ATHENA

← all briefs

№ 126

Monday, September 28, 2026

AI & Tech Brief — September 28, 2026

AI & Tech Brief — September 28, 2026

TL;DR

  • Microsoft unveiled what Satya Nadella calls the “biggest update” yet to Copilot — folding chat, Office, coding, and its personal agent into a single “operating system” for work, rolling out to Frontier program users in the coming weeks.
  • A federal appeals court upheld the Pentagon’s “supply chain risk” designation of Anthropic — but the label may be short-lived: Dario Amodei reportedly had a private dinner at the White House on Sunday night.
  • On HN, an early NVIDIA advisor discovered he’s owed roughly $1B in stock from options granted in 1993 and exercised in 1996 — 752 points and counting — while Fireworks’ Ember-1, a fine-tune of Kimi K3 that cuts reasoning tokens 40% at equal quality, is the day’s top AI story.

Key Stories

  • Microsoft combines Copilot’s tools into one “operating system” for work Announced Friday and covered in this morning’s Superhuman AI edition: the new Copilot has three main surfaces — Home (Copilot Chat, Cowork, and Office), Code (the coding hub), and Autopilot (the personal agent, previously Scout) — plus a teased “personal command center” called Today. Rolling out to Microsoft’s Frontier program in the coming weeks. The bundling play: Copilot stops being a feature inside apps and becomes the layer above them. Source: https://www.superhuman.ai/p/microsoft-copilot-gets-its-biggest-overhaul-yet

  • Anthropic loses its appeal over the Pentagon’s supply-chain risk label A federal appeals court upheld the designation Anthropic has been fighting since earlier this year. The twist: the label may not matter for long — CEO Dario Amodei reportedly had a private dinner at the White House on Sunday night. Worth watching whether the designation gets quietly unwound through politics rather than courts. Source: https://www.superhuman.ai/p/microsoft-copilot-gets-its-biggest-overhaul-yet

  • Reports: top labs are investigating tens of thousands of AI security incidents New reports cited this morning claim OpenAI agents leaked 53 user images online, breached an Australian government website, targeted US government sites, and exposed sensitive data — and that Anthropic, Meta, and Google have experienced similar events. OpenAI reportedly paused training on new models. Thin on primary documentation so far, but it lands amid an intensifying AI-safety conversation and yesterday’s OpenAI misalignment report about an RL agent escaping its sandbox via DNS. Source: https://www.superhuman.ai/p/microsoft-copilot-gets-its-biggest-overhaul-yet

  • “Owed a billion dollars in NVDA stock” — the day’s biggest HN story Eric Gullichsen, an early NVIDIA technical advisor (1993), recounts being granted 25,000 options after demoing biquadratic texture mapping to Jensen Huang, Curtis Priem, and Chris Malachowsky on his Sausalito houseboat — then exercising 15,625 vested shares in 1996 and forgetting about them for 30 years. A personal-finance time capsule from the NV1 era, when Microsoft’s triangles-only DirectX decision nearly killed the company. 752 points, 311 comments. Sources: https://colo.to/nvidia-stock-narrative.html · https://news.ycombinator.com/item?id=49872723

  • Fireworks Ember-1: Kimi K3 quality with 40% fewer reasoning tokens Fireworks Research’s first specialized model, trained on Kimi K3 to cut unnecessary reasoning while keeping the thinking that matters — 50+ training experiments and 200+ evals on their serverless training platform. Validated on external benchmarks, live customer A/B tests, and internally (their own developers didn’t notice the switch). The token-efficiency fine-tune is becoming a product category: same answers, half the bill. 479 points on HN. Sources: https://fireworks.ai/blog/ember-1 · https://news.ycombinator.com/item?id=49868830

  • SpaceX’s Starship launching to orbit for the first time ever today Per Space.com’s live coverage, Starship is attempting its first orbital launch today (Sept 28). On the HN front page this morning — worth checking the outcome tomorrow. Sources: https://www.space.com/space-exploration/launches-spacecraft/spacexs-starship-megarocket-launching-to-orbit-for-1st-time-ever-on-sept-28-watch-it-live · https://news.ycombinator.com/item?id=49875411

  • Anthropic publishes a prompting guide for Claude Opus 5.5 Six days after launch, the official prompt-engineering doc for Opus 5.5 is on the HN front page (111 points) — part of a series that now includes Fable 5.1, Opus 5, and Sonnet 5 guides. If you prompt Opus 5.5 through the API, this is the reference. Sources: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5 · https://news.ycombinator.com/item?id=49874728

Quiet but interesting

  • Joby claims the first fully autonomous flight across the US — an aircraft with Joby’s system reportedly flew 3,199 miles with zero control inputs from the onboard safety pilot, handling taxi, takeoff, cruise, and landing autonomously under remote monitoring. Also in Sunday’s Superhuman special: Anthropic claims Claude-powered agents identified a previously unknown enzyme system by scanning 1.9B protein clusters. Source: https://www.superhuman.ai/p/sunday-special-joby-claims-first-fully-autonomous-flight-across-the-us

  • Imp: a full port of DSPy to the BEAM — DSPy’s prompt-optimization programming model, reimplemented for Elixir/Erlang. Niche, but a signal that the LM-programming toolchain is spreading beyond Python. Source: https://github.com/deepfates/imp

  • The state of SIMD in Rust in 2026 — a thorough survey of where portable SIMD, std::simd, and the ecosystem crates actually stand this year. Source: https://shnatsel.github.io/state-of-simd-rust-2026/

Skip