The Morning Wire

AI NEWS REPORT

MOST OF WHAT AI TELLS YOU TO BUY IS WRITTEN BY THREE WEBSITES NOBODY READS: THEY CHURNED OUT 215,000 MACHINE-WRITTEN 'BEST SOFTWARE' PAGES JUST TO GET QUOTED, AND PERPLEXITY NOW CITES THEM AS ITS THIRD-BIGGEST SOURCE OF EVIDENCE. CHECK THE SOURCE BEFORE YOU BUY
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

📚 LearnAlibaba quietly previewed Qwen4's architecture inside an open-weight model called Qwen3.8-Flash-Next
The 125-billion-parameter model activates only 6 billion parameters per token, plus a new n-gram embedding layer that scales without the cost of adding more experts. Native context runs to 262,000 tokens, stretchable to 1 million, and it is released under an open license on Hugging Face.
📚 LearnPerplexity now splits your AI tasks between the cloud and a small model on your own Mac, so sensitive files never have to leave the machine
The cloud model handles research and planning while a compact local model, running on Apple silicon, processes tax returns, medical records or client files on device. A privacy gate decides what is allowed to leave the Mac, masking or blocking sensitive details by default. Live now for Pro, Max and Enterprise subscribers.
Anthropic open-sourced its playbook for building shopping and merchant agents on Claude
The free repo ships working code for a shopping agent and a merchant agent, with guardrails against manipulative upselling and a human-approval step before a merchant agent changes a price or launches a promotion. Anthropic says retail partners have seen carts up to 35 percent larger, though that number is the company's own and not independently audited.
ChatGPT can now pull a patient's real chart from Epic, read-only, so a clinician can ask what changed since the last visit
The integration surfaces authorized lab results, medications and specialist notes inside ChatGPT for Healthcare without writing anything back to the record. A companion Healthcare Public Data plugin adds direct access to PubMed, DailyMed and CMS coverage data. UCSF Health is the pilot partner.
Fireworks made it easy to keep training a model past what the frontier labs shipped you, without building your own training stack
The Training API lets a team write its own training loop in Python while Fireworks runs the distributed infrastructure underneath, with serverless LoRA training billed per token or dedicated full-parameter training for bigger jobs. Early customers Harvey, Vercel and Heidi Health report 2 to 4 times more training iterations per week.
📚 LearnA two-day-old open-source project shows swarms of Claude agents building explorable, camera-matched 3D cities from public map data
The agents pull OpenStreetMap and USGS elevation data, then research, model and quality-check a real place with almost no step-by-step human direction. Two demo worlds are live now, a 453-building recreation of San Francisco's Union Square and a hand-painted walking route through Kyoto. MIT licensed, runs in a browser.

📺 Watch · latest videos

📚 LearnAI SDK Software Factory
Featured by Latent Space.
Lars Grammel
📚 LearnIntroducing GPT-5
Featured by Latent Space.
OpenAI
📚 Learn🔬 Why Transformers Hit a Wall the Moment Physics Shows Up — Anima Anandkumar, Caltech
Featured by Latent Space.
Latent Space
Use Fable 5.1!!
Latest from Matthew Berman.
Matthew Berman · 2.5K views · 75 likes
AI Software Factories Are the Next Big Thing (And I'm Building You One)
Latest from Cole Medin.
Cole Medin · 17.6K views · 340 likes
How Anthropic ACTUALLY Prompts Fable 5.1
Latest from Nate Herk.
Nate Herk · 36K views · 527 likes
★ Must-WatchHow the Claude Code team uses Claude Code
A year ago, using Claude Code meant prompting, giving feedback, and accepting permission prompts. What does it look like now? Thariq Shihipar, Sid Bid
Claude / Anthropic · 57.3K views · 648 likes
How to Think Clearly In The Era Of AI: Full Course (5 Hours)
Latest from Nick Saraev.
Nick Saraev · 46.6K views · 1.3K likes
From coding to Knowledge work agents — Karan Vaidya, Composio
Latest from AI Engineer.
AI Engineer · 2 views
The mystery is solved... and the answer is 40x cheaper than Claude
Latest from Fireship.
Fireship · 1.4M views · 20K likes

💻 Builder Desk · Hacker News

Google avoids a breakup of its ad tech business
427 pts · 299 comments
Audacity 4.0
362 pts · 80 comments

🗣 Voices & Blogs

📚 LearnSimon Willison caught Anthropic quietly rewriting Claude's system prompt to refuse song lyrics, right after two music publishers sued
Claude now declines to reproduce lyrics, poems or book passages 'in whole or in part, including the last lines, a chorus or hook,' and keeps refusing even a reworded version of the same request. Willison tracks every change to Anthropic's published system prompts in a public GitHub repo, so anyone can see what shifted and when.
Richard MacManus found that the biggest open-source AI projects are quietly closing the door on human pull requests
Vercel's AI SDK now has its own agents write 25 to 35 percent of the code it merges and close most incoming issues automatically. The 50,000-star project tldraw goes further and auto-closes outside pull requests entirely, arguing contributors are more useful reporting problems than submitting code an AI can write faster.
📚 LearnA commerce consultant read Anthropic's new shopping-agent code and found the real bottleneck: your product catalog, not the AI
Daniele Vella shows Claude's shopping agent can only filter on facts that exist as clean, structured fields. In his test catalog, 63 real product facts sat only in spec-sheet text and were invisible to the agent, and one item's weight appeared in five different, incompatible formats. The fix is catalog data governance, not a better agent.
The team behind MyClaw's hosted OpenClaw service admits the open-source upgrade broke plenty of self-hosted setups this week
OpenClaw 2.0 makes personal AI agents more capable and safer on long-running tasks, but the newsletter is candid that self-hosters hit broken gateways and failed migrations getting there. Their pitch: a managed platform absorbs that pain so you do not have to debug it yourself.

