Trump's war on science.
Max Kozlov: The Trump administration has downsized US science by historic margins — but it's not just via grant or workforce cuts.
Our new @nature analysis reveals the government has cut more than 100 scientific advisory panels across all major science agencies.
kennethlou
the AI loop that's been rewiring how i think about company design.
sat in a @ycombinator talk this week where the framing finally clicked on what's reeally happening.
old pitch: make engineers 20% more productive. add copilots. ship more software with AI. all true. all also a faster-horse upgrade.
actual move: one person more powerful than old structures. Building a queryable company. agent-native software. different category entirely.
5 layers:
1/ sensors + data. every signal from the outside world. customer emails, support tickets, cancellations, product events, code changes. if it's not captured, it didn't happen to the company.
2/ policy layer. the rules. what the system can do alone, what needs human sign-off, what must be logged. guardrails that make the loop trustworthy.
3/ tool layer. the deterministic stuff. SQL, API calls, calendar lookups. things that live in code, not english. @garrytan 's framing: figuring out what belongs in markdown vs what belongs in code is 90% of the battle.
4/ quality gates. safety checks. human review for high-stakes calls. the escape hatch back into judgment.
5/ learning mechanism. the unlock. Monitoring agent watches every query, sees where it fails, writes the fix overnight, opens the merge request, ships it. The same query that failed yesterday works tomorrow. company gets better while you sleep.
most teams have 1 through 4. almost nobody is running 5 across every function yet.
that's the next 6 months. we're 5 people at @usemitohealth across two cities. everyone touches code. revenue per employee at a level i wouldn't have believed in my fintech days. headcount as a feature, not a bug.
humans aren't getting replaced. we're going deeper. the orchestration, the taste, the high-stakes calls - that layer is expanding.
the middle is what's compressing... if you're operating today, the question isn't whether to use AI but around whether the shape of your company makes sense.
@JessicaBRiedl: - Comey indicted for tweeting a number.
- Trump FCC threatens ABC's broadcast license.
- Trump defacing more govt instit…
The billionaire asset seizure tax will send billions of dollars in tax revenue to all the other states that are not California
Christopher Lochhead 🏴☠️ ✍🏼🎙: @garrytan Garry, the startup ecosystem in California will be cooked if the (so-called) "billionaire tax" is passed. Most people in Silicon Valley still do not know.
California wants to change the constitution so they can tax illiquid startup stock.
It's not "billionaire tax".
It is an
I haven't kept up to date with the latest @openclaw updates - is live low-latency calling with your claw now possible?
> be me
> "the internet is polluted by ai slop, we need low-background tokens"
> "wouldnt it be cool if we could time travel and see what our ancestors 100 years ago would say to us"
> all the existing vintage models are like <4B
> we need a chat tuned 13B vintage model
> assemble avengers of ML incl the GPT-1/2 guy
> need vintage tokens
> train new vintage OCR model for old books, newspapers, periodicals, scientific journals, patents, and case law
> need vintage RLHF but cant use chat
> synthesize RLHF pairs from historical texts with regular structure eg etiquette manuals, letter-writing manuals, cookbooks, dictionaries, encyclopedias, and poetry and fable collections, shove it into ChatML
> train it
> future knowledge still got in somehow
> dammit.jpg
> train new SOTA document-level n-gram-based anachronism classifier
> meticulously curate hundreds of billions of pre-1931 tokens (public domain)
> train it
> ok! it checks out vs our FineWeb baseline!
> release it
> it's the most confidently racist model ever released by humankind
> mfw
Nick Levine: New work with @AlecRad and @DavidDuvenaud:
Have you ever dreamed of talking to someone from the past? Introducing talkie, a 13B model trained only on pre-1931 text.
Vintage models should help us to understand how LMs generalize (e.g., can we teach talkie to code?). Thread:
Sherry Jiang
we're down to the last 150 tickets before we sell out of @aiDotEngineer singapore!
so stoked to be doing this in 2 week's time 🔥
see you all there! 🇸🇬
https://www.ai.engineer/singapore
@swyx
GPT-5.5 party on 5/5:
Sam Altman: GPT-5.5 is going to have a party for itself. it chose 5/5 at 5:55 pm for the date and time.
if you'd like to come, let us know here: https://luma.com/5.5
codex will help the team pick people from the replies. 5.5 had some good ideas/requests for the party, which we'll do.
