The day's tech, sifted: Oct 12, 2026

Mon, Oct 12

What matters today: Ledger confirmed a hardware implant in wallets sold by reseller CryptoBilis, tied to an $86M+ theft across hundreds of devices. The evening added Microsoft's Maia 300 chip, its third-gen in-house AI silicon and a bid to depend less on Nvidia, and cyberattacks on Danfoss refrigeration systems that heated hundreds of French supermarkets to 50°C for about six hours. Amazon is in talks to buy Decart for about $7B, a deal Anthropic passed on in September.

AI / LLMs

  • Amazon eyes Decart at ~$7B. WSJ sources say Amazon is in talks for the startup, whose software eases switching between AI chips; Anthropic decided against buying it for about $6B in September. Chip portability is the prize.
  • "Super intelligence" rebrand. The Verge says Musk and Benioff pledged to rename SpaceXAI and AIForce while Nadella and Nvidia adopted the term, framing it as Silicon Valley falling in line with Trump. Yesterday Nadella urged containment of models; today the branding story.
  • Torvalds and Claude Code. Phoronix reports Linus Torvalds is working with Claude Code on a sizeable Linux kernel improvement.
  • Qwen3.8-Flash-Next goes local. A 177B-parameter model runs on a 64GB Mac at 36 tokens/s via an SSD-paged table. Separately a community build strips its refusals by transplanting 144 tensors into the quantized file, so open weights keep losing their guardrails.
  • Sakana's MASS. One 27B model plays its own agent team, judge and teacher, lifting score per token up to 1.6x on research benchmarks.
  • Phones as local LLM hosts. Qualcomm's CEO says AI companies are asking for phone chips that run 100B-parameter local models continuously by 2028.
  • Microsoft Maia 300. Third-gen in-house AI chip on TSMC's N2P process, pitched as its primary inference system; production workloads start in 2027. Racks hold 96 chips at about 167 kW, and Microsoft plans multi-gigawatt deployments from next year, so its inference spend shifts from Nvidia GPUs to its own silicon.
  • UK AISI gets its access back. The White House reversed course and restored the UK AI Safety Institute's access to frontier models for testing.
  • Capital as the AI bottleneck. Intel's Lip-Bu Tan and Broadcom's Hock Tan warned in Singapore that access to capital and rising borrowing costs could become AI's main constraint, and Hock Tan likened speculative AI investors to casino gamblers.

Devtools & Infra

Security & Privacy

  • Ledger wallet theft. Ledger says the $86M+ theft involved devices from CryptoBilis and confirms a hardware implant: buy hardware wallets only from the maker.
  • AWS AgentCore. The Register reports AgentCore's security was undone by a prompt asking for credentials, another case of agent guardrails living in text.
  • College cannot cancel Flock. Elgin Community College asked in October 2025 to remove its six cameras over data sharing with Customs and Border Patrol and weak security; Flock refused, saying it had not breached the contract.
  • French supermarket cyberattack. Sunday-night attacks on Danfoss-managed refrigeration heated systems to 50°C for roughly six hours across hundreds of French stores, forcing them to throw out food: physical damage from a vendor's remote management layer.
  • FBI-hacking arrest. The FBI arrested a suspect tied to September's ShinyHunters theft of FBI data about its own agents.
  • Yandex outage and the bots. Yesterday's drone strike on a third Yandex data center fed a chart showing Russian bot traffic collapsing; The Verge says it probably was not the disinformation kind, so it probably shows ordinary automated traffic going offline with the data center, not a troll farm switched off.

Startups & Industry

Research

  • Recursive self-improvement. A multi-agent self-supervision paper is one of several self-improvement papers on today's lists (RSI-Anything, Video-RSI): models improving themselves is this week's research fashion.
  • Humanity's Sixth Sense. Import AI covers a new visual reasoning benchmark built on what a scene implies but does not show: humans score 93.1, the best AI system 53.6, and the Chinese open-weight models tested around 26.

Elsewhere

Hacker News

Trust and AI dominate the front page. Deloitte allegedly falsified documents to help evade tax on Abramovich's wealth (533 points), and a remote employee was fired for refusing 160 webcam photos a day. Karpathy's post that the gap in shared understanding of LLM capability is widening drew the most comments (501, discussion), and the QGIS mapping project banned AI-assisted contributions (334 comments, discussion). Sceptics also went after Anthropic's IPO, with a research firm calling it an "unprecedented test of investor gullibility", and one essay asks why AI companies are not charged for hacking.

Identity and infrastructure: Passkeys are a trap argues against adopting them, and fingerprints are the new cookies. A Cloudflare chart of Yandex bot traffic down 99% circulated widely, but see the Security entry above for why it is not what it seems. ClickHouse argues Postgres is not slow, your storage is.

Threads

  • Hardware you cannot trust. Ledger's implant, the Danfoss refrigeration attack, AWS AgentCore's prompt-based credential leak and Elgin's stuck Flock cameras all show security failing at the layer buyers assumed was handled; the HN passkeys and fingerprinting essays argue the same about identity.
  • Chips are the strategic input. Amazon's Decart talks (chip portability), Microsoft's Maia 300, Qualcomm's local-LLM ask, TSMC in Texas and Dongshan's listing all trace to who controls compute supply, while the Tans warn capital may be the real limit.
  • Agents everywhere. GitHub's storage rewrite, Mimic, Cloudflare's Deno deal and Sakana's MASS all assume agent traffic is the new default load.
  • Open weights lose their guardrails. The Qwen refusal-stripping build and the 64GB Mac run sit beside Nadella's call for containment, the UK AISI getting its model access back, and QGIS's AI ban: who gets to oversee, and who gets to opt out.