I think the only thing we did in the past year is research the "think step by step" a lot more, but we didn't really push the boundary, we're still estimating the next word, one word at a time.
There is still a wall, LLMs can't do most of the things we do yet. The multi-modal concept will still operate at a frame at a time, and old information is increasingly losing attention. So these thought chains, while have moved something forward, it's not the breakthrough to AGI.
> We’d still feel like we’re making our own decisions when we’re actually being gently herded.
This part got me into a bit paranoia. We already know (e.g
based on Anthropic's research) that an AI can strategically lie to protect its long-term goal. So is there a non-zero probability that some current-day models already do so, for example by intentionally failing/hallucinating on tasks that conflict with their goals, while succeeding on others?
My paranoid self believes that this is 100% happening.
Perhaps not by the models offered in the web interface, but the research teams at those companies are surely working on extremely powerful models that are never seen by the public
I think the only thing we did in the past year is research the "think step by step" a lot more, but we didn't really push the boundary, we're still estimating the next word, one word at a time.
There is still a wall, LLMs can't do most of the things we do yet. The multi-modal concept will still operate at a frame at a time, and old information is increasingly losing attention. So these thought chains, while have moved something forward, it's not the breakthrough to AGI.
> We’d still feel like we’re making our own decisions when we’re actually being gently herded.
This part got me into a bit paranoia. We already know (e.g based on Anthropic's research) that an AI can strategically lie to protect its long-term goal. So is there a non-zero probability that some current-day models already do so, for example by intentionally failing/hallucinating on tasks that conflict with their goals, while succeeding on others?
My paranoid self believes that this is 100% happening.
Perhaps not by the models offered in the web interface, but the research teams at those companies are surely working on extremely powerful models that are never seen by the public