Jake Schwartz

The first experimental evidence of recursive self-improvement (RSI). Autoresearching the autoresearch agent for eight days. The result beats the harness we hand-tuned for two years, on held-out benchmarks: 🧵(1/7)

Stepped on sidewalk today. Got a red card. Must sit out tomorrow’s walk. Clear over reaction. Lack Messi’s street cred #USMNT

last night’s fable still reading out this AM on just 3 (!) of 10 multi-level dynamic workflows cloning the Reactor core into OpenProse OS (PrOS it called it), an upgrade (remote-first, agent-native) to the Reactor rebuilt from ground up fable is only 24% (!) of the way into top thread context (uncompacted) after 8 hours of work. RLMs alleviate context rot on long tasks! fable is an absolute unit of a mind. spiky, but where it’s good it sings we’ll see how well it does with PrOS, may be a total slop fest. but the scale and the degree of planning, iteration, and coordination being undertaken in the dynamic workflows is kind of breathtaking (and expensive) — I think the reaction to Anthropic was overly harsh yesterday and misunderstands some things, especially the way steering vectors and other techniques have been applied for years on all the modern models you’re otherwise using and liking it’s true that using steering to amplify a capability is different than using it to suppress one, but also we’ve also been doing _that_ for years too on all the models you’re using and liking in part I think it was a framing misstep from Anthropic, it’s not so much as censoring the model, it’s changing the model. this seems like it should be fair game and even highly useful, to the extent that it makes it an acceptable model for the task at hand the one discussion this does force to a head is the drum that @jeremyphoward is beating. he is absolutely right to do so and I join him in it as a researcher he feels it most viscerally, but he’s a canary for your field too: this is what it feels like to have your sovereignty stripped. to know that the insiders have access to capabilities that will crush your current line of work, and you have none. this is the moment it switches from tool that 100x’s you to adversary that can crush you on cost in some ways, Anthropic has some answering to do on this. in other ways, they’ve already answered for it. this is what Dario has been telling you, whether or not you wanted to hear it Anthropic is the least hated now as it will be again for quite some time, but they should listen to what the universe is reflecting back and then and learn to not speak bad futures that nobody wants into existence — praying this morning that the leaders at the labs walk the next 12 months with love and grace _and_ hope fable ushers us into a new chapter here

@contextconor 🫡 Gotta get investment exposure to the 5′7″ founder segment. 💰💰

announcing 5'7 ventures fiveseven.vc

@contextconor Life is great down here! 📈

credit to @Jake_Schwartz for creating the website (and launching the fund) btw he's 5'7 himself

@jaiyagill fiveseven.vc

before you invest in a startup always ask how tall the founders are

That's right. LLMs architect and research better than you. This might hurt to hear, but it's true. Coding agents are _nearly_ as good as you at driving business outcomes. This is why I need you :)

To the Full Stack Engineer of 2023 I say this: It's early 2026 and LLMs can not only write code 10x faster than you, they program more accurately and better than you.

I'm hiring for a new role: the Full-Stack AI Architect. blog.jakeschwartz.com/the-full-stack…

@xdenuzzo @modaflows @danylo_dev @pranavbedi @legalos @tanagram_ Came for the meat, stayed for the bubbles

i'm hosting a yc darty next saturday. we'll have a private chef, dj, games, and a massive fucking bubble machine. last time I personally grilled >40 lbs of meat. dm for an invite. w/ @modaflows @danylo_dev @pranavbedi @legalos @tanagram_

@Aiswarya_Sankar @EntelligenceAI So many talented folks @ Cal!

Memories of working out of Soda hall at Cal all night for project demos will stay with me forever. Excited that @EntelligenceAI will be cohosting a hackathon with HKN next weekend Saturday, May 9th! We're looking to hire more Berkeley students and would love to chat :)

@anvisha Agent design tools are sorely needed!

Introducing the Moda API: the design layer for agents, apps, and workflows. Embed presentations, reports, and design creation right into your product or agent. See what people are building with it 👇

@thatguybg Good perspective, seeing the M&A deals

It’s 2026. If you’re writing code with your brain, you’re too slow.

@ericzhang98 They've been doing it for at least a year 😅

Has the Pentagon tried running --dangerously-skip-permissions yet?

@DamiDina I think so man! I imagine we're actually above 100k economically productive agents as of this month, if you consider individuals and companies. Exponential from here.

@Jake_Schwartz 3 years for chatgpt...based on that, could be in 1 year or less..

How long until we have 1 billion AI agents doing real work in the world?

@conor_ai Outside, what's that, is that like a new moltbook environment?

claude is down do we… go outside?

Vercel, the backbone of many web companies, is down in Dubai right now vercel-status.com