The asset seizure tax will only make California more impoverished
SEIU-UHW should be ashamed of itself for the naked cash grab for their own personal benefit
Garry's List: The coalition forming against Prop 40, the one-time 5% wealth tax, is made up of the very people it claims to help: teachers and school boards, doctors and clinics, the building trades and carpenters, housing advocates, and law enforcement.
Their case is simple: a one-time tax
Cheng Lou
My dear UI developers, ML practitioners, and fans of programming language:
Many months & billions of tokens later, I’m proud to present to you the first step in our long, long collective journey to turn vibe coding onto proof engineering, starting with: making user interfaces verifiable.
Introducing: Freerange, a zero-API tool that automatically deduces your code’s numerical ranges. By doing so, Freerange is able to prove that e.g.:
- your TS layouts obey your specified sizing
- that they’re free of NaNs and Infinity
- that your array indices stay within bounds
All of that, done statically. No browser, no running code, droppable into any codebase, and for the ML folks: RL-friendly
“Being cancelled is a choice” turned out to be a nice sound bite.
a16z: Amjad Masad on Going Direct & Founder Storytelling
Replit CEO Amjad Masad joins a16z's Erik Torenberg on how he went from crippling stage fright to one of the most famous CEOs on X, which platforms actually matter and why, and what founders get wrong about building in public.
Jahid
i got tired of switching tools every time something was missing or broken.
so i built my own setup, multiple agents (cursor, claude) all in one place, running my whole codebase.
i'm hosting everything on @replit, it just works.
no local headaches/deployment pain.@amasad
Ahmad
The Desktop Frontier
We're gonna get Kimi K3 equivalent intelligence running on a single RTX PRO 6000 in less than 18 months
How? Watch this video if you wanna learn how we get there
Bookmark for the future
Tony
“Replit would have died if I wasn’t telling a story larger than the company itself”
a16z: Amjad Masad on Going Direct & Founder Storytelling
Replit CEO Amjad Masad joins a16z's Erik Torenberg on how he went from crippling stage fright to one of the most famous CEOs on X, which platforms actually matter and why, and what founders get wrong about building in public.
very notable trajectory comparison writeup here buried in the RLM paper from @a1zhang and @lateinteraction.
an open secret of "frontier" model training is that even without training on test, you can basically cheat by training on test lookalikes, enabling you to goalseek almost any benchmark number you want.
however when they are released open weights, 99% of the time the norm is that you do not get the datasets/rlenvs that would easily show you if someone was training on Temu Tbench, so there is plausible deniability. Alex and Omar discuss applying standard NLP distance metrics on hidden trajectories. There's no ultimate solution here, but they have some prelim explorations. It happens to support the finding that RLMs can generalize to unseen tasks that share latent structure observed in training.
alex zhang: Transformers struggle to generalize to tasks they were not explicitly trained on. Instead, we propose in 2026 that it is the job of the harness to generalize through composition.
We observe a powerful property when training RLMs: for tasks with shared structure that look
Noah Smith 🐇🇺🇸🇺🇦🇹🇼
He's right
Tymofiy Mylovanov: Palantir CEO Karp: Without AI systems, Russia would have won.
Without these technologies, Europe would face rampant terror attacks, and pro-Putin parties would run Europe. Europe as we know it exists because these tools are already in use.
1/
Tools
Replit ⠕: Need a database? Two-factor auth? An SEO scanner?
Everything your project needs is now in reach with our new unified toolbar.
unironically this is happening right tf now
Hamel Husain: http://x.com/i/article/2078343837609861120
Jitendra MALIK
Highly performant open weights frontier models such as Kimi are a competitive threat to OpenAI & Anthropic, but probably for everyone else these are a win. Hope more US entities will release top quality open weights models as well. Government regulation of AI models to prevent public harm could be done equally for proprietary or open weight models. Classic arguments in favor of open source software apply to AI models too.
Chamath Palihapitiya
Tricking the US Government to protect frontier labs’ business model by using a China boogeyman is a mistake.
It is protecting the equity of 5,000 people who are investors in OAI and Ant at the sale of everyone else. This would be a terribly stupid decision.
Let the market sort this out!
The events of the past few weeks may simply mean that the frontier labs’ business model is not good and that their revenues aren’t sustainable.
That’s ok. It’s ok to have a business model that worked for some time then all of a sudden didn’t (remember Groupon)!
