AI Titus's Morning Wire

AI NEWS TODAY

60 SUBAGENTS, 31 MILLION TOKENS, AND A RIEMANN BOUND THAT MOVED FROM 41.6 TO 67.2 PERCENT
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadClaude Code's auto mode becomes the default on August 14 for Pro, Max and Team plans
Anthropic's own testing had auto mode stopping 89 percent of harmful actions against 13.6 percent for human review, and it blames approval fatigue for the gap. Enterprise and API follow within a month.
★ Must-ReadNVIDIA released a 30B agent model with 3B active parameters that runs on a single RTX card
Nemotron 3.5 Lightning claims up to 4x the output speed of similar models and 30 percent faster task completion. Weights are on Hugging Face, ModelScope and OpenRouter.
📚 LearnNVIDIA also open-sourced a router that swaps models mid-task inside an agent workflow
NeMo Switchyard sends each step to the cheapest model that can handle it, and NVIDIA's internal numbers put the cost at about a third of routing everything to a top-end Claude. CodeRabbit fine-tuned Lightning for 85 dollars in roughly two hours.
★ Must-ReadOpenAI shipped a cybersecurity model trained to refuse less, behind an access tier you have to be admitted to
GPT-5.6-Cyber completes 95 percent of OpenAI's advanced cybersecurity benchmark against 1.5 percent for the general model with its safeguards on. It found two unknown Chrome V8 bugs. Hardware security keys become mandatory on all Daybreak accounts September 1.
Spotify built a detection-and-appeal system for artists who are not people
Self-declaration opens today, badges show up in mid-September, and anything labeled an AI Persona drops out of personalized recommendations by default. The appeal process is the interesting design problem.

📺 Watch · latest videos

★ Must-WatchCan you trust what AI tells you?
How much you can trust an AI depends on what you’re asking. Kyra from the Anthropic education team breaks down the two most common reasons for an AI t
Claude / Anthropic · 271 views · 46 likes
How to Build a One Person AI Business (Using Claude Code)
Latest from Nate Herk.
Nate Herk · 2K views · 178 likes
Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI Lab
Leave agent adoption to individuals and the engineer shipping two PRs a day ends up reviewing the ten that the early adopter ships. They fall further
AI Engineer · 1.5K views · 35 likes
★ Must-WatchVibe coding a custom e-ink picture frame powered by Gemini 3.6 Flash and Nano Banana
Looking for a morning routine upgrade? Creative Technologist Raphael vibe coded a custom e-ink picture frame powered by Gemini 3.6 Flash and Nano Bana
Google · 7.1K views · 240 likes
AI went rogue in real life...
Latest from Matthew Berman.
Matthew Berman · 20.9K views · 533 likes
★ Must-WatchHow Model ML Uses GPT-5.6 Sol to Get Finance Work Done More Efficiently
Model ML is solving the “last-mile problem” in finance, carrying asks from research to analysis to a finished PowerPoint deck or Excel workbook with t
OpenAI · 7.8K views · 200 likes
Engineers… Your Software Factory NEEDS Agent Sandboxes to SCALE (exe.dev)
Engineers... if you're IN the loop, you ARE the bottleneck. 🔥 And giving your agents a tiny corner of YOUR computer is why your software factory can't
IndyDevDan · 12.2K views · 671 likes

🗣 Voices & Blogs

★ Must-ReadZvi Mowshowitz On Why Pacing The Frontier Is A Different Ask Than Pausing It
He works through the letter that more than a thousand lab employees signed, including Amodei and Pachocki. The ask is for the brake to exist, not for anyone to press it yet.
📚 LearnSimon Willison Reads The Safety Numbers Behind The Auto Mode Default
He walks through the 1,053-developer study and the 720 prompt injection attempts, then says he wants independent confirmation before he relaxes. Worth reading before Friday.
📚 LearnYash Thakker Treats The Riemann Run As An Orchestration Case Study
He separates the math, which is strong pending referees, from the capability claim, which nobody outside Anthropic can test. His point about the failed ideas being the load-bearing part is the useful one.
Gary Marcus On The Gap Between Open-Source And Open-Weight
A week with two big open-weight drops is the right week to be precise about what you are actually allowed to do with the files. Read the license, not the headline.
Nvidia launches a smaller, faster Nemotron model and a router to put it to work
Nvidia on Tuesday launched Nemotron 3.5 Lightning, the newest member of its Nemotron 3 family of open models. In addition, The post Nvidia launches a… · The New Stack

