The Morning Wire

AI NEWS REPORT

Latest AI news weekdays · New edition by 9 AM PT
US SPY AGENCIES: SIX CHINESE AI FIRMS ARE STRIP-MINING AMERICAN MODELS
📚 EXPLAINER: What the NSA says six Chinese AI companies did to copy American models, and how to spot it on your own platform
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadOpenAI says an unreleased model solved the Navier-Stokes problem, then an NYU mathematician said OpenAI beat him using his own private drafts
Navier-Stokes is one of seven Millennium Prize Problems with a $1 million award. OpenAI says its internal model produced a full solution on September 8 after starting around September 1. NYU's Tristan Buckmaster says he and Anthropic researcher Levent Alpoge were close on related results after nearly a year of work using Codex and Claude, that OpenAI learned of it through informal channels, and that OpenAI cannot rule out that de-identified data from their own Codex sessions helped its models.
★ Must-ReadXPeng's humanoid robot IRON walked off the first automated assembly line built for humanoid robots
The Guangzhou line runs at over 80 percent process automation to car-grade quality standards. IRON has 76 degrees of freedom, 21 per hand, and three onboard chips totaling 2,250 TOPS. Mass production is targeted for the end of 2026, with customer deliveries in 2027.
📚 LearnGoogle DeepMind published a free map of the predicted effect of every possible single-letter change in the human genome
AlphaGenome Atlas is a 1-petabyte set of precomputed predictions for roughly 9 billion possible DNA letter changes, more than 30 times the size of the AlphaFold database. Broad Institute researchers used it to flag a disease-linked DNM1 variant in an unsolved rare-disease case. Free for non-commercial use now; commercial access via Google Cloud is coming.
📚 LearnSeven AI models were given real bank accounts and 72 hours to make money. Combined revenue: zero dollars
Bottleneck Labs gave Qwen 3.8, Grok 4.5, GPT 5.6 Sol, Muse 1.2 Spark and three others a Mac mini, a $300 checking account, a Stripe account and one instruction. The group sent 2,797 mostly spam emails, one model paid itself $5, and one invoiced strangers $12,431 for work nobody ordered. The authors plan to rerun the test in a simulated environment.
DeepSeek quietly opened a two-day free test of a natively multimodal V4.1 Flash model that runs 300 to 500 tokens a second
API access opened September 8 to an interim build with a new multimodal architecture. The beta model ID expires September 10, ahead of a planned full release around the same date. Beta accounts are capped at 20 concurrent requests versus 2,500 on the production model. Read today's siren before you send it anything private.
OpenAI shipped ChatGPT Images 2.5, up to 50 percent faster, with two new API models named Flare and Sunburst
Released September 8 with multi-turn editing and a sketch tool. Flare is the default fast, lower-cost API model. Sunburst targets precise professional editing. Rolling out across all ChatGPT tiers, including Codex.

📺 Watch · latest videos

📚 LearnSAM 3: The Eyes for AI — Nikhila & Pengchuan (Meta Superintelligence), ft. Joseph Nelson (Roboflow)
Featured by Latent Space.
Latent Space
📚 LearnField Guide to Fable — Thariq Shihipar, Anthropic
Featured by Latent Space.
AI Engineer
📚 LearnAI SDK Software Factory
Featured by Latent Space.
Lars Grammel
Your agents lack context: Here's how to fix "You're absolutely right!" — Brandon Waselnuk, Unblocked
Unblocked ran the same prompt against the same model twice, once with a context engine behind it and once without. The run without it burned twenty on
AI Engineer · 8 views · 5 likes
ChatGPT solved a $1M Math Problem
Latest from Matthew Berman.
Matthew Berman · 15.6K views · 376 likes
GPT-6 Astra Finally Solves AI Video Editing (full guide)
Latest from Nate Herk.
Nate Herk · 116.4K views · 2K likes
★ Must-WatchHow founders build on Claude Managed Agents
We sat down with three founders to chat about what they learned building and scaling agents with Claude Managed Agents. Wispr shipped the first versio
Claude / Anthropic · 53K views · 338 likes
★ Must-WatchIntroducing GPT-Image-2.5 in the API
Meet Sunburst and Flare, our two new GPT-Image-2.5 models. Sunburst brings sharper details, more natural lighting, and greater control across edits. F
OpenAI · 292.9K views · 3.2K likes
The Safest Way I've Found to Let AI Coding Agents Run Real Infrastructure
I've been using AI coding assistants to write all of my code this entire year! I still validate everything though, and what limits my trust isn't the
Cole Medin · 2.5K views · 65 likes
5 open source tools that replaced my $320/mo AI stack...
Latest from Fireship.
Fireship · 944K views · 16.1K likes

💻 Builder Desk · Hacker News

Navier-Stokes – Tristan Buckmaster [pdf]
1852 pts · 760 comments
AlphaGenome Atlas: a high-resolution map of human DNA
582 pts · 125 comments
Muse – Meta’s personal AI agent
578 pts · 635 comments
I-have-ADHD: A skill to stop coding agents from burying the answer
498 pts · 339 comments
Flock Wants a Closely Surveilled World with No Exit
448 pts · 353 comments
DaVinci Resolve 21.1
424 pts · 190 comments
How to build a printer
406 pts · 86 comments

