Re Video is now live! Watch here: https://youtu.be/H29laIK8q7M
Deirdre Sommerkamp
60 seconds of Vibecon day 1.
SO much to do and see… this is just a glimpse.
-Met and created with @Replit builders, fellow Ambassadors, and technologists
-Met inspiring artists and learned what's possible in art with AI and robotics
-Finding and connecting with the ridiculously talented @volvoxlabs 🤯
-Hearing authentic discussions with un-conventional questions and brilliant insights
-Many great conversations and new friends
Thank you @HayaOdeh for the intentional design and curated spaces and thanks to the whole @Replit team for making this happen.
Excited for day 2!
noam is one of the people I have most wanted to work with since the very beginning of openai.
only took 10 years.
i think it will be worth the wait!
Noam Shazeer: I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there.
It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together. It has been an honor and a pleasure to
We offer no explanation as to why Noams are so good at AI; we attribute their success, as all else, to divine benevolence.
Noam Brown: I'm always thrilled to have more Noams at @OpenAI, but I'm especially thrilled to welcome @NoamShazeer!
Paper review:
You Don’t Need Strong Assumptions: Visual Representation Learning via Temporal Differences
https://temporal-difference-vision.github.io/
https://temporal-difference-vision.github.io/static/pdfs/tdv.pdf
@AlexiGlad @ninaddaithankar
The premise is that the more data you can use, the fewer inductive biases you should have. Starting with strong priors is helpful with limited data, but eventually, architectural priors will hinder learning true knowledge buried in sufficiently large datasets. That sounds correct.
Concretely, the ad hoc image cropping / masking / augmentations used in self supervised representation learning all make assumptions about what is important in the images, and appendix A gives examples where they can be harmful.
Figure 3 looks very compelling for this argument, but if you notice that the X axis is log scale, it is sketchier; the anchoring values on the left are relative values from experiments on 0.1% of Imagenet, which I would expect to be quite high variance.
Instead of making multiple augmentations of an image for self supervision, this work uses sequential video frames as related image pairs for representation learning. They train two separate models: a frame encoder, and a “motion encoder” that takes the RGB subtraction between the sequential video frames to produce a delta vector. The models are jointly trained so that the first frame’s representation vector, added to the delta vector, will equal the second frame’s representation vector.
They use a fairly substantial 0.25 second stride between images in the pair, noting that too small of a stride results in near-zero differences in slow moving scenes, while too large of a stride gives incoherent pixel jumps.
It looks to me like the model should be stride independent, and they could simultaneously train on many different strides, increasing the dataset diversity.
I don’t like the DINO EMA teacher approach for avoiding collapse, I think SigReg would have been more direct.
The LeWorldModel work uses sequential video frames and SigReg, but it just minimizes latent distances between neighboring frames; you really want to predict (the ‘P’ in JEPA) from one latent to the next. Linear extrapolation based on the previous frame kind of works, but some level of conditioning on the current latent should be better.
Still, I’m unsure about the soundness of using frame subtraction to create the delta, since it has both frames entangled in it, so it really isn’t doing any kind of causal prediction. The architectural prior here is “only represent things that can be disentangled from a delta frame”, and I’m not sure that is universally valuable.
In their limitations section, they note that scaling to larger video datasets did not help their performance, but they expect better datasets and hyperparameter tuning will.
Re Guys are you telling me that /goal build me an amazing game doesn’t work? 🤣
Rust is great
Charlie Marsh: At OpenAI, we're continuing to bet on Rust as the future of systems programming.
I'm proud to announce that we're making a $600,000 commitment to the Rust Foundation, which combines our Platinum membership with additional support for maintainer efforts across the Rust ecosystem.
Rust is great.
We’re making a $600,000 commitment to the Rust Foundation:
Charlie Marsh: At OpenAI, we're continuing to bet on Rust as the future of systems programming.
I'm proud to announce that we're making a $600,000 commitment to the Rust Foundation, which combines our Platinum membership with additional support for maintainer efforts across the Rust ecosystem.
Watched Mortal Kombat 2 and it was terrible. The Furious has set the bar too high for action movies.
a much better way to build iOS apps
Blaida: OpenAI just killed the worst part of building iOS apps with AI.
The new Codex "Build iOS Apps" plugin can now:
→ run your app in an in-app browser
→ open SwiftUI previews
→ hot reload edits
all without leaving Codex.
No more copy-paste-build-screenshot loop. The agent sees
Julia Kempe
I am excited to share that I am joining @amilabs as Director of Research, Paris, working with @ylecun and an exceptional founding team.
Further progress in machine intelligence will require not only scaling foundation models and large-scale engineering, but also new ideas and breakthroughs.
This is what makes AmiLabs such a unique and exciting place to build. Its focus on world modeling — systems that can learn richer representations of the real world, reason, plan, and learn from interaction — is a long-term research direction I am deeply excited about.
