Leo Gao
@nabla_theta
working on AGI alignment. prev: GPT-Neo, the Pile, LM evals, RL overoptimization, scaling SAEs to GPT-4, interp via circuit sparsity. EleutherAI cofounder.
608 Following 16.4K Followers
the cope will increase until morale improves
we choose to believe these things, not because they are pleasant, but because they are true
"it's crazy how there can exist something so incredibly intelligent and yet hobbled by maladaptive post-training", remarked the burnt-out bay area ai researcher,
x dot com (the everything app) is a casino for people who care more about status than money
your (lack of) circuits makes me nervous
happy petrov day - may we yet have many more years of not destroying the world
@corsaren if you think the external naming is bad, the internal naming would drive you crazy
the two most valuable skills before the singularity might be the ability to think clearly about conceptually tangly problems and high-stakes decisionmaking under pressure
it's crazy how much of human behavior is actually secretly avoidance of unpleasant emotions
the capacity to pause is very convergently useful.
even making an aligned AI that is superhuman at alignment research is not a sufficient win condition. it may successfully create a smarter AI, and then need a long time to align that AI, but race dynamics force yoloing sth half baked.
Show more
just as capabilities is spiky, alignment is also likely to be spiky. models will probably be very well aligned in some domains and extremely misaligned in other domains at the same time
right now i mostly tweet memes and takes on the Discourse. should i tweet more random alignment takes that might be more in the weeds technically?
"hackable software is less useful. so really if you think about it usefulness and security are actually kind of the same thing. by shipping 100 new features this month, we have made our most secure product ever"
*gets instantly hacked due to enormous vulnerability surface area*
Show more
"wait, it's all cope ..?"
"always has been"
project hail mary: science, fuck yeah! everything always works out
hpmor: human nature is full of biases like motivated reasoning and tribalism, but we can overcome that and achieve greatness
three body problem: you really thought we could have good things? lol. lmao even.
Show more
why are constant nonsensical blabberings called bullshit and not Lipschitz constants?
oh man i’m sure that accenture will have lots of expertise at, uh, *checks notes* redteaming models, assessing alignment, and testing safeguards. after all, their website talks about embracing the power of change to create 360° value and shared success for clients, people, shareh
Show more
We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years.
Show more
“we must go faster to beat china [we invent novel algorithmic improvements and immediately accidentally leak them to china via poor opsec, its not clear what kind of move we were trying to do]“
codex, please explain this paper but replace all of the math with pytorch notation
New blog post!
Making AI go well will require international coordination. Humanity has risen to the occasion once before: the ozone crisis bears a striking resemblance to AI risk. Understanding the past can help us navigate the future better.
Show more
