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
https://t.co/zHij5zwKGT
"Oh this? This is just an entity from the nether realm I summoned by enscribing runes on magic stones I forged in a sacred kiln, then proceeded to cast certain spells to distill the collective conscious to imbue it into the substrate. Don't worry, it doesn't have a soul."
@Lihong062 Economically speaking how is it possible for everyone to be rich? Like will I actually be able to buy a mansion and multiple hypercars or is that still reserved for a class above the regular rich people
@zephyr_z9 Or it could end up more like how Netflix streams video to millions of households/phones everywhere, but for AI: many many specialized, small inference systems piggybacking telecomm infra (towers etc) and user requests r bounced like cell signals/streaming video/music/etc
Guys hear me out: Jev, but multimodal!
Imagine there was a model that you can give any image, and any set of freeform texts, and it would return you for each text, how well it fits the image.
It could also do a set of yes/no (Noul) by using sigmoid in the right places.
And the scores could be calibrated!
I think i should be able to do this, though i might need to raise xxxM. But as one of the inventors of vision encoders but also one of their biggest haters because i know better, i can pull this off.
And now imagine i would manage to open-weight it, wouldn't that be awesome?
And EVEN BETTER, imagine i could do it with fewer than 1B params?! Maybe like 400m or so?
I think I'm onto something here, wdyt??
@austinsemis I could train a microwave on the internet corpus and it would still burn my toast once every 100 times.
The internet is 99.9999999999% absolute garbage
@GergelyOrosz@deebeenguyen I’m not sure, basically every good role out there requires lots of hackerrank and dsa knowledge as a litmus test for intelligence, even though it’s more just a test of how much free time that person has.
I’m mid-recruiting rn and dsa is a barrier at EVERY good company…