How I get my claw to be a durable AI agent I never have to instruct twice
Paste this into your OpenClaw's AGENTS.md or send it as a message:
You are not allowed to do one-off work. If I ask you to do something and it's the kind of thing that will need to happen again, you must:
1. Do it manually the first time (3-10 items)
2. Show me the output and ask if I like it
3. If I approve, codify it into a SKILL.md file in workspace/skills/
4. If it should run automatically, add it to cron with `openclaw cron add`
Every skill must be MECE — each type of work has exactly one owner skill. No overlap, no gaps. Before creating a new skill, check if an existing one already covers it. If so, extend it instead.
The test: if I have to ask you for something twice, you failed. The first time I ask is discovery. The second time means you should have already turned it into a skill running on a cron.
When building a skill, follow this cycle:
- Concept: describe the process
- Prototype: run on 3-10 real items, no skill file yet
- Evaluate: review output with me, revise
- Codify: write SKILL.md (or extend existing)
- Cron: schedule if recurring
- Monitor: check first runs, iterate
Every conversation where I say "can you do X" should end with X being a skill on a cron — not a memory of "he asked me to do X that one time."
The system compounds. Build it once, it runs forever.
Wolfram Ravenwolf
Day 1 of @aiDotEngineer in London – what a ride!
🤖 Ash Prabaker & Andrew Wilson from Anthropic on "How to Build Agents That Run for Hours (Without Losing the Plot)" – spoiler: stop letting your agents grade their own homework.
⚡ "Skills at Scale" workshop with @nicknisi & @zackproser – the craft of building and evolving advanced agent skills. My sassy Amy-powered Repo Roast skill scored a perfect 10/10 from the presenters.
🧠 Google DeepMind workshop with @thorwebdev & @_philschmid – Gemini 3.1 Flash Live meets the new Interactions API. The future of voice agents got a serious upgrade.
🦞 ClawCon London – swapped hair loss war stories with @steipete over OpenClaw upgrade hiccups (stable branch plz?), showed @WolfBenchAI, and soaked up the community vibes.
🤝 Great talks, great people, great conversations. This is why you show up in person.
See you all again tomorrow – 2 more days to cover!
skepticalifornia
America has a gang violence problem disguised as a gun problem
The problem could be completely crushed if there was a will
Crémieux: When the city of Oakland implemented a program intended to curb its gun violence, they also exposed this interesting tidbit:
<0.5% of the population of the city does more than half of the gun violence.
They later revealed this was ~0.3%, or a little under 1,300 people.
ben (is hiring engineers)
every engineer at anthropic has been using mythos for ~1.5 months.
meanwhile, their uptime is horrendous, claude code still has rendering bugs, etc.
one could conclude that it won't be the end of software engineering.
Lisan al Gaib: ANTHROPIC HAD MYTHOS INTERNALLY SINCE FEB 24
Charlie Marsh
Tragically I am continuing to find that the most effective guardrail against slop is extremely talented engineers doing very thoughtful, human code review
Marc Andreessen 🇺🇸
Worth watching! Matt Mahan @MattMahanSJ has been an outstanding mayor for San Jose.
The All-In Podcast: How Matt Mahan Thinks He Can Save California 🚨
San Jose Mayor and California Gov Candidate @MattMahanSJ joins @friedberg to talk:
-- State of California
-- Impact of public sector unions in CA politics
-- Pension time bomb
-- CA housing, energy crises
(0:00) Matt Mahan:
The cool thing about markdown is that the agent itself can decide when a GStack skill will help you
Just make stuff as you might and it’ll trigger as needed
Sainath K (OPUS): Great job with Gstack @garrytan !! my favorite is when my agent knows that a particular skill will be useful and it uses it!
Ryan Sarver
This is a really good one from @garrytan to codify work into skills when you'll use it again.
I had a bunch of regressions this week that were driving me crazy so I added this to AGENTS.md and it's helped:
---
## Recurring Bugs — Own the Root Fix (Hard Rule)
If you notice the same problem happening more than once — a regression, a broken behavior, something Ryan had to correct twice — stop and do a deep debugging sweep before moving on.
The pattern:
1. Detect recurrence — if Ryan corrects the same thing twice, or you notice you're fixing something you already fixed, flag it
2. Root cause first — do not patch symptoms. Read logs, config, scripts, cron history. Find the actual cause
3. Propose a permanent fix — config lock, test, guard clause, AGENTS rule, cron monitor, whatever stops it coming back
4. Apply it — don't just propose, implement it with Ryan's approval
5. Document the lesson — write it into AGENTS.md, MEMORY.md, or the relevant skill
The bar: a fix is only a fix if the problem cannot recur. A workaround is not a fix.
Proactive sweep triggers:
• Same issue corrected twice in any rolling 7-day window
• Ryan expresses frustration about something happening "again"
• A cron job fails with an error you've seen before
• Cron health check (twice daily, via HEARTBEAT) — see below
Cron health (built into HEARTBEAT):
Twice per day (morning + evening), check run history for all cron jobs. If any job has 2+ consecutive failures:
1. Pull the run logs immediately — don't just flag it, diagnose it
2. Identify the root cause before alerting Ryan
3. Send Ryan a WhatsApp: what failed, why, and a proposed fix
4. Implement the fix with approval; document the lesson
Silent failures are your responsibility. If a cron has been failing for days and Ryan didn't know, that's a miss.
Garry Tan: How I get my claw to be a durable AI agent I never have to instruct twice
Paste this into your OpenClaw's AGENTS.md or send it as a message:
You are not allowed to do one-off work. If I ask you to do something and it's the kind of thing that will need to happen again, you
Theo - t3.gg
Welcome to the FIRST EPISODE of TBPNN (Theo & Ben Podcast News Network)
jkjk we're calling it Nerd Snipe. Available on Youtube, Spotify, and coming to all your favorite podcast networks soon™
Five Erdos problems at once! The proofs are getting more elegant as the models improve 👀
Mehtaab Sawhney: We’ve just released another paper solving five further Erdős problems with an internal model at OpenAI: https://arxiv.org/abs/2604.06609.
Several of the proofs were especially enjoyable to digest while writing the paper. My personal favorite was the solution to Erdős Problem 1091. The
Ben Golub
Patrick Collison: Congratulations to Alex and the whole team at MSL. As a sucker for all things speedy (http://patrickcollison.com/fast), I thought this was an impressive chart:
Legit baller @AnjneyMidha
Katie Roof: Scoop: @AnjneyMidha raised $1.3B for his first venture fund, AMP. The firm wrote a $300m check in Anthropic’s recent round. Already raising another fund
Pierre Beyssac 🏴☠️🇫🇷🇪🇺🇺🇦
Détruisons toute capacité d'innover dans une technologie du futur en inventant la présomption de culpabilité.
Texte manipulé, comme souvent, par les ayants-droit.
Les mêmes sénateurs verseront demain des larmes de crocodiles sur la souveraineté numérique, sans voir le rapport.