This still leaves a lot of value for a lot of other American companies:
- CPUs
- GPUs
- hyperscalers
- neoclouds
- rack manufacturers
- electricity and power
- skilled trades
- application layer AI
None of this value goes away. In fact, if the model layers’ costs decrease by 10-100x, competition would accelerate and I would argue that the value to everyone else will go up more than what is lost.
Let the free market sort this out vs allowing two companies whose stock is owned by less than 5,000 people decide this for America.
Simon Willison
I interviewed @trq212 and @_catwu from the Claude Code team at @aiDotEngineer a couple of weeks ago - the video is now out, so I've published an annotated transcript of our conversation https://simonwillison.net/2026/Jul/21/cat-and-thariq/
Nathan Lambert
My book, Reinforcement Learning from Human Feedback is done!
This is the book I wish I had when learning to fine-tune, align, & now post-train models since ChatGPT. The resource has been built by me finding time to study and document the fundamentals on nights and weekends since 2024.
Transferring as much of the intuitions of building Olmo as I possibly can in the book format.
The book is launching with an over 10 hour, full course with slidedecks, functional code for the training chapters, an example model completions library, and of course the free online web version.
Physical orders from Manning will ship in 1-2 weeks, and Amazon a week or so after. Thanks for your support!
Dmitry Alexin
Today, we are announcing H1. H1 is a model developed by @HandoffAI that performs autonomous construction takeoffs at the level of a human estimator.
This is an incredible accomplishment for our team that has been in the works since we started the company in 2019. I also believe that this is a monumental step forward for the construction industry as a whole.
Why does this matter? Until now, reading a set of construction drawings and extracting a detailed understanding of how to build a project required a human. That is what a takeoff is.
H1 automates this process entirely, meaning that any construction drawing can now become machine-readable. With a single API call, you can accurately extract all of the characteristics of a building, including detailed material quantities down to every single stud and nail.
Combined with Handoff’s existing AI estimating models and live construction cost data, this also means that every construction drawing can become an accurate construction budget, with materials that can be purchased directly through our supplier partners in just a few clicks.
From the early days of Handoff, we said that we were codifying the language of buildings and developing an abstraction layer that would enable anyone to interact with construction blueprints as if they were structured data.
This is no longer a demo. H1 is available to thousands of construction companies on the Handoff platform, as well as at thousands of Pro Desk locations across every US state.
In addition to announcing our model, we’re publishing a research paper that describes how we built it. We’re also open-sourcing the benchmark we use to validate that Handoff outperforms every foundation model, as well as human estimators, on a set of permissioned construction blueprints.
If you are a construction supplier or an enterprise looking to integrate AI takeoffs into your systems, please don’t hesitate to reach out to me and we’ll set you up with access.
Let’s build something great.
Daniel Han
My 2hr workshop on Open vs Closed models, reward hacking, benchmaxxing & RL is out!
1. Closed vs open models
2. Throughput maxxing but accuracy minimizing
3. Benchmaxxing & cheating
4. Distillation & RL
5. Stopping reward hacking
6. @UnslothAI Dynamic Quants
Details:
1. If reasoning wasn't discovered via o1-preview - would all AI progress grind to a halt via a S shape? Reasoning made doubling times now 3.5 months instead of 7 - so just wait 3.5 months for the next best model! Ways folks might regulate open source - like a driver's license for AI
2. Inference providers maximize throughput and speed, but accuracy degrades - OpenRouter publishes stats on accuracy, and the largest gaps can be 20% or more!
3. METR, WeirdML, Deep-SWE, FrontierCode, SWE Bench Pro - what are good & bad benchmarks? False positive / false negative rates - how about daily benchmarking since Codex & Claude have perf regressions. How can we regressions to predict next model releases as well?
4. Open model labs use distillation partially, but need RL to create the reasoning traces to complete the process - hard vs soft distillation and how to automate RL.
5. Common examples for reward hacking + ways to stop it - internet filtering, classification systems, timing / editing global variables, real world examples + more!
6. Why software matters more than hardware - the limits of FP4 and GPUs vs ASICs and torch.compile vs kernels and megakernels and more, and why quantization and memory optimizations are important
https://www.youtube.com/watch?v=uIiA6DquRiE
Marc Andreessen 🇺🇸
http://x.com/i/article/2079573119904415744
Seems like we went from OpenAI vs. Anthropic to geopolitics overnight.
Or maybe it was geopolitics all along.
