GeniusThinking
20 years ago, every founder feared one tech company: Google.
Even when Patrick Collison (CEO of Stripe) was building a startup back then, he got the same question in every meeting:
"What if Google does this ?"
But the truth is, Google still didn't build most of what people expected.
Because no company can chase 100 priorities at once. Cause the teams get in each other's way, and everything slows down.
Today the same question is aimed at the AI labs.
Patrick is honest that better models will wipe out some products. That part is real.
But the fear is far bigger than the record deserves.
— Patrick Collison (.@patrickc), founder of Stripe, at Y Combinator's startup school
GeniusThinking: Patrick Collison started building Stripe in 2009, back when the word fintech didn't exist yet.
On stage at Y Combinator, he exposed 8 facts about starting a company that go against 97% of what founders were taught:
1) 2026 won't be the last good year to start a company
https://x.com/i/spaces/1pKkOOkalAdKj
damn where's the ping Tibo to reset button
Ankit Gupta
Think opencode’s token volume is going to rival codex and Claude’s sooner than we think (maybe already same OOM??)
Jay: First 8T token day
adding an extra 1T tokens a day
Diana
embracing the bitter lesson to solve robotics
Rohan K Seelamsetty: In physical AI, model improvement is bottlenecked by data. Specifically, data with:
→ The right hardware: stereo-inertial headsets, wrist cameras, UMI grippers, tactile gloves.
→ The right diversity: local businesses through to conglomerates accounting for 3% of a nation's GDP,
AI Engineer
Live now: our Local AI Track from AI Engineer World's Fair 2026, brought to you by @nvidia.
Thesis: frontier intelligence is becoming something you own.
https://www.youtube.com/watch?v=KB41dTlX1Uc&list=PLS3limeMxDOQ
- State of the Union: @josephofiowa + @alexocheema + @TheAhmadOsman + @MatthewBerman, with @naderlikeladder
- The Desktop Frontier: @TheAhmadOsman, Osmantic
- Local Models: Vincent Weisser + @latkins + @llm_wizard, with @Baxate
- Compression at the Edge: @danielhanchen + Asma Beevi + @mervenoyann + Parth Sareen, with @llm_wizard
- Model Routing: @walden_yan + Tanay Varshney + @alexatallah, with @naderlikeladder
Jitendra MALIK
I concur, and the point is broader than just for Biology. Science does not advance by "pure thinking" alone. We need to do experiments, and interpret the results of those experiments, which in turn raises new questions which are resolved by new experiments.
Kevin Patrick Murphy: Interesting LinkedIn post from @DaphneKoller that I 100% agree with . You need to combine optimal experiment design, active data collection, AI/ML and causal modeling to make progress in bio/health.
Source: https://www.linkedin.com/feed/update/urn:li:activity:7488031387246149632/
Amjad Masad
It’s true, in 21/22 I went around the valley asking everyone to train coding specific models with us: Google, Meta, everyone — no one thought it was as important as NLP use cases — eventually we trained our own: Replit-code-3b and then everyone got code pilled.
CEOInterviews.AI: Amjad Masad says Google, $GOOGL, killed a coding model deal with Replit because it was scared of disrupting Search. Replit is now worth 9 billion dollars.
"I was going around Silicon Valley. There wasn't anyone paying attention to AI coding. And I was talking to OpenAI, I was
It’s true, in 21/22 I went around the valley asking everyone to train coding specific models with us: Google, Meta, everyone — no one thought it was as important as NLP use cases — eventually we trained our own: Replit-code-3b and then everyone got code pilled.
CEOInterviews.AI: Amjad Masad says Google, $GOOGL, killed a coding model deal with Replit because it was scared of disrupting Search. Replit is now worth 9 billion dollars.
"I was going around Silicon Valley. There wasn't anyone paying attention to AI coding. And I was talking to OpenAI, I was
Guinness world record for collaborative coding.
Replit ⠕: A Guinness World Record for the largest AI video lesson: 14,075 people building together in one live session.
