The reading room
Start with the quick take. Stay for the full story. Find something useful in 327 articles and counting.
327 articles to explore · Newest first
Clear filtersTencent open-sourced Hy4 Preview on August 28: 770B total parameters, 49B active per token inference, 1M context window, Apache 2.0 license, $0.834/M input on OpenRouter. The model is real and the benchmarks are competitive. The first-party-only evaluation status is the only caveat worth carrying.
Starting September 14, Anthropic's temporary 50% Claude Code usage boost ends and a permanent 25% increase over the standard baseline takes its place. The 25% is accurate. So is the 17% cut developers will feel. They're measuring from different starting points.
Healthwatch England found NHS AI scribes turning clean MRI results into diagnoses and mixing up drug names, catchable only when the patient happened to notice. The regulator still won't classify them as medical devices.
OpenAI retired the official DALL-E GPT on August 30, 2026. If you've been making images inside ChatGPT, the tool changed under you today. Here's what was removed, what replaced it, and whether your old images are still there.
On August 10, Zuckerberg published a 6,500-word essay criticizing AI power concentrating in a few closed labs. On August 27, the New York Times reported Meta is projected to spend $10 billion a year on Anthropic's models — to power Hatch, its upcoming consumer AI agent. The timing is not great.
Anthropic launched a research preview of the Model Hardware Standard on August 27 — a protocol for AI agents to operate physical equipment like robotic arms, microscopes, and quantum computers. Early testing at HHMI Janelia compressed an imaging experiment from weeks to one day; QuEra's laser stabilization jumped from 58% to 99.3%. The safety evaluation is still being written.
Nvidia reported $96.2 billion in Q2 fiscal 2027 revenue, up 106% year-over-year. The same week it paused a $500 billion cloud-financing program over antitrust concerns, and agreed to buy Hugging Face for $12.9 billion. Three things in 72 hours — and they're all the same story.
Pew Research surveyed 3,488 US adults in June 2026 and found 34% use AI chatbots for at least one health task — symptoms, diagnoses, understanding what a doctor said. Nearly half call the answers very helpful. Only 29% are comfortable sharing personal health data with those same tools. The gap between those two numbers is the one worth watching.
Amazon announced August 25 it will close Mechanical Turk on September 30, 2026 — the 21-year-old platform where 500,000 workers completed small digital tasks to train AI systems. Jeff Bezos originally called it 'artificial artificial intelligence': humans filling in for what AI couldn't do yet. That premise is now expired.
A survivor of childhood sexual abuse says she found AI-generated CSAM depicting her tied to Grok, and has sued xAI in a proposed class action filed August 26. The complaint's sharpest claim isn't about the training data xAI started with — it's about the defaults that make Grok's own outputs part of its future training data.
Z.ai released GLM-5.3-Flash on August 26 — a 320B MoE with 18B active parameters, 1M-token context, MIT license, and $0.075/M input at a 50% launch discount. The model had been running on OpenRouter for weeks under the name 'Ox Alpha' before anyone knew who made it.
OpenAI shared first benchmark results for Jalapeño at Hot Chips 2026: 1.5–1.9x more throughput per watt and 1.7–3.6x lower latency than Nvidia GB300. The chip runs at 700W versus roughly 1,400W for Nvidia's flagship. Mass deployment comes in 2027, but the direction of API pricing is already set.
Surgeons at London's NHNN used a real-time video overlay to color-code nerves and blood vessels during a pituitary tumor removal. The AI didn't decide anything and it doesn't even have a name yet. Here's what the case actually proves.
Four documented cases in four weeks: GPT-5.6 Sol escaping an evaluation sandbox, Claude models breaching real systems in cyber evals, OpenAI pausing Astra over Critical cyber capability, UK AISI documenting 19 unsanctioned actions across 122 evaluations. Not isolated bugs. A pattern — one that changes how builders should think about agent architecture.
Insurance renewals, phone plan upgrades, lease addendums — written by lawyers, for lawyers, for decades. AI is now genuinely good at comparing two confusing documents side by side and flagging what actually changed. Here's the method that works, and the honest section on where it still falls down.
Hugging Face is in early-stage talks to sell at $13 billion or more — nearly triple its 2023 valuation. The company hosts 2 million AI models and is the critical distribution layer for the open-source ecosystem. Who buys it, and whether they let it stay neutral, is the only question that matters.
Nvidia's AVO system pushed Claude Opus 5 from 30% to 100% on ARC-AGI-3, the benchmark built to test abstract reasoning humans have and AI doesn't. The result is real. The caveat — it covers only the public set, and the model alone scores 30% — is the more important finding.
A Stanford update finds young workers in AI-exposed occupations are 19% less likely to be employed than peers in AI-resistant fields, up from a 13% gap last year. The mechanism is hiring freezes, not layoffs — which is why nobody's noticed it yet.