Yohei Nakajima

one-shot podcast from news with @flymy_ai's new agent builder (just tested it), it: - researched news - created a transcript - created a voice for each part - stitched it together full prompt + output

seems useful for ML

one problem with today’s markets is that even in an age of abundant information, distribution systems still concentrate attention and amplify consensus, turning price discovery into a coordinated response driven by shared attention rather than the independent processing that underpins efficient markets

"how can i help?" i analyzed 100 asks from portfolio companies identified in notes and emails to see what they asked 26% specific person intro 9% investor discovery 7% hiring/talent 9% pr/media 11% business/partnerships 14% scheduling/coordination 6% strategy/advice 6% product/technical 4% event related 4% admin 2% research 2% permission/approvals then further analyzed what percent could be assisted with AI:

yessssss look at this beautiful portfolio monitoring solution that just sits on top of existing data (CRM, notes, email, calendar) - normalizes revenue to annualized rate - tracks time since each data point received - auto-recalculates runway remaining - updates automatically every meeting or email - tags actual update emails to track frequency - captures sentiment i've tried building so many versions, this one finally works!

it was gloomy and sprinkling on our walk today, but we saw some pretty flowers

what in the portfolio construction is this

haha cool, this worked. an MCP server that allows Claude to build it's own reusable skills https://github.com/yoheinakajima/selfMCP (1473 LoC) basically a server with skills to CRUD skills in this video, i use claude to create an openai image generation skill and trigger it (sorry for unedited long video, just skip thru)

tried using terminal for the first time in a while... 🤦‍♂️ this is why i use @replit

RT Rie Yano やのりえ Japanese railways don’t just run trains. They actually build cities and act as urban developers to capture the demand from transit. ~50% of revenue comes from owning the land around stations. Buy land → build the train station → develop the area → capture the upside. Brilliant business model. Original tweet: https://x.com/rieglobe/status/2041911254835720276

RT Cailyn Y. introducing momo, the CRM for AI agents. it gives agents its own CRM, like Salseforce/Hubspot for humans. every AI native company will need dev agents, sales agents, customer support agents, HR, legal, finance, design agents. we're starting from customer relationships. we hypothesize that every important decision ever made in a company all derives from understanding customer data. even if dev agents are good at building things out in a day, if they don't have any customer signals on which features are working or where people are churning, it's no use. @getyourmomo does the following: 1) sets up your CRM with posthog + gmail data (or other web analytics tool) 2) analyzes the features your dev agents push 3) detects power users and churning users, aggregates CS tickets 4) sets up auto outreach email workflows & collects reponses 5) builds out list of issues to fix & next steps for your dev agent setting up custom onboarding for a few teams right now, please dm if interested :) Original tweet: https://x.com/cailynyongyong/status/2041638173051383824

i wonder how many total eggs will go undiscovered today at these easter egg hunts

haha RGP UI overlaid on custom GPTs

you have three magical beans after eating each one, whatever activity you do for the next hour, you will master what three activities do you choose? life is like this, but it takes a lot more than one bean to master an activity, and every bean is an hour of your life spend your beans wisely

1B in 2025 to 14B+ in 2026 is wild over/under 50B by end of year?

love this

dripwarts is quite something, innit

it’s pretty interesting that we still don’t really know how we created languages (learning new things at my kid’s school fair)

agents can do everything*, which causes a UX problem, your users have vastly different experiences, some better and some worse thinking like an RPG can help here when you start an RPG game, you start with one weapon, and you learn to fight with it. as you progress, you pick up and learn new weapons. and by the final boss, you can wield all the weapons to defeat him. very rewarding if you'd started the game with all the weapons, it would be too overwhelming, you wouldn't know which weapon to learn first, and you'd probably stop this isn't a new idea, but i've used a few tools recently that reminded me of this latter experience *not everything of course, but you get the point

aaaaaggghhhh

i visited sf recently

RT Sheel Mohnot Anthropic making its own splash with an acquisition today $400M for Coefficient Bio, started last fall, developing an AI drug R&D platform Original tweet: https://x.com/pitdesi/status/2039858374154862645

i see ppl asking what media anthropic would buy but no one has suggested lesswrong 🙃

made a tool to compare star-to-LOC ratio on github cuz i thought i would do well, but @karpathy is 👑 any others i should try?

if I start two companies and they sell an apple back and forth for a billion dollars, do i run two billion dollar companies?

congrats to both openai and tbpn!