GPT-5.5 party on 5/5:
Sam Altman: GPT-5.5 is going to have a party for itself. it chose 5/5 at 5:55 pm for the date and time.
if you'd like to come, let us know here: https://luma.com/5.5
codex will help the team pick people from the replies. 5.5 had some good ideas/requests for the party, which we'll do.
Well we had a good run of almost a year being healthy but my kid has made me sick again
Matt Ridenour
"How to Build the Future" Demis Hassabis + Garry Tan (@ycombinator x @GoogleDeepMind) video just dropped.
Recommend a watch for all things AGI, Gemini & Gemma, the 1000x Engineer, Future of Agents, Scientific Discovery, and more. 🚀 @demishassabis / @garrytan
https://youtu.be/JNyuX1zoOgU?si=lC6aYDNx2dkalDR_
DevDay 2026
OpenAI: OpenAI DevDay is back.
San Francisco
September 29
CHOI
This GPT Image 2 prompt is going insanely viral right now.
“Redraw the attached image in the most clumsy, scribbly, and utterly pathetic way possible. Use a white background, and make it look like it was drawn in MS Paint with a mouse. It should be vaguely similar but also not really, kind of matching but also off in a confusing, awkward way, with that low-quality pixel-by-pixel feel that really emphasizes how ridiculously bad it is. Actually, you know what, whatever, just draw it however you want.”
Almost any decision in business and life can be decided by:
1/ Are we lucky? Great.
2/ Otherwise what is the smart gritty thing to do?
Lucky things happen to lucky people. But after that most of the manifold of outcomes requires true grit and smarts.
a tale of some fun ML debugging
OpenAI: We’re talking about Goblins.
https://openai.com/index/where-the-goblins-came-from/
Matthew Berman
Demis says he wants to see a Western open source AI stack and that we’re losing to China.
He also says Google doesn’t have enough compute to build two frontier (open and closed) models, which is why Gemma is a smaller family of models.
Watch this incredible clip.
Shout out @ycombinator and @garrytan for the fantastic interview.
Matthew Berman: American open source AI is in trouble.
China is eating our lunch.
This is a bigger problem than people realize.
we're starting rollout of GPT-5.5-Cyber, a frontier cybersecurity model, to critical cyber defenders in the next few days.
we will work with the entire ecosystem and the government to figure out trusted access for cyber; we want to rapidly help secure companies/infrastructure.
Saikat Chakrabarti participated in the creation of tech wealth by being early at Stripe. Now he wants to pull up the ladder on the whole industry.
It’s craven. It’s selfish. It is entirely driven by wanting power over us.
He is unfit for any public service. Tell your friends.
Anthropic shouldn’t do this and doesn’t have to
Would be great if it stopped
Theo - t3.gg: Fun fact - if you have a recent commit that mentions OpenClaw in a json blob, Claude Code will either refuse your request or bill you extra money.
This is an empty repo, I'm just calling Claude Code directly. Insanity.
alignment failure
Theo - t3.gg: Fun fact - if you have a recent commit that mentions OpenClaw in a json blob, Claude Code will either refuse your request or bill you extra money.
This is an empty repo, I'm just calling Claude Code directly. Insanity.
When future historians write about Silicon Valley, they’ll have an entire chapter dedicated to the Ron Conway way: how he turned generosity, warmth, and showing up for founders into a winning strategy.
Paul Graham: "Ron discovered how to be the investor of the future by accident. He didn't foresee the future of startup investing, realize it would pay to be upstanding, and force himself to behave that way. It would feel unnatural to him to behave any other way."
https://paulgraham.com/ronco.html
Amjad Masad
When future historians write about Silicon Valley, they’ll have an entire chapter dedicated to the Ron Conway way: how he turned generosity, warmth, and showing up for founders into a winning strategy.