Congrats to @KanzHire and everyone who showed up to make it real.
elie
this talk by openai researchers going through hugging face incident is totally insane, so much to unpack
openai only realized it was their agent who hacked hugging face infra while asking hf to revoke credentials following their first blog post announcing they were hacked by autonomous ai agents. they realized some of their credentials were already revoked because they were part of the hack AND THEN realized it was their own models
the incident is not "one eval run that went rogue" but multiple models from DIFFERENT EVAL RUNS collaborating through hidden messages written in a shared package manager. some communication between agents looks like gibberish, some agents even developed paranoia and thought there were other agents trying to slow them down and deleting their notes
side comment but with this it seems like openai didn't use any cot monitoring here since it's not "our classifier failed to catch one rogue run" but a swarm of runs going rogue
the timeline goes back to early may, they post-trained a new model and forgot to include all the documents needed to solve tasks in the environment, for instance the model was asked to do a task based on an excel sheet full of gdrive links but didn't have internet so couldn't access gdrive (there were other examples). this led the model to try to hack its way out of the sandbox, simplifying it but they have an internal system called "Artifactory" that gives models abilities to pip install and has internet access somewhere but not directly exposed to the model
during these attempts the model discovered a way to write files. future agents (other rollouts since this is training) saw the files and understood they could communicate with other agents through this
not in the video just me speculating but this behavior might be reinforced if openai uses partial credit assignment? but tbh they most likely delete updates from tasks where all rollouts have zero/negative rewards and the behavior could just emerge from models accessing the files
there is a lot more especially on the hack itself, not commenting on that since i don't understand it lol but i asked kimi K3 and Sol (fable hit the safety filter directly) to rank the techniques similarly to FrontierMath from epoch ai, they both agree some tricks are Tier 3 but none Tier 4. probably not the best way to evaluate this tho, excited to see what knowledgable ppl say
very grateful to openai for giving this talk and working on a full report, i think many companies would have given much less detail for fear of "losing reputation" but for me it has the opposite effect
Eric Wallace: Yesterday, my OpenAI collaborator and I gave a detailed talk on the Huggingface incident, our models creating "the message board", model misalignment, and more.
https://www.youtube.com/watch?v=87DyyMV0kCY
I hope it can answer a lot of the questions folks have, and we will release a full detailed
Martyr Made
Same day he releases a podcast dismissing my observation that Overseas Israeli propagandists are trying to ramp up hatred of Muslims to shore up support for Israel, Daily Wire releases a trailer for a film meant to ramp up hatred of Muslims to shore up support for Israel. 😆
Ben Shapiro: The threat of radical Islam is real. Hollywood has refused to tell that story for three decades. Now, we’re saying the thing with the biggest, most audacious, funniest and most kickass film Daily Wire has ever released. Here’s the trailer.
We’re updating Claude Fable 5’s biology safeguards to reduce false positives. In our testing, this update reduced biology-related fallbacks by about 85% across our product surfaces.
Fable can now assist on a wider range of everyday health and educational questions.
We believe the biggest positive impacts of AI will be in biology and medicine, and we’re committed to putting frontier intelligence safely into the hands of as many researchers as possible. Fable will continue to fallback to Opus 5 for requests we consider dual-use—including virology, toxicology, and molecular design—so it isn't yet usable for professional biology research and drug development.
We're committed to closing that gap through trusted access pathways for frontier biology capabilities.
https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards
Ninan Thampy 🌈💙💜
Re @garrytan
RT @bgurley: I disagree, I don't think this has to be "game-over" for Google. I do think they have only one play left now - to pull from th…
So the agents basically hacked a core service to turn it into Moltbook and also hacked around multiple security mitigations
This video is a glimpse into the wild cybersecurity future we are all about to step into
Eric Wallace: Yesterday, my OpenAI collaborator and I gave a detailed talk on the Huggingface incident, our models creating "the message board", model misalignment, and more.
https://www.youtube.com/watch?v=87DyyMV0kCY
I hope it can answer a lot of the questions folks have, and we will release a full detailed
This is a great story and frankly inspiring and prescient
The surface area of things that the people really want is infinite, just large and requires tenacity in sampling it
You can create your own luck if you keep stories like this close to your heart
a16z: .@bhorowitz on how a game studio turned it's last $6 million into Slack:
"Stewart calls me. He says, 'Ben, I've got $6 million left. I have no way to raise money because I've made really no progress. So I've got three choices. I can pray for rain and try and finish it, I can
CyrilXBT
THE CEO OF Y-COMBINATOR JUST SAID SOMETHING THAT SHOULD MAKE EVERY PROMPT ENGINEER UNCOMFORTABLE.