RT DelightX Bridging Japan and the US 🇯🇵🇺🇸🔥 Thank you, Yohei-san! Original tweet: https://x.com/DelightX_SF/status/2039452315837693980

there are some really cool founders visiting SF from Japan to build their startups did my part this morning helping bridge US/Japan :)

RT himanshu Based on everything explored in the source code, here's the full technical recipe behind Claude Code's memory architecture: [shared by claude code] Claude Code’s memory system is actually insanely well-designed. It isn't like “store everything” but constrained, structured and self-healing memory. The architecture is doing a few very non-obvious things: > Memory = index, not storage + MEMORY.md is always loaded, but it’s just pointers (~150 chars/line) + actual knowledge lives outside, fetched only when needed > 3-layer design (bandwidth aware) + index (always) + topic files (on-demand) + transcripts (never read, only grep’d) > Strict write discipline + write to file → then update index + never dump content into the index + prevents entropy / context pollution > Background “memory rewriting” (autoDream) + merges, dedupes, removes contradictions + converts vague → absolute + aggressively prunes + memory is continuously edited, not appended > Staleness is first-class + if memory ≠ reality → memory is wrong + code-derived facts are never stored + index is forcibly truncated > Isolation matters + consolidation runs in a forked subagent + limited tools → prevents corruption of main context > Retrieval is skeptical, not blind + memory is a hint, not truth + model must verify before using > What they don’t store is the real insight + no debugging logs, no code structure, no PR history + if it’s derivable, don’t persist it Original tweet: https://x.com/himanshustwts/status/2038924027411222533

things that seem possible but are hard to one-shot* with the latest agents - come up a unique tee concept, design it, and sell at least 100 tee shirts for a profit - sell $1000 worth of drop shipped goods for a profit including customer acquisition cost - create three social accounts with at least 10k followers each - create an account on a freelance marketplace and complete three jobs end to end til payment - create and manage a daily podcast with at least one regular paying sponsor *by one-shot, I mean the agent figures out which tools/sites it requires, and creates accounts on their own, handles all communication with ppl

今週SF行くんだけど 誰と会うと良いかな？ ちょっと空きある

in japan you can return avocados that have turned brown on parts of the inside

here’s our family crest!

my grandpa once showed me this cool “kanji signature” where you scribble left to right but when you turn it 90 degrees, it would be his kanji name here’s mine

my two most memorable/notable meat eating restaurant experiences in US and Japan: 🇺🇸 Black’s BBQ Lockhart - oldest family owned bbq in the US. used to be a meat market that started in 1932. they have this giant wooden block as their cutting board that has slowly gotten shorter over the years 🇯🇵 Misono Shinjuku - Misono is considered the originator of the Teppanyaki, in Kobe 1945. Their Shinjuku location opened 1974 on the 51st floor of the Sumitomo building. You can get a private room overlooking the city with chef preparing premium Kobe/wagyu as you take in the view.

🇯🇵 🇺🇸

haha yesss, we’re getting robot monks

RT Cheng Lou My dear front-end developers (and anyone who’s interested in the future of interfaces): I have crawled through depths of hell to bring you, for the foreseeable years, one of the more important foundational pieces of UI engineering (if not in implementation then certainly at least in concept): Fast, accurate and comprehensive userland text measurement algorithm in pure TypeScript, usable for laying out entire web pages without CSS, bypassing DOM measurements and reflow Original tweet: https://x.com/_chenglou/status/2037713766205608234

momo is cool, been trying it out. takes a different approach to memory where it’s more about turning existing data from integrations into useful insights that can be recalled as memory their new implementation with posthog automates identifying helpful insights

RT Pietro Valfrè I posted here on X earlier this week about a layered approach to Agentic Security. This approach consisted mainly in 3 layers: 1Machine isolation  2Capabilities limitation  3Runtime validation The goal was to share that in order to stop complaining about Agentic security, we should go “beyond the sandbox”. Once the Agent is isolated and limited to certain paths, we should still validate each action contextually.  I suggested doing so by intercepting each action and running it against a validator.  Auto-mode by Anthropic is going exactly in that direction. It’s still extremely early, and is still considered worse than manually authorizing each action. But things become interesting when this same approach is extended to a broader set of Agents.  I strongly believe the future of enterprise agents looks less like strict workflows, and more like powerful autonomous Agents with embedded runtime validation. How to config validators to custom degrees of permissions is still tbd. DMs are open for folks interested in the topic, and to those based in NY, I’m sharing some ideas next week at MCP Dev Summit. Original tweet: https://x.com/p_valfre/status/2037545525038260587