Paul Graham: "Ron discovered how to be the investor of the future by accident. He didn't foresee the future of startup investing, realize it would pay to be upstanding, and force himself to behave that way. It would feel unnatural to him to behave any other way."
https://paulgraham.com/ronco.html
Amjad Masad
When future historians write about Silicon Valley, they’ll have an entire chapter dedicated to the Ron Conway way: how he turned generosity, warmth, and showing up for founders into a winning strategy.
Paul Graham: "Ron discovered how to be the investor of the future by accident. He didn't foresee the future of startup investing, realize it would pay to be upstanding, and force himself to behave that way. It would feel unnatural to him to behave any other way."
https://paulgraham.com/ronco.html
Daniel Jeffries
These folks are trying to ban open source.
They're looking to take away your freedom to choose.
They're also looking to take away the rights of businesses like Cursor to fine tune and make their product faster and cheaper.
They must be stopped.
Their premise is pretty simple "China = bad."
That's it. There is no other technical reason at all. Zero. None.
There is no reason or evidence for this probe. If these models are running on American hardware in American datacenters fine tuned by American companies then there is no ground for them to stand on other than fake "Manchurian candidate" nonsense (aka the model might be trained to do something nefarious but you won't know, like exfiltrate your data secretly, Muhahahahah.). This is absurd at every level.
It coveys a complete lack of basic IT understanding.
As if professional IT admins would not notice network traffic to a random IP address through just basic networking logging and users would not notice anything weird in its outputs or thought patterns.
There is no ground to stand on.
It's just a push to take away your rights and freedoms and to take away the rights of businesses.
Don't stand for it.
(Article in links)
Beff (e/acc)
SF Broz are always the leading indicator of what's going to happen to the normies.
Rn SF bros are vibe coding most waking hours because work feels like trading time for value but with 100x leverage.
There's never been higher value per person and thus demand for competent folks
zerohedge: Apollo's top economist says AI is about to spark a job-market boom
LOL
I'm that guy
roon: people are walking around with their laptops slightly ajar to keep their agents running
Gustaf Alströmer
Y Combinator changed my life. So it meant a lot to have the co-founders of YC @paulg and @jesslivingston in Stockholm last night. Thank you 🙏
Want/need -- would also like a ventilated backpack
Muad'Deep - e/acc: @garrytan Image gen has a hard time with this but you get the idea.
Jean-Rémi King
✨🧠 Tribe v2, our latest model of human brain responses to sound, sight and language can now be (partly) explored on your phone📱:
▶️demo: https://aidemos.atmeta.com/tribev2/
📄paper: https://ai.meta.com/research/publications/a-foundation-model-of-vision-audition-and-language-for-in-silico-neuroscience/
💻code: https://github.com/facebookresearch/tribev2
Browser Use
Browser harness just raised the SOTA bar
Alexander Yue: Been cooking up something amazing lately
New highest scoring browser agent of all time
Jarvis
The downside of Gawker being gone is that a lot of kids who would have become harmless bloggers are now big city mayors.
Brandi Kruse: INSANE. Seattle's Socialist Mayor responds to exodus of wealth from Washington state by saying "BYE" ... then laughing. We're doomed.
chronicle gives codex passive memory over what you’ve been doing with your computer, which unlocks surprising use cases
Andrew Ambrosino: it's still experimental so we hide it a bit, but in the codex app, try:
> what have i been doing very inefficiently on my computer (according to Chronicle). make some recommendations. be direct. tell me what i need to hear.
Datacenters are good actually
https://www.city-journal.org/article/loudoun-county-virginia-data-centers-construction
Abhishek Bhardwaj
10 years ago I took Justin's C++ 11 class at Google. He is a gifted teacher.