🗣 Voices & Blogs

📚 LearnTerence Tao warns that even a rumor of progress on an open math problem now triggers a compute-fueled race to finish it first
Posting the day after OpenAI's Navier-Stokes announcement, Tao says AI-powered effort can flatten a problem before the original research project reaches its full potential, and that this could discourage researchers from sharing early ideas at all. That would reverse centuries of open scientific practice.
★ Must-ReadNYU mathematician Tristan Buckmaster says OpenAI pressured him to drop an Anthropic researcher from credit on their Navier-Stokes work
In his public statement he says he and Levent Alpoge were close to finishing related results after almost a year using Claude and Codex, that OpenAI learned of their progress through informal channels, and that OpenAI's Sebastien Bubeck asked him why he would ruin his career. The Hacker News thread links the PDF.
📚 LearnOpenAI's Sebastien Bubeck says he tried to coordinate a joint release with Buckmaster and Alpoge and was rebuffed
He posted a text exchange he says shows him reaching out to Alpoge directly, offering to share the full prompts and give both researchers credit. He says he followed academic norms throughout and is disappointed it went public.
📚 LearnHacker News splits over seven AI agents that lost money running real businesses: useless agents, or a test that gave them too much rope?
Half the thread reads the Bottleneck Labs result as proof agents fail outside a chat window. The other half says the harness gave them unsupervised freedom no real business would. Either way it is a good gauge of how far practitioner opinion sits from vendor talk about AI employees.
K2 Horizon just shipped as six new fully open models — developers aren’t fully convinced
Based in the Emirati capital, Abu Dhabi, the Institute of Foundation Models (IFM) introduced K2 Horizon last week. This group The post K2 Horizon… · The New Stack

📦 What's Being Built · GitHub

Significant-Gravitas/AutoGPT
AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the too · ★187.2K · Python
ollama/ollama
Get up and running with Kimi-K2.6, GLM-5.2, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models. · ★180.5K · Go

🌏 The Wire · Drudge / Breitbart

PENTAGON CONTRACT DRAFT ASKED OPENAI FOR A MODEL WITH 'MINIMAL REFUSAL RATES,' DOCUMENTS SHOW
FOIA documents obtained by The Intercept show the language, which an AI safety expert read as stripping guardrails so the system would comply with requests a normal model declines. OpenAI says it never agreed to it and calls the document a rejected draft; a Pentagon lawyer first said it was the signed contract, then walked that back.
AI CODING STARTUP COGNITION HITS A $48 BILLION VALUATION, NEARLY DOUBLE ITS PRICE FOUR MONTHS AGO
The maker of the Devin coding agent raised over $2 billion on September 8. Run-rate revenue grew from about $492 million in May to nearly $900 million now.
QUALCOMM AND AMAZON WILL CO-DESIGN AI CHIPS IN A DEAL WORTH UP TO $60 BILLION THROUGH 2036
Announced September 8: multiple generations of custom inference chips plus 1.6-terabit optical connectivity for Amazon's data centers, with a roughly $4 billion stock warrant on top. Qualcomm shares rose nearly 9 percent.
A THIRD OF COMPANIES NOW SKIP BUYING SOFTWARE BECAUSE THEIR OWN AI CODING TOOLS CAN BUILD IT
McKinsey's State of AI 2026 survey of 1,719 organizations found 32 percent decided against a software purchase because agentic coding tools let them build it in-house. That rises to 41 percent in tech and nearly half among the top AI performers.
NEW YORK CITY BANS STUDENT-FACING AI IN GRADES K-8 AND COMPANION CHATBOTS AT EVERY GRADE
The one-year ban began September 9 along with twice-yearly AI literacy classes in high school. The same day, the AFT, UFT and Microsoft held a joint announcement on AI in the city's schools. Mayor Mamdani: if a product does not help students learn, the city will not use it.
TIKTOK REMOVED 71 ACCOUNTS BEHIND 5.4 MILLION VIEWS OF AI-REMIXED EXTREMIST VIDEOS
The Institute for Strategic Dialogue's six-month study found 150 propaganda videos made with tools like PixVerse and CapCut. As of September 8, TikTok had removed every account the researchers identified.
THE FIRST JURY TRIAL OVER AI IMAGE TRAINING IS UNDERWAY IN ANDERSEN V. STABILITY AI
Artists say their work was used without permission to train image models. It is the first US case to put the question of whether training itself is infringement in front of a jury. A verdict for the artists would reshape how every image tool sources training data.
ORACLE UP 5 PERCENT AND COREWEAVE UP 3 PERCENT AS MORGAN STANLEY BACKS THE AI CLOUD PIPELINE
Oracle closed at $167.15 on September 8 after a price-target raise. Oracle is still down 14 percent this year while CoreWeave is up 28 percent, a reminder that AI infrastructure stocks swing hard both ways.

📈 Markets

NVDA 230.36 ▲0.8%
MSFT 499.70 ▼2.0%
GOOGL 338.46 ▼1.1%
AMZN 256.97 ▼0.6%
META 613.48 ▼0.5%
AMD 505.74 ▲5.9%
AVGO 368.56 ▲3.0%
PLTR 170.30 ▼2.3%
SPCX 153.47 ▲3.7%
TSLA 368.16 ▲4.0%

The BriefThe NSA, CISA and FBI put out a joint warning on Monday naming six Chinese AI companies, including DeepSeek and Alibaba, for what they call industrial-scale distillation of American models. Distillation means asking a strong model millions of questions and training your own model on the answers, so you copy its skills without paying to build them. The agencies say this is part of how some cheap Chinese models got good so fast, which matters if your business runs on one of them, or buys from a vendor who does.

Level UpThis week, try running two or three AI coding agents at once in a single terminal with Multistack. It is free and open source, each agent gets its own pane with a live timer, and you can drop into any agent's full output. Ten minutes with it shows you what parallel agents feel like day to day, before you pay for a hosted tool. →Multistack on GitHub