Come and learn about «Anatomy of Massive Activations and Attention Sinks», 📆 today (Thu) at 🕥 10:30, 📍 HALL A #3002. #ICML2026

There you have it.
DSA Watch: DSA Members: "The most important thing we can do is take America down from within"

i don’t love that anthropic and openai are starting to release models much later publicly than they do privately to employees, friends, and influential people.
as these models become increasingly capable, it creates a dangerous precedent. how is it going to play out when a handful of individuals and companies had gpt 10 five months before you?

If “All Housing is Recovery Housing,” why do 26% of all drug overdose fatalities in S.F. happen in Permanent Supportive Housing?
It’s time move beyond ghoulish drug-tolerant mandates that are failing! Come to City Hall Thursday at 10:30 a.m. to speak for Drug-Free Housing options.
Here are the details...
Drug-Free Supportive Housing Ordinance Hearing
Public Safety and Neighborhood Services Committee
10:30 a.m. – Thursday, July 9, 2026
Board Chambers, Room 250
San Francisco City Hall

Fun conversation with @swyx on our journey building the cloud for true elastic inference, sandboxes, and more. And of course, how we're evolving Modal's dev experience to be better for agents.
Latent.Space: Modal's Agent-Native Cloud: DX→AX, sandboxes, elastic inference, and 100,000 rollouts https://www.latent.space/p/modal2026
@modal CTO @akshat_b explains why developer experience is becoming agent experience, why agents need infra they can operate instead of YAML they have to reason through,

I really enjoyed this piece. It’s a rare article that explains, at an intuitive level, how “self-evolving agent products” actually work today.
My takeaway is this: a system needs to build an effective data feedback loop from real user interactions by controlling context; move beyond static benchmarks by creating accurate end-to-end behavior evals that can become automated optimization loops for agent iteration; and use the model’s own reasoning and coding capabilities to drive product improvements at scale (i.e. automatically cluster large volumes of production logs to identify systemic failures, then have AI write the patch, run tests, and submit the PR on its own.)
Through mechanisms like these, developers can build agent systems that can self-repair and continuously improve.
Hopefully one day I’ll be able to share my learning and observation like this one too.
Amjad Masad: Many are asking how Replit is improving so rapidly—we closed the loop and the agent is self-improving.
Technical details here:

I will present tomorrow our recent work on Crys-JEPA, where we use Joint Embedding Predictive Architectures to learn crystal distributions for materials discovery.
https://www.linkedin.com/posts/zhongpc_ai-energymaterials-batteries-share-7464854007417581570-drdn

Trump’s cuts in funding for higher education is reducing the number of PhD candidates and “raising fears that the nation’s capacity to produce new science could be diminished.” https://trib.al/TzuHUm6

Personally I don’t think @sfgov, a government that spends >$2B a year, half of Denver’s entire budget, on homelessness but still manages to have some of the worst homelessness in the country, should run a bank.
ConnectedSF: At the First Reading for the "Charter Amendment - Municipal Finance Corporation or Public Bank" at today's Board of Supervisors meeting. The Board voted 10-1 in favor to file a motion for this item be read a second and final time at the July 7 meeting. Bravo @alankennywong for

Re @KenRoth The biggest risk of AI is the concentration of power in a few dominant providers of proprietary AI assistants.
The only solution to AI sovereignty is open source foundation models.

All problems have been solved by OpenAI!

an interesting property of traditional capitalism is that businesses often have to pick between scale and culture
but with AI, small teams are scaling revenue exponentially without losing their culture
a new pareto frontier is emerging in this regard