Watch this talk where he explains the 500B AI build out from 1st principles. One gpu to world wide racks and why some of the big semiconductor names have moats.
https://www.youtube.com/watch?v=cyJU32ivIlk
How to Build the Future with @demishassabis
https://youtu.be/JNyuX1zoOgU
You’re not gonna believe this guys but I have found my new favorite X account 👀
Philo Groves: just cut a 7k loc file down to ~300
This sort of the AI age’s equivalent to all those malware search toolbars that would make you use an obscure search engine instead
Matthew Berman: My mom just told me she paid for ChatGPT
This shouldn't be legal
World Labs
60 million Gaussian splats. One massive dark fantasy world ready to explore! ⚔️
Created entirely with Marble, this persistent world is brought to life in-browser via our Spark 2.0 LoD system and Three.js
Fly through it yourself and learn more about how it was made 👇
Kane 謝凱堯
Silicon Valley congressman @RoKhanna kissing the ring of the Teamsters who are trying to ban Waymo.
Ro Khanna: I stand with @Teamsters. We do not have planes without pilots. We need drivers on trucks for safety, edge cases, and inclement weather.
AI should be for the people, not just billionaires.
I will continue to fight for legislation to stand up for truck drivers that was vetoed.
Fireside chat at Sequoia Ascent 2026 from a ~week ago. Some highlights:
The first theme I tried to push on is that LLMs are about a lot more than just speeding up what existed before (e.g. coding). Three examples of new horizons:
1. menugen: an app that can be fully engulfed by LLMs, with no classical code needed: input an image, output an image and an LLM can natively do the thing.
2. install .md skills instead of install .sh scripts. Why create a complex Software 1.0 bash script for e.g. installing a piece of software if you can write the installation out in words and say "just show this to your LLM". The LLM is an advanced interpreter of English and can intelligently target installation to your setup, debug everything inline, etc.
3. LLM knowledge bases as an example of something that was *impossible* with classical code because it's computation over unstructured data (knowledge) from arbitrary sources and in arbitrary formats, including simply text articles etc.
I pushed on these because in every new paradigm change, the obvious things are always in the realm of speeding up or somehow improving what existed, but here we have examples of functionality that either suddenly perhaps shouldn't even exist (1,2), or was fundamentally not possible before (3).
The second (ongoing) theme is trying to explain the pattern of jaggedness in LLMs. How it can be true that a single artifact will simultaneously 1) coherently refactor a 100,000-line code base *and* 2) tell you to walk to the car wash to wash your car. I previously wrote about the source of this as having to do with verifiability of a domain, here I expand on this as having to also do with economics because revenue/TAM dictates what the frontier labs choose to package into training data distributions during RL. You're either in the data distribution (on the rails of the RL circuits) and flying or you're off-roading in the jungle with a machete, in relative terms. Still not 100% satisfied with this, but it's an ongoing struggle to build an accurate model of LLM capabilities if you wish to practically take advantage of their power while avoiding their pitfalls, which brings me to...
Last theme is the agent-native economy. The decomposition of products and services into sensors, actuators and logic (split up across all of 1.0/2.0/3.0 computing paradigms), how we can make information maximally legible to LLMs, some words on the quickly emerging agentic engineering and its skill set, related hiring practices, etc., possibly even hints/dreams of fully neural computing handling the vast majority of computation with some help from (classical) CPU coprocessors.
Stephanie Zhan: @karpathy and I are back! At @sequoia AI Ascent 2026. And a lot has changed. Last year, he coined “vibe coding”. This year, he’s never felt more behind as a programmer.
The big shift: vibe coding raised the floor. Agentic engineering raises the ceiling.
We talk about what it
lisan say more mean things about us you're being too nice
Lisan al Gaib: GPT-5.5 is on par with Claude Mythos
- GPT-5.5 average pass rate of 71.4% (±8.0%)
- Mythos Preview 68.6% (±8.7%)
- GPT-5.5 solved a task that takes a human expert ~12 hours in under 11 minutes at a cost of $1.73
This is the the quote I've been citing a lot recently.
kache: you can outsource your thinking
but you cannot outsource your understanding
Ole Lehmann
i can't believe they're really gonna do it again
the last time stripe shipped a primitive this big, it literally birthed modern saas.
shopify, substack, gumroad, all your fav indie hackers...
they exist only because stripe's payments api let them easily accept money with a few lines of code
yesterday stripe shipped the spending version.
anyone can now get an agent to spend money on their behalf.
if stripe's payments api created saas, this spending api creates autonomous commerce:
a new category of businesses that run on agents buying, booking, restocking, and paying on your behalf
to make that concrete for you, here's some cool ideas you can build now:
1. ai ad managers. you connect your meta, google, and tiktok accounts, set a monthly budget cap, and an agent runs your entire paid strategy.
