OpenAI's unreleased Astra model solves ten open problems in mathematics
An internal version of OpenAI's next major model, Astra, solved ten previously open problems across mathematics and theoretical computer science, publishing formal, machine-checkable Lean proofs on GitHub for about $2,000 in compute. Chief among the results is an explicit construction of a non-sofic group, settling a question open since 1999, plus new sphere-packing bounds. Fields Medal winner Timothy Gowers said he would recommend one of the proofs for a top journal without hesitation.
Why it matters: A frontier model producing fully verified, zero-"sorry" formal proofs of decades-old open problems is a concrete data point in the debate over whether AI can do genuine novel mathematics, not just pattern-match known results.
EU AI Act's high-risk system rules become enforceable, AI Office gains authority
The EU AI Act's governance framework became operational on 2 August, with the AI Office assuming enforcement authority over general-purpose AI models. Providers of high-risk AI systems must now implement seven safeguards, including risk management, data quality controls, activity logging, documentation, deployer notification, human oversight and technical robustness standards. The AI Office can request technical documentation, evaluate models, require corrective measures and issue fines.
Why it matters: This is the EU moving from AI Act text to active enforcement, a full year before the equivalent US federal framework has even published its own promised deliverables.
Amazon tops $3 trillion market cap after completing its $50 billion OpenAI investment
Amazon's market value passed $3 trillion on 3 August as its stock extended a post-earnings surge, with AWS revenue hitting $42.2 billion, well above the $40.54 billion analysts expected. The company also completed its full $50 billion investment in OpenAI, paid across three tranches through the year, and raised its 2026 capex forecast to $220 billion, up from $200 billion in February.
Why it matters: Amazon joining the trillion-dollar AI capex race on both the compute-buyer and OpenAI-investor sides shows how deeply the cloud giants are now cross-invested in each other's AI bets.
Supabase open-sources a benchmark that grades AI coding agents on real backend work
Supabase released Evals, an Apache-2.0 benchmark that runs Claude Code, Codex and OpenCode through real Supabase development tasks like constructing an application schema, repairing broken projects, or fixing row-level-security bugs, rather than isolated code-generation prompts. The project powers a public leaderboard and an internal regression suite monitored daily.
Why it matters: Testing coding agents on real infrastructure work rather than synthetic benchmarks is a more honest signal of which agent actually helps in production, and it's directly relevant to anyone running Supabase-backed tools.
Africa Deep Tech Foundation launches a challenge to build offline AI for 8GB laptops
The Africa Deep Tech Foundation launched the Laptop LLM Challenge on 29 July, asking engineers across the continent to build language models that run entirely offline within an 8GB RAM limit, with a $20,000+ prize pool and submissions due 25 August. Supporting organisations include Nduzi Foundation, AI Africa XPrize, Future Africa and Udu Technologies' Africa GPU Hub. The goal is capable AI for professionals without reliable cloud access, power or internet, such as "a bank teller in Lagos, a nurse in Arusha, or a student in Dakar."
**A thin section again today.** Roughly a dozen search queries across frontline-deployment, funder-deadline, US/AU and disability-accessibility angles turned up nothing else that cleared the recency and dedup bar. Notable drops, all stale-on-verification rather than already-covered: Current AI's AlphaChat (implied fresh, actually launched mid-July); HIAS's "Annie Moore" refugee resettlement matching tool (implied a 2026 pilot result, actually deployed since 2018, the cited employment-outcome stats trace to 2017 data); Aira's Meta AI glasses support (implied current, actually launched 20 April). This is the same shape as yesterday's briefing and yesterday's network scan — logged for the ongoing gate-vs-supply investigation rather than treated as independent noise.
**Metrics logged:** stale_item_rejections=3, briefing_np_geo_diversity=1 (Bucket B only), briefing_type_diversity=4/5 (missing: product/consumer), briefing_source_tier=4/4 tier-1.
Why it matters: Most AI tools assume unlimited cloud access; a challenge explicitly designed around the infrastructure constraints African users actually face is a direct test of whether AI capability can decouple from cloud dependency.
AI Impact Hub