Socialists imagine a class struggle. In their made-up fantasy the CEO is in competition with low level workers, the wealthy entrepreneur is stealing from the underpaid nurse.
In reality, workers do not compete vertically they compete horizontally.
Entrepreneurs compete with entrepreneurs. Investors outbid each other. CEOs are benchmarked against other CEOs. Nurses are hired from a pool of nurses. Etc.
The CEOs pay has no correlation to the entry level workers. The Football star on £300K a week isn’t linked to the person selling drinks in the stadium. A biotech entrepreneur raising VC capital isn’t paid relative to a cleaner.
What is linked is the demand and supply dynamic of each role.
If a company places an ad for a qualified truck driver and 150 people apply for the role, then the company knows it does not need to increase wages for that role. If the company has an open role for months, it is forced to look at the compensation package.
Same for a CEO. A board representing shareholders would like to hire a CEO for a lot less if they could. Their dream scenario would be to hire a CEO who brings in institutional investors, attracts top executives, drives innovation and growth, keeps margins steady and is a good public face for the business even under pressure. It turns out there aren’t a lot of these people looking for work and if you want one you have to pay more than other companies are offering.
The class struggle isn’t vertical it’s horizontal. CEOs are in competition with CEOs. Retail workers are in competition with retail workers. Demand and supply dynamics set the price.
Sure you can say that a CEO want’s profitability and would like wages to be lower BUT it’s not up to the CEO - demand and supply tension sets the price of workers. An Airline like RyanAir would like free pilots if they could get them but they can’t… so they pay the market rate.
The reason incomes are rising at the top and falling at the bottom is not class warfare. It’s technology and globalisation.
Technology makes basic jobs simple, remote or fully automated. At the same time tech makes executive roles more leveraged, more important and more valuable.
A CEO used to run a smaller organisation. Today a CEO who’s 2% better on a $5B company is generating $100M more. Seems sensible to try and pay a few million to get $100M.
Globalisation has put workers from all over the world in completion with each other - downward pressure on wages. Globalisation has given CEOs more market opportunities to explore - upside opportunity to unlock.
The rich are not very interested in buying houses that poor people own. The poor are not buying up the homes the rich want. They are separate groups living separate lives. Try finding the genuinely rich people whose strategy is to hoard normal residential homes - it barely exists as a thing. About 85% of landlords are people who own 1-4 properties. Super-landlords (100+ properties) are 0.2% of landlords and own a tiny fraction of the 30M homes in the UK… and they’re heavily taxed.
Class warfare isn’t real. It’s an imagined war in the minds of socialists.
Demand and supply dynamics are real. To the degree it is measured in class, it’s a horizontal competition not a vertical one.
Gary Stevenson: There's a difference between normal people spending money and really rich people spending money. And it explains why our economy is failing.

I am coining a new word - Yanntificate. To espouse and communicate sensible AI policy.
Yann LeCun: @KenRoth The biggest risk of AI is the concentration of power in a few dominant providers of proprietary AI assistants.
The only solution to AI sovereignty is open source foundation models.

(1) Today we're releasing Muse Spark 1.1 -- a strong agentic and coding model at a very low price. It's available through our new Meta Model API and in Meta AI.

Most founders obsess over dashboards and aggregate metrics, but some of the best product insights come from understanding how individual users actually use their product.
In this episode of Startup School, YC's @dflieb walks through one of his favorite tools for better understanding your users, the dot plot. It's a simple two-dimensional grid that reveals usage patterns no aggregate chart can show you.
He’ll cover why it gives founders a better sense of product health, what patterns to look for, and real-world examples of how dot plots helped teams at Google Photos and PayPal.
00:52 — Why DAUs Lie to You
01:39 — What is a Dot Plot and How Does it Work?
02:50 — Picking the Right Event to Track
03:34 — Reading Patterns in the Dots
05:17 — Tracking User State & Attributes
06:16 — The PayPal Fraud Insight
07:59 — Dot Plot vs. DAU Graph
08:56 — Finding Features That Drive Retention
10:30 — Scaling Dot Plots to Billions of Users
11:13 — The $80K Contract That Churned
11:57 — Common Dot Plot Mistakes
12:41 — Dot Plots + Cohort Curves

Best District Attorney in maybe the entire country. 👇
Brooke Jenkins 謝安宜: Four years ago, I was sworn in as San Francisco’s District Attorney. While there is still more work ahead, I’m proud of the progress the @SFDAOffice has made in standing up for victims, reducing crime, and building a more balanced justice system that holds offenders accountable.

