The day's tech, sifted: Oct 28, 2026

Wed, Oct 28

What matters today: The Anthropic S-1 keeps generating second-day readings: the WSJ says it implies a yearly market of $40T and up, and Ars catalogs its stranger risk factors. In security, a 9.5-severity flaw in EDK II, the open source base of most server boot firmware, runs code before Secure Boot can protect anything. Chip money keeps circling AI: AMD puts $1B into Meta's stock raise a day after Nvidia's $3B.

AI / LLMs

  • Anthropic's market sizing: The WSJ reads the S-1 as implying a $40T+ annual addressable market, based on the 2025 economic value of knowledge work and above SpaceX's $28.5T figure, without the filing naming it. GeekWire's Etzioni offers a napkin-math calculator for testing a $2T valuation, so the debate has moved from revenue to how much of all human work is up for grabs.
  • S-1 risk factors: Across 93 pages and 70,000+ words, the filing's risk section covers hundreds of failure modes, including its own AI blackmailing people or becoming sentient; yesterday's margin readings looked at the upside, this is the downside list investors will read.
  • Claude for government: Anthropic's expanded GSA deal gives federal, state, local, tribal and territorial governments a 40% discount on Claude, estimated to save $1B, which makes state and city agencies eligible for the same pricing for the first time.
  • Claude in drug labs: Anthropic's early progress report covers six biotech partnerships and in-house programs, with early wins in CAR-T (engineered immune cell therapy), allergies and regulatory paperwork; AlphaSignal's summary cites an 89% cut in review time.
  • Hiring in the Claude era: An Anthropic hiring manager says he stopped giving coding interviews because Claude does them better, and tells juniors that "hunger beats tenure", a signal of what the interview loop tests next.
  • OpenAI's results, the reaction: Yesterday's three claimed breakthroughs now draw pushback: researchers tell The Verge they are tired of OpenAI gatekeeping its findings, after hundreds of preprints landed on mathematicians in early October.

Devtools & Infra

Security & Privacy

  • Boot firmware flaw: CVE-2026-97631, rated 9.5, is a memory-corruption bug in the ICMP (network ping) component of TianoCore EDK II, used by AMI, Insyde, Phoenix and Microsoft's Project Mu; it can run code before Secure Boot and the OS defenses start, and while the payload does not persist, it can install ones that do.
  • Next.js security release: Next.js schedules a security release for November 3, so teams should plan the upgrade window now.
  • FISA ruling fight: Two Democratic senators allege the administration is withholding a secret FISC ruling on search filters that let analysts reach Americans' messages while it presses Congress on surveillance law.

Startups & Industry

Research

  • Seven electrons, 17,895 lines: Vals AI describes ten Claude Sonnet 5.5 agents producing a Lean proof for seven electrons on a sphere, the latest AI-math claim in a week of them.
  • Benchmarks against cheating: Artificial Analysis moves its Intelligence Index to 60% private tests, so scores are harder to game by training on public questions.

Elsewhere

Hacker News

The front page leans practical. Discord's plan to make public server messages findable through Google topped the privacy conversation, and Tailscale posted an SSH investigation and remediation update. Amazon's reported AI shake-up and Claude's drug-development progress (covered above) drew the biggest threads. Local and open tooling had a good day: Survival LLM is an offline assistant and emergency library for your own computer, and Liquid AI introduced LFM3, a multimodal model family. For the hobbyist crowd, device trees hint at a Raspberry Pi Zero 3W and Pi 4 Model A, while Git for US Legislation shows the legal code as version-controlled history.

Threads