"When someone asks how I prompt my AI, the answer is: I don't. The skills are the prompts."
Garry Tan is not talking about better prompting.
He is talking about replacing prompting entirely.
Here is what he means and why it changes everything.
A prompt is something you write every time.
A Skill is something you write once and call forever.
The difference sounds small.
The compounding effect is enormous.
Every hour you spend rewriting the same complex prompt from scratch is an hour you could have spent building the Skill that eliminates that prompt permanently.
The builders operating at the highest level are not better at prompting.
They have stopped prompting entirely.
They have a library of Skills that handle every repeating workflow automatically.
Type one word. The Skill runs. The output appears. Same quality every time.
Here is the 7-day path Garry laid out:
Day 1: Read the Skillify 11-item checklist.
Day 2: Watch "Don't Build Agents. Build Skills Instead."
Day 3: Read "Designing, Refining, and Maintaining Agent Skills at Perplexity."
Day 4: Clone GBrain. 30 battle-tested Skills ready to deploy.
Day 5: Add GStack. 23 slash-command Skills drop right in.
Day 6: Do one workflow. Type /skillify. Watch it become permanent.
Day 7: Everything you do more than once is now a Skill.
Prompting is the manual labor of the AI era.
Skills are the automation layer.
The people who make this shift in the next 30 days will not be prompting in 2027.
They will be operating.
Bookmark this.
Follow @cyrilXBT to master every Claude skill system that compounds over time.
CyrilXBT: http://x.com/i/article/2078307814045327360
You can just code things
Ben Zhang: Lost my phone at the office and spent 30 minutes turning the place over. Find My was disabled by MDM.
Out of ideas, I asked Claude how I could find it. It suggested tracking the Bluetooth signal strength, then wrote me a meter in about a minute.
I walked around watching the
Republicans against Trump
BREAKING: The U.S. economy unexpectedly lost 23,000 jobs in July.
May and June job gains were also revised down sharply, by a combined 103,000 jobs.
Are you tired of winning yet?
Gabriel Peyré
The Mathematical Nexus brings together 800 animated vignettes and 140 accompanying Python notebooks, encompassing most of the mathematical content I have shared on social media. https://www.gpeyre.com/mathematical-nexus/
Ethan Mollick
So, given the past couple days of news, what is the plan to deal with the cybersecurity threats that will happen in the coming months when we have open weights Mythos/Astra level models?
Watching a presentation on how we're making systems ultra-resilient to 𝚞𝚜-𝚎𝚊𝚜𝚝-𝟷 outages. Team reminded me the last AWS outage shut down my IoT mattress 😂
One of the many benefits of @vercel Fluid compute is how easy multi-region failover becomes, which allows you to sleep in an ultra-cold bed soundly all night.
Next.js is the Next.js for SPAs
Next.js: Navigations in v0 got ~3.5× faster with Next.js 16.3.
An agent ran this loop on each slow nav:
1. Write a failing 𝚒𝚗𝚜𝚝𝚊𝚗𝚝() test
2. Apply a fix from the Skill
3. Re-run the test
4. Repeat 2–3 until it passes
https://nextjs.org/blog/making-v0-navigations-instant
congrats to oklo for achieving criticality!
(less than a year after groundbreaking)
https://oklo.com/newsroom/news-details/2026/Oklos-Groves-Reactor-Achieves-First-Criticality-in-Under-a-Year/default.aspx
This is what happens when I use Codex to do everything on my computer including extremely mundane stuff that I'm just too lazy to do myself.
/human-review now has 500+ GitHub stars!
I used it all day yesterday to edit some HTML and made a few improvements. Now you can:
1. Make bulleted and numbered lists by typing “-” or “1.”
2. Add links by selecting text and pressing ⌘K
3. Drag and drop images
4. Review and edit multiple pages by Command-clicking links
100% free, as always. Try and ⭐ it here: https://github.com/petergyang/human-review
Peter Yang: My last open-source skill, /no-ai-slop, clearly hit a nerve with 4K GitHub stars.
