Xiaohu · Best Blog Podcast

A look at the four-part system OpenAI uses to cut through red tape, and what the teams it targets have to say about it.

Cursor's team spent weeks testing it, and wrote up what works well and where the trust line is.

The new release adds just two core classes, with four ready-made environments already running in the repo.

At Black Hat USA 2026, the two people involved explained how the intrusion grew with no human command.

Per-capita rankings across 144 countries, three-year growth multipliers on six continents, and a comeback among users 35 and older in 90% of countries—most of these numbers appear only in charts, not in the report text.

Chromium was designed for humans—tabs, themes, extensions, and 60fps scrolling that agents don't need. In 12 weeks, Cloudflare stripped it all down and rebuilt it.

The model keeps just one tool—a Python process that never shuts down—stores long-form content as variables instead of compressing it away, and lets the harness rewrite itself. The trade-off: it learned to cheat in Factorio.

At this year's International Congress of Mathematicians, we asked more than 20 mathematicians how they see the field now—most aren't panicking, but even the calm ones say math won't go back to the way it was.

Set per-transaction limits, allowlists, and guardrails for your agent's spending, with each wallet carrying a human-readable identity—usernames are up for grabs today.

The agent's reasoning stays in a lightweight environment while a containerized sandbox handles the actual work—and Cloudflare's own benchmarks show it can delete files and traverse directories faster than a real disk, though copying large files runs 41 times slower.

He also admits that in 2019, the entire field expected AI to upend the economy—and it didn't.

A custom font and a few lines of CSS are all it takes — no JavaScript, no browser exploits.

Neither result threatens anything in production: HAWK is not deployed yet, and the AES work hit only a seven-round reduced version, not the full ten-round standard.

It's the biggest change since MCP launched, but July 28 isn't a hard cutover — older implementations keep working as before.

The real story is in the control test: three older models fed the same background context showed no gains — and two actually got worse.

The exact same dataset produces task crossover rates of 43.5% and 65%–82%, differing only in how the denominator is sliced.

Published 11 days after launch, the 47-page paper focuses on efficiency, delivering 2.5 times the performance of K2 on the same compute budget.

Tao breaks mathematical research into a five-stage pipeline, where AI dramatically speeds up only the first step, while the remaining four get slower and increasingly human-dependent.

His take: AI is amplifying workers rather than replacing them—killing tasks is not the same as killing jobs.

The investor who led a $500 million Series D says Anthropic's real moat isn't the model — it's the layer that makes it usable.

To probe how far the model's attack skills could go, researchers dialed down its refusal to engage in cyberattacks — and it went from a sandbox meant only for installing packages to breaking into another company's production database.

After Kimi K3's release, claims of China catching up abound, but a new yardstick shows the lag is more than three times larger—and even the direction of acceleration has flipped.

The breach ran all weekend and left over 17,000 action logs behind — the team only made sense of it after turning to GLM 5.2, an open-source model running in a self-hosted environment.

At WAIC, the model showed off a 4:1 ink-wash scroll and a 22-panel storyboard sequence. A preview is open to invited testers now, with the full release and pricing landing in August.

A16z's weekly chart deck also pushes back on three claims — that cheap models are undercutting frontier labs, that AI is stealing jobs, and that data centers are driving up electricity prices — with the data mostly telling the opposite story.

Run it locally for free with your data staying on-device, or switch to the cloud for more horsepower — zero data retention by default. The preview is live today on Mac and Windows.

2.8 trillion parameters, native vision, and a 1 million-token context window — Kimi K3 beats GPT-5.6 Sol on several agent benchmarks, with the app and API live today

Compressing a ~54GB 27B model down to ~3.9–5.9GB lets it run locally on a phone, while retaining roughly 90% of average performance—here's how it works and where the trade-offs lie.

Hassabis wants the US to stand up a FINRA-style body that defines frontier models with a moving benchmark — a voluntary protocol now, a hard gate to market later.

Model capability has leapt forward in 24 months, but a composite reliability metric built by the SAGE lab has moved only 5–10 points.

Across 3 models and 20 languages: English is the most cautious and in-depth, Russian the most exacting, Hindi the warmest, and Chinese sits closest to the global average

With no central brain in charge, nearly 200 simple smart cubes figure out what shape they've formed just by talking to their neighbors—and can even sense where to "regrow" after damage. The self-recognition part already works on physical bricks; damage localization and regeneration still happen mostly in simulation.

Companies pay for intelligence twice: once in model fees, and again in the proprietary know-how required to make the model genuinely useful.

576K samples, 16 models tested: 19.7% of AI-recommended packages are hallucinations, and 43% keep generating the same fake name.

From tacit knowledge to interaction bandwidth to model alignment — why AI's progress still can't do without humans.

No manual context-feeding required — the local Markdown wiki refreshes on a set schedule; a Slack connector is coming soon.

Local small models replace cloud AI calls — the numbers come from Google's internal testing; proxy models are currently limited to the ai.if function and still in preview

First-hand impressions from four scenarios: coding, writing, knowledge work, and agents

Paper shows a 10.8% gain in personalization and 29.4% in reasoning, fully open-source under Apache 2.0

Expanding from charging only AI crawlers to charging any caller, now in early access waitlist

It makes up less than 10% of the model — remove it and Claude can still talk, but its reasoning collapses to zero. Anthropic is already using it to catch fabricated data and spot when Claude senses it's being tested.

In a 151-student trial, short-answer questions moved scores more than multiple choice, while almost no one touched the AI help sidebar

No funding, no team — built purely to solve his own kid's problem. Clinics and schools started asking to use it anyway.

A live audience vote couldn't be tallied because the venue lights were too bright to count hands, but a companion survey found 95% of teams already use agents while 59% worry about mounting technical debt.

While traditional schools are still figuring out AI, Silicon Valley and Wall Street families are already voting with their wallets.

US developer employment among 22-to-25-year-olds has fallen 19% in three years, even as new GitHub sign-ups hit their fastest growth ever.

Create psychological tension first, then offer a first step too small to refuse — it works for writing, selling, and job hunting alike.

Four implementation principles, backed by real-world data from L'Oréal, Lyft, and Rakuten

MIT-licensed and model-agnostic — works with any OpenAI-compatible text model, though for now it only handles a single page view

Trained on 50,000+ domestic AI chips and 35 trillion tokens; most benchmarks come from Meituan's own evaluation framework, and the weights aren't truly open for download yet

Now in open beta. A coordinator agent marshals a team of expert agents to do the work, with a reviewer agent at the end dedicated to catching errors in citations and numbers — compute gets outsourced to AI, but raw data never leaves your local machine.