Who wants Drug-Free Supportive Housing? San Franciscans do — and by an overwhelming margin!
It’s time to center the needs of supportive housing residents THEMSELVES, by offering them the choice to have the same lease provision banning illicit drug use found in every standard residential lease in S.F.
Speak out today at City Hall at 10:30 a.m.
Information below...
Matt Dorsey: If “All Housing is Recovery Housing,” why do 26% of all drug overdose fatalities in S.F. happen in Permanent Supportive Housing?
It’s time move beyond ghoulish drug-tolerant mandates that are failing! Come to City Hall Thursday at 10:30 a.m. to speak for Drug-Free Housing

We stand at a critical crossroads in the debate over AI governance in the United States, and it feels like we are inching closer to a very serious battle over whether or not open source models will even be allowed in an environment where a new de facto licensing regime has been taking shape.
Lacking formal congressional statutory frameworks or clear administration rules (like the diffusion rule revision), we appear to be left with a sporadic, arbitrary, non-transparent process for model review. The fiction of “voluntary” agreements hangs over this debate, and some large model developers are already showing an incredible willingness to bend over backwards to accommodate national security-related officials / orders that the rest of us are not privy to. It's a very opaque process. And those model developers are expected to play ball with those officials, or else their models get pulled from the market or held up for long periods. Or they will lose any government procurement contracts they have. There is nothing “voluntary” about it when that Sword of Damocles hangs in the room.
As this mess worsens, at some point the question of how to handle open source models will come into sharper focus because it will have to. I've even heard some rumors lately that something may be coming from the admin on this front to address this.
Needless to say, if this informal new AI model review regime expands and takes on more pre-vetting characteristics / requirements, it is hard to see how open source players could comply with such quasi-licensing of AI models. Specifically, if this ambiguous new regime is accompanied by a general presumption of ‘restrict-until-permitted,’ then that would spell doom for open source. That is a very dark path for our country.
Worse yet, of course, would be a move by national security officials to more directly restrict open source models and capabilities. If that happens, then we would be right back in the thick of a Clipper Chip-like battle along the lines of what we saw in the late 1990s. That is a much darker path for America.
Meanwhile, open source developers have no “golden shares” or other goodies to offer the government to make their problems go away.
Let’s be clear: If our government takes the dark path, it will become the single most important battle over computational freedom of modern times. It is time for people to make a stand in defense of open source before it is too late.

Ex-NVIDIA engineer who built Unsloth explained RL, kernels, reasoning, quantization, and agents in 2 hours 42 minutes - better than $5000 fine-tuning bootcamps.
pick the base model -> write triton kernels for 2x faster fine-tune -> quantize to 4-bit -> run GRPO/DPO -> ship a reasoning model on your single GPU.
That loop is why Unsloth is the default way to fine-tune Llama, Qwen, Gemma, and Phi on hardware you already own.
Unsloth + Triton kernels + 4-bit quantization + GRPO/DPO + single-GPU fine-tuning - that's the stack.
Watch and save it, then fine-tune your first model tonight.
h100envy: http://x.com/i/article/2068798868930920448

🚀 We open-sourced Sighthound today.
Sighthound is a Rust-based static vulnerability scanner for source code. It runs locally or in CI, uses Tree-sitter parsers, and supports pattern-based detection plus taint flow analysis. It also comes with all it's rulesets with no paid account or signup needed.
https://github.com/Corgea/Sighthound/
Why build another static scanner in 2026?
Semgrep and OpenGrep are useful, but we kept running into a few constraints: the OCaml core raises the contribution bar for many developers, adding language support is not as simple as we wanted, and some higher-signal rule content in the ecosystem sits behind accounts or paid offerings.
We wanted something fast, inspectable, easy to extend, and fully open.
A few implementation details:
- Rust binary for local and CI use
- Tree-sitter parsing
- Pattern matching and taint analysis by default
- Cross-file taint propagation and dependency tracking
- Parallel file discovery and scanning
- Text, JSON, and CSV output
- Rules written in RON and deserialized into typed Rust structs/enums
- MIT licensed
Current support includes Python, JS/TS, Java, Go, C#, HTML, PHP and Ruby.
It focuses on source-code vulnerability classes like command injection, SQL injection, XSS, path traversal, code injection, unsafe deserialization, and crypto issues. It is not a secrets scanner.
We're still have a lot of work to getting to where we want it to be so, we would love the community to test it, break it, report issues, and tell us where it falls short. PRs for rules, language support, fixtures, and false-positive tuning are very welcome.