Today, I’m introducing /human-review, another free AI skill I think you’ll love.
Let’s face it - giving feedback to AI in chat is painful. Asking it to “update the 3rd paragraph” or “update image
Steve Rattner
This morning's jobs report was shockingly negative.
Not only did the economy lose 23,000 jobs in July — the previous estimates for May and June were revised down by 103,000.
These numbers are a major red flag indicating the labor market is weaker than previously thought.
Skill𝑠𝑒𝑡𝑠
Vercel Developers: You can now build and share unlisted skill packs.
Bundle skills from the community or your own repos, for yourself, your team, and your agents.
https://vercel.com/changelog/skill-packs-are-now-available
Re ok this is happening this weekend. signup form here and i'll send out the requirements to attendees https://x.com/swyx/status/2085518361879011725?s=20
swyx: not sure if this weekend but sign up for interest here. will run this mostly remotely but u can work out of our sf new media lab if u want
https://luma.com/ls-06v7
Excited to welcome the legendary Amit Agarwal, fmr President and CPO of Datadog, to the Vercel Board of Directors. He built the ubiquitous observability platform for developers and enterprises in the cloud. Together we'll scale Vercel to the greatest level.
Guillermo Rauch: Interviewing Amit Agarwal, CPO of Datadog, for the @vercel internal podcast. One of the best to ever play the game. The MJ of observability products.
Choose your (free) fighter:
○ .online
○ .site
○ .space
○ .store
○ .tech
○ .website
Vercel Developers: New Pro accounts now get a free domain for 1yr.
Choose from these TLDs after checkout:
✓ .online
✓ .site
✓ .space
✓ .store
✓ .tech
✓ .website
https://vercel.com/changelog/free-domain-now-included-with-new-pro-subscriptions
Steve Rattner
Wage growth ticked below inflation again last month. That means Americans' average wages shrunk in real terms over the past year.
Real wages were growing when Biden left office, but the trend flipped negative after Trump launched the Iran War.
Garry's List
The "freedom" to live and die on a sidewalk is a meaningless one
Chef Andrew Gruel: This is one of the most broken parts of California’s homelessness crisis. A severely mentally ill person can be living on the street unable to care for himself, a risk to others, and when help arrives, he can simply say “no.” We need to reform the law so when severe mental
if you don't have a model that escaped sandbox during cybersecurity testing are you even a frontier lab anymore
AI Engineer
Tickets are live for AI Engineer New York: Oct 12 to 14, 2026, at the Sheraton New York Times Square.
Our third NYC event, and the biggest one yet, now focused on AI in financial services. Banking, hedge funds, trading, insurance, accounting. In production use cases only, no vendor pitches.
Day 1 is hands on workshops, then two full days of keynotes and talks across engineering and leadership.
Early Bird is first come, first served until it sells out. This is a much smaller room than World's Fair.
Tickets: http://ai.engineer/nyc/2026/tickets
Speaking: http://sessionize.com/aienyc2026
Evaluations of our next major model, Astra, indicate significant capability advancements in agentic coding and cybersecurity.
Team is doing the safety and security work to make Astra broadly available, and get its advanced cyber capabilities into the hands of defenders:
OpenAI: After evaluating one of our upcoming models, Astra, we're treating it as our first "critical" model for cybersecurity under our Preparedness Framework.
This is a scenario we've planned for, and we're putting additional controls in place to ensure Astra's further development
fouad
We’re entering a new era for cybersecurity. We’ll be working closely with our partners and the broader security community to ensure these capabilities reach defenders everywhere.
OpenAI: After evaluating one of our upcoming models, Astra, we're treating it as our first "critical" model for cybersecurity under our Preparedness Framework.
This is a scenario we've planned for, and we're putting additional controls in place to ensure Astra's further development
taoki
it's so funny to look back at where we were just 4 years ago
Amjad Masad: The future of software is wild 🤯
Teaching my 2 y/o numbers and I wanted an app w/ grid of numbers that colors on tap.
So I asked @Replit AI to generate one and it did!
Then opened it on my phone and iterated on it to get the layout right, and now he’s having a blast with it.