2. ai procurement agents for ecommerce. they monitor supplier prices, auto-order when costs hit your threshold, send you a morning summary of what they bought and why
3. ai travel agents that actually work. they search, compare, book, and pay within your budget rules. no more toggling between 6 tabs to save $40 on a flight
4. ai bookkeeping agents. they handle the recurring operational payments your business already makes every month (contractor invoices, ad account top-ups, subscriptions, etc)
the 18-month window after a new primitive ships is historically when the category-defining companies get built
if i was building right now i'd pick one of these, find the narrowest possible version of it, and ship before it closes
generational opportunities here imo. good luck!
Stripe: Today, we’re launching the @link wallet for agents. It lets you securely empower agents to spend on your behalf. Your payment credentials are never exposed and you approve every purchase.
https://link.com/agents
Judge Glock
Many people want to stir up local opposition to data centers as a sort of proxy war against AI generally.
But as I try to show in this piece, data centers can bring wildly positive returns to the communities that host them.
Garry Tan: Datacenters are good actually
https://www.city-journal.org/article/loudoun-county-virginia-data-centers-construction
We treat Replit, inc as customer number zero of Replit. It goes beyond mere "dogfooding." We expect our usage to have insane ROI. This is an example:
明德: On @Replit one highly-leveraged builder can match the per-week output of an 8-person agency team.
I just spent 22 weeks proving it: 12 production projects, ~$1.4M of agency-equivalent work, all shipped solo. 🧵 for the full breakdown.
The ceiling for what one builder can do
OpenAI
It's never been easier to do everyday work with Codex.
Choose your role, connect the apps you use every day, and try suggested prompts.
Codex helps with everything from research and planning to docs, slides, spreadsheets, and more.
The Wall Street Journal
From @WSJFreeEx via @WSJOpinion: All the cool kids think killing people is OK now. It gives them a kind of edgy thrill to talk about how violence in the service of a good cause might be necessary, writes @jamesbmeigs.
https://on.wsj.com/4tL0rjv
Ari Weinstein
Computer Use runs this use case 42% faster in today's Codex app update.
Ari Weinstein: This is the first time I've ever seen an LLM operate a GUI as fast as a person, and it's surreal.
Secure your ChatGPT account with Advanced Account Security:
OpenAI: Now available for ChatGPT accounts: Advanced Account Security, a new opt-in setting for people at higher risk of digital attacks, with stronger protections including phishing-resistant sign-in and more secure account recovery.
https://openai.com/index/advanced-account-security/
Codex is for everyone, for any task done with a computer
OpenAI: It's never been easier to do everyday work with Codex.
Choose your role, connect the apps you use every day, and try suggested prompts.
Codex helps with everything from research and planning to docs, slides, spreadsheets, and more.
Prompt ➡️ LLC
Arjun Mahadevan (Mr. LLC 🇺🇸): The last tab a founder ever opens to start a business has been closed.
@doolaHQ is integrated with @claudeai and @Replit. You can now form a US LLC without leaving the AI chat you’re already in.
First business formation platform to do it.
Start one in your next prompt.
https://agent-s.app
More powerful than OpenClaw.
So simple my mom uses it daily.
Richard L ₿urton
Sitting in at coffee shop then I glance over seeing someone looking at the GStack Github page by @garrytan - I asked the guy if he's tried it yet, but he's just starting to look at it. When I mentioned GStack another guy in front of me quickly turned around. Shit man, Garry built something viral.
it does seem cool
Max Weinbach: Codex goal feature seems cool
Looks like you can give Codex a goal and it’ll continue to work, plan, and test until it’s done?
I’m just reading the commits here but that’s what I think it is?
big upgrade for codex today!
try it for non-coding computer work.
