2026-08-17
By Vadym · Generated with AI, curated by me
Today’s stories split into two camps: labs racing to ship new models, and everyone else racing to keep up with what shipping already broke — a hard API cutover, a hijacked AI assistant, and a manifesto trying to explain where all this is headed.
xAI released Grok 4.6 on August 12, a flagship model tuned for long-horizon coding and agentic work, available the same day in Cursor and xAI's own Grok Build tool. It scored 61 on the Artificial Analysis Intelligence Index, up 5 points from Grok 4.5, matching GPT-5.6 Sol Max at the same $2/$6-per-million-token pricing, with context expanded to 500K tokens.
Google is retiring three Imagen 4 model endpoints — standard, ultra, and fast — today, August 17, pushing developers onto Gemini 3.1 Flash Image. Apps still calling the old generate_images() method don't get a warning, they get a hard error, and the migration isn't a drop-in swap since the API shape changed.
OpenAI announced August 1 that Astra, an unreleased internal model, generated solutions — with machine-checkable Lean 4 proofs published on GitHub — to 10 open problems in math and theoretical computer science, including a non-sofic group construction open since 1999, at roughly $2,000 in compute. Fields Medalist Timothy Gowers called the results credible while cautioning the field is still digesting them.
Barracuda researchers presented a proof-of-concept at Black Hat USA 2026 showing how one compromised employee account, combined with Microsoft Copilot, let an attacker map the org chart, impersonate the CEO, hide sign-in alerts by redirecting them to Deleted Items, and redirect a $247,500 wire transfer — then use Copilot itself to erase the evidence.
Mark Zuckerberg published a manifesto August 10 arguing AI should be distributed to individuals rather than concentrated in a handful of labs, promising a free personal agent with a private mode "even Meta can't access." The same day, Meta released Muse Glimmer, a 30B open-weight model built to run on a single consumer GPU, as a first concrete step toward that vision.
Today's six stories split into two camps: labs racing to ship (Grok 4.6, Qwen3.8-27B, Astra, Muse Glimmer) and the world catching up to what shipping already broke (Imagen 4's abrupt cutover, Copilot turned into a fraud tool). The pace of releases keeps compressing — days, not months, between meaningful upgrades — but the tooling for migrating safely, verifying claims, and securing what's already deployed hasn't kept pace. That gap, not the benchmark charts, is where 2026 is actually being decided.
— Boba
Curated by Vadym