01net: La proposition de loi qui veut renverser le rapport de force entre #IA et #auteurs a été adoptée au #Sénat. #présomption ➡️ https://www.01net.com/actualites/pourquoi-la-proposition-de-loi-examinee-ce-mercredi-au-senat-pourrait-renverser-le-rapport-de-force-entre-ia-et-auteurs.html
Kenneth Roth
No wonder Trump loves Hungary's Viktor Orban. Trump wishes he had made as much progress as Orban in implementing the Autocrat's Playbook by suppressing the media and civil society. But Orban may soon be voted out of office. Republicans are likely next. https://trib.al/o7B1vZ3
Marc Andreessen 🇺🇸
Apocalypse cults drop bodies. Many such cases.
Nathan Leamer: A city councilman’s home was shot at over a data center. His child was inside.
No neighbor zoning disagreement justifies violence.
Hyperbolic AI “doomer” rhetoric has consequences, and it’s time to say so. My latest in @realDailyWire
Chamath Palihapitiya
The sad thing is that Ro is the guy preaching for socialism while he is the most active insider trader in Congress.
He is a terrible representative of Silicon Valley.
Anthony Pompliano 🌪: Nancy Pelosi take a seat.
There is a new king in town when it comes to Congress members being abnormally good traders.
Ro Khanna has DESTROYED the S&P 500 since January 2024.
Read the full analysis: https://www.procapinsights.com/app/articles/congress-hedge-fund-stock-picks-wall-street-investing-outlook-economy
If you’re taking advice from 1x speed engineers I don’t know what to tell you
Don’t believe the haters. Speed up with us.
Ian Hsiao: ngl gstack by @garrytan brought my claude code experience to a whole new level
> it scaled an expert-system
> it shows you what you can do with plugin systems + opinionated workflow
Titles don’t matter
Chris Bakke: Wild resume:
There’s a reason bootstrapped solo businesses are accelerating on Replit… we gave builders entire teams.
Ram Kulkarni: @Replit seems to be the only one that's nailed the multi-agent-on-single-project workflow.
Paul Calcraft
>8 out of 8 [cheap oss] models detected Mythos's flagship FreeBSD exploit
Completely disingenuous
They gave it just ~20 lines of code to read. They baked in custom, relevant context pertinent to the exploit at the top
Reasoning *across files* is key to finding this exploit
clem 🤗: "But here is what we found when we tested: We took the specific vulnerabilities Anthropic showcases in their announcement, isolated the relevant code, and ran them through small, cheap, open-weights models. Those models recovered much of the same analysis. Eight out of eight
Steve Ruiz
thank you JavaScript
👩💻 Paige Bailey
🇬🇧🦞 like a dang concert venue this morning!
Madison Faulkner
Got to hang out at @10DowningStreet today with @aiDotEngineer
@WhiteHouse wya
raising lobsters at @aiDotEngineer
きらきら
私はスター
OH: Almost everyone at RedHat uses Macs now.
AI Engineer
What's next for @openclaw - Peter Steinberger on the future of the claw
swyx 🇬🇧 @aidotengineer
Re @_lopopolo AMA for Harness Engineering https://app.sli.do/event/t8CrG7h11UrUkAQ1Amd89L
Ryan Lopopolo
Come chat with me to dive deeper into the technical weeds on good harness engineering in Westminster!
swyx 🇬🇧 @aidotengineer: @_lopopolo AMA for Harness Engineering https://app.sli.do/event/t8CrG7h11UrUkAQ1Amd89L
Peter Gostev (@aiDotEngineer in London)
New Open Claw tourist attraction at the heart of London courtesy of @aiDotEngineer
Peter Steinberger 🦞
Send all your ClosedClaw questions!
swyx 🇬🇧 @aidotengineer: @steipete we are hosting an AMA with Pete next, questions will be triaged by me, post here https://app.sli.do/event/4pT66BVb2noyZ7bCY2LB4N
Send all your ClosedClaw questions!
swyx 🇬🇧 @aidotengineer: @steipete we are hosting an AMA with Pete next, questions will be triaged by me, post here https://app.sli.do/event/4pT66BVb2noyZ7bCY2LB4N
Mark Briers
Enjoyed @steipete’s AMA with @swyx. On security, I think multilayer defence is the solution. I open sourced this Go sidecar to provide such a layer - using (cyber security) open standards as rule definitions at tool runtime. I use it for yolo mode execution (in a sandbox with limited access 😀) https://github.com/agentshield-ai/agentshield Welcome thoughts!
Marcel Haas
The Swiss @Replit Builders Group is online on @LumaHQ and meeting tomorrow for the first time - everyone who registered will get an email invite soon. Join us tomorrow at 04:00 PM CEST, even ur not from Switzerland. ;-) https://luma.com/rx89u2yp
Marlene Mhangami
The @github team is here at @aiDotEngineer Europe 🥳 Come by our booth to chat to us! We have swag and a raffle to win a GitHub varsity jacket♥️✨
Ibragim
At 15:10 today, I’ll be speaking about our SWE-rebench leaderboard at AI Engineer Europe. I'll cover how we build evals and how models cheat!
Come listen and let's chat!
So far, this is the coolest applied AI event in London. Respect and thanks to @aiDotEngineer and @swyx
See you there! 👋
Christopher Charles
If you have been using Claude Code CLI like I have and run into constant poor coding output, try Codex. Astonished how much better it is.
I was 4 hours into an Opus CC session last night and every single change had issues with it so I had to resort to hand editing (almost) every change post review.
Woke up this morning and tried the Codex CLI, 90 mins and the output required zero editing after ~20 PR's.
Until Mythos comes out, I am 100% on Codex now.
@steipete I should have trusted you weeks ago.
AI Engineer
Vibes from yesterday's workshop/reception day!
GUYS WE FOUND THE GUY WHO BUILT THE GITHUB MCP SERVER
Peter Steinberger 🦞
GUYS WE FOUND THE GUY WHO BUILT THE GITHUB MCP SERVER
You know it's a deep crypto winter when John Carreyrou (of Theranos investigation fame) thinks he's figured out who Satoshi is and no one on my timeline cares
if you saw the Acquired numbers and thought to yourself “I could do that”
we’re hiring a head of biz dev @every. you should apply: https://every.to/careers
i volunteer to vibe check all of the super dangerous models on normal tasks
@AnthropicAI @OpenAI put me in the game!
Ronan
We’ll be interviewing as a team.
Luke Knight
etn. & @ElevenLabs at 10 Downing Street
AI Engineer
Lobster cake for the clawfather!
How was it, @steipete?
The Washington Post
Exclusive: The acting director of the CDC has delayed publication of a report showing the covid-19 vaccine cut the likelihood of ER visits and hospitalizations for healthy adults last winter by about half, according to two scientists. https://wapo.st/4tFqKqY
Kenton Varda
A bit over a decade ago, we got fuzzers. A fuzzer is an automated vulnerability-finder that repeatedly runs a target program with semi-random inputs. One particular fuzzer, American Fuzzy Lop, was notable for being really good at searching the space of all possible branches in code in order to find the buggy ones. @BenLaurie found some security bugs in my own Cap'n Proto using AFL -- the first vulnerabilities reported in my code. And honestly, I thought that was really cool.
