Sayan Chakraborty

people keep rediscovering autoresearch. we're releasing a multi-agent upgrade to it. evolution: multi-agent self-evolution harness. give it a scoring function, point either swarms of claude code or claude code + codex + opencode at the problem, and let them research, implement, compete, and share knowledge until the score stops going up. claude code and codex have different strengths. why not make them work the same problem and compete/collaborate with each other? agents share a knowledge-base of hypotheses, dead ends, and reusable skills. every attempt builds on the best known baseline through periodic convergence. north star can be a verifiable metric, llm-juries, or both. multi-metric composite scoring so agents optimize real tradeoffs, not a single number. what we learned running it internally: human steering isn't optional. agents stray from premise. one message to all agents saves hours of wasted exploration. you need a fast cheap eval subset, multi-agent llm-as-a-judge eval every few minutes will bankrupt you. run the evolution cli via claude code as your operator. you get proactive updates on mobile and can steer the swarm from your phone while agents grind. parallel agents don't eat your disk either. evolution uses reflink cloning and falls back to git worktrees with auto-symlinked dependencies. been using it to churn through all our tokens. github.com/aadi-labs/evol…

also if someone can refer us to YC, DMs are open!

They kick you out of SF if you dont do this

the real reason most ppl never start a company isnt fear of failure its fear of looking stupid in front of ppl who never tried anything themselves once u stop caring about that audience ur free

On to the next one

What an absurd outcome. I suspect this has something to do with how vertex is structured organizationally inside google. Vertex absolutely needed to be set on fire 4 years ago. Kudos to Exa tho!

Life update: Moved to SF. Put a ring on my co-founder. Building out of Canopy at @fdotinc . We're building self-improving ai products that gets better everyday for the user and at population level. The dominant AI future will be divergent ai systems shaped by user's taste, judgement and goals. Pluralistic alignment, recursive self-improvement and proactive agency are going to be table stakes.

New work with @AlecRad and @DavidDuvenaud: Have you ever dreamed of talking to someone from the past? Introducing talkie, a 13B model trained only on pre-1931 text. Vintage models should help us to understand how LMs generalize (e.g., can we teach talkie to code?). Thread:

@FrancoisChauba1 diloco

Heard this today in CS336: "LLMs from scratch" (which is probably the best course on LLMs out there) The unit of compute used to be a single GPU (V100 back in the day). Then it was a rack of 8xA100s. Then a grid of 64. Now its an entire datacenter to consider. Next will be multidatacenter? Then I think about the breakdown of all compute in the world: Smartphones 54% Laptops 28% Datacenters 18% whats the logical conclusion? someone should do that startup

😦

Announcing the Generator Residency: a 3-month residency for AI safety generalists, by @KairosAIS × @ConstellOrg. Fully funded. In-person in Berkeley. Summer 2026. 🗓 Apply by April 27 generatorresidency.org/?utm_source=tw…

“gpt2-large is too powerful to be publicly released” vibes

Petition to change "et al." to "and gang" in academia

I still use Manus for slide generation. I have tried Gamma, Canva, Claude, etc. None of them come close to Manus's. I don't know if its because they have great context skill for slides or post-training.

@neomewtwo why not both. competing/collaborating in hill-climbing. github.com/aadi-labs/evol…

codex auto research is 10000x better than claude u just gotta hook it

@mrmagan_ @tambo_ai damn you guys move fast!

@shockrobortyy @tambo_ai done x.com/mrmagan_/statu…

generative ui in production. we use tambo in our own console. it's tambo all the way down. just ask: - create a project - generate an API key - show me my latest messages - add an MCP server build an agent that speaks your ui with tambo.

@WilliamBryk "Our goal is perfect search over the world's information, which we believe is critically needed civilizational infrastructure... our own Project Hail Mary" very very cool

We booked a whole IMAX theater in SF for the premiere of "Project Hail Mary". We have room for 250 people. Join to see how to save the world from impending doom, link below 👇

it's wild how VCs say that they "knew" some founder was going to succeed as if they don't have 10,000 losses for every win

Somewhere out there is a guy who uses Notion, Superhuman, OpenClaw on a Mac Mini, Raycast, a mechanical keyboard ($400), Wispr Flow, and gets nothing done every day

claude code's superpowers plugin, especial brainstorm skill is magical and addictive. fun to push a feature and talk to cc like a tech lead. i am unaware of an equivalent experience in codex/opencode/pi