Many such cases
Richard L ₿urton: Sitting in at coffee shop then I glance over seeing someone looking at the GStack Github page by @garrytan - I asked the guy if he's tried it yet, but he's just starting to look at it. When I mentioned GStack another guy in front of me quickly turned around. Shit man, Garry built
Asked @v0 what it would look like if @vercel shipped @github 😁 (2 prompts)
Y Combinator
We’re bringing Jensen Huang to Startup School for a fireside chat with @garrytan!
From co-founding @nvidia in 1993 to building the backbone of the AI era, Jensen helped turn GPUs into the engine of modern computing.
Apply to attend: https://events.ycombinator.com/startup-school-2026
Kane 謝凱堯
Incredible story about progressive politics:
Criminals keep stealing copper and destroying lights. @LACityCouncil does nothing.
Man comes up with way to stop theft. LA bans it.
Meanwhile, LA is trying to raise taxes for continuous repair instead of punishing criminals.
California Post: Inventor creates ingeniously simple device to end LA's copper wire theft. The city immediately shot it down https://trib.al/4wVGciD
Kevin Patrick Mahaffey
More people would be pro datacenter if every one had a beautiful open-to-the-public heated pool.
Tibo
You can now keep codex going for days.
With GPT-5.5 it will build an entire OS kernel for you if you ask, or find critical bugs in a codebase, or optimize your database schemas, or… the options are endless.
Felipe Coury 🦀: /goal also lands in Codex CLI 0.128.0.
Our take on the Ralph loop: keep a goal alive across turns. Don't stop until it's achieved.
Built by my co-worker and OpenAI mentor Eric Traut, aka the Pyright guy. One of the GOATs I get to work with daily.
codex app becoming incredible
The Economist
When he entered the crowded governor’s race in January, Matt Mahan was a long shot. But an uninspiring field of candidates and a scandal has made the election unpredictable http://econ.st/48zIY50
Photo: Reuters
Tanin
gstack from @garrytan is so incredibly well though out and powerful. My fav commands are
👨💻/qa - headless browser is genius. It's light weight, gets the job done so I won't have to manual test, one more human bottleneck removed
👨💼/office-hours - already finished 1 session. Extremely thorough, considering all facets of a startup. Better than chatting the normal opus 4.7
And I love how at the end it promotes YC. To me this is another classic value-generation playbook executed in a new domain, build something of value and signpost them to find out more. Like a book pointing to the services the author provides.
My friend found the best use-case for his agent
Matt Shumer: https://agent-s.app
More powerful than OpenClaw.
So simple my mom uses it daily.
Browser Use
Introducing: Browser Use Box 🚀
We gave your agent its own computer.
> Runs 24/7, access from web and mobile
> Supports all CLI agents
> browser-harness + stealth browsers w/ persistent auth
Try it for free ↓🔗
Liz4SF
The Ethnic Studies Mandate is literally driving kids OUT of SFUSD, even when they have been accepted to Lowell. SFUSD high school enrollment declined by 14% (2000-2023), while SF private school enrollment grew by 19%. The rigidity of mandating for 2 semesters ONLY in the 9th grade disrupts all other courses that have a track series.
https://open.substack.com/pub/sfeducation/p/enrollment-trends
Liz4SF
Harvard lost at SCOTUS bc they discriminated against students like Jiang bc of race. The Students for Fair Admissions case took 10yrs; the discovery opened the floodgates to a treasure trove of data that was undeniable. Guess what? It’s still happening and not just at ivy leagues, but at UCs. How else would UCs have such low acceptance rates but many are unable to do basic middle school & h.s. math? And why would transfer acceptance rates be double of first year admissions at schools like UC Berkeley, while over 90% of admitted transfers are from community colleges.
https://www.bestcolleges.com/news/students-attempting-uc-to-uc-transfer-enroll-in-community-college/
Redd: He had a GPA score of 97.3%, SAT score is 1560, enrolled in a top high school, and does lots of extracurricular work.
He got rejected by multiple colleges.
They call this equity and inclusion.