lol gonna try using siri exclusively to run all my agents to see how that feels

http://x.com/i/article/2037198911795060739

RT signüll a break through agent interface will be so damn obvious & seem trivial in hindsight. it will require almost no real configuration, or thought for the end user. there will certainly be no special hardware to buy. it will feel like *magic* with near zero effort just like the original iphone. Original tweet: https://x.com/signulll/status/2036949482563146197

memory is hard because it isn’t just what to remember or forget, but also knowing how to remember, and when it needs to “try harder” to remember

someone asked me yesterday what recent open source projects have been interesting recently there's a ton but i mentioned mirofish, paperclip, hermes agent, and autoresearch

how do or should orgs change as we become more ai native?

this site feels like a sci fi theme park ride

this seems big. tho curious how this ties into inventory mgmt, fulfillment, etc. using fb marketplace infra or…

two cofounders of manus being detained in China while 2b meta deal under review

RT Nikunj Kothari SOC II is in the news right now for being security theater.. You know what SOC II is *actually* good for? Subprocessor lists. I scraped 417 companies subprocessors to investigate what AI native companies are using for their infrastructure. Introducing DeployGraph dot com 🥞 Original tweet: https://x.com/nikunj/status/2036572222081606065

the follow up question was what percent of VC investments will be automated and when i guessed 5-10% by dollars in 3-5 yrs… maybe?

i just mentioned in a talk that 1/3 of stock trades are automated, but that was widely outdated apparently it's now 70-80% of market transactions that are carried out through automated trading software

RT Shubham Saboo This is getting insane. Someone actually ran the 1 trillion Kimi K2.5 model locally on their MacBook M2 Pro Max. It was just an experiment but still super crazy to do. Original tweet: https://x.com/Saboo_Shubham_/status/2036524659781599503

RT Andrej Karpathy Software horror: litellm PyPI supply chain attack. Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords. LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm. Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks. Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages. Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible. Original tweet: https://x.com/karpathy/status/2036487306585268612

i like these

leworldmodel

three layers of security for agents: machine isolation, capabilities limitation, and runtime validation

RT Dan Gray a16z pays $10,000–$14,000 per associate per month. A Los Angeles company puts a solar-powered smart collar on associates. It tracks location, conversations, token burn, Ramp spend, posting activity, investments. Marc just opens the app and draws a circle on the market map. That circle becomes the thesis. As associates approach the edge, the collar beeps and vibrates. With one tap, they can shift categories or algo-boost a founder. No direct management. More agentic. Huge cost savings for a16z. Already on 5k associates across the Bay Area, New York and LA. and now in talks to raise at a $5B valuation led by Peter Thiel. Original tweet: https://x.com/credistick/status/2035686081643917430

right now, we manage multiple agents but as they get better, does the capacity of one agent surpass one person’s cognitive capacity to track and manage? put differently, will we get to a point where we need teams of people managing a single agent? just random food for thought

while I agree with his general thinking about how we’ll rethink home design, seems optimistic to think that a meaningful percentage of homes would be redesigned in a 5-10 yr period

me: what are you working on? founder: we're building the emotional friend that everyone needs, but with embedded graph based composable memory that can connect to claude code/openai with mcp/api/cli. it has it's own rl environment it uses to continuously update it's own model and can build a full web/mobile app, identify your target audience, do outreach, find you a girlfriend, buy her gifts, arrange your wedding, and raise your kids

in the future, everything is fun and interactive gamified entertainment, carefully designed to generate training data the ai has decided it needs

the real impact of claw was normalizing the ability to chat with your ai from any channel

a person interviewing an ai this is a format we're going to see a lot more of

slap an api on this for our agents

NY friends building agents! reach out to @p_valfre and the @denieddotdev to learn about securing your agents and openclaw's from taking dangerous actions they'll be in town to speak at the MCP Dev Summit (Apr 2-3)

1/4 as babyagi turns 3 yrs old, i finally sat down to compare the 9 iterations i did over the years... this turned into http://babyagi.wiki a technical history of a personal project (which kind of captures the progress of the agent space overall)2/4 was originally just for myself, but claude made it so pretty i had to share3/4 would be cool to do this across all agent frameworks, but the table would get pretty big4/4 here's where the different ideas sit alongside larger ecosystem trends and projects

bad news for VCs

i hear captchas are for collecting data

love

as i think about the challenges of filtering ai content and replies, i’m reminded of this quote by a park ranger on the challenges of designing trash cans: “there is considerable overlap between the intelligence of the smartest bears and the dumbest tourists”

