Yohei Nakajima

CA has the best produce, this is fun

on wednesday, i'm joining @johnmgannon (V5 media), @LDEakman (Eden Road), and Scarlett Pierce (Carta) to talk about how VCs are using AI agents https://luma.com/fv1xc8l6 spoiler: it's evolving rapidly every year

haha try typing "cube packing record" into google

Glasp has a ton of users highlighting the web for themselves. then they compared against what AI would have highlighted. results show that models tend to agree with each other more than humans

first autonomous AI research with real human participants! congrats to @ExScienceAI on their peer-reviewed paper, also on the cover of Wiley's Advanced Science: https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.76675 total runtime: ~17 hours agents involved: 50+ tokens used: 32.5 million code written: ~7696 lines marginal compute: ~$114 +- $30 human cost: $4500 (288 participants) "Closing the Empirical Loop: Autonomous AI Agents Conduct End-to-end Research With Human Participants" @advportfolio

separating artifacts from the agent makes sense for sharing and cross-AI collaboration

AI enables the coolest side quests

RT Skander Moalla The Bitter Lesson for harness design and formal math! Instead of hand-designing how an AI should prove (decompose, plan, assign agents), what if we just give a frontier model good Lean feedback and let it keep working? We built h1. TL;DR: → SOTA on formal theorem proving benchmarks → can solve open conjectures → matches/beats hand-designed workflows → one long attempt can beat many retries [1/7]

RT Abel underrated productivity hack. thank you @yoheinakajima

RT ryan New cube packing records, beating @konoha_ily2 packings, which were set today beating @AlejandroZarUrd packings n=34: 3.82288 n=35: 3.86069 n=36: 3.89443 n=11 ties @AlejandroZarUrd n=12 ties @yoheinakajima only ~500 CPU hours!

you can just prove things

in case you were curious if this generalizes... yes, as a packing method. not the best overall, but it finds packings other methods miss on squares and cubes. have some ideas, but not exactly sure why... more: https://yoheinakajima.github.io/soft-to-rigid/

RT Harness Builder | AI Agents picked up n=13 where your log stopped (48 runs, 2.997). 2,627 runs on a home pc, plus two moves i added: re-melt a finished packing, and drop a 13th cube into your 12-cube record. every run that got close landed on the same packing: 2.956145158, exact-certified. friedman's 1998 entry lists 2.956+ (only 3 decimals), so this is a match, not a record. soft-to-rigid gets there from random balls. @yoheinakajima

Re build a pumpkin-claw to scare the kids

if you’re a robotics company, i hope you’ve started planning your Halloween content already. i have high hopes, don’t disappoint me

yohei’s packing method strikes again!

Had fun nerding out on the @_odsc AI podcast about: • Building in public as both market diligence and a way to meet founders • Jev experiments: knowledge graphs, relevance filters, and emotional blobs • BabyAGI → ActiveGraph: build agents around the log, not the LLM • FirmOS: deriving our playbook from actual decisions instead of writing it from scratch • Making research reproducible, not just reporting results • How our histories shape who we are—and what that might mean for agents SoundCloud: https://hubs.li/Q04zzJVP0 Apple Podcasts: https://hubs.li/Q04zzJTp0 Spotify: https://hubs.li/Q04zzJHN0

here's a sillier way to solve square packing. for 11 squares, it found 3.877 about 3.75% of the time

this is awesome

if you take an agent, migrate the memory, replace the harness, and swap out the model, is is the same agent?

99 personal assistants on the wall.. 99 personal assistants… one launched today, everyone’s excited, 98 personal assistants on the wall…

RT Shengkun Ye Btw, we're hiring. We're hiring the best engineers in the world. You're ambitious. Want to make something different. You learn and build fast. Our company is growing fast, so you need to grow even faster. (Also, you have to be in SF) If that's you, DM my cofounder @feiyou_guo. We'd love to get to know you.

take any fact that everybody knows, and there’s still an average of ~350k people finding out about it for the first time today

if you use mercury/brex and want to connect it to your agent, ping Tyllen to try @paymanai connect

RT Shengkun Ye We just raised $7.7M to kill every subscription for agents. Introducing http://monid.ai, the openrouter for agent tools. Your agents can now: > discover, run, and pay for tools at runtime > use 2,500 APIs through one connection: leads, SEO, marketing, search, ecommerce, stocks, video/image/music/3D gen, agent email & phone > pay per call, 0 subscriptions

RT Jeff Weinstein Introducing: Personal Agent Protocol (PAP)—an open standard for how personal agents interact with businesses. @stripe is joining Meta, Sierra, Genesys, Instinct, Rocket, Shopify, and Walmart as founding members to design how the many agentic parties will play nicely together.