I am particularly looking forward to working with extraordinary co-founders @sainingxie, @lxbrun, @michaelrabbat, @pascalefung, @mavenlin, @laurentsolly, and the amazingly talented AmiLabs team, to helping build the research organization and to the journey ahead.
Florian Brunner
Yann Le Cun porte un constat juste : les IA génératives, comme ChatGPT, Claude ou Gemini et autres, sont extrêmement habiles pour manier les langages, mais elles ne sont pas connectées directement à la réalité physique. Les IA génératives n’ont accès que partiellement à une dimension essentielle de notre univers.
Le projet de Yann Le Cun, de créer des world models, avec sa nouvelle start-up, se révèle particulièrement intelligent. Ces modèles inédits seraient en capacité de cerner et modéliser la réalité de notre monde, avec davantage de précision et de discernement. Il en résulterait une meilleure faculté de traiter l’information et surtout une lecture immédiate des dynamiques entropiques, qui façonnent et structurent notre réalité.
Pour une fois, nous pouvons engager une révolution technologique. L’Europe peut s’imposer dans le domaine prometteur des world models. Le succès sera au rendez-vous si l’Union européenne se montre enfin capable de mobiliser de puissants investissements.
RT @FrHalper: A #Vivatech, le chercheur français @ylecun Yann Le Cun défend le droit des États à être maîtres de leur destinée technologiqu…
http://vercel.com/design.md
liemandt
This Monday, 100 students from Stanford, MIT, and UT start our Alpha School summer internship.
Their mission: fuse learning science with AI to build apps that transform how kids learn.
@gdb @demishassabis @elonmusk @karpathy - we'd love early access to your frontier models to put into the hands of these students.
Matt Shumer: While we wait for Claude Fable 5 to come back, here's another wild demo I built for Alpha School.
A 3D game that generates itself on the fly, fusing what a kid has to learn with what they're actually into.
Just the first prototype, done in two days with Fable.
(sound on!)
For people who don’t have a clear sense of the future they want, AI is just another mechanism of control
But in the hands of someone with agency, AI is the breaker of chains, something that lets you do things no humans can do alone.
AI can be a liberator if you choose agency.
Recall Jackie Fielder
Jackie Fielder tried to ban research labs in the Mission and got completely humiliated.
In December 2025 she pushed an 18-month freeze on new labs by forcing every one through a subjective “conditional use” process that would have given her unilateral power to block companies. She claimed she was protecting blue-collar PDR jobs from AI and life sciences firms “destroying” the neighborhood. She called opponents “entitled techno absolutists” and “techno fascists.”
The tech and investment world exploded. GrowSF and others slammed it as “technofascist” and a direct threat to San Francisco’s economic recovery. The backlash was so intense that Bilal Mahmood and the Mayor’s office gutted the bill. It ended up applying only to “outdoor” labs — which don’t exist in the Mission. Every indoor lab was carved out. The whole thing became a total nothingburger.
She waged an ideological war against the city’s most important job-creating sector, lost spectacularly, and proved she can’t build coalitions or deliver for working people. This is what happens when academic theory meets real governance.
Jackie Fielder is hostile to jobs, hostile to growth, and hostile to the future of her own district.
She must resign.
Here’s the actual skill I use to set up my personal AI advisor.
I like to keep the SKILL md concise and clean so that I can share it with others. It tells Claude Code or Codex:
→ What role to play
→ Which context files to read
→ How to give advice
→ When to save new learnings
But the real magic with an advisor skill is the personal context you give it. That lives in files like:
→ plan md: Goals, principles, energy, life
→ learnings md: Patterns from past chats
→ eval md: A checklist for AI to give better advice
The more context your advisor has, the less generic the advice gets.
📌 Full tutorial: https://www.youtube.com/watch?v=H29laIK8q7M
Peter Yang
Here’s the actual skill I use to set up my personal AI advisor.
I like to keep the SKILL md concise and clean so that I can share it with others. It tells Claude Code or Codex:
→ What role to play
→ Which context files to read
→ How to give advice
→ When to save new learnings
But the real magic with an advisor skill is the personal context you give it. That lives in files like:
→ plan md: Goals, principles, energy, life
→ learnings md: Patterns from past chats
→ eval md: A checklist for AI to give better advice
The more context your advisor has, the less generic the advice gets.
📌 Full tutorial: https://www.youtube.com/watch?v=H29laIK8q7M
Ankit Gupta
in china, there are ~10 frontier AI labs, about half of them the offshoot of a money printer (Tencent, Bytedence, Alibaba, etc) and the other half of them new startups.
America has the startups, but has very few money-printer-attached labs. It's really just Meta, Google, and sort of nvidia.
why is an IBM or Cisco or Netflix or amazon frontier AI lab not a thing (w/ actually shipped products not just a research team).
