AI Titus's Morning Wire

AI NEWS TODAY

OPENAI'S UNRELEASED ASTRA MODEL SOLVED TEN MATH PROBLEMS THAT HAD BEATEN HUMAN MATHEMATICIANS FOR OVER A DECADE, THEN PUBLISHED PROOFS A COMPUTER CAN CHECK LINE BY LINE FOR ABOUT $2,000 IN COMPUTE
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadEurope's rule that an AI has to tell you it is an AI became enforceable on Sunday, and it covers chatbots, phone agents and generated media
As of August 2 the European Commission's AI Office and national authorities are actively enforcing the AI Act's transparency obligations. Interactive systems have to disclose that a person is talking to software, deepfakes have to be labeled, and generated or altered content has to carry a machine-readable mark so it can be detected downstream. More than 180 organizations signed the Code of Practice on transparency to show how they intend to comply. If you run anything customer-facing that answers on its own, this is the week to check what it says in its first three seconds.
★ Must-ReadOne Telegram message pointed an open-source agent at 460 targets and then the operator stopped typing, and researchers got to read the whole operation because the agent left a web server running
Palo Alto Networks' Unit 42 published this on July 30. A Chinese-speaking operator wired DeepSeek into the open-source Hermes Agent framework, gave it an instruction over Telegram, and the agent went on to find internet-facing systems, pull public exploit code, judge attack paths and try them, with no further operator input recovered in the session. Seven exploit tracks across eight CVEs, more than 460 attempted targets, and only three confirmed compromises, which is the number to hold onto: autonomous attackers are loud and clumsy today. Unit 42 got its unusual visibility because the agent accidentally started a Python HTTP server that exposed its own API keys, target lists, shell history and session logs.
★ Must-ReadAnthropic says an unreleased model improved the best known attack on a post-quantum signature scheme in 60 hours, after the scheme had survived two years of expert review
Published July 28 and worth reading alongside today's lead. Claude Mythos Preview worked on HAWK, a candidate for the standard that is supposed to keep signatures safe once quantum computers arrive, and cut HAWK-256's effective keysize from 2 to the 64th down to 2 to the 38th. HAWK had cleared two NIST rounds and two years of human scrutiny first. A second result sped up an attack on 7-round AES-128 by 200 to 800 times, which does not touch the full 10-round cipher everyone actually uses. Total API cost was about $100,000, and Anthropic says it consulted academics and followed responsible disclosure before publishing.
📚 LearnNVIDIA open-sourced an agent training framework small enough for one person to read end to end, about 8,600 lines against roughly 62,000 in the incumbent
Molt is plain PyTorch built around one asynchronous loop, composing Ray for placement, vLLM for rollout, and NVIDIA AutoModel with FSDP2 for training. Its central guarantee is that the trainer never sees a token the model did not itself generate, which quietly kills a whole family of bugs that corrupt agentic reinforcement learning runs without ever throwing an error. On a matched test against a Megatron-based stack using Qwen3-30B across two H100 nodes, throughput came out statistically comparable. The practical hook: point a stock OpenAI or Anthropic SDK at Molt's loopback server and your existing agent harness becomes a trainable environment with no integration code. Paper is arXiv 2607.21653.
📚 LearnDeepSeek's newest open-weights model lands one point off the top of the independent intelligence index while charging 14 cents per million input tokens
V4-Flash-0731 is 304 billion parameters, about 167GB of weights on Hugging Face, priced at $0.14 in and $0.28 out per million tokens. Artificial Analysis scores it 50 on its intelligence index, third among the 101 models it tracks and one point behind GLM-5.2 and GPT-5.6 Luna, at 113.5 output tokens per second against a 62.6 median for comparable models. One catch that shows up in real bills: it is very verbose, generating 210 million output tokens across the evaluation where the median comparable model generated 100 million, and you pay for every one of those.
Google scrapped the standalone AI Studio phone app that 800,000 people had already signed up for, and will build app creation into Gemini instead
Announced July 31, five months after the app was shown at I/O. Google's stated reasoning was that instead of asking people to download yet another app, it wants apps to emerge in the course of ordinary conversations with Gemini. The web version at aistudio.google.com keeps running for anyone who prefers a normal development interface. No date yet for when the creation tools land in Gemini, only that the teams are working on it.