📦 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 · ★187.1K · Python
ollama/ollama
Get up and running with Kimi-K2.6, GLM-5.2, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models. · ★180K · Go

🌏 The Wire · Drudge / Breitbart

GOOGLE'S GEMINI TOLD THREE HIKERS THE MOUNT SHASTA CLIMB WOULD TAKE 8 HOURS. IT TOOK 16. THEY GOT STRANDED OVERNIGHT AND NEEDED FOREST SERVICE RESCUE RANGERS TO GET THEM DOWN
The AI also told the group to pack simple carbs instead of fat and missed a no-camping zone near a protected spring. Search and rescue says never rely solely on AI to plan a climb like this; call the ranger station instead. All three hikers made it out safely, one with a knee injury.
THE JUSTICE DEPARTMENT TOLD A FEDERAL JUDGE IT SIDES WITH OPENAI, NOT THE NEW YORK TIMES, IN THE FIGHT OVER WHETHER TRAINING ON NEWS ARTICLES IS FAIR USE
It is the first time the federal government has taken a side in any of the AI copyright suits. The filing argues a competitive AI industry serves national security, and that forcing labs to pay for every training article would hand the biggest tech companies an even bigger edge.
A STARTUP THAT TELLS YOUR SECURITY TOOLS WHICH AI AGENTS ARE REAL AND WHICH ARE ATTACKERS RAISED $27 MILLION FROM BLACKSTONE
Huskeys already screens more than a trillion web requests a day for customers including TikTok and Hugging Face, and can patch a known vulnerability at the network edge within minutes while a permanent fix is still being written. The round pushes its valuation past $100 million.
DELL BOOKED A RECORD $60.9 BILLION IN AI SERVER ORDERS LAST QUARTER AND RAISED ITS FULL-YEAR REVENUE FORECAST BY $25 BILLION
AI server revenue hit $16.4 billion, double the same quarter a year ago, and the order backlog now sits at $95 billion. Dell also became the first vendor to ship rack systems on Nvidia's newest Vera Rubin platform, and its AI customer base grew by 1,500 companies in a single quarter.
HIDDENLAYER RAISED $100 MILLION TO BUILD SECURITY SPECIFICALLY FOR AI MODELS AND AGENTS, NOT JUST THE SERVERS THEY RUN ON
The company already protects a frontier lab serving more than 700 million weekly users and just launched a tool to secure AI coding agents at runtime. Gartner expects companies to spend $2.83 billion on AI security tools this year, 83 percent more than last year.
THE ENGINEERS WHO BUILT FACE ID RAISED $165 MILLION TO GIVE ROBOTS BETTER EYES
Lyte's custom sensor chip measures motion directly instead of stitching it together from mismatched cameras and lidar, the kind of sensing gap that makes warehouse and factory robots miss things. The Series C values the company at $1.6 billion and funds production scaling for its industrial customers.
SOUTH KOREA'S WOORI BANK IS REBUILDING ITS CHATBOT SO A CUSTOMER'S CONTEXT FOLLOWS THEM FROM THE APP TO A HUMAN BANKER TO A PHONE CALL
KT's new Agent Connector links the bank's chatbot, its AI Banker advisory tool and live agents so a customer never has to repeat themselves switching channels. KT already runs similar AI contact-center systems for four of Korea's five biggest commercial banks.
A NEW GATEKEEPER CHECKS EVERY MOVE AN AI AGENT TRIES TO MAKE BEFORE IT HAPPENS, NOT AFTER, SO A ROGUE AGENT GETS STOPPED MID-ACTION INSTEAD OF LOGGED AFTERWARD
JetStream's Clearance engine sits in front of every tool call an agent makes and can block one dangerous step, like a delete command, while still allowing the safe ones from the same tool. It reaches general availability this fall, aimed at businesses running large fleets of customer-facing agents.

🤖 Trending Models · Hugging Face

zai-org/GLM-5.3
text-generation · ★1.6K · 151K dl
Qwen/Qwen3.8-Flash-Next
image-text-to-text · ★4.8K · 263.3K dl
tencent/Hy4-preview
text-generation · ★407 · 4.4K dl

📈 Markets

NVDA 224.41 ▲3.2%
MSFT 496.82 ▼0.8%
GOOGL 337.12 ▲0.6%
AMZN 254.98 ▲0.0%
META 592.85 ▲2.5%
AMD 457.06 ▼0.6%
AVGO 367.24 ▼0.7%
PLTR 169.46 ▼5.8%
SPCX 140.71 ▼1.1%
TSLA 357.01 ▲0.3%

The BriefA new report found that when Perplexity answers questions about which software to buy, nearly 60 percent of the sites it cites barely get visited, and almost a quarter don't rank in the top million websites on earth. Three connected domains, wifitalents.com, worldmetrics.org and gitnux.org, share the same template and the same DNS setup, and together they wrote about 215,000 machine-generated 'best software' pages built to get quoted by AI, not read by people. If you or your team ask an AI which tool to buy, that answer may be leaning on content written for the AI, not for you, so check what it actually cites before you spend money on its say-so.

Level UpThis week, before you act on an AI's software recommendation, ask it which sources it used, then run each one through a free site authority checker. A site with almost no real backlinks or traffic behind it is a sign the 'review' was written to get quoted by AI, not to help an actual buyer. Ahrefs: free Website Authority Checker