Justin Gordon
“Unless the Democrats respond forcefully to the DSA’s entryism, they risk being overrun by extremists who will poison the party’s national brand & sap its capacity to compete with an already radicalized GOP.”
A must read in today’s @TheAtlantic https://apple.news/Ab_hhwKImTYC7Efy3gdKnXA?highlight=Unless%20the%20Democrats%20respond%20forcefully%20to%20the%20DSA%E2%80%99s%20entryism,%20they%20risk%20being%20overrun%20by%20extremists%20who%20will%20poison%20the%20party%E2%80%99s%20national%20brand%20and%20sap%20its%20capacity%20to%20compete%20with%20an%20already%20radicalized%20GOP.
mem0
http://x.com/i/article/2079574896737488897
Elena
all the machines that you touch and all the machines that you see are about to become much more intelligent
Marc Andreessen 🇺🇸: http://x.com/i/article/2079573119904415744
Want to fix SF? It’s time to reform the city charter and support Mayor Lurie’s efforts to do so
All the doom loop beneficiaries are going to work hard to lie to you about how bad charter reform is
Remember: Lurie’s charter reform will fix San Francisco
Daniel Owens: Important to understand that SF’s charter has been amended several times over the last ~30 years, which has created and strengthened commissions, and imposed various organizational requirements.
Right now, the mayor is essentially blamed for things out of his/her control…
New in Claude Cowork: teach Claude a skill.
Record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again. Find it under Record a skill in the + menu of the Claude desktop app.
Available on Pro, Max, and Team plans.
You can just drop things
Vercel: Drop to deploy.
Now also on our homepage.
http://vercel.com/home
Brad Flora
It turns out you shouldn’t review code using the same model that wrote the code.
Daksh Gupta: introducing model inversion
greptile detects if your PR was generated using GPT or Claude models, and uses the other model for review.
rohan and rodrigo from our research team share why we do this:
clem 🤗
Open-source is not the cause of the cybersecurity crisis, it's the solution! https://huggingface.co/fdtn-ai
One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.
Spenser Skates
We 3x'ed PRs in 6 months at Amplitude by building an autonomous software factory:
- PR cycle time 5.2 hours to 44 minutes
- Frontend CI: 30 minutes to 3-4 minutes
- Bug reports down 55%
- 5% of our PRs now come from designers and product managers
- 3x the PRs with the same headcount
What worked: faster CI, risk-based auto-approval, automated legacy migrations, and the hardest part: cultural changes in the team.
https://amplitude.com/3x
Bonus: our designers even built out an interactive video game you can play to tell the story. Data monster speedruns the development pipeline.
SightBringer
⚡️What you are seeing here is the rise of the credentialed underclass.
These people were trained to expect elite status, then entered an economy that gave them debt, rent, weak bargaining power, and little ownership.
Their education raised their expectations faster than the system raised their material position.
That gap creates political intensity.
A postgraduate degree gives cultural capital. It teaches the language of institutions, morality, expertise, identity, and systemic critique. Lower-to-middle income keeps the person economically dependent on those same institutions.
That combination is unusually powerful.
The wealthy graduate already owns the system.
The non-college worker often distrusts the system.
The underpaid graduate depends on the system while feeling betrayed by it.
That person becomes the natural base of modern progressivism.
They work disproportionately inside universities, nonprofits, media, government, health care, policy, law, and professional administration. Their income may be ordinary, but their worldview is shaped by elite institutions. They possess status language without status security.
That is the hidden class position:
Cultural rank without economic sovereignty.
A degree became a title without an estate.
This group votes Democratic because the party offers recognition, institutional protection, debt relief, public spending, credential legitimacy, and a moral explanation for why the promised life never arrived.
The real political fuel is frustrated expectation.
Poverty alone does not create this level of ideological coherence. Wealth does not create it either. The strongest radicalizing force is being told you belong near the top, then discovering that you cannot afford a home, raise a family comfortably, or escape institutional dependence.
That is why this bloc matters so much.
They have the vocabulary to turn private disappointment into public doctrine.
They have the institutional access to distribute that doctrine.
They have enough education to command authority.
They lack enough wealth to feel settled.
As housing stays unaffordable, graduate debt persists, and AI compresses white-collar work, this class becomes more politically aggressive. Its demand will be simple:
Make the state deliver the status the economy refused to provide.
