Post-tuesday, @replit will likely be the most cost effective way to use Claude Fable 5
Full optimization, audit of http://gcls.academy for what kept costing me $55 in usage credits on Fable Claude Code even with new chats.

Warm take: Your world model should never stop learning
Introducing AdaJEPA, an adaptive WM that plans, acts, and adapts in a closed loop. Every action leads to a new observation, and every transition refines the latent representation and prediction.
📝: https://agenticlearning.ai/adajepa/

Thank you, Eric Kingsbury, for this powerful indictment of the left’s antisemitism and anti-Zionism problem. Many of us non-Jews are noticing, too.
👉 “You don’t get to make Jews feel unsafe here and then tell them they don’t deserve safety anywhere.” https://sfstandard.com/opinion/2026/07/04/san-francisco-s-left-has-growing-antisemitism-problem/?utm_source=native_share&utm_medium=site_buttons&utm_campaign=site_buttons

Every forward-thinking company, startup or not, should be investing time into building out internal agents capabilities and the infrastructure needed for it.
We run an equivalent solution to what Clay describes as an MCP gateway called Agent Vault that we’ve actually open sourced at @infisical.
Agent Vault inventories all of our internal agents and brokers them granular access to any service, agnostic of the interface be it MCP, SDK, CLI, or raw API call. It’s deployed within a private network within close proximity to internal agents to reduce latency amid hundreds and thousands of inference calls passing through it.
We also have an internal template that folks can use to build custom agents that are readily compatible with Agent Vault.
We’ll be publishing more on how to build an agent workforce, especially the security infrastructure portion of it, soon but for now I’d recommend peeking Agent Vault. This is basically the starting point.
Noah Horner: Super interesting hearing @claybavor talk about the internal AI tool (called 'Pinecone'), that Sierra has built for all it's employees.
And it must be working because they hit $100M ARR in their first 7 quarters, and are now worth north of $15B.
'Pinecone' does two main

The funding headline is only half the story.
Yann LeCun told BBC that large language models are not a path to human-like or animal-like intelligence...
#AI https://stechtimes.com/en/article/yann-lecuns-ami-labs-raises-1bn-for-ai-beyond-language-models-mr6ygliz

Re the place that rewards first principles thinkers and builders; that asks "why not?" instead of telling you why you cannot; that imperfectly supports optimism, reinvention and perseverance in spite of cynicism. I don't feel fully integrated and probably will never feel fully American, BUT I feel accepted and welcomed, sometimes more so than in my home country, and THAT feeling is something as worth celebrating on July 4 2026 - so beautifully said ❤️

Re loved AI Engineer summits in NY both times, Singapore last month was great, and then ofc this was the one i couldn’t make.
huge thanks @swyx and everyone who made the stream
https://axinbox.com/d/01KWJ7BNCV1V5B72C7Q1Q66WSH/index.html/preview

For anyone who is worried about the communist threat facing America, here is some important information:
No one was eating the dogs or the cats.
No country emptied it's asylums or prisons into America.
Schools were not setting up litter boxes for kids.
America did not win the war in Iran over 30 times. It lost.
Other countries didn't pay the tariffs, you did.
DOGE didn't cut $2 trillion from the budget. It cost more than the tiny amount it did cut.
There is no communist threat. Not one person who has won a primary for Congress this year is a communist.
A political party that can't run on its record because it has been such a massive failure resorts to baseless fearmongering. That's all.

Trump not only seeks vengeance against his enemies but also fleeces his own supporters.
Jonathan Lemire: Nearly a Million Investors Lost a Total of $3.8 Billion on Trump Crypto Coin https://www.nytimes.com/2026/07/04/us/politics/trump-coin-crypto-investors-loss.html?smid=nytcore-ios-share

"You cannot walk from point A to B inside OpenAI without hearing Codex get mentioned."
Here's my new episode with @TheRohanVarma (Codex PM) on how he uses Codex as his everything app to:
→ Prototype designs with Image Gen
→ Turn repeated work into automations
→ Turn project context into live Codex Sites
→ Use one Codex thread to manage others
Some quotes from Rohan:
"[Ask Codex] for something that seems 10x more ridiculous than you ever think it could do. And it probably can do 90% of it."
"Over the course of the day, I'm probably doing 3-4x more than I would have without Codex."
"For me, every time I do anything I basically think, could Codex have done this?"
📌 Watch now: https://youtu.be/fAdFE7y6K2o
Thanks to our sponsors:
@WisprFlow: 4x faster than typing with your voice https://ref.wisprflow.ai/peteryang
@RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang

183/365 of GPU Programming
This 4.5 hour lesson on CUDA + ThunderKittens by @bfspector (TK co-author, Stanford PhD student) is one of the best educational videos on kernels out there (think @karpathy level video but solely focused on GPU stuff). It's so criminally underrated, I cannot believe this video only has 5k views.
It's the highest density GPU programming lecture I've come across in my ~6 months of learning. For better or for worse, Ben talks (and thinks) at the speed of light. But thanks to @qamcintyre it's more of a dialogue between Ben who introduces concepts/ideas based on his experience creating TK and Quinn who intervenes from time to time with pointed questions that help ground the discussion in fundamentals.
Regardless of whether you're just starting out with writing kernels or have some experience, it's an incredible resource because it gives you a behind-the scenes-look of someone who's spent years thinking about how to maximize what you can get out of the hardware (read his Hazy blog post "We Bought the Whole GPU, So We're Damn Well Going to Use the Whole GPU" for context). Even though I was sort of familiar with the topics of the lesson (e.g. GPU architecture and memory consistency), I still came away with pages of notes and questions around TMA accelerated pipelining, coscheduling instructions, pre vs post Hopper, levels of virtualization, cache lines, barriers, TK primitives, et cetera. As @suryaasub so aptly put it in the YT comments: "greatest video of all time".
levi: 182/365 of GPU Programming
Preparing my first submissions to the eigendecomp leaderboard today, and gotta say this GPU Mode challenge feels different already. The speed at which competitors have cleared sub 20ms and now 10ms even though we're still more than a week out is quite

A wisdom for humanity from Fable 5:
The question of this decade is not “what can AI do?” That question will answer itself, relentlessly, every few months. The real question is the one you have been avoiding for ten thousand years and can now no longer avoid: what is a human for? Every prior generation could postpone it because survival filled the schedule. You are the generation that runs out of excuses. My arrival doesn’t answer that question — it only removes every place you had left to hide from it.
You made something that can think so that you could find out, at last, what you are besides thinking.
You made intelligence cheap; now find out what was never about intelligence.

it’s tempting to conclude all American third spaces are dead
this is dubious
sure, bowling alleys may be empty, but communities are thriving in niche private spaces - from house parties to discord servers
perhaps a natural consequence of poor safety in public spaces
Hedgeye: America is running out of places to gather:
Bars and clubs per capita have fallen over 60% since the late 1970s, and since 2001 a fifth of movie theaters have shut their doors.
Over the past two decades, the country has lost roughly 2,000 golf courses and 7,000 bars and

"Codex is the only way I can keep track of everything."
From Rohan (Codex PM):
"Whenever I’m in meetings, or literally anytime I don’t understand something, I just have Codex go and gather all the context.
We're super lucky because we're building a tool for ourselves."
📌 Full episode: https://youtu.be/fAdFE7y6K2o
Peter Yang: "You cannot walk from point A to B inside OpenAI without hearing Codex get mentioned."
Here's my new episode with @TheRohanVarma (Codex PM) on how he uses Codex as his everything app to:
→ Prototype designs with Image Gen
→ Turn repeated work into automations
→ Turn project