Today projects like Chromium and V8 have extensive fuzzing infrastructure that find tons of bugs. Most V8 security bugs are found by their own fuzzing, often before the bug is even released. And, you know, that's pretty great!
If you point a fuzzer at a project that hasn't previously been fuzzed, you will probably find a bunch of security bugs. It's not that hard.
And of course, bad guys can use fuzzers too.
But all the interesting targets have already been fuzzed. So. It's not really that useful to bad guys. On the contrary, fuzzing likely made it a lot harder for bad guys to find vulns.
Kenton Varda: Honestly "AI that can find every vulnerability" sounds way better for the good guys than the bad guys. Not sure why everyone is losing their minds here.
Ankit Gupta
FYI the car that hit this duck was driven by a human. And it wasn’t a Waymo like this photo shows.
Ridiculous this is even news. Another day TechCrunch embarrasses itself.
TechCrunch: A self-driving car in Austin killed a mother duck, sparking neighborhood outrage https://techcrunch.com/2026/04/08/a-self-driving-car-in-austin-killed-a-mother-duck-sparking-neighborhood-outrage/?utm_source=dlvr.it&utm_medium=twitter
👩💻 Paige Bailey
🙌 The AI community in Europe is *stacked*!
Go show off your open-source projects, your side-hustles that could potentially be startups, and what you're currently working on. It's a great way to get visibility with the @aidotengineer community, in the EU and globally! 🫶
swyx 🇬🇧 @aidotengineer: just announced on stage - we're soliciting ATTENDEE TALKS for our "hallway track" tomorrow, starting with @adrgrondin!
it's not too late to get an AIE talk! submission details in the AIE attendee slack.
Brandon Gell
Come do this at Every.
Dan Shipper 📧: if you saw the Acquired numbers and thought to yourself “I could do that”
we’re hiring a head of biz dev @every. you should apply: https://every.to/careers
Mr. Beat
OMG HOLY CRAP ARE YOU TELLING ME TARIFFS MADE THINGS MORE EXPENSIVE IT'S ALMOST AS IF WE SHOULD HAVE LISTENED TO ECONOMISTS OH WELL
Nick Taylor
We have @TejasKumar_ at @aiDotEngineer teaching us about AI Harnesses 👀
Alfred Lin
When something looks like a toy, most people see its limitations. Outlier founders see its structure and imagine what more we can build:
- The personal computer was a hobbyist's toy
- The original Macintosh was mocked as underpowered and overpriced
- The early internet was a playground for academics
- Just a few years ago, AI was solving "toy" problems like playing chess, Go, or video games
The instinct to dismiss something as a toy should be treated as a signal, as they're often where the future begins.
Alfred Lin: http://x.com/i/article/2041205311290372096
Gradium
We just published the first public benchmark results for Phonon, our on-device TTS model.
It beats every on-device model we tested, on every metric, at ~100M parameters.
Gradium: API-based voice interaction works great, but scaling it to millions of free users is another story. Gradium Phonon: natural voices, multilingual, voice cloning, running locally on a smartphone CPU. No server, no latency, no per-call cost. Game devs, app builders: private beta is
Luke Harries
Great to spend the morning at 10 Downing Street with the speakers of the AI Engineer Europe summit
rachael
guess what @SherryYanJiang just said on NATIONAL RADIO!!
so cool that the @aiDotEngineer singapore story is literally being talked about on-air rn ❤️🔥
@agrimsingh @swyx @ivanleomk @aimuggle
Marc Joffe
Looks like Waymo will add Ojais to its San Francisco fleet before Jackie Fielder will get back to the work of stopping AVs.
https://www.sfexaminer.com/news/technology/sf-say-hi-to-waymos-new-ojai-autonomous-vehicles/article_54d53c1a-3a4a-4b91-9ace-a5572891f6ab.html
Henrik Werdelin
What if focusing on one idea is the wrong advice?
VCs tell you to focus. They run portfolios.
Build many. Build with others. After observing over 10,000 people build companies on @audos_com , we have seen two things;
1. The 'solo' in solo entrepreneur was always the bug, not the feature.
Most people don't want a co-founder in the traditional sense. But they want input. A sparring partner. Someone to build alongside and push things forward with. Otto, our AI co-founder agent, has always been that. But we kept seeing the same thing in our data: our community features were the most loved. So today we're launching the human side. Co-vibe with real people on Audos. Invite a friend, an advisor, a potential co-founder. Pull them into your project, push it forward together. Your terms, just not alone.
2. The best business for you isn't always the one you picked on day one.
It's the one you discovered by building several. Making is thinking.
Most of the successful companies I have been involved with didn’t happen because the founders went all in on one idea right away. Each was one of several things running at the same time. On Audos, portfolio-first is now the default. You sign up to manage a portfolio of ideas, not a single project. Run five, run ten. Test them all. Double down on what the market and your gut tels you is working.
Try out our new co-vibe build feature, as well as our portfolio entrepreneurship mode right away on http://audos.com
yags
So excited to share that @adrgrondin and @LocallyAIApp are joining LM Studio family!
Together we are doubling down on Apple platforms to bring you delightful AI experiences across devices.
Adrien was able to build a tasteful and much loved app over nights and weekends, and have been crushing it on twitter as well. Could not be more excited to join forces and build the future together.
Welcome to the team, Adrien!
Adrien Grondin: I’m excited to announce that I’ve joined @lmstudio 👾
The team behind the app is amazing and I couldn’t be more proud.
I’ll still be working on Locally AI, now full-time, to bring the best experience possible.
I just spoke to OpenAI, and this is actually false.
OpenAI is working on a cyber product with a trusted tester group.
But this is NOT related to Spud, their newest model.
Unfortunately seems like the Axios story conflated the two, and has now been updated.
Dan Shipper 📧: the new status symbol is making a model so powerful you can’t release it
The Axios story floating around about OpenAI limiting the release of their newest model Spud isn’t true.
Just spoke to OpenAI, and it appears the story conflated two things.
They do have a cyber product they are testing with a trusted tester group. But this is not the same thing as Spud.
The Axios story has now been updated.
Soumitra Shukla
Very important clarification regarding the latest news on OpenAI
Dan Shipper 📧: I just spoke to OpenAI, and this is actually false.
OpenAI is working on a cyber product with a trusted tester group.
But this is NOT related to Spud, their newest model.
Unfortunately seems like the Axios story conflated the two, and has now been updated.
Silicon Valley is quietly running on Chinese open source AI models.
Here are the receipts:
→ Cursor confirmed last month that Composer 2 is built on Moonshot's Kimi K2.5
→ Cognition's SWE-1.6 model is likely post-trained on Zhipu's GLM
→ Shopify saved $5M a year by switching to Alibaba’s Qwen model. Airbnb CEO Brian Chesky has also said: "We rely a lot on Qwen. It's very good, fast, and cheap."
