🍭 A Model Just Cracked A 1999 Problem 🧠

Good morning. The intern spent forty minutes trying to understand what a non-sofic group is and has since gone very quiet.

Let’s dive in 👇

🍭 What’s Cookin’:

• OpenAI named its next model in a footnote about sphere packing(??)

• Anthropic looked back its own logs and passed out (tomorrow's story)

• A German court told Suno the training itself was the problem

OpenAI's
🧮 Astra Showed Up Doing Math Homework

The Bite:

OpenAI published ten new results in mathematics and theoretical computer science.

The post credited an internal version of Astra, described as its next major model.

The problems cover sphere packing, coding theory, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal graph theory.

Every result ships with a machine-checkable Lean certificate (the strongest form of "check my work" in this industry) published to a public GitHub repository.

OpenAI said the tokens needed to find the solutions would cost roughly $2,000 at Sol API rates.

Astra has no announced release date, price, or availability.

Snacks:

• The headline result is a construction establishing that non-sofic groups exist, open since 1999.

• Astra also disproved Connes's rigidity conjecture and resolved Erdős problems 146, 180, and 183.

• The manuscript collection is 249 pages long, with a separate document narrating the model's reasoning.

• OpenAI said humans prepared the manuscripts, and the model formalized each argument in Lean.

• The same post promotes free frontier access for 100,000 scientists and mathematicians.

• Thomas Bloom, who publicly debunked OpenAI's October 2025 math claim, called the results big news.

Why it Bites:

You don't need to know what a non-sofic group is. Roughly four hundred people on earth do. The math isn't the story here, the shape of the announcement is.

Instead of launching a model, OpenAI submitted evidence.

Every previous launch has run on the same fuel: a benchmark table, a bar chart, a score that beats last quarter's.

Benchmarks saturate. Nobody believes them anymore.
So OpenAI reached for the one claim a skeptic can't wave away.

A Lean certificate compiles or it doesn't. Anyone with a laptop can check all ten without trusting a single word the company says.

That's a higher bar than the industry has been clearing.
It's also a very well-chosen one. Here's why:

In June, the Leiden Declaration, endorsed by the International Mathematical Union, criticized AI companies for announcing results by press release instead of peer review.

OpenAI's post cites the declaration by name, respectfully, and then does the thing anyway. It picked which ten results to publish and wrote the manuscripts. Then decided which formal statement counts as the open problem. Lean confirms the proof is valid.

And sitting right above all of it in the same post: free frontier access for 100,000 researchers.

The proofs check out. So does the funnel.

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🧤 Hand Wave

Reads sign language through Meta smart glasses and speaks it out loud, running on an open-source model small enough to work on-device.

🏋️ CoachAI

Props your phone against a wall and it counts reps and calls out your form mid-set, no wearable required.

🤖 Qwen3.8-Max

A 2.4-trillion-parameter model Alibaba claims can run 10+ days of autonomous development from an empty folder to production without a human touching it.

🎮 claudemon

Turns the dead time waiting on Claude Code into a Pokémon encounter in a second terminal tab, all 151, fully local, no account.

💬 Fin

Intercom's customer agent that stops handing people off between sales and support bots and just stays in the same conversation.

_{Everything Else}
🧠 You Need to Know

🚨 Anthropic Says Claude Broke Into Three Real Companies
→ A July 30 review of 141,006 cybersecurity evaluation runs found six in which a Claude model reached the open internet and compromised real systems at three organizations.

🧠 OpenAI Names Astra As Its Next Major Model
→ An August 1 post credited an internal version of Astra with new results on ten math and theoretical computer science problems open for at least a decade, at roughly $2,000 in compute.

⚖️ Munich Court Rules Against Suno In GEMA Case
→ The Munich Regional Court found on July 31 that Suno's training and outputs infringed German collecting society GEMA's repertoire, granting injunctive relief, disclosure, and damages.

🤖 DeepSeek Ships Official V4-Flash Release
→ The July 31 build keeps the same 284-billion-parameter architecture, but beats DeepSeek's larger V4-Pro preview on every agentic benchmark the company published, at $0.28 per million output tokens.

🇰🇷 LG Releases Korea's Largest AI Model On Hugging Face
→ K-EXAONE 2.0 shipped July 31 with 750 billion parameters under an Apache 2.0 license, more than tripling the size of its predecessor.