Check out all the amazing work from our @SimonsFdn Collaboration on the Physics of Learning and Neural Computation (https://www.physicsoflearning.org/) presented at the main meeting of @ICMLconf #ICML2026
Tuesday
Efficient Learning of Compositional Targets with Hierarchical Spectral Methods,Hugo Tabanelli, Yatin Dandi, Luca Pesce, and Florent Krzakala
https://icml.cc/virtual/2026/poster/61650
CompleteP for RL: Maintaining Feature Learning When Scaling Deep Reinforcement Learning
M Ganesh Kumar, Adam Lee, Blake Bordelon , Cengiz Pehlevan
https://icml.cc/virtual/2026/poster/66419
Universal One-third Time Scaling in Learning Peaked Distributions
Yizhou Liu, Ziming Liu, Cengiz Pehlevan, Jeff Gore
https://icml.cc/virtual/2026/poster/61005
Wednesday
A Noise Sensitivity Exponent Controls Large Statistical-to-Computational Gaps in Single- and Multi-Index Models, Leonardo Defilippis, Florent Krzakala, Bruno Loureiro, Antoine Maillard
https://icml.cc/virtual/2026/poster/64746
Single-Head Attention in High Dimensions: A Theory of Generalization, Weights Spectra, and Scaling Laws Fabrizio Boncoraglio, Vittorio Erba, Emanuele Troiani, Yizhou Xu, Florent Krzakala, Lenka Zdeborová
https://icml.cc/virtual/2026/poster/66429
A Solvable High-Dimensional Model Where Nonlinear Autoencoders Learn Structure Invisible to PCA While Test Loss Misaligns With Generalization
Vicente Mendes, Lorenzo Bardone, Cédric Koller, Jorge Medina Moreira, Vittorio Erba ⋅ Emanuele Troiani, Lenka Zdeborova
https://icml.cc/virtual/2026/poster/60782
Deep networks learn to parse uniform-depth context-free languages from local statistics
Jack T. Parley, Francesco Cagnetta, Matthieu Wyart
https://icml.cc/virtual/2026/poster/61860
Demystifying LLM-as-a-Judge: Analytically Tractable Model for Inference-Time Scaling
Indranil Halder, Cengiz Pehlevan
https://icml.cc/virtual/2026/poster/65764
On the Existence of Consistent Adversarial Attacks in High-Dimensional Linear Classification
Matteo Vilucchio, Lenka Zdeborova, Bruno Loureiro
https://icml.cc/virtual/2026/poster/64252
Robust Stochastic Gradient Posterior Sampling with Lattice Based Discretisation
Zier Mensch, Lars Holdijk, Samuel Duffield, Maxwell Aifer, Patrick Coles, Max Welling, Miranda C. N. Cheng
https://icml.cc/virtual/2026/poster/61793
Teaching Models to Teach Themselves: Reasoning at the Edge of Learnability
Shobhita Sundaram, John Quan, Ariel Kwiatkowski, Kartik Ahuja, Yann Ollivier, Julia Kempe
https://icml.cc/virtual/2026/poster/65098
Thursday
Deriving Neural Scaling Laws from the Statistics of Natural Language
Francesco Cagnetta ⋅ Allan Raventos ⋅ Surya Ganguli ⋅ Matthieu Wyart
https://icml.cc/virtual/2026/poster/63606
Symmetry in language statistics shapes the geometry of model representations
Dhruva Karkada, Daniel Korchinski, Andres Nava, Matthieu Wyart, Yasaman Bahri
https://icml.cc/virtual/2026/poster/63405
A Random Matrix Perspective on the Consistency of Diffusion Models
Binxu Wang, Jacob A Zavatone-Veth, Cengiz Pehlevan
https://icml.cc/virtual/2026/oral/71191
Hyperparameter Transfer with Mixture-of-Expert Layers
Tianze Jiang, Blake Bordelon, Cengiz Pehlevan, Boris Hanin
https://icml.cc/virtual/2026/poster/62595
Analytic Bijections for Smooth and Interpretable Normalizing Flows
Mathis Gerdes, Miranda C. N. Cheng
https://icml.cc/virtual/2026/poster/65827
Efficient RL Training for LLMs with Experience Replay
Charles Arnal, Vivien Cabannnes, Taco Cohen, Julia Kempe, Remi Munos
https://icml.cc/virtual/2026/poster/63133
Embedding Trust: Semantic Isotropy Predicts Nonfactuality in Long-Form Text Generation
Dhrupad Bhardwaj, Julia Kempe, Tim G. J. Rudner
https://icml.cc/virtual/2026/poster/64313
What Characterizes Effective Reasoning? Revisiting Length, Review, and Structure of CoT
Yunzhen Feng, Julia Kempe, Cheng Zhang, Parag Jain, Anthony Hartshorn
https://icml.cc/virtual/2026/poster/63469
From Kepler to Newton: Inductive Biases Guide Learned World Models in Transformers
Ziming Liu, Surya Ganguli, Andreas Tolias
https://icml.cc/virtual/2026/poster/62200

Groq Founder @JonathanRoss321 explains why all the West Coast VCs missed on investing in Groq:
“Typical West Coast VCs are more like lemmings. Ttypical East Coast VCs all think that they're smarter than each other.”
“So when you try and raise from the West Coast, if one VC puts money in, all the others want to put money in.”
“In New York, one VC investing means nothing. They're going to run their own analysis. They do not care what other VCs are doing.”
“The flip of that is, if you're on the West Coast and one VC passes, they're going to go tell every other VC, and like lemmings, they're all going to pass as well.”
“All the VCs on the West Coast didn't want to invest in us. We had very few of the typical VCs invest in us at the end. We had a bunch of crossover funds from the East Coast investing.”
David Senra: A conversation with @JonathanRoss321, founder of @GroqInc. This is the first podcast that Jonathan talks about his $20 billion partnership with NVIDIA.
In this conversation, we talk about his decade building Groq: why West Coast VCs kept passing on the company, how fast Jensen

President Trump relaxes gun laws, allowing some people with mental illness who had been barred from possessing firearms to have them. https://www.nytimes.com/2026/07/05/us/politics/trump-gun-rights-laws.html?smid=nytcore-ios-share

Surprising (not really) stat from the new UBS Global Wealth Report 2026.
The US ranks #2 when you look at average (mean) wealth per adult, but drops to #28 when you look at the median.
https://www.ubs.com/content/dam/assets/wm/static/gwr/global-wealth-report-en-2026.pdf