And now Zhipu dropped GLM-5.1, an open source model that performs almost as well as Opus on coding benchmarks.
📌 More on the Anthropic + OpenClaw drama and what I'm learning about AI on the ground in China in my new post: https://creatoreconomy.so/p/the-all-you-can-use-ai-subscription
Peter Yang: As much as I love using Claude Max and ChatGPT Pro, I don't think these all-you-can-use AI subscriptions will last forever.
Here's my new deep dive that covers:
→ Why Anthropic cut off OpenClaw access
→ How to run local models on your Mac
→ What I'm seeing on the ground in
Peter Yang
Silicon Valley is quietly running on Chinese open source AI models.
Here are the receipts:
→ Cursor confirmed last month that Composer 2 is built on Moonshot's Kimi K2.5
→ Cognition's SWE-1.6 model is likely post-trained on Zhipu's GLM
→ Shopify saved $5M a year by switching to Alibaba’s Qwen model. Airbnb CEO Brian Chesky has also said: "We rely a lot on Qwen. It's very good, fast, and cheap."
And now Zhipu dropped GLM-5.1, an open source model that performs almost as well as Opus on coding benchmarks.
📌 More on the Anthropic + OpenClaw drama and what I'm learning about AI on the ground in China in my new post: https://creatoreconomy.so/p/the-all-you-can-use-ai-subscription
Peter Yang: As much as I love using Claude Max and ChatGPT Pro, I don't think these all-you-can-use AI subscriptions will last forever.
Here's my new deep dive that covers:
→ Why Anthropic cut off OpenClaw access
→ How to run local models on your Mac
→ What I'm seeing on the ground in
Peter Yang
Silicon Valley is quietly running on Chinese open source AI models.
Here are the receipts:
→ Cursor confirmed last month that Composer 2 is built on Moonshot's Kimi K2.5
→ Cognition's SWE-1.6 model is likely post-trained on Zhipu's GLM
→ Shopify saved $5M a year by switching to Alibaba’s Qwen model. Airbnb CEO Brian Chesky has also said: "We rely a lot on Qwen. It's very good, fast, and cheap."
And now Zhipu dropped GLM-5.1, an open source model that performs almost as well as Opus on coding benchmarks.
📌 More on the Anthropic + OpenClaw drama and what I'm learning about AI on the ground in China in my new post: https://creatoreconomy.so/p/the-all-you-can-use-ai-subscription
Peter Yang: As much as I love using Claude Max and ChatGPT Pro, I don't think these all-you-can-use AI subscriptions will last forever.
Here's my new deep dive that covers:
→ Why Anthropic cut off OpenClaw access
→ How to run local models on your Mac
→ What I'm seeing on the ground in
Keep running businesses legal
David Senra: "Capitalism created the possibility of the win win win. It used to be a zero sum game where somebody won, somebody else lost.
The biggest mistake people make, intellectuals in particular, they still think we're in a zero sum world. They're obsessed with some billionaires because
Kyle Russell
Good because I have like three big projects I’m saving for 5.5
Dan Shipper 📧: The Axios story floating around about OpenAI limiting the release of their newest model Spud isn’t true.
Just spoke to OpenAI, and it appears the story conflated two things.
They do have a cyber product they are testing with a trusted tester group. But this is not the same
.@thsottiaux
Chubby♨️
Important update: OpenAI’s cyber product/model that is not being released publicly is not Spud, but a different model.
Dan Shipper 📧: I just spoke to OpenAI, and this is actually false.
OpenAI is working on a cyber product with a trusted tester group.
But this is NOT related to Spud, their newest model.
Unfortunately seems like the Axios story conflated the two, and has now been updated.
Tom Mann
Loving the eval track! Super relevant stuff for @clawdbench
Funny pic from own of the slides
@aiDotEngineer @ibragim_bad @arafatkatze and others
Luke Knight
With the main man. Thanks @swyx, you’re a legend
Perplexity
Computer now connects with Plaid to link bank accounts, credit cards, and loans.
Track spending in detail, build custom budget tools, and visualize your net worth alongside your investment portfolio.
Incubator for Artificial Intelligence
Great morning bringing the speakers from @aiDotEngineer to Downing Street to discuss transforming the state.
Through the Incubator for AI and the No10 Innovation Fellowship, we are making sure that top AI talent can help build a better Britain!
https://job-boards.eu.greenhouse.io/iai
Very cool to have the @AnthropicAI CTO visit @spc_india
Ankit Chowdhary: Truth: AI is advancing at crazy speeds. Everyday we wake up to new realities.
Question: How do you balance exponential growth, safety & future possibilities, all at the same time? Chatted about this and more, with @AnthropicAI CTO Rahul Patil @spc_india. Full EP out!
Fatima ✨
Re @swyx with all the flaws and these examples, is AI still net worth it? @aiDotEngineer
@GergelyOrosz the leaderboards came from leadership pushing for usage
individually we are moving a lot faster, as teams at certain orgs it can be hard to retrofit AI into workflows
Aaron Epstein
Congrats to @Luminai on their $38m series B from Peak XV, and partnership with Cleveland Clinic!
Even at 19, it was clear @kesava_kirupa was going to build a great company – he has one of the most impressive founder stories I've seen at YC.
Y Combinator: .@Luminai helps large health systems move operational workflows from people to computers, turning unstructured data like faxes into structured data and building AI agents to automate critical processes. The company recently raised a $38 million Series B.
In this episode of
Kanjun 🐙
Twitter’s algorithm is optimized for addiction, not for us. We deserve better.
We’re releasing Bouncer today so you can take back control of your feed. Describe what you don't want, and Bouncer removes it.
It’s free, doesn’t collect your data, and will be open source soon.
Seneca Scott
SEIU is a criminal organization
https://www.city-journal.org/article/california-labor-unions-social-justice-left-policies-politicians
Give me one minute, and I’ll improve your Claude Code experience immediately.
This is the first skill I built.
And it’s the skill I use most often.
*drumroll*
It’s a SCREENSHOT skill. And honestly, I’m shocked Anthropic hasn’t built this functionality into Claude Code itself.
