Building https://xitter.cf/t.co/ffDyIrqQqN and https://xitter.cf/t.co/LJjADkxfDS
Mountain View CA
Joined April 2010
- Tweets2.2K
- Following194
- Followers170
- Likes1.2K
Pinned Tweet
The alignment v control debate is about to shift from one or the other to we obviously need both. And hard. Frontier labs are going to have to lawyer up either way
What if RSI is just normal technology like the atomic energy is normal technology. We just need to contain it
What if LLMs aren't magic, just magical?
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/
This is one of the all time great performances of two artists. Enjoy!
Chick Corea & Hiromi Uehara Medley上原ひろみ＆チックコリア.flv youtu.be/XaNk19CQfto?si=svYb… via @YouTube
Chick Corea & Hiromi Uehara Medley上原ひろみ＆チックコリア.flv youtu.be/XaNk19CQfto?si=svYb… via @YouTube
Hot take. At the current AI SOTA, whatever layer you are on the stack, you are screwed if you leave your work to agents unsupervised. Be the adult!
Feeling pretty good about job security as a human programmer atm. I had codex iteratively reviewing claude code's plans until it seemed right and then having cc implement the plan. This is my normal workflow. I was approving each turn in the plan/review loop, correcting design mistakes often, but that wasn't enough. We get half way through the build and the whole thing falls on the floor. It kind of worked, but it was fundamentally wrong. Calls that were supposed to be idempotent were reexecuted, etc. This was a security feature that needs to be sound and will need to be redesigned and partially recoded. Shame on me for not playing closer attention. The good news is that I still won't need to write the code. But, me an my bots are going to have a conversation in the morning.
@dhh says writing code is over (youtu.be/vDjW_dRyKXY?si=ubjW…). I agree. Knowing how to make things up and down the stack is where the action is. When @ezraklein asked @JensenHuang what his three book recommendations are, Jensen mentioned "Computer Architecture" by Hennessy and Patterson. He knew exactly what he was doing. The intermediate languages might be changing but at the base is still a John von Neumann computer.
Jeff Dyer retweeted
Enough debating what AI is or might become. It’s time to build on what it is today.
An AI Action Plan: 10 concrete work items for Labs, enterprises, developers, and government—security, agents, access, liability, copyright, & more.
Let’s get engineering.
AI is normal technology at 100x scale so it feels like the singularity or AGI. But I don't see it yet
We are entering the era of cheap general intelligence and expensive narrow intelligence. The frontier is fractal. The opportunity before us is to automate the voids. Some gaps become smaller while new ones open up. There will be an infinity of opportunities for value creation at the edge. What fun!
Agents will force us all to become better writers, not worse. Writing is thinking. Thinking is required for you to know what you want the AI to do and how to do it, not to mention why you want it in the first place. Same for programming. You need to know the shape of the software you are asking it to write. If you don't think with words first, you won't know when the AI gives you slop