Open source powers everything in America. We must fight any and all attempts to restrict it in the age of AI.
We must never surrender to the short-sighted safetyists and hawks.
Open source is the entire reason American tech dominates across the globe. If you can't see this it is largely because it is invisible but your blindness to it is no excuse. You are protected by a smoothly running invisible shield the powers everything and that shield is open source.
It powers every major cloud. The router in your house. Your phone. The servers at your work. The NASDAQ and NY Stock Exchange. Radar systems. Just about every website you know and love. The SaaS you're reading this on right now, called X.
Restricting it in the AI era is idiotic, short-sighted and foolish at an almost mind boggling level. It is civilizational suicide.
Take Microsoft. In the early days of Linux they pushed hard against it and called open source "communism" and "cancer." Imagine if their short-sighted stupidity had won the day? Their Azure cloud now predominately runs Linux.
They would have destroyed *their own* future revenue with their foolishness and greed.
In this fight we must never surrender.
We shall fight on the beaches, we shall fight on the landing grounds, we shall fight in the fields and in the streets, we shall fight in the hills.
We shall never surrender!
Adam Thierer: We stand at a critical crossroads in the debate over AI governance in the United States, and it feels like we are inching closer to a very serious battle over whether or not open source models will even be allowed in an environment where a new de facto licensing regime has been

Today. 10am PT.

In San Francisco, the public school district (55% reading, 46% math proficiency) will block downtown commercial revitalization projects unrelated to schools.
San Francisco Chronicle: NEW: The prospective buyers of the shuttered San Francisco Centre mall have walked away from the pending deal to buy the downtown shopping center after months of due diligence, according to sources familiar with the talks. https://www.sfchronicle.com/realestate/article/sf-centre-mall-downtown-sale-dead-22337271.php?taid=6a4e7befeafdd900014de862&utm_campaign=trueanthem%2B3988&utm_medium=social&utm_source=twitter

OpenAI just made GPT-5.6 available to everyone, so I wanted to answer the most burning question:
Is GPT-5.6 better than Claude Fable 5?
I tested both models across 6 real use cases:
→ Frontend design (5.6 has closed the gap)
→ Build a 3D Star Fox-style game
→ Edit and publish clips with browser use
→ Add a feature to a mobile app
→ Get life and business advice
→ Improve my personal AI OS
📌 Watch my full head to head comparison here: https://www.youtube.com/watch?v=8mY9wx_iMSU

OpenAI just made GPT-5.6 available to everyone, so I wanted to answer the most burning question:
Is GPT-5.6 better than Claude Fable 5?
I tested both models across 6 real use cases:
→ Frontend design (5.6 has closed the gap)
→ Build a 3D Star Fox-style game
→ Edit and publish clips with browser use
→ Add a feature to a mobile app
→ Get life and business advice
→ Improve my personal AI OS
📌 Watch my full head to head comparison here: https://www.youtube.com/watch?v=8mY9wx_iMSU

GPT-5.6 Sol sets a new SOTA on ARC-AGI-3: 7.8%
Sol is the first verified frontier model to ever beat an ARC-AGI-3 game
It is the best model at orienting in a situation it's never encountered

Congrats to @jmorgan, @mchiang0610 and @ollama on their $65M Series B!
They built the easiest way for developers to get up and running with open models, and it's become the leading platform for exactly that, with 8.9 million developers and 85% of the Fortune 500 using it today. All with just 14 employees.
https://ollama.com/blog/all-aboard-open-models

Reminder that Rep. Ro Khanna defended Hasan Piker—a man who celebrates the murder of the wealthy—even while quietly running one of the most profitable stock portfolios in Congress.
Kane 謝凱堯: Rep @RoKhanna has been playing populist w performative policies like banning Waymo.
Instead of Congress's electronic financial disclosures, he hand-files papers that are not searchable.
I burned some tokens and OCR's all >1000 pages and >$300M of his finances! Site shortly.