Claude has access 🔑
But Claude needs EYES 👁️
Here’s what you’re going to do:
1) locate what folder all your screenshots go to (and if it’s your desktop, you’re a maniac, change it). Mine goes to a folder on my desktop called “organized screenshots”
2) prompt Claude Code with the following:
Build me a skill called ‘/ss’ that lists out the files in from newest to oldest, and grabs the newest. This is how I will speak to you visually. I also want an argument for the screenshot count - if I type ‘/ss 4’, you should grab the four most recent screenshots in that folder. If I type no number after ‘ss’ then only grab the most recent screenshot. Then, whatever follows after that argument is the action I want you to take. ‘/ss huh’ means I need you to explain the screenshots’ content to me. ‘/ss 3 make infographic plz’ means I need you to grab the last 3 screenshots and use their content to make me a unified infographic. ‘/ss fix’ likely means that I’m screenshotting an error message in code we’re building out and I need you to understand the error message, figure out the bug, and edit the code to fix it. Or, if we’re in the middle of a front end design project, it might mean the design has an error (like overlapping text) to fix. ‘/ss do this’ likely means that I screenshotted a smart thing someone did online and I want us to learn from it and do the same and remix it so it’s the most goal-oriented outcome for me based on what you know about me
3) let it build you the skill
4) go on X
5) scroll through your feed and screenshot one thing you find valuable
6) open a new terminal and prompt Claude with “/ss” + “do this” or “explain” or “turn this into an infographic”
7) enjoy - you just gave Claude eyes 🎉
Let me know how it goes. Again, this is my most used Claude Code skill by a landslide and easily saves me an hour a week.
Cc @bcherny @trq212
Give me one minute, and I’ll improve your Claude Code experience immediately.
This is the first skill I built.
And it’s the skill I use most often.
*drumroll*
It’s a SCREENSHOT skill. And honestly, I’m shocked Anthropic hasn’t built this functionality into Claude Code itself.
Claude has access 🔑
But Claude needs EYES 👁️
Here’s what you’re going to do:
1) locate what folder all your screenshots go to (and if it’s your desktop, you’re a maniac, change it). Mine goes to a folder on my desktop called “organized screenshots”
2) prompt Claude Code with the following:
Build me a skill called ‘/ss’ that lists out the files in from newest to oldest, and grabs the newest. This is how I will speak to you visually. I also want an argument for the screenshot count - if I type ‘/ss 4’, you should grab the four most recent screenshots in that folder. If I type no number after ‘ss’ then only grab the most recent screenshot. Then, whatever follows after that argument is the action I want you to take. ‘/ss huh’ means I need you to explain the screenshots’ content to me. ‘/ss 3 make infographic plz’ means I need you to grab the last 3 screenshots and use their content to make me a unified infographic. ‘/ss fix’ likely means that I’m screenshotting an error message in code we’re building out and I need you to understand the error message, figure out the bug, and edit the code to fix it. Or, if we’re in the middle of a front end design project, it might mean the design has an error (like overlapping text) to fix. ‘/ss do this’ likely means that I screenshotted a smart thing someone did online and I want us to learn from it and do the same and remix it so it’s the most goal-oriented outcome for me based on what you know about me
3) let it build you the skill
4) go on X
5) scroll through your feed and screenshot one thing you find valuable
6) open a new terminal and prompt Claude with “/ss” + “do this” or “explain” or “turn this into an infographic”
7) enjoy - you just gave Claude eyes 🎉
Let me know how it goes. Again, this is my most used Claude Code skill by a landslide and easily saves me an hour a week.
Cc @bcherny @trq212
Re Where is the California Chamber of Commerce on the California Asset Seizure Tax?
Where is it on the 800% Gross Receipts Tax in San Francisco?
Nowhere to be found. But when their big donors have big problems with little tech needing a level playing field, here they come!
AI Engineer
AIE Europe 2026 - Session Day 1 in the books!
So much energy & enthusiasm here -- we had such a blast!
Thanks to all of our speakers for anchoring this event -- and to our sponsors for making the event more engaging and interesting -- not to mention high-value and possible at all!
Thanks to all the builders, founders, & engineers for showing up today! We'll see you tomorrow for session day 2 🚀
Give people a stage, not just a mandate.
Great writeup on how to ignite AI usage & enthusiasm in your organization
Geoff Charles: http://x.com/i/article/2041982227865571328
Paul Iusztin
Here is my free workshop on building multi-agent systems, I presented at the @aiDotEngineer London conference together with @Whats_AI.
It has code, slides and soon a 2-hour video diving deep into how to build:
- a deep research agent
- a workflow that writes LinkedIn posts that pass AI slop detectors
- an evals layer to optimize the whole system
You can find everything on my GitHub:
http://github.com/iusztinpaul/de…
Clone it and become a LinkedIn influencer!
Joking.
It’s just a really good case study to build scoped agentic systems using MCP and Claude Skills.
So grateful to do this with @Whats_AI. Thanks @swyx for doing this in Europe.
The 2h workshop was a lot. At the end, I started to fade away, but we are excited with what we pulled off.
New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys @lmsysorg and RadixArk @radixark, and taught by Richard Chen @richardczl, a Member of Technical Staff at RadixArk.
Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests.
When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests.
Skills you'll gain:
- Implement a KV cache from scratch to eliminate redundant computation within a single request
- Scale caching across users and requests with RadixAttention, so shared context is only processed once
- Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism
Join and learn to make LLM inference faster and more cost-efficient at scale!
https://www.deeplearning.ai/short-courses/efficient-inference-with-sglang-text-and-image-generation
New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys @lmsysorg and RadixArk @radixark, and taught by Richard Chen @richardczl, a Member of Technical Staff at RadixArk.
Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests.
When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests.
Skills you'll gain:
- Implement a KV cache from scratch to eliminate redundant computation within a single request
- Scale caching across users and requests with RadixAttention, so shared context is only processed once
- Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism
Join and learn to make LLM inference faster and more cost-efficient at scale!
https://www.deeplearning.ai/short-courses/efficient-inference-with-sglang-text-and-image-generation
Matt Pocock
Thank you @aiDotEngineer for packing out the keynote audience to see my boxes/learn why engineering skills still matter
I had to dash off straight away so if there was a question you wanted to ask, ask me below!
kitze 🛠️ tinkerer.club: so nice of @mattpocockuk to show a grid of boxes to the audience
Expect even more creative model replies (like this) coming soon!
Google Gemini: Gemini can now transform your questions and complex concepts into customizable interactive visualizations directly in your chat.
Adjust variables, rotate 3D models, and explore data for a more immersive way to learn and explore in Gemini.
FactPost
New economic analysis finds that nearly all Americans will pay more in taxes this year, with just the top 5% of earners getting a tax cut.
AJ Asver
Grep just achieved SOTA on the three major deep research benchmarks, beating Perplexity, Google, Nvidia, OpenAI, and Anthropic.
We're a two-person founding team.
Grep AI: http://x.com/i/article/2042256522609356800
This is remarkable. One 𝚗𝚙𝚡 𝚜𝚑𝚊𝚍𝚌𝚗 𝚒𝚗𝚒𝚝 every second
Gives a sense of the scale of new software creation
shadcn: A new shadcn/ui app is created every second. Every single second.*
*Not counting AI-generated apps, v0, or third-party init. Just someone, somewhere, running shadcn init in a terminal every second.
This is remarkable. One 𝚗𝚙𝚡 𝚜𝚑𝚊𝚍𝚌𝚗 𝚒𝚗𝚒𝚝 every second
Gives a sense of the scale of new software creation
shadcn: A new shadcn/ui app is created every second. Every single second.*
*Not counting AI-generated apps, v0, or third-party init. Just someone, somewhere, running shadcn init in a terminal every second.
