of data at the problem. Training data is limited and expensive, and I believe we are mostly tackling AI-assired driving with rule-based approaches.

I believe self driving works just fine in simulations where data is effectively unlimited. But then it doesn't generalise to the real world where context matters.

Well, I would say language isn't a single task but a system, process or tool that's flexible enough to aid in many tasks at many different levels. It be used to signal social status and education or to guide someone through fixing a flat tired and often its used on these multiple level simultaneously.

One can be succeed at one level of using language without succeed at another level. But we humans expect another human to succeed or to fail fairly uniformly - or we call them "a bullshit artist". These expectations may not be met with large language models.

In that sense language models have a somewhat complete training dataset. Yes this requires an understanding of the material, rather than just parroting, but the signal is there if you can separate it from noise.

--> I'd agree that there's conceivably more that can be done with a language dataset. But the training process of transformers-based models isn't really oriented to engaging in the process you describe. It treated language as data and it's fundamentally a very sophisticated database that only appears to engage in such logic. As it will tell you.

I guess I'd say that a sufficiently (!) accurate language model is indistinguishable from a AGI limited to text. The question is, is that level of accuracy achievable. I'm not claiming, btw, we are close to that.

By contrast it feels like driving requires an understanding of effects that are really hard to distill from pure driving data. Not just analytically hard, but requiring an understanding of external context.

But it's certainly a fine point.

I think the problem of self-driving cars is that the driving problem isn't entirely sealed and success requires a long string of successes (most of which seem trivial). I think the situation with language is that it too isn't entirely sealed from other parts of reality but here each immediate success is judged a victory. A lot of self-driving process are straightforward, "just" adaptive control. Minimally self-driving vehicles have existed for a long time. Corner cases are the problem - distinguish a newspaper floating into an area from a load of bricks falling off a truck in front of you, etc.

For language, a problem to consider for a language using system is "talking a person through" a task. The thing about that is it involves two entities sharing a common model of reality and each updating their model as they listen to the other. And here I think corner cases of reality are basically as likely to show up.

If there were enough cars fitted with sensors that can learn from their driver's actions and reactions I am pretty sure an AI system can learn from this huge pool and be a good driver. A good driver also makes mistakes. And this is still nothing related to intelligence. Our networks are as of now what one calls universal approximates. Actually replying to so many different queries whether it is medicine, computer science, history , it almost seems like a better interface to the net than google

> Driving a car requires an understanding of a much wider context …

Yeah, it struck me that perhaps a better approach to self-driving would be the road infrastructure and all the vehicles using it cooperating to build a local activity map - that way each individual vehicle doesn’t have to detect, classify and route around every object in real time, most of the work will have been done by previous vehicles and / or stationary traffic cameras.

Self-driving cars need the neural networks to run on computer that are orders of magnitude slower with a tiny percentage of the memory though.

If each self-driving car had its own supercomputer, things might look different.

Is amount of compute an the issue preventing self driving cars from being viable? I don't think so. Putting supercomputer in the car won't bring self driving cars.

Compute limits what you can run realtime. Not saying self-driving cars would necessarily be here today with more compute, but comparing with language models, self-driving car computers could run only models like GPT-2. There's quite a big difference from that to ChatGPT. Is self-driving an easier problem that requires less than language models? Perhaps. language models somehow try to compress all the knowledge on the internet (not entirely successfuly yet), so maybe that somehow needs much more than self-driving, a single human can drive a car after all.

But self driving cars are on the roads, and they are expensive hardware that can kill lots of people. Mistakes need to be exceptionally low so it's a lot of order of magnitude improvements but minor visible changes before it hits major thresholds.

I'd also argue that LLMs and image ai has grown far beyond linearly over a fairly short time horizon.