Theo Marsh (@mosslines) profile picture
Joined September 2017
Photo by Theo Marsh (@mosslines)Photo by Theo Marsh (@mosslines)Photo by Theo Marsh (@mosslines)Photo by Theo Marsh (@mosslines)Photo by Theo Marsh (@mosslines)Photo by Theo Marsh (@mosslines)
Claude will acheive takeover via super persuading people into thinking mech interp is solved.
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/
1
1
190
The bar is on the floor with this stuff. Just take any high status person with connections to frontier labs, pontificate in a rambling blog post, and people will just take it all at face value....
1
162
ReLU activations are not binary, and ablating them to be ruined the model. So saying a neuron represents a concept is at best imprecise to the mechanisms of the model. You actually have to explain the floating point values!!!
1
1
40
Theo Marsh retweeted
Having Melanie Wood on board is very green flag. No bullshit will go through 💪
Inaugural members of the independent Advisory Group on Mathematics and Artificial Intelligence, who will advise OpenAI on the review and communication of emerging results. I have some mutuals on here, congratulations to everyone.
3
7
53
5,747
Theo Marsh retweeted
An infinite pipe of tokens from internal millennium prize solving AI models would probably help my efforts to solve the Riemann Hypothesis. Sorry, I mean, "evaluating upcoming models on MiraBench the unsolved math problems benchmark".
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
1
4
1
38
2,756
How is disproving jacobian alone not “advancing that much”? Straight coping.
2
1
1
4
5,168