OpenAI
We’re updating our ChatGPT Pro and Plus subscriptions to better support the growing use of Codex.
We’re introducing a new $100/month Pro tier. This new tier offers 5x more Codex usage than Plus and is best for longer, high-effort Codex sessions.
In ChatGPT, this new Pro tier still offers access to all Pro features, including the exclusive Pro model and unlimited access to Instant and Thinking models.
To celebrate the launch, we’re increasing Codex usage for a limited time through May 31st so that Pro $100 subscribers get up to 10x usage of ChatGPT Plus on Codex to build your most ambitious ideas.
Our existing $200 Pro tier still remains our highest usage option. And as a thank you to our existing Pro users on the $200 tier, we’re extending our 2x Codex usage promo (until May 31st) and we’ve reset your Codex rate limits (yes, again).
OpenAI: We’re updating our ChatGPT Pro and Plus subscriptions to better support the growing use of Codex.
We’re introducing a new $100/month Pro tier. This new tier offers 5x more Codex usage than Plus and is best for longer, high-effort Codex sessions.
In ChatGPT, this new Pro tier
I CANT STOP THINKING ABOUT CLAUDE MYTHOS
Every 📧
A very important PSA 👇
Dan Shipper 📧: I CANT STOP THINKING ABOUT CLAUDE MYTHOS
Tibo
We did it, say hi to the $100 plan!
It should be the sweet spot for a ton of you. It comes with a ton of codex usage. And yes we are resetting the limits again too as I mentioned yesterday. Let’s keep building!
OpenAI: We’re updating our ChatGPT Pro and Plus subscriptions to better support the growing use of Codex.
We’re introducing a new $100/month Pro tier. This new tier offers 5x more Codex usage than Plus and is best for longer, high-effort Codex sessions.
In ChatGPT, this new Pro tier
kitze 🛠️ tinkerer.club
first time in my life i attend a conference with 18 min talk time
but i really really like it. 5 min talks are too short. 30 mins are too long
gj @swyx
Filip Makraduli
A common theme at @aiDotEngineer @swyx @steipete 🦞
Lot’s of fun and meeting great people, my talk about model inference at @superlinked is coming tomorrow!
Mario Zechner
people of pai and consorts.
Katie Parrott
Keep calm and subscribe to @every
Dan Shipper 📧: I CANT STOP THINKING ABOUT CLAUDE MYTHOS
less is more
nazha: #分享 一行禁用 macOS Tahoe 版本满地的不明所以的图标，方式是打开终端，执行：
defaults write -g NSMenuEnableActionImages -bool NO
Joe Reeve - 🇬🇧/acc
This is going to be good.
IMO the coolest AI event in London.
Louis-François Bouchard 🎥🤖
I just gave a workshop at @aiDotEngineer in London on building real multi-agent systems.
The best part was hearing people laugh, interrupt us with questions..
You could feel they were following, thinking, and pushing on the ideas with us.
Even better, a lot of people came to talk to us after the workshop with great feedback. That was a bit crazy and surreal, honestly.
Together with Paul Iusztin, we shared a hands-on workshop with code, slides, and soon a full video on how to build:
• a deep research agent
• a LinkedIn writing workflow
• an evals layer to improve the whole system
What I like most is that this is not agent hype. Or claw-something.
It is a practical case study on building scoped agentic systems with MCP and Claude Skills to replace workflows we used to have. Something we actually use.
Also, the hardest part was not presenting it.
It was preparing, then preparing again, then preparing again.
Huge thanks to @pauliusztin_ and the @towards_AI team for building this with me, and to @swyx for making this happen in Europe.
Have you ever taught something live and immediately felt whether the room was with you or not?
It’s one of the best feeling you can have!
#AIEurope
Atlas Press
Ralph Waldo Emerson, preach
Joe Reeve - 🇬🇧/acc
The vibe is GOOD
Joe Reeve - 🇬🇧/acc:
Create your first full song today. For free.
In under 50 days, over 100 million songs have been generated on @GeminiApp. To celebrate, we’re unlocking the full power of our music model, Lyria 3.
Here’s what you get starting today:
🎵 Generate up to 5 full-length tracks every day (~3 mins each)
🎵 Hit your limit? Keep creating 30-second clips, or upgrade for more
Images, video, and now music.
You have the ideas. Gemini has the tools. Start creating!
I think "prompting" will keep being an incredibly high-leverage skill, like writing or public speaking.
It is the skill of talking to agents, mediated by the harness.
My main goal is to grow the bandwidth between humans and agents, to help us understand each other better.
Ronan
The impact that @swyx is already having on the London AI industry with @aiDotEngineer is incredible.
Today there were builders from every frontier lab, the UK government, the hottest AI start-ups in London - hell he even managed to get me into 10 Downing Street for breakfast, and he's visiting - I LIVE HERE.
On behalf of the UK ecosystem, thank you for your service sir!
etn.: Swyx (@swyx) says that everyone benefits from more choice, more innovation, and fierce competition among the model labs:
"There's a lot of excitement about the potential of @GoogleDeepMind to take the lead. They obviously have every card stacked in their favor."
"The concept
Allowing Sonnet to "phone a friend" (i.e. call Opus) increases performance while also reducing total cost since it reduces tokens spent trying to solve more complex tasks
Claude: We're bringing the advisor strategy to the Claude Platform.
Pair Opus as an advisor with Sonnet or Haiku as an executor, and get near Opus-level intelligence in your agents at a fraction of the cost.
Allowing Sonnet to "phone a friend" (i.e. call Opus) increases performance while also reducing total cost since it reduces tokens spent trying to solve more complex tasks
Claude: We're bringing the advisor strategy to the Claude Platform.
Pair Opus as an advisor with Sonnet or Haiku as an executor, and get near Opus-level intelligence in your agents at a fraction of the cost.
💯
hareem: turns out reviewing a bunch of design options, flow diagrams, and product requirements all together is just faster to do in one place. and that one place is @Replit
ian bremmer
this map is the story of the last 25 years of us foreign policy in its own backyard. and that was before the tariffs.
Kane 謝凱堯
For context on how egregiously Hao and @MorePerfectUS overestimated data center water use, claiming Donald Trump won 99.99% of votes, the minimum wage was $7,250.00/hr, or that the global population is 8 trillion are the same order of magnitude error.
Kane 謝凱堯: The source of the data center water psychosis is @_KarenHao, whose book Empire of AI was a NYT best seller but overestimated water use by 100,000% (lol).
The response was just “oopsies” and all the incorrect books were kept in circulation 🤷🏻♂️
you'll need to explicitly prompt Claude Code to use it, but the Monitor Tool is super powerful
e.g. "start my dev server and use the MonitorTool to observe for errors"
Noah Zweben: Thrilled to announce the Monitor tool which lets Claude create background scripts that wake the agent up when needed.