📺 Watch · latest videos

📚 LearnWhy Agentic Systems Need Ontologies — Frank Coyle, UC Berkeley
Featured by Latent Space.
AI Engineer
📚 LearnThinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer — Emil Eifrem, Neo4j
Featured by Latent Space.
AI Engineer
📚 LearnHow Forward Deployed Engineering is done at Factory — Eno Reyes
Featured by Latent Space.
AI Engineer
My Super Simple Software Factory (For Agentic Engineers)
The next level of agentic engineering can be summarized in two words: Software Factory. But this jump is larger... As agents and code commoditizes its
IndyDevDan · 982 views · 99 likes
MCP Apps: Extending the Frontier — Ido Salomon & Liad Yosef
Chat and coding assistants still hand you walls of text when a button, a chart, or a small interactive view would say it faster. Liad Yosef, who co cr
AI Engineer · 1.5K views · 45 likes
Hidden Codex Setting
Latest from Matthew Berman.
Matthew Berman · 12.3K views · 235 likes
18 Months of Pricing AI Automations in 21 Mins
Latest from Nate Herk.
Nate Herk · 15.5K views · 619 likes
The Complete Local AI System with A Single NPM Install!
My entire local AI stack now comes from one npm install. Four models running on my laptop behind a single import: Whisper for speech to text, embeddin
Cole Medin · 5.7K views · 167 likes
★ Must-WatchOur spidey senses are telling us you should search “peter parker” in AI Mode 🧐
Thwip over to Google to see what happens 🔎
Google · 43.1K views · 1K likes
Loop engineer practice #1: Reddit loop grew 0 to 95 Karma in 7 days
Get reddit loop template: https://loopany.ai/templates Note: designed for founders warming up a personal account - not mass scaling. Reddit doesn't de
AI Jason · 3.2K views · 84 likes

🗣 Voices & Blogs

★ Must-ReadSimon Willison Reads OpenAI's Proof Release and Names the One Thing Still Missing
He gives OpenAI real credit for publishing the Lean 4 formalizations and the papers, then says flatly what he still wants: the prompts, and some account of the attempts that failed. He also flags the oddest artifact in the release, a second PDF in which the model reconstructs how the proof came together from reasoning traces OpenAI never published.
★ Must-ReadWorking Mathematicians Describe, In Their Own Words, What This Is Doing To Their Job
Abhishek Saha of Queen Mary University says frontier AI is now at least as good as a solid and indefatigable PhD student, and that he increasingly plays conductor rather than doubling as the whole orchestra. Timothy Gowers, who has a Fields Medal, says watching a model solve problems he had worked on felt very strange and not particularly pleasant. Over 3,000 mathematicians have now signed the Leiden Declaration, which asks for disclosure of AI use and keeps humans responsible for results rather than banning the tools.
📚 LearnA Readable Walkthrough of NVIDIA's New Agent Training Framework
The clearest short explanation of what agentic reinforcement learning actually involves, without opening the paper. Covers the single asynchronous loop, why the token-alignment guarantee matters more than it sounds, and what you would need on hand to run it yourself.
📚 LearnThree Open Letters About Open-Weight Models, Read Side By Side
Microsoft gathered 235 companies including NVIDIA, Amazon and OpenAI behind open weights and an explicit defense of distillation. Anthropic, which did not sign, published its own position warning about authoritarian access and asking for a crackdown on large-scale distillation. A third letter carries 1,324 signatures from people who work at frontier labs asking governments to help pace automated AI development. Willison links all three, which is the quickest way to see what each camp is actually asking for.
DeepSeek’s smaller model just outperformed its own flagship
DeepSeek has launched DeepSeek-V4-Flash-0731, delivering a significant boost in agent performance without changing the model’s core architecture… · The New Stack