GPT 5.6 has closed the gap with Fable on frontend design.
I asked both models to make a travel site for my upcoming Japan trip:
→ GPT 5.6 created a beautiful site with a 3D Torii gate
→ Fable 5 made a nice site too with animated snowflakes
Tip: Ask AI to “add a 3D WebGL element to the hero” when you want animated 3D graphics like this.
📌 Watch me walk through both sites live here: https://youtu.be/8mY9wx_iMSU?si=4Mu5bZHRyWHbnQ3y&t=116
Peter Yang: OpenAI just made GPT-5.6 available to everyone, so I wanted to answer the most burning question:
Is GPT-5.6 better than Claude Fable 5?
I tested both models across 6 real use cases:
→ Frontend design (5.6 has closed the gap)
→ Build a 3D Star Fox-style game
→ Edit and

Massive day for us @OpenAI:
- GPT-5.6 SOTA at ~everything & by far most token efficient
- Agents for everyone in the new ChatGPT app Work and Codex modes
- Work mode available on desktop (most powerful), web & mobile
- Ultra mode
- Sites out to everyone
- Artifact templates

🆕 Congrats to @OpenAI on the highly anticipated launch of GPT 5.6 Sol, Terra, and Luna!
https://youtu.be/pMggiOb18tc
Here's @romainhuet, @embirico, and @steipete on why this is the dawn of the Golden Age of AI Engineering:

OpenAI Build Week is here.
The challenge starts July 13. Across the week, join live sessions and community events with builders around the world.
Registration is open.
Bring the idea sitting in your backlog.
Build it with Codex.

The Future of Meta Superintelligence: A 1 Year Progress Update
A top tier RL environment startup spawns out of thin air,
the most aggressive compute ramp we've ever seen,
2000km+ scale-across, and some advice for Google DeepMind
https://semianalysis.substack.com/p/the-future-of-meta-superintelligence

NEW: The New York Times reports that "China, Russia and, to a lesser extent, Iran have sought to use state media outlets to turn the controversy over data centers in the United States into 'a domestic fracture point.'"
Between Jan. and June, state media from these 3 countries mentioned data centers 700 times.
The NYT report cites BPI's research and highlights the work of Chinese propagandists and "covert Russian information operations" to foment anger against data centers.

I asked both GPT 5.6 and Fable to build a 3D Star Fox-style level complete with enemies, power-ups, and even a boss battle.
Here are the results. It’s incredible that both models can one-shot a working game from a simple prompt in about 10 minutes.
However, I have to give the slight edge to Fable on this one for remembering to support barrel rolls 😅
📌 Watch me play each game here: https://www.youtube.com/watch?v=8mY9wx_iMSU&t=344s
Peter Yang: OpenAI just made GPT-5.6 available to everyone, so I wanted to answer the most burning question:
Is GPT-5.6 better than Claude Fable 5?
I tested both models across 6 real use cases:
→ Frontend design (5.6 has closed the gap)
→ Build a 3D Star Fox-style game
→ Edit and

Ran for almost a week.
Looks right in line with METR’s time-horizon graph.
Matt Shumer: GPT-5.6-Sol one-shotted this voxel-based Manhattan.
Just look at the precision... it's insane.
It ran for almost a week, completely autonomously, to get the job done.

YC is coming to NYC next week! 🗽
We're hosting a happy hour for designers with YC's @aaron_epstein, @raphaelschaad, and @eve_bouff.
If you're a designer who wants to build, come hang.
RSVP: https://events.ycombinator.com/nydesignhh-summer26

Technical writing is the #1 top-of-funnel motion at a $12B company
@philipkiely broke down exactly how he does it, his full writing process, the hindsight 20/20 lessons, and how one article drove 500K+ views
This [technical] Write and Learn workshop #5 is part of our Independent Studies series, hosted with @swyx @KernelLabs_ai.
All sessions are recorded. Our next one will be in two weeks :)