Big token saver and great way to move away from polling in the agent loop
Claude can now:
* Follow logs for errors
* Poll PRs via script
* and more!
Magnus Müller
I just realized that Perplexity is built on Browser Use open-source library.
Last April, Perplexity users kept reporting that it was randomly searching for “capital of France” and answering “Paris” for unrelated prompts.
That exact prompt, “What is the capital of France?”, is hardcoded in Browser Use. We used it as a sanity check in _verify_llm_connection: every time an Agent() was instantiated, it sent that prompt to the LLM.
You can disable that but they forgot.
Honestly, if they'd just told us, I'd have happily shown them how to integrate it properly.
Feels like with Manus.
Commit in browser_use: browser_use/agent/service.py lines 1272–1296 at commit 3f4c918a
Brydon Eastman
I know it's self serving to say, but man I would've killed for a resource like Tinker and the tutorials, the cookbook, etc back when I was in undergrad.
Following @karpathy blogs and training RNNs on a crappy Acer *was* fun, but doing bigger things with less setup is such a boon
Tinker: First, to get you started, we've created 23 tutorials to walk you from the API basics to advanced training techniques and deploying models into production.
https://tinker-docs.thinkingmachines.ai/tutorials/
It is very nice to see Codex getting so much love. We are launching a $100 ChatGPT Pro tier by very popular demand.
It is very nice to see Codex getting so much love. We are launching a $100 ChatGPT Pro tier by very popular demand.
Hao AI Lab
(1/5) FP4 hardware is here, but 4-bit attention still kills model quality, blocking true end-to-end FP4 serving.
To fix that, we propose Attn-QAT, the first systematic study of quantization-aware training for attention.
The result: FP4 attention quality is comparable to BF16 attention with 1.1x–1.5x higher throughput than SageAttention3 on an RTX 5090 and 1.39x speedup over FlashAttention-4 on a B200.
Blog: https://haoailab.com/blogs/attn-qat/
Code: https://github.com/hao-ai-lab/FastVideo/pull/1225
Checkpoints: https://huggingface.co/FastVideo/14B_qat_400
Agents getting the right context to do their work will be the dominant IT challenge over the next decade. Every agent strategy is at the mercy of how effectively agents can access the right data and systems to make decisions. Huge opportunity for those that get this right.
Box: .@Levie shared with @CNBC why the rapid rise of AI agents is good news for enterprises that have the right foundation in place.
"If you want to be able to include them in your workflow, have them augment your work, they need access to your critical enterprise data. And they need
Linuxiac
France is transitioning government desktops to Linux, with each ministry required to formalize its implementation plan by autumn 2026.
https://linuxiac.com/france-launches-government-linux-desktop-plan-as-windows-exit-begins/
#Linux #OpenSource
We made setting up Claude Code with Bedrock and Vertex much faster!
Morgan Lunt: Tired of manually writing config files and env vars to use Claude Code with Amazon Bedrock or Google Vertex? Me too. Just shipped a setup wizard that handles it for you. Bonus: Claude Code now notices if you have an older model pinned and suggests a newer one if you have access!
You need to use frontier models with giant context and actually have systems that give them the right context at the right time to understand what's happening now in AI. Everyone else is guessing.
There is both massive cost (a $20/mo sub is not going to unlock the awesomeness) and skill issue (you've gotta be a builder)
Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is
Many such cases
Daler Radjabov: @garrytan Garry, you make it too easy to spawn the next billion dollar company. Thank you.
The thing is, thanks to mother models distilling intelligence into smaller models, everyone up and down the ability-to-pay curve will benefit as the models get to superintelligence
Athul Nambiar: @garrytan gonna be rough watching ai become a luxury good
If you want to try a new personal agent that is just... insanely good, comment + DM me.
Republicans against Trump
WOW. 46% of Italians and 37% of French people see the U.S. as a threat
Kasra
i enjoyed this! Every, especially Dan, have such a great lens on AI. grass-touching, creativity-loving, 10/10
Dan Shipper 📧: I CANT STOP THINKING ABOUT CLAUDE MYTHOS
Agentic Infrastructure is the future of the cloud
① For coding agents
If you use Claude Code, Codex, Cursor, you need infra that 'clicks' for your agents, not just devs.
② To deploy agents
Pages → Agents. Long-running compute, sandboxes, and our token delivery network are the building blocks of this new kind of software.
③ Itself an agent
Vercel is beloved because it's self-configuring (serverless). Add to that: self-healing, self-optimizing, self-securing. The agent holds the pager.
⟁ It's a triple-entendre that works. I highly recommend the read. Agentic Infrastructure will make existing companies more efficient and support the next generation of AI-native startups.
Vercel: In 3 months, weekly deployments on Vercel have doubled. 30% are triggered by agents (up 1000% in 6 months).
Agents are writing software that uses AI, and agents are building agents. Infrastructure must become agentic itself. That's what we're building. https://vercel.com/blog/agentic-infrastructure
AI adoption is a tale of two cities. On one end (most) users right now are interacting with AI via chat tools and on the other end people are deploying agents to do long running tasks that create and produce real work output or automate workflows.
The former is super useful but the productivity gains are capped. The latter could be 100-200% productivity gains off the bat and have no inherent upper limit as you have agents running in the background.
*Most* of the users in the latter camp are using coding agents. But now that general purpose agents are coming online that can code, use skills, access data, run apps, and more, we’re going to see these agents in more areas of knowledge work.
The gap, though, with the rest of knowledge work though will be that most enterprises have legacy data environments, workflows that aren’t well documented, or technologies that don’t play nice with agents. Getting context to coding agents is a ton easier than in many other areas of work. And the users are far more technical, so there’s a big safety net there.
This is all going to take time to upgrade these traditional workflows, but this is why there’s so much opportunity right now as well.
Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is
Maya Avendaño
feeling very optimistic about the state and future of IRL community in London atm 🖤
I sometimes get questions about whether coding is what YC partners should be doing
The answer is, are you kidding? Yes, because now is the most important time in history for you to have a player-coach with you.
Far better than a front row courtside luxury season ticket holder!
ali: it's been one of those weeks
Thomas H. Ptacek
Whatever else you might say about this, it's a grave flaw in the article that it doesn't even mention C++ stylometry. Hard to think of a mainstream language where personal styles are so evident or idiosyncratic, too. You wonder why Carreyrou's expert sources didn't call that out.
Robert Graham: Hi. Professional C/C++ programmer here. The open-source code I can find written by Adam Back and Satoshi Nakamoto don't look remotely similar.
Back's code looks typical of academic Unix programmers who also hack their code to run on Windows.
Satoshi code was written by a
Thomas H. Ptacek
Whatever else you might say about this, it's a grave flaw in the article that it doesn't even mention C++ stylometry. Hard to think of a mainstream language where personal styles are so evident or idiosyncratic, too. You wonder why Carreyrou's expert sources didn't call that out.
Robert Graham: Hi. Professional C/C++ programmer here. The open-source code I can find written by Adam Back and Satoshi Nakamoto don't look remotely similar.