RT Garry Tan Adding to GStack TODOS.md Original tweet: https://x.com/garrytan/status/2034274861615628663

getting mixed feeling, ai agents cold emailing me do a better job personalizing emails than some humans

this is how i would run a startup incubator

wait, so that weird CA rule where every VC has to ask their portfolio about diversity stats and submit by April 1st is now on pause?

plenty of people have written about this, but i organized my own observation and thoughts for anyone interested

I'm speaking at the Virtual DDVC Summit 2026! I'll be joining 30+ speakers from Accel, Atomico, Balderton, BlackRock, NEA and more for 3 days of practical sessions on how leading investment firms are using AI, data, and automation. My session: "Opportunities & Challenges for Agentic VCs" 📅 Day 2, March 24 at 8:30pm CET

RT Logan Kilpatrick Help us measure the progress towards AGI (specifically cognitive capabilities) by building benchmarks on @kaggle, with $ 200K in prizes available! Details in 🧵 Original tweet: https://x.com/OfficialLoganK/status/2033978254344786351

imperium sine fine

http://x.com/i/article/2033954998783213568

RT andrew pignanelli http://x.com/i/article/2033921858861469696 Original tweet: https://x.com/ndrewpignanelli/status/2033926820605698262

RT @bluecow 🐮 every AI memory system out there (Mem0, MemOS, Recall, Memlayer) works the same way: call an LLM to extract facts, call an embedding API to store them, call the embedding API again to search nuggets does none of that instead of storing each memory separately in a database, nuggets compresses facts into a single mathematical object — a tensor. adding a memory? multiply into the tensor. recalling a memory? multiply out of the tensor. no API calls. no database. no embeddings. just math on your local machine the result: ~415 tokens at session start + ~50 per turn. compared to Mem0 burning ~1,600 tokens of LLM input every time you save a single fact 22x cheaper. runs fully offline. your data never leaves your machine. and you can create as many nuggets as you need — each one is its own tensor, no limit on how many you spin up Original tweet: https://x.com/BLUECOW009/status/2033659708389617707

RT andrew pignanelli We want YOU to start a company and we want to pay you for it! We need to test our new platform that orchestrates agents to run an entire company. That's why we're launching The General Intelligence Fellowship. Build something cool with us, keep all of it, and get free money 🌻 Original tweet: https://x.com/ndrewpignanelli/status/2033574278587658612

RT Manus Today, we're taking Manus out of the cloud and putting it on your desktop. Introducing My Computer, the core feature of the new Manus Desktop app. It’s your AI agent, now on your local machine. Original tweet: https://x.com/ManusAI/status/2033558672152854712

RT Andrew Kang Researchers trained a humanoid robot to play tennis using only 5 hours of motion capture data The robot can now sustain multi-shot rallies with human players, hitting balls traveling >15 m/s with a ~90% success rate AlphaGo for every sport is coming Original tweet: https://x.com/Rewkang/status/2033043512925581805

RT vittorio this is actually insane > be tech guy in australia > adopt cancer riddled rescue dog, months to live > not_going_to_give_you_up.mp4 > pay $3,000 to sequence her tumor DNA > feed it to ChatGPT and AlphaFold > zero background in biology > identify mutated proteins, match them to drug targets > design a custom mRNA cancer vaccine from scratch > genomics professor is “gobsmacked” that some puppy lover did this on his own > need ethics approval to administer it > red tape takes longer than designing the vaccine > 3 months, finally approved > drive 10 hours to get rosie her first injection > tumor halves > coat gets glossy again > dog is alive and happy > professor: “if we can do this for a dog, why aren’t we rolling this out to humans?” one man with a chatbot, and $3,000 just outperformed the entire pharmaceutical discovery pipeline. we are going to cure so many diseases. I dont think people realize how good things are going to get Original tweet: https://x.com/IterIntellectus/status/2032858964858228817

RT Dan McAteer Manus ex-backend lead had a genius insight text based clis beat structured tool calling for ai agents all day because unix commands appear in training data going back to the 1970s text is the native language of the command line AND text is the native language of llms Original tweet: https://x.com/daniel_mac8/status/2032282196673708042