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
Training tiny models for special purpose use cases works so incredibly well if you have a great self improving recursive flywheel. Shopify ML team is on fire.
finetuned 0.8b model beats GPT 5.6-sol xhigh in this very specialized task.
@aidenybai You need to solve UI architecture before you can solve UI testing.
Foldkit Scene tests are pure, fast, and test what you actually care about. No browser, no JSDOM, no mocks.
https://t.co/EzQqThm0nZ
been trying out https://t.co/l2agiAlNM8 today, it's really interesting.
initial thoughts:
- agent really liked it. cursor w/ 5.5 one-shot a bunch of stuff that was super annoying before that i was trying to (keep a persistent music player playing across different pages, share state, load state back from local storage, etc).
- it's all effect, which i like. dunno how well it integrates further down, gotta try some more httpapi integration stuff but should be fine w/ httpapi client
- seems to handle a bunch of stuff that normally felt really finnicky with getting right with effect, schema stuff, state thing,s blah blah
- the dev mode is insanely cool, can replay mesages and everything, super sweet
idk probably more to think of later but been enjoying it a lot
ty @devinjameson love the direction
Slavery in this country lasted 246 years. Jim Crow lasted around 100 years.
Our attempt to remedy the impact of those crimes against humanity was met with white terrorism & savagery.
The laws we finally got passed to address these harms were relentlessly challenged & lasted only 61 years before they were all finally stripped away.
This country has decided that the idea of race can be used to multigenerationally oppress us, but any remedy must be race neutral, which means there can be no remedy.
Never again will I vote for a presidential candidate who is unwilling to admit that America—a country that wouldn’t exist without chattel slavery—is a racist nation. That’s the beginning for me, before we even get to policies & legislation.
Today’s Supreme Court decision effectively guts a key pillar of the Voting Rights Act, freeing state legislatures to gerrymander legislative districts to systematically dilute and weaken the voting power of racial minorities - so long as they do it under the guise of “partisanship” rather than explicit “racial bias.” And it serves as just one more example of how a majority of the current Court seems intent on abandoning its vital role in ensuring equal participation in our democracy and protecting the rights of minority groups against majority overreach.
The good news is that such setbacks can be overcome. But that will only happen if citizens across the country who cherish our democratic ideals continue to mobilize and vote in record numbers - not just in the upcoming midterms or in high profile races, but in every election and every level.
I will NEVER forget the Trump administration blaming a plane crash that killed a plane full of kids last January on a fictional Black person with Karoline Leavitt saying “When you are flying on an airplane with your loved ones, which every one of us in this room has, do you pray that your plane lands safely and gets you to your destination, or do you pray that the pilot has a certain skin color? I think we all know the answer to that question, and as President Trump said yesterday, it's common sense."
This man will go down in history as one of the best pilots to ever do it.
early in my career when i was learning a new tech or language i would tinker and google whenever i hit a roadblock
eventually i realized books had all the information i needed pre-googled for me
i think this is happening again with LLMs - sometimes i waste so much time letting the LLM keep taking swings instead of reading something
hope the industry doesn't abandon producing good reading material