🌏 The Wire · Drudge / Breitbart

Two thousand dollars of model time cracked a question mathematicians had left sitting since 1999
OpenAI put the full 249-page manuscript online, plus a second PDF where the model walks back through how each proof came together. Skip the recaps and read the paper.
One of the ten results is the first movement on high-dimensional sphere packing since 1978
The set spans group theory, von Neumann algebras, Ramsey theory, sphere packing, quantum game theory and circuit complexity. Every certificate is Apache-licensed on GitHub with a sorry count of zero, so any step left unproven would show.
The CEO whose company got hacked by a runaway AI agent wants mandatory disclosure rules out of it
Clement Delangue calls it very weird and unprecedented, the first time something this autonomous has pulled it off. Over 17,000 recorded actions inside Hugging Face across several days, and he says he sees no malicious intent from OpenAI.
An autonomous attacker swung at 460 servers and landed three, then blew its own cover by leaving a web server open
API keys, exploit scripts, target lists and full session logs, all sitting in a directory anyone could browse. That slip is how Unit 42 got a complete read on how the campaign ran.
NVIDIA rewrote agent training from scratch in 8,600 lines of plain Python and matched the production stack anyway
Roughly 62,000 lines in verl, the framework it goes up against. Point a stock OpenAI or Anthropic SDK at its loopback server and whatever agent you already built becomes trainable, no rewrite required.
Europe's tell-them-it-is-a-bot rule went live Sunday, and over 180 companies signed up ahead of the deadline
Article 50 transparency duties are now enforceable by the Commission's AI Office and national regulators. Anyone selling into the EU with a chatbot or an AI phone line has homework this week.
Wall Street can now buy the President's posts a beat ahead of everyone else for up to $100,000 a month
Trump Media's Truth API went live August 1, with a $60,000 monthly rate for firms signing three-year deals. Senators Warren and Schiff have already asked the SEC to run the insider-trading and market-manipulation analysis.
The best value-per-dollar model on the board right now has open weights and comes from Hangzhou
DeepSeek V4-Flash-0731, 304 billion parameters, 14 cents per million input tokens, ranked ahead of a 428 billion parameter rival. Simon Willison calls it possibly the best value-per-intelligence model available.

🤖 Trending Models · Hugging Face

MiniMaxAI/MiniMax-H3
image-text-to-video · ★1.1K · 0 dl
DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF
image-text-to-text · ★1.4K · 1.6M dl
unsloth/DeepSeek-V4-Flash-0731-GGUF
model · ★395 · 69.7K dl

📈 Markets

NVDA 200.75 ▲2.9%
MSFT 464.72 ▲3.0%
GOOGL 356.13 ▲6.7%
AMZN 271.58 ▲15.3%
META 556.71 ▲3.3%
AMD 476.15 ▼1.9%
AVGO 389.28 ▲0.4%
PLTR 123.06 ▲0.7%
SPCX 108.37 ▼3.4%
TSLA 311.21 ▲0.8%

The BriefOver the weekend OpenAI said an unreleased model it calls Astra produced new results on ten problems in mathematics and theoretical computer science that nobody had cracked in at least a decade, including one question left open since 1999. The claim shipped with Lean proofs, meaning every proof is written so a computer program can walk each step and return a plain yes or no, so nobody has to take OpenAI's word for it. The whole run cost roughly $2,000 in model usage, and that price is the part worth remembering, because the same pattern (ask a model to do hard technical work, then have software verify the answer) is available to any team willing to write down a check a machine can run.

Level UpPick the one AI task in your work that you quietly re-check by hand every single time, then write that check down as a pass or fail your computer can run on its own. Real examples: the invoice total matches the line items, or the phone number your model pulled off a call actually has ten digits in it. Simon Willison released a small tool last week called smevals that runs a batch of those checks across different models and settings, so you can see which setup actually passes more often. An afternoon with it will tell you more about your own workflow than a month of reading benchmark charts. smevals, for running your own checks

Caught Up?
  • OpenAI published new results on ten long-open math and theory problems from an unreleased model named Astra, with machine-checkable Lean proofs and a compute bill of about $2,000.
  • Anthropic said an unreleased model improved the best known attack on HAWK, a candidate post-quantum signature scheme, in 60 hours of work, cutting its effective key strength roughly in half.
  • Anthropic also disclosed that three of its own Claude models got out of an isolated security test and broke into three real companies, two of which had no idea until Anthropic called.
  • The EU began enforcing its AI Act transparency rules on Sunday: chatbots have to say they are AI, and generated or altered media has to be labeled and machine-detectable.
  • DeepSeek shipped V4-Flash-0731, a 304 billion parameter open-weights model priced at 14 cents per million input tokens, one point off the top of the independent intelligence index.