Marijan Cipcic
Looking at the sponsor wall for @aiDotEngineer World’s Fair SF (June 29–July 2), it’s hard not to be impressed.
Huge credit to @swyx and the entire team for building what has become the AI conference you simply can’t miss, arguably our industry’s version of the Super Bowl. 🏈📷
@Daytona is joining again this year as a Gold Sponsor, and excited to see our CEO, @ivanburazin, speaking again at the conference.
See you all there!
Ethan Bloch
Scheduled tasks + Finances = ChatGPT can help optimize your money, 24/7/365.
Here are some of my favorites:
- Monitor my checking accounts for too much or too little cash and tell me what I should do
- Every month give me an update on my subscriptions: 1) what am I paying for 2) have any changed in price?
- Send me a portfolio update every weekday at 4pm
- Every Sunday morning give me a full briefing about my finances
- Tell me once a quarter what tax moves I should consider before year-end
ChatGPT: New in ChatGPT: a better way to schedule tasks.
Scheduled tasks are faster, more reliable, and easier to manage from the new Scheduled page.
The new scheduled tasks experience is rolling out to Go, Plus, Pro, Business, and Enterprise users on web and mobile.
Karan Singhal
One of the ways scaling test-time compute can benefit people most: have reasoning models think really hard about rare undiagnosed diseases.
Today we’re sharing published evidence that this can work, in some of the most difficult pediatric cases!
OpenAI: Together with researchers at Boston Children’s Hospital and Harvard, we published a study in NEJM AI showing how o3 Deep Research helped clinicians revisit previously unsolved rare pediatric disease cases, and find answers for families who had waited years.
Movez
Creator of Claude Code:
"At Anthropic, almost 100% of our engineers are running 100+ agents with self-improving loops
self-improving loops help agents become better with each run."
in a 1-hour podcast, Boris explains how they build agents loops from sratch.
Claude + loops + routines + dynamic workflows - that’s the secret.
Watch the talk, then read how to apply the same playbook to quant trading below.
Horizon: http://x.com/i/article/2067608522959847424
OpenAI for helping find 18 new diagnoses across 376 previously unsolved medical cases.
Includes diagnosing Kyra, who has been trying to understand her muscle weakness since age 9, with a rare form of myofibrillar myopathy shortly before her 28th birthday.
OpenAI: Together with researchers at Boston Children’s Hospital and Harvard, we published a study in NEJM AI showing how o3 Deep Research helped clinicians revisit previously unsolved rare pediatric disease cases, and find answers for families who had waited years.
Nicolas Dessaigne
I met a team this week selling "cut your AI spend." I encouraged them to flip it. Buyers don't get promoted for spending less, they get promoted for shipping more. Frame the same product as "do 3x more with what you already pay," and you go from a line-item nobody defends to a budget nobody questions. Selling savings caps your ceiling at the customer's current spend. Selling output has no ceiling.
jason
come by! https://www.ai.engineer/worldsfair/2026
completely unprompted wow moment from today - asked @DevinAI to make us a @tbpn style breaking news style announcement card for our AIEWF speakers drop tmr, FULLY expecting it to fail at a heavily visual task
and it oneshotted the WHOLE DAMN THING
Ryan Carson: @theo Honestly just use Devin. It's really really good now
Replit ⠕
Replit, now inside @SlackHQ.
Tell Slackbot what you want to build and it'll pull in Replit to make it happen, straight from the conversation.
https://replit.com/partners/slack-app
Latent.Space
Long Live Outputmaxxing: AI compute grids, Anthropic’s coding takeoff, data center backlash, & frontier systems https://www.latent.space/p/anj
@amppublic founder @AnjneyMidha explains why 95% GPU utilization was considered an outage at Google, why the AI race is no longer just about buying more GPUs, how AMP is trying to make FLOPs flow like megawatts, why data center backlash could become one of AI’s biggest bottlenecks, how Anthropic cracked coding through culture and preparation, why DeepMind research hoarding creates a market failure, and why the next frontier may belong to teams that can “output max” across compute, capital, culture, and science.
We've collaborating with hundreds of physicians across 60 countries, 49 languages, and 26 specialties to make ChatGPT great at health-related questions for everyone:
OpenAI: GPT-5.5 Instant is now on par with our frontier Thinking models for health-related questions.
Every week, more than 230 million people turn to ChatGPT with health and wellness questions, and GPT-5.5 Instant is better at recognizing when urgent care may be needed, asking for
Replit ⠕
That’s a wrap on day one of vibecon ⠕
See you again tonight.
http://vibecon.ai
Replit in slack!
Slack: Introducing Slackbot's MCP client, now available with more than 20 partner apps!
