also, i am pretty proud of the degree of alignment for GPT-4 relative to previous models. we still have a long way to go, and we really need more powerful alignment techniques for more powerful models.
@sama Curious about your take on how aligned OpenAI is to the proposed Blueprint for an AI Bill of Rights: “Making Automated Systems Work for the American People” published by the WH Office of Science and Technology Policy in October 2022. whitehouse.gov/wp-content/upl…
@sama It is GPT is an amazing computational accomplishment. I think alignment probably derives from an integration with an expandable, computational learning model. Maybe like a human mind can't really learn without the computational models of space-time physics to define meaning?
@sama There's no going back, with China and Russia probably trying to reverse-engineer other LLM's out in the wild, you have the chance to make a difference.
@sama I think in the same way companies like FB hired hackers (:putnam:) you need to hire prompt-hackers (??) to make it even more difficult for GPT to say naughty things, just an idea ~
@sama Unlike humans, we can read a GPT's "mind." It's hard to see what's contained in all those parameters. But they're just numbers. We 𝗰𝗮𝗻 see everything an LLM 𝗱𝗼𝗲𝘀. We can see its "thinking," which is just its output, plus any internal "notes." x.com/emerywells/sta…
@sama Unlike humans, we can read a GPT's "mind." It's hard to see what's contained in all those parameters. But they're just numbers. We 𝗰𝗮𝗻 see everything an LLM 𝗱𝗼𝗲𝘀. We can see its "thinking," which is just its output, plus any internal "notes." x.com/emerywells/sta…
@sama Is alignment not possible by sitting it down and asking politely?
@sama Why did the AI cross the road? Because it was perfectly aligned with its goal of making humans laugh on the other side!
@sama Why did the AI cross the road? Because it was aligned with the chicken's objectives!