Comment by godelski
To the best of my knowledge every major autonomous vehicle and robotics company is integrating these LVLMs into their systems in some form or another, and an LVLM is probably what you're interacting with these days rather than an LLM. If it can generate images or read images, it is an LVLM.
The problem is no different from LLMs though, there is no generalized understanding and thus they can not differentiate the more abstract notion of context. As an easy to understand example: if you see a stop sign with a sticker that says "for no one" below you might laugh to yourself and understand that in context that this does not override the actual sign. It's just a sticker. But the L(V)LMs cannot compartmentalize and "sandbox" information like that. All information is equally processed. The best you can do is add lots of adversarial examples and hope the machine learns the general pattern but there is no inherent mechanism in them to compartmentalize these types of information or no mechanism to differentiate this nuance of context.
I think the funny thing is that the more we adopt these systems the more accurate the depiction of hacking in the show Upload[0] looks.
[0] https://www.youtube.com/watch?v=ziUqA7h-kQc
Edit:
Because I linked elsewhere and people seem to doubt this, here is Waymo a few years back talking about incorporating Gemini[1].
Also, here is the DriveLM dataset, mentioned in the article[2]. Tesla has mentioned that they use a "LLM inspired" system and that they approach the task like an image captioning task[3]. And here's 1X talking about their "world model" using a VLM[4].
I mean come on guys, that's what this stuff is about. I'm not singling these companies out, rather I'm using as examples. This is how the field does things, not just them. People are really trying to embody the AI and the whole point of going towards AGI is to be able to accomplish any task. That Genie project on the front page yesterday? It is far far more about robots than it is about videogames.
[1] https://waymo.com/blog/2024/10/introducing-emma/
[2] https://github.com/OpenDriveLab/DriveLM
Many large companies have research departments that do experimental work that'll never get to the product. This raises prestige, increases visibility and helps hire smart people.
Things like Waymo's EMMA is an example of this. Will the production cars use LVLM's somewhere? Sure, probably a great idea for things like sign recognition. Will they use a single end-to-end model for all driving, like EMMA? Hell no.
Driving vehicles with people on board requires an extremely reliable software, and LLMs are nowhere close to this. Instead, it'd be usual layered software - LLM, traditional AI models, and tons of hardcoded logic.
(This all only applies to places where failure is critical. All that logic is expensive to write, so if there is no loss of life involved, people will do all sorts of crazy things, including end-to-end models)