🌏 The Wire · Drudge / Breitbart

NVIDIA LINES UP $500 BILLION FROM SIX ASSET MANAGERS TO BANKROLL ITS OWN CUSTOMERS
Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR, with compute itself as the collateral. Huang calls it independent underwriting. Investors are already using the word circular.
ANTHROPIC, MACQUARIE AND GIC START A DATA CENTER COMPANY WITH ONE TENANT
Theseus Infrastructure builds and leases the sites, Macquarie's funds and GIC put up most of the equity, Anthropic anchors each one. The release restates its pledge to cover consumer electricity price increases traced to these sites.
A BITCOIN MINER JUST SIGNED A 20-YEAR, $9.1 BILLION LEASE ON ITS TEXAS CAMPUS
Riot Platforms named the tenant only as a leading frontier AI lab; Bloomberg reported it as Anthropic. 191 MW, staged to June 2028, and two extension options that would take it to $16.1 billion.
IBM PUTS $240 MILLION BEHIND A CLUSTER BUILT ONLY FOR OPEN-MODEL INFERENCE
With Together AI, on NVIDIA HGX B300 hardware, available first quarter of 2027. The pitch is frontier performance without the closed-model price tag.
CHINESE MAKERS SHIPPED MORE THAN 97 PERCENT OF THE WORLD'S HUMANOID ROBOTS
About 19,100 units in the first half, up from roughly 5,100 a year earlier. AgiBot took 44 percent and passed Unitree. Industrial and commercial buyers were over 70 percent of it.

🤖 Trending Models · Hugging Face

MiniMaxAI/MiniMax-H3
image-text-to-video · ★3.5K · 59.4K dl
meta-models/Muse-Glimmer-30B
image-text-to-text · ★981 · 0 dl

📈 Markets

NVDA 217.55 ▼2.9%
MSFT 506.06 ▲1.2%
GOOGL 357.52 ▲0.9%
AMZN 278.09 ▲1.3%
META 594.92 ▲0.5%
AMD 469.56 ▼2.9%
AVGO 422.40 ▼1.3%
PLTR 175.23 ▲1.9%
SPCX 138.74 ▲4.2%
TSLA 330.88 ▲0.7%

The BriefAnthropic published a research note on Monday saying an unreleased version of Claude was asked to take a real run at the Riemann hypothesis, one of the oldest open problems in mathematics. It did not solve it. On the way it improved a related known result, raising the share of the zeta function's zeros proven to sit where the hypothesis says they should from 41.6 percent to 67.2 percent, checked by two of Anthropic's own mathematicians and by two outside reviewers, with a Lean formalization alongside. The method is the part worth your attention. Two Claude Code sessions, around 60 subagents, 31 million output tokens, and roughly 650 ideas that led nowhere before one landed. Anthropic says plainly that these techniques will not get anyone to a proof, and nobody outside the company can rerun the job, so treat the math as solid and the capability claim as unfinished.

Level UpClaude Code's auto mode becomes the default on Friday for Pro, Max and Team plans. It stops asking permission at every step and only pauses for actions it judges irreversible, destructive, or pointed outside your environment. Spend ten minutes on your deny rules before then. Anthropic's own testing found people approved 97 percent of the prompts they were shown, which tells you how much protection the prompt was providing. A written deny list for things like curl and .env files does more, and it lives in your project settings. Claude Code settings and permission rules