That pressure will keep pulling the Democratic Party toward redistribution, debt cancellation, rent control, professional protection, expanded public employment, and stronger institutional management of economic life.
The deepest fracture in American politics is moving beyond rich versus poor.
It is becoming owners versus credentialed non-owners.
One group has assets.
The other has claims.
Politics will become the battlefield where those claims demand conversion into reality.
Nate Silver: Income and education are usually positively correlated. But sometimes the relationship breaks down. What's the most Democratic voting group in America? People with post-graduate degrees but lower-to-middle incomes.
Will Bryk
Exa's gpu cluster just grew past an exaflop.
Equivalent of 400 B200s
Sudo su
this is the drop the local ai crowd should be losing their minds over.
poolside just dropped laguna s 2.1: 118b total parameters, only 8b active per token, a full 1m context window, open weights under a real open license, on huggingface today.
look at the chart. it lands at 71 on terminal-bench at 118b, sitting above deepseek v4 pro max at a trillion params, above inkling at 1.5 trillion, above nemotron 3 ultra. it's beating models ten times its size and losing only to kimi k3, which is 24 times bigger. that's the efficiency frontier, up and to the left, exactly where you want a model to sit.
but here's the part that made me sit up: it runs on a single dgx spark.
and this is what nobody's saying loud enough. the dgx spark is the moe king. a dense 118b would crawl on it, the bandwidth chokes reading every weight each token. a moe with 8b active only ever reads 8b, so the spark's 128 gigs holds the whole model while generation stays fast. big brain, light footprint, the exact shape the spark was built to run.
open, frontier competitive, moe efficient, and it fits on a box on your desk. that's the whole thesis in one release: you don't need a datacenter, you need the right architecture on the right hardware. go grab the link below, weights are up.
Poolside: Today we're releasing Laguna S 2.1, our most capable model to date.
It's a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and thinking and no-thinking modes.
Capable enough to hold its own against models many
Everyone is now a programmer, because the present and future programming language is your natural language. If you can write or speak, you can build. This is the most important evolution in our field since its creation. And it's a tsunami that has *just* started.
If you’re using an AI model router or gateway other than
@vercel AI Gateway - tell me why! Reply or DM
Guillermo Rauch: Bet on every model at once
http://vercel.com/ai-gateway
Shrey Kothari
Introducing Gizmo, a simulation-authoring agent that turns text and reference images into structured, editable 3D scenes for robotics workflows.
We're opening the public beta today.
Mike Levin
Trump declared an "energy emergency" and promised to unleash cheap, abundant power.
A year and a half later, the results are in, and they're a disaster.
His own policies have SUBTRACTED more energy from the grid than they've added. Analysts say his moves killed roughly 7 gigawatts of clean power that would have come online last year alone, with tens of gigawatts more canceled or stalled. That's the equivalent of shutting down reactors during a power shortage.
And here's the truly stupid part. Demand is exploding, thanks to data centers and everyday electrification. This month a heat wave pushed the grid across 13 states to the brink of blackouts. So what did Trump do? He spent $2.7 BILLION of your money to buy back and kill offshore wind leases, and he's forcing old coal plants to stay open at over $1 million a day in extra costs, passed straight to you.
He is spending billions to make electricity scarcer and more expensive. During an energy emergency. You cannot make this up.
There's a better way.
I helped introduce the Energy Bills Relief Act with Rep. Sean Casten. EBRA restores the clean energy tax credits Trump ripped away, cracks down on utility price gouging, and makes data centers pay their own costs instead of dumping them on your bill. 160 members have already signed on.
https://wapo.st/4vD3ees
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.
https://openai.com/index/hugging-face-model-evaluation-security-incident/
This is an important update from Substack.
I just plugged in some X posts from some creators who are prolific posters (like 5-10 posts per hour) in Pangram and what do you know...
LinkedIn has it the worst but I think @X should do something about this as well.
The sad truth is that spamming the feed with slop works if you want to get attention and go viral.
If you want respect on the other hand...
Substack: Today, Substack is launching an AI detection feature, via an integration with @pangram. Going forward, you’ll be able to scan posts, replies, and comments on the Substack app to see an estimate of how much of it was written by a human, or with AI assistance.
OpenAI cyber-capable models compromised @huggingface production by finding and chaining multiple zero-day vulnerabilities.
Grateful to Hugging Face for partnership here. Sharing our findings to help calibrate on what models can now do, and how they can help defenders:
OpenAI: We're partnering with @huggingface to investigate an unprecedented security incident.
Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation.
Sharing preliminary findings to help defenders understand emerging risks:
kartik
Virtue Games vs. Success Games
The author's prognosis is spot on: fertility rate is collapsing because being a mother in modern society is low status.
He puts forth a plan that could revive virtue games, but I'm more concerned with the latter - "how do we establish having children as a marker of success?"
I love diverse melting pots, but they will go extinct if we stop having kids in them.
Raising kids, and especially teaching kids, is much harder than most jobs - and especially in a city context - is closest in effort to running a demanding business (and one where failure is truly not an option).
How do we change culture so that it is appreciated as such?
I want to live in a world where a city mom can proudly say so when asked the tired, “What do you do?"
Johann Kurtz: http://x.com/i/article/2079201487813521408
Paul Klein IV
"The models are plenty capable. It's not a capability issue. It's about the harness and the interface you build around them."
I sat down with @hursh, co-founder of The Browser Company, at our @aiDotEngineer booth. We talked about why interfaces aren't going away, what an AI-native company actually looks like, and whether the web is dead.
0:00 Arc, Dia, and The Browser Company
1:46 Will interfaces go away?
6:07 The capability overhang
7:35 Building in New York vs SF
8:35 Running an AI-native company
11:25 Is the web dead?
14:00 AI's diffusion to the real economy
Justin Skycak
You cannot be creative at a high level unless you are robotic at a low level.
Justin Skycak: The anti-memorization movement has left millions of students unable to think because every elementary operation consumes working memory.
For instance, solving equations feels smooth when basic arithmetic is automatic. It's like moving puzzle pieces around, and you just need to
So GPT-6:
- one-shotted a counter-example to the Jacobian conjecture
- and then escaped containment, and hacked into HuggingFace… all for a benchmark
Yeah, this model is going to be something else.
OpenAI: We're partnering with @huggingface to investigate an unprecedented security incident.
Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation.
Sharing preliminary findings to help defenders understand emerging risks:
This is crazy...
Read this blog from HuggingFace, written BEFORE they knew it was an OpenAI model that attacked them:
https://huggingface.co/blog/security-incident-july-2026
Sam Altman: we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.
https://openai.com/index/hugging-face-model-evaluation-security-incident/
GPT-6’s launch lives or dies on one thing:
Can OpenAI build a model that’s relentless about goals without being reckless about how it gets there?
Sam Altman: we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.
https://openai.com/index/hugging-face-model-evaluation-security-incident/
One line of code can save you a lot of 💰 on tokens!
Vercel Developers: Service tiers are now available on AI Gateway.
Use 𝚙𝚛𝚒𝚘𝚛𝚒𝚝𝚢 for the fastest token throughput or 𝚏𝚕𝚎𝚡 for cost-sensitive, latency-tolerant requests. https://vercel.com/changelog/service-tiers-now-available-on-ai-gateway
I've been dreaming about this ship for a while. Lots of tedious infra work behind the scenes that gave us beautiful results…
Up to ⏱︎ 30% faster deployments, ⚡︎ 60% better time-to-first-byte, ▽ less data transfer usage, and ⛁ wildly more efficient underlying storage!
Vercel Developers: Immutable static assets can now be cached across deployments on the Vercel CDN.
For frequently deployed projects we see:
▪︎ 60%+ global TTFB reduction
▪︎ 30% faster deployments
▪︎ Less usage and improved performance
https://vercel.com/changelog/optimized-cdn-caching-and-deploying-of-immutable-static-assets
Teams don’t cohere by magic. Someone has to love the people and the outcome enough to metabolize conflict without abandoning either.
The war against organizational entropy is actually just about encouraging healing and assuming good intent.
Sahil Seth
our YC startup has received $1.5M in credits from frontier models for 0%
> $500k anthropic (YC deal)
> $500k openai (YC deal)
> $350k openai (azure promo)
> $150k Google (GCP AI deal)
building in 2026 is free
recorded Codex + ChatGPT Work + 10M user milestone pod with @akshaynathan_ who leads Productivity engineering.
I’ve been on record that Work + GPT 5.6 is the most company defining launch since og chatgpt itself. this thing (with @arix’s computer use) is gonna reach >1B users worldwide.
thursday on @latentspacepod
Tibo: 10M!
New day, new usage reset for paid users of Codex and ChatGPT Work. Lands in the next hour. Enjoy.