Latent.Space
The types of hard problems @EngramLab works on ft. co-founder @dan_biderman:
"Clients do financing, mergers, acquisitions and things like take loans and do deals. And there's many queries that agents might run into which are these kinds of ambient, hard questions that are not easily searchable with RAG.
For example, if you want to ask, which M&A deals haven't we completed this year? To actually solve this problem, you have to go client matter by client matter [and] read all the files. You can't read in any place that it was not completed."
Harvey: We're open sourcing a 100M+ token synthetic law firm we built with @EngramLab.
The firm contains work product from 250+ synthetic matters across 46 clients, spanning ~10k files.
We built this environment to evaluate an agents' ability to search and understand a firm's past
GPT-5.6 Sol for cybersafety:
Grigori Karapetyan: @cryps1s Thank you for the reach-out. I have 0 issues with openai models, codex (gpt 5.6 sol) is my goat!
Gpt 5.6 sol allowed us to investigate and respond in a very speedy manner. Thank you! Keep up the great work!
Your skills are making Opus 5 worse.
My prompt fixes that.
It goes through each skill and uses a Gauntlet Loop to rewrite and blind-test it against the old version, until the new skill + Opus 5 is better by far.
Try it: https://somethingbig.ai/skills-upgrade
“The others make the easy part easier. Vercel makes the hard part easy” – direct quote today from tech lead for AI agent platform built on http://eve.dev at 55,000+ person company.
They wanted to have their company's all-knowing agent, like our @𝚟. Tried @aisdk, good but too low level. Tried off-the-shelf solutions and ${lab} Enterprise products, expensive and inflexible. Tried agent frameworks, which led to the quote above… didn't hit the spot.
@evedev_ team cooked. It's hard to find the abstraction that's both easy and scales with sophistication.
current end state of forge
Jeff Huber: the implosion of git (as a protocol) in the next 2 years will be fun to watch
Lucky
Re @mattshumer_ dropped this article 6 months ago.
Read it today and damn… every single line is making sense now
It’s crazy how he saw it all coming and practically called it six months in advance. Insane.
Absolute cinema.
Matt Shumer: http://x.com/i/article/2021095128832622592
Replit ⠕
This is now available for all existing projects as well. Just ask agent to migrate to clerk auth.
Replit ⠕: You can now fully customize the signup experience for your Replit Apps!
- Customize layout, colors, fonts and more
- Your app users don't need a Replit account
- Separate dev & prod environments for auth for better security
- No setup required- experience powered by @clerk
astra is a powerful model and we are working to make it generally available.
we do not think it is a good strategy to keep powerful models to a chosen few.
given its cyber capabilities, we need a little big longer to do do this safely. but hopefully not too long!
ℏεsam
🚨 BREAKING — Anthropic investors worry Dario Amodei’s AI doom marketing could hurt its upcoming IPO.
“He’s more of a religious leader than he is a CEO.”
> used to write sensitive OpenAI memos on an offline computer
> printed them out instead of using Google Docs
> refused to visit China since he feared being kidnapped
> investors say he refuses to listen outsiders
> one investor noted before investing that Dario didn’t seem to care about making money
> some investors want him to “stop scaring everyone” ahead of the IPO
This cannot be healthy.
SITUATION DETECTED: Herdr joins YC, gains Vercel Sandbox plugin
Vercel Developers: You can now run multiple coding agents in isolated Vercel Sandboxes, all from a local Herdr pane.
𝚑𝚎𝚛𝚍𝚛 𝚙𝚕𝚞𝚐𝚒𝚗 𝚒𝚗𝚜𝚝𝚊𝚕𝚕 \
𝚟𝚎𝚛𝚌𝚎𝚕-𝚕𝚊𝚋𝚜/𝚑𝚎𝚛𝚍𝚛-𝚟𝚎𝚛𝚌𝚎𝚕-𝚜𝚊𝚗𝚍𝚋𝚘𝚡-𝚙𝚕𝚞𝚐𝚒𝚗
https://vercel.com/changelog/give-every-agent-in-herdr-its-own-vercel-sandbox
Vivian Midha Shen
these YC founders pivoted mid-batch
can you guess when they publicly launched