With Slackbot MCP, you can:
🚀 Build an app with @Replit
📊 Pull a live @Amplitude_HQ chart
✅ Update a @linear ticket
🎨 Create a @canva design
...and a whole lot more, all without leaving the
Amjad Masad
Re @itamarbengvir @JDVance Have you said thank you once?
David Lieb
Reminder: If an investor tells you they were "an operator", what they are actually saying is "I was an employee".
Manling Li
Planning with the views:
Can VLMs predict how each camera move changes the view, and plan many such moves ahead?
We introduce ViewSuite with 6 DoF camera control and ~165K task instances, testing:
Path-to-View
View-to-Path
Interactive View Planning
A sharp Planning Gap emerges:
+ can roughly "track" how camera action changes views
- cannot "compose" a plan towards a target view at all
We then try to teach VLMs with Reinforcement Learning. - RL cannot teach VLMs such planning ability, only 2.5% success rate with Qwen2.5-VL-7B.
+ With View Graph Distillation (our RL-Graph-SFT framework), 2.5% → 47.8%
Below, we answer these questions:
Q1. What are the failure modes?
Q2. How can we make RL work?
Q3. What has the model learned? Can we open up the model to see before/after? Can such spatial priors transfer to other view related tasks?
Led by @James_KKW, great to work with @LINJIEFUN @zhengyuan_yang @shiqi_chen17 @wzenus @drfeifei @jiajunwu_cs Leonidas Guibas, Lijuan Wang.
A joint efforts with @StanfordAILab @StanfordSVL @MSFTResearch.
Replit ⠕
Replit’s vibecon is the coolest! here’s what you missed:
> blind self portraits
> custom fragrances
> robotic arm art
> mini games
> talks with the top creatives
can’t wait for what’s next <3
you can now teach Codex by demonstration:
OpenAI Developers: Show Codex a workflow once. Reuse it as a skill.
Record & Replay lets you show Codex a recurring task, like filing an expense report or submitting a time-off request.
Codex turns that demo into an inspectable, editable skill.
You control when recording starts and stops.
Your thoughtful reflection is so inspiring and encouraging @smallfly ! As everyone talks about AI and automation, human creativity, story telling and productivity are even more important and essential to our society. @theworldlabs is founded on the premise of empowering human ingenuity and productivity. We are very grateful to be able to work with people like you! 🙏🌐
Hugues Bruyère: @FastCompany just published a great piece on @theworldlabs , @drfeifei , Marble, and the idea that spatial intelligence / world models may be one of the next big shifts in AI.
I was happy to be quoted in the article, but I also wanted to share more context about my own
We did this years ago with our Agent-1 CUA model at @HyperWriteAI
Just a lil early :)
https://x.com/mattshumer_/status/1761118083270180957?s=20
OpenAI Developers: Show Codex a workflow once. Reuse it as a skill.
Record & Replay lets you show Codex a recurring task, like filing an expense report or submitting a time-off request.
Codex turns that demo into an inspectable, editable skill.
You control when recording starts and stops.
Anjney Midha
there are very few people who are as thoughtful about their questions as @swyx
it was fun to sit down to talk through the reality of frontier AI lab creation with him
Latent.Space: Long Live Outputmaxxing: AI compute grids, Anthropic’s coding takeoff, data center backlash, & frontier systems https://www.latent.space/p/anj
@amppublic founder @AnjneyMidha explains why 95% GPU utilization was considered an outage at Google, why the AI race is no longer just about
Garry's List
At UC Berkeley Law, 37.5% of students claim a disability.
Students don’t have to disclose this to employers, and accommodations come with no downside.
Those who don’t opt in are competing at a disadvantage on the exams that determine their futures.
the reasoning paradigm unlocking medical progress for humanity
Noam Brown: When we announced @OpenAI o1 some researchers from other labs told me we made a strategic mistake and should have kept it secret so we could accelerate ourselves and pull farther ahead of the competition. Studies like these make me confident we made the right choice.
Anything you do on a computer codex can do too
OpenAI Developers: Show Codex a workflow once. Reuse it as a skill.
Record & Replay lets you show Codex a recurring task, like filing an expense report or submitting a time-off request.
Codex turns that demo into an inspectable, editable skill.
You control when recording starts and stops.
نـايـف آل طـليان الغامدي
Replit
افضل AI اخترعته البشرية
50 في واحد
Jay Owen
One of the biggest critiques I hear related to AI is that it kills creativity.
If there’s one thing I’ve taken away from VibeCon this week with @Replit is that it actually unlocks creativity.
The future is bright if we use this technology for good.
Great work @amasad and team. The first VibeCon was a hit with the @businessbldrsfl team!
AI Engineer
🚨 BREAKING
(are we doing X right?)
Thariq: I'll be giving the the Day 2 opening keynote at AI Engineer World's Fair on July 1st at 9am.
Hope to see you there, here's a quick preview
x.com
