Maybe I am going to lose cred in the LLM whisperer community for this, but I thought Chiang's article was quite good. Chiang, perhaps of all people, should know how powerful a story is. But many of the other potential objections to Ted's frame are covered.
Someone needs to make a benchmark called "ReviewBench" where basically it teaches LLMs to stop approving PRs that humans rejected. "But this benchmark is unfair, no one wrote down X constraint." Yes. That is EXACTLY the point.
Some discourse around Erdos was about a "hint book", where each hint was effectively one bit of information for the LLM. "Look for a counterexample." "Generalize the best known counterexample." This idea feels very important for elicitation beyond math. https://t.co/NLE8qfwUqL
Importantly, this is even after post facto we discovered already publicly available models could have made this discovery! From this I infer there is a built in slow down in capability diffusion predicated simply on elicitation ability
Importantly, this is even after post facto we discovered already publicly available models could have made this discovery! From this I infer there is a built in slow down in capability diffusion predicated simply on elicitation ability
This makes the "if anyone can make a bioweapon" xrisk argument less scary. It does NOT make the "extremely motivated adversary" (crime, governments, etc) xrisk less scary, but this scariness feels more "priced in" in terms of traditional geopolitical risk (eg nuclear)
Here's an argument why LLM based biological xrisk will have a warning lead time. If any random joe could elicit model capability, we would have seen major AI math breakthroughs from random cranks. But instead OAI got there first.
Importantly, this is even after post facto we discovered already publicly available models could have made this discovery! From this I infer there is a built in slow down in capability diffusion predicated simply on elicitation ability
The problem with token efficiency maxxing is you spend all your time building harnesses to overcome the model problems and not enough time actually getting shit done
Many strange things happened when you scaled things like "make it easy for people to talk to each other" or "tell people about things they might want to buy". Would you have predicted the rollout?