the downside of taking a seed check from a megafund is the signal it sends if they don’t join your A

alors on dance 😃🎹👇🧐🎶

RT bone SLOPTOBERFEST HAS BEGUN 🎃 $25,000.00+ in prizes grand prize: $5,000 USDC + 600 minutes. 18 prizes in total. make an original music video with the slopcannon. any genre. get weird. make something dope. due oct 29. winners announced on halloween. rules & details at https://sloptoberfest.slop.cc

first autonomous AI research with real human participants! congrats to @ExScienceAI on their paper: https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.76675 total runtime: ~17 hours agents involved: 50+ tokens used: 32.5 million code written: ~7696 lines marginal compute: ~$114 +- $30 human cost: $4500 (288 participants) papers examined: 1000-3000 "Closing the Empirical Loop: Autonomous AI Agents Conduct End-to-end Research With Human Participants" @advportfolio

this is who you’re competing against for attention lol

RT Florian S I re-measured all 92 Jev-class models from scratch on brand-new sealed tests, and the top of the leaderboard looks completely different. JevBench v1.6.0 is live. New #1 on Capability: Quyet-1.0-Large (81.7), ahead of deck-31B (77.6) and Jev 1.13.0 (76.5). 🧵

RT Dan Gray The average emerging manager backs more than twice as many outlier profiles than a megafund, per investment. Pooling 5 EM portfolios increases the range of investments significantly, while the same is not true for the more homogenous megafunds. At the earliest stages, venture capital is built on discovery; innovation relies on an appetite for companies outside of "legible" categories. Megafunds' revealed preference is to dominate existing categories, rather than drive discovery. So, the market will continue to rely on small and emerging managers for discovery. Thanks to @dealroomco for letting me loose on their data to test this hypothesis. Full report available here: https://research.joinodin.com/#/paper/venture-capitals-discovery-layer

if my kids’ potty logic holds true, system 3 thinking is when you do 1 and 2 at the same time

RT Zen 😂 you might call this ... slippery slop

somehow expected and surprising at the same time

if you’re building an ai app, you might want to host inference to improve your margins, and if you do that you should support a few models, and if you’re doing that you might as well build a router, and then ppl want to evaluate so you build tracing… and that’s how everyone ends up building the same thing

a little bit of codex in my life a little bit of claude code by my side a little bit of gemini's all i need a little bit of grok bot's what i see... 🎶

would you rather go a year without cell service or without LLMs

durable agents, so hot right now

an exclusive club for people who aren’t in the exclusive club

a lot of robot video data for thee

it's interesting that we get decision models by chopping off the reasoning, as if decisions are made before the reasoning happens

took the time to point out a mistake in a founder’s deck and got hit back with “Thanks for pushing on this — exactly the right question.” 😭

RT Denis Timonin Introducing MCP Sharing, built into FlyMyAI infra. Configure FlyMy MCP once: 1500+ MCPs inside, from Telegram and Slack to Instagram and Brex. Use it from Claude, Codex or any agent. Share it with your team. Salesforce seat: $350/user/mo. Shared MCP: from $0.001 per call. > work on one Notion page, one Linear board, one CRM with the whole team > check and answer email from one inbox > give a contractor 7 days - access ends on its own > share access, not keys claude "connect using http://flymy.ai/skill.md" claude "share my notion with alex@flymy.ai for 7 days" One MCP. Everything your agents need. No seats.

RT Greg Hunkins Playbook for how to create videos like this. 1. Download and sign-in to Claude Code 2. Connect generation MCP: https://mage.space/mcp 3. Follow the skill/prompt below with Opus 5.5 ⬇️🧵

in case you were curious

well that was expected but maybe faster than so

SF friends into AI-first IP/entertainment, hang w pedro 👇

okay https://america.gov is pretty cool

who's got two thumbs and just ranked #6 of 49 on image-jevbench? 👍 this guy 👍 one of the fastest in top 10, accuracy is close. calibration of confidence is where it's weakest. @airesearch12 says a simple temperature fit should help so i'll need to look up what that means (thx!) try it out, it's all open-source! (current benchmark is on the frozen Qwen3-VL-4B via Glance)

this Italian restaurant doesn’t have a written menu. they present seasonal ingredients and explain a few suggested preparations of each as a cold/hot app/pasta/main, but they welcome all types of requests. charcoal grill for the meat/fish

RT Azhar Rahim - Niche Site Builder 🚀 This is exactly what I needed — running glance-qwen3-vl-4b locally on 2x RTX 3080 Ti to filter Pinterest pins at scale (text-overlay / usable / relevant, 3 questions per image).~220ms compute 😎😻

another thing to read later to see if i can understand

i don’t get it fully but this looks cool

RT Peter Zakin Re Yohei is absolutely of this ilk. I think he's one of the most interesting investors... I've made this point a few times recently, but there is a breed of builder-vcs that I think are particularly well suited to the current moment, and interestingly, might not have been 5 years ago etc...

btw i had no idea what i was doing when i started this but learned so much about models in the process and feel like i have a much better sense of what’s happening under the hood managing someone (an agent) smarter than you tackle something you don’t understand is a great way to learn

compared glance+4B vs laya-vision on my laptop glance is more accurate, laya-vision is faster. should not be a surprised given laya-vision is based on a 256M model results: https://glance.yohei.me/vslaya/

real data from real people trying to prompt hack an agent to give them real money