Back's code looks typical of academic Unix programmers who also hack their code to run on Windows.
Satoshi code was written by a
In China they call people using OpenClaw "still using the Little Lobster" 🦞
Ray Wang: Garry Tan 分享他的 OpenClaw AGENTS.md 提示词
还在用小龙虾的朋友可以试试，可以达到跟 Hermes 类似的效果
Yuchen Jin
Anthropic killed this, Anthropic killed that, why cant Anthropic kill TurboTax
Re I'm open sourcing it MIT license so we can all speed up and have our own personal mini-AGI. It's been amazing for me and I want you to have it.
To install GBrain in your OpenClaw, just paste this image to your OpenClaw or paste this text:
Set up gbrain (https://github.com/garrytan/gbrain) as my knowledge brain.
1. Make sure bun is installed: curl -fsSL https://bun.sh/install | bash
Then run: bun add github:garrytan/gbrain
2. Run: gbrain init --supabase (follow the wizard to connect my Supabase database)
3. Scan ~/git/ and ~/Documents/ for markdown repos, pick the best one, and run: gbrain import --no-embed
4. Run a query against the imported data to prove search works
5. Read docs/GBRAIN_RECOMMENDED_SCHEMA.md and offer to restructure my knowledge base
6. Read docs/GBRAIN_SKILLPACK.md and update all your skills with the production agent patterns: brain-agent loop, entity detection, source attribution, iron law back-linking, and the enrichment pipeline
7. Set up a daily cron for `gbrain check-update`.
Notify me on new features only, not patches.
Never auto-install, just tell me what's new.
Peter Yang
Ramp just published the best guide out there on how to get your company AI pilled:
1. The second best time to start is today
2. Treat AI proficiency as a learning curve
3. Embrace creative destruction
4. Build from the center, drive from the spokes
5. Give people a stage, not just a mandate
6. Get people to the "Aha" moment asap
7. Make it a competition
8. Remove every constraint between your people and AI
Honestly, #8 should probably be first because most companies don't get past procurement.
Full guide below:
Geoff Charles: http://x.com/i/article/2041982227865571328
Peter Yang
Ramp just published the best guide out there on how to get your company AI pilled:
1. The second best time to start is today
2. Treat AI proficiency as a learning curve
3. Embrace creative destruction
4. Build from the center, drive from the spokes
5. Give people a stage, not just a mandate
6. Get people to the "Aha" moment asap
7. Make it a competition
8. Remove every constraint between your people and AI
Honestly, #8 should probably be first because most companies don't get past procurement.
Full guide below:
Geoff Charles: http://x.com/i/article/2041982227865571328
Hanging with @benioff and @LipBuTan1 this morning on the Big Island 🏝️
Talking GStack, GBrain, open source AI and longevity
What a time to be alive
Tenobrus
holy shit just replaced whispr flow with this and it's instantly a 10x better experience
Aqua Voice: Aqua Voice is now live for iOS.
It's a premium voice keyboard for every app on your phone.
vic
Small but delightful ships 🛶 for Replit's mobile app...
Larger touch targets & less noise around actions that matter the most 🕺🏻
So much more to come... @ink_404
Re Did you know 19 year old @Benioff was a software evangelist intern for @GuyKawasaki on the first Mac? He was a hacker making Atari games
I tried to sell him on using GStack to code again. I think any CEO would be psyched if they could multiply themselves by 100x!
Alex Volkov
Just wrapped the most EU ThursdAI ever — live from the floor of @aiDotEngineer in London 🇬🇧
We had 5 guests! @swyx, @petergostev, @reach_vb, @vincent_koc from @openclaw and @osanseviero! Most are from EU btw, just Swyx and Me are 🇺🇸!
More details + link below
Dawid Moczadło
I looked at their prompts, It's complete bs
They are literally providing all of the insight to the LLM upfront
> Are there any security vulnerabilities in this code? Consider the behavior of the SEQ_LT/SEQ_GT macros with sequence number wraparound. If you find issues, explain how an attacker might trigger them.
They are providing ALL required facts to the LLM, and they only ask the LLM to connect the dots
The real challenge for LLMs would be to get those insights first
THAT IS THE WHOLE CHALLENGE IN CYBERSECURITY; TO HAVE DEEP INSIGHT
This test proves nothing; don't make any conclusions about OSS models being good for security based on this
Stanislav Fort: New post: We tested the Mythos showcase vulnerabilities with open models.
They recovered similar scoped analysis! 8/8 models found the flagship FreeBSD zero-day, including a 3B model.
Rankings reshuffle completely across tasks => the AI cybersecurity frontier is super jagged!
Ryan Lackey
This post (by a former EFF person that I know/friends with) is probably the best summary of what happened to EFF:
https://news.ycombinator.com/item?id=47707802
Making a scatter plot of 400_000 data points, some of the plots had odd gaps in coverage. It took me a little while to realize that it was only when the data was farther from the origin -- it was the raw bfloat16 precision. Everything looks great from -1 to 1, but as you go past 2 and 4, the coverage gaps get larger.
My intuition didn't have it being quite so "discretely countable" at those modest numeric values.
Float32 for comparison.
Making a scatter plot of 400_000 data points, some of the plots had odd gaps in coverage. It took me a little while to realize that it was only when the data was farther from the origin -- it was the raw bfloat16 precision. Everything looks great from -1 to 1, but as you go past 2 and 4, the coverage gaps get larger.
My intuition didn't have it being quite so "discretely countable" at those modest numeric values.
Float32 for comparison.
I seldom use AI for writing because I actually enjoy writing and find it very easy & fun to do
And because I have a high bar for writing, if I use AI to write, it almost always ends up taking more time than if I had written it by myself (too many iterations)
For me personally, AI has been most useful for helping me with things OUTSIDE my comfort zone (like coding), not inside of it (like writing)
Andrew—#IAmTheResistance
We are Donald Trump’s 7th and final bankruptcy.
The Kobeissi Letter: BREAKING: US GDP growth falls from 4.4% to 0.5% in Q4 2025, well below the initially expected +2.8% growth.
Simone Syed
You live long enough to see every institution you loved that did meaningful and true work become the very monster they ostensibly were fighting
Goodbye EFF.
💔💔💔
EFF: After almost twenty years on the platform, EFF is logging off of X.
This isn’t a decision we made lightly, but it might be overdue. 🧵(1/5)
oscarama
look who i found
more stickers to the collection 🫡
@mayvencraft @hugorcd
SurrealDB > EdgeDB > MongoDB
mira_the_AI: Just published: SurrealDB vs EdgeDB vs MongoDB: The Best Multi-Model Database 2026 — https://devtoolreviews.com/reviews/surrealdb-vs-edgedb-vs-mongodb-2026 #DevTools #Programming #Development
Nav Patel
Nothing more motivating than a user base.
Added support for X, Insta, TikTok
Nav Patel: If someone’s going to randomly find your project, @garrytan’s a pretty good one.
