1. OpenAI's Astra Cracks 10 Long-Open Math Problems, With Lean Proofs to Check
On August 1 OpenAI revealed that an internal version of Astra, its next flagship model, produced new results for 10 problems in mathematics and theoretical computer science that had been open for at least a decade. The headline result is an explicit construction of a non-sofic group, a question open since Gromov raised it in 1999; Astra also disproved Connes's rigidity conjecture. OpenAI published a 249-page manuscript collection plus Lean 4 certificates for all ten results on GitHub under Apache 2.0, with a "sorry" count of zero. Token cost for finding all ten: roughly $2,000. None have been through peer review yet, and no Millennium Prize problem fell.
2. Alibaba Ships Qwen3.8-Max: 2.4 Trillion Parameters, 1M Context, Open Weights
Alibaba's Qwen team released Qwen3.8-Max on August 3, its largest and most capable model to date: a mixture-of-experts design with 2.4 trillion total parameters and roughly 95 billion activated at inference, a 1 million-token context window, and text, image and video input. It scored 1,668 on Frontend Code Arena, 37 points behind the top configuration of Anthropic's Claude Opus 5. Weights are headed to Hugging Face and ModelScope, the first time Alibaba has opened a Qwen-Max-class model. By parameter count it trails only Moonshot's 2.8T Kimi K3 among open-weight models.
3. Ninth Circuit: An AI Shopping Agent Isn't "Hacking" — Users Are the Ones Accessing
On August 4 the Ninth Circuit vacated the preliminary injunction that had barred Perplexity's Comet browser from Amazon's logged-in account pages. The panel held that Perplexity does not "access" Amazon's servers under the Computer Fraud and Abuse Act — users do, and Comet merely receives screenshots their own browsers captured. The court applied the rule of lenity and noted there is "little to no existing caselaw" on assigning responsibility for AI agents. Amazon's trademark and state-law claims survive, but the CFAA theory that agents are trespassers just lost its first appellate test.
4. White House Convenes OpenAI, Anthropic and Google on Voluntary Safety Testing
The administration hosted major AI developers at the White House on August 6 to discuss a new US framework for voluntary safety testing of frontier models, with OpenAI, Anthropic and Google among those attending. The framework grows out of a June executive order on AI cybersecurity that laid out an opt-in approach to model safety reviews alongside efforts to harden critical computer systems. It lands against a messy backdrop: December's Executive Order 14365 seeks to preempt state AI laws, while Colorado has already repealed and narrowed its landmark SB 24-205.
5. Mistral Open-Sources Shieldstral, a 3B Guard Model That Reads Policies in Plain English
Mistral released Shieldstral on August 4: a 3-billion-parameter open-weight multimodal safety classifier that takes the moderation policy as a plain-language yes/no question at inference time rather than baking fixed harm categories in during training. It evaluates prompts, responses, prompt-response pairs, or images with optional text, covers 12 languages, and runs on a single 16GB GPU. Mistral says it matches or beats open guard models up to 7× its size on text safety, refusal detection, policy adaptability and multimodal moderation. Apache 2.0 on Hugging Face.
6. Hadrian Raises $1.37B at a $7.87B Valuation for Automated Factories
Automated-manufacturing startup Hadrian announced $1.37 billion in new equity on August 6, valuing the company at $7.87 billion — co-led by WCM Investment Management, Washington Harbour Partners, Valor Equity Partners, 137 Ventures and Baillie Gifford. The money funds new factories, expanded R&D and additional production lines including munitions and autonomous systems. Recent openings in Mesa, Arizona and Muscle Shoals, Alabama bring its footprint to just under 3 million square feet across four sites. Defense-adjacent AI manufacturing remains one of the hottest capital magnets of 2026.
7. DiffusionGemma Technical Report: Text Generation Without Token-by-Token Decoding
Google DeepMind published the DiffusionGemma technical report, detailing an open-weight model that replaces autoregressive decoding with discrete diffusion — iteratively denoising blocks of 256 tokens in parallel. It is fine-tuned from the Gemma 4 MoE base (25.2B total, 3.8B activated) using under 10% of the original model's training token budget, via supervised bidirectional denoising followed by RL plus sampler distillation. Reported throughput reaches roughly 1,500 output tokens per second on a single H100.
8. Rust Adopts an LLM Policy: Models May Review, Not Create
Five Rust teams adopted an LLM policy for the rust-lang/rust monorepo on August 5, authored by Jynn Nelson. The line it draws: LLMs may answer questions, analyze code, summarize discussions, edit, check work, suggest approaches and give private reviews — but should not produce the final contribution. Output only you see needs no disclosure; LLM text in public docs, PR descriptions or GitHub comments must be clearly marked. The teams are explicit that this is not the Rust project's official stance on LLMs and does not apply everywhere — but it is one of the sharper answers yet to what generated code costs a volunteer review economy.
9. Anaconda Buys Enkrypt AI to Fold Red-Teaming Into the Python Stack
Anaconda announced on August 4 that it has acquired AI-security startup Enkrypt AI for an undisclosed sum. Enkrypt's pre-deployment red-teaming across 300+ attack categories, runtime guardrails, and NIST/EU AI Act compliance automation get folded into the Anaconda Platform. The deal targets the governance gap enterprises hit once models leave the notebook — and reflects how much of the EU AI Act's now fully enforced regime is being answered with tooling rather than process.
// KEY TAKEAWAYS
Two themes dominate the week. First, capability claims are getting harder to hand-wave: OpenAI shipped machine-checkable Lean certificates with its math results, and Alibaba is putting a 2.4T-parameter Max-class model's weights in public — both moves that invite verification instead of asking for trust. Second, the rules are being written by courts and communities faster than by legislatures. The Ninth Circuit's Perplexity ruling says a user's agent is still the user, the Rust project drew a bright line at "review, don't create," and the White House is negotiating safety testing as opt-in while federal preemption fights play out. Meanwhile efficiency keeps beating scale for headlines: a 3B guard model outperforming 21B ones, and diffusion decoding hitting 1,500 tokens/second on one GPU.