Comment by dyauspitr
5 hours ago
Do you agree that if physical LLMs are as good at replicating human behavior as text based LLMs are, they would be able to do a lot of the vast majority of what we already do in the real world?
5 hours ago
Do you agree that if physical LLMs are as good at replicating human behavior as text based LLMs are, they would be able to do a lot of the vast majority of what we already do in the real world?
I'm not entirely sure what a physical LLM is. Would you care to elucidate how they work and when they're likely to be doing my gardening?
Sure physical LLMs are when an LLM is trained on data where it outputs control bits for actuators on an actual humanoid robot instead of text like you see in a chat bot.
How likely are they to do your gardening if you ask me I would optimistically say within the next six months. For examples of working instances of this look up Figure robotics they’ve had live streams where it sorts mail and cleans houses although the cleaning of the houses is very early stage, and it kind of only does contrived situations. Look up Sunday robotic’s live stream where they literally have a 99% accuracy on folding nine different types of garments. Look up Unitree where they’re making progress on a person asking a robot to do something and it coming up with a novel action on the fly.
Self driving cars started appearing after Darpa 2005. Even after 21 years we don't have fully self driving cars around the world. And Driving is a "easier" robotics problem compared to other robotics problem.
Human body is far more complex than human brain. The real world is far more complex than our artificial world of "concepts" that these LLMs are dealing with. Robotics company are showing a robot working in a clean environment not the edge cases. Real world is all about the edge cases. I think humans(animal) can perform in real world is because we have biological instincts(survival, reproduction etc) which are missing from these machines. And maybe machines need this instinct to flourish in real world.
Code is a easy problem as it can be verified. I would be amazed with AI can write a good short story.
1 reply →
LLMs are inherently text-based. For physical labor, some other deep-learning model will be used.
Tokens don't need to represent text.
They are sequence to sequence models.
No, they are input and output based. They work with tokens/bits which could be anything. It can be a paragraph of text or it can be a bitstream for actuators. This isn’t science fiction. It’s happening right now.