Not in all conditions. The devil really is in the details.

A truly 100% autonomous vehicle requires a much higher level of intelligence than a self driving vehicle with a driver able to take the wheel when necessary.

Take the case when some work is happening on the road and workers make signs with their hands to tell you to go this or that way. But on the same road there is also someone who is dressed up as a policeman cause it’s Halloween, and he’s waving at some friends.

I agree. However I think there is a gap of meta learning that someone may figure out how to fill. Imagine you could take the state of the art driving AI, deploy it, and then wait until it makes a mistake. And then... suppose you could just explain to it in english what it did wrong like you would interacting with a multi-modal LLM. The missing component (for now) is that AI taking your feedback and adjusting the weights in the driving model to fix that mistake. Not just adding additional training data, but correcting based on more fundamental understanding and abstraction of what just occurred and the key take-aways, etc. and then making sure not to repeat that mistake. Just like a new human driver would learn.

It's possible someone might figure out a way to create a training loop using a multi-modal LLM to generate synthetic training data based on the situation you just explained and then updating the driving model by training on this new data until its performance improves on the task.

> Not in all conditions. The devil really is in the details.

I imagine we will reach a place where fully autonomous vehicles will pull off to the side of the road in certain weather. Which I wish we could force for humans, but seems infeasible to implement.

Right, it's not about the conditions, it's about the ability to perceive those conditions. However this asserts that we won't change how we construct, manage and maintain the routes of travel. Is it unlikely roads will become retrofitted in an effort to enhance the needs of autonomous vehicles?

Seems likely to me; we built EV's before we built the infrastructure to support them and conversely we created in infrastructure for petroleum vehicles before they entered mass production at the largest of scale in the period (1950's) after the fact.

I don't see how autonomous vehicles are not going to become a reality. Perhaps not in my lifetime, but, absolutely likely and possible.

Based on what? All the evidence I see suggests there is no clear path to full autonomy. It's within the realm of possibility, but certainly not inevitable.

> 100% autonomous ability is well within reach.

I don't doubt that, but the timeframe is unknown. 5 years? 10 years? Within our lifetime?

We have a problem with data - even when we apply our newest advanced technologies to put data in ever increasingly small environments (even down to the atomic level), entropy still exists.

So what we have is the ability to input data, but not yet a delivery system and retrieval system that can fit on say, a small chip, or light array, or other small systems.

It's a giant part of reason we'll see diminishing returns with data being applied in classical material approaches. New materials (currently being workd on, like graphite and others) will be needed to harness the compute power to enable large scale data capabilities at increasingly smaller and smaller levels (already a well known issue to be resolved).

Similar in physical approach but different in application would be TinyLM.

Timeline? Not sure but we created ION drives over 30 years ago. Seems to me we're limited not by the science, but the material needs to continue technological advancement. Seems to me autonomous driving is within reality in under 25 years. If I had to put a guestimate on it.

Hello my dear old friend, automatic car

I’d like to see the strong case for autonomous driving being the harbinger of generalized AI laid out. Specialized AI, with varying hardware support, has yet to solve the last-mile of 5-10% of autonomous driving in favorable conditions. In snow, sleet, hail, dusty, smoky, foggy or some rainy conditions, progress has not been commensurate. Yet somehow this is the success template for generalized AI applications. I’m missing the chain of logic here.

A lot of this breathless talk surrounding this turn of AI is so uncomfortably reminiscent of what I’ve seen before in the mainstream the last turns around the 1970’s and 1980’s, and the potential failure mode might not be so different: solving the last 5-10% is tantalizingly close but remains stubbornly out of reach of calls for the “more cowbell” of each era or call to action by the sales legions (currently cowbells look like NVIDIA boards and various counts of AI models be it tokens or what have you), and the last 5-10% is the necessary advance to cross the chasm.

I love and use the tech myself every hour, but it has deep gaps I don’t see being resolved even incrementally between versions or competitors.