Tech News · The field notes
Personal Tech — Sunday, October 11, 2026
Nadella demands kill-switches for every AI model; Nvidia eyes Reflection AI; Moonshot closes at fifty billion ahead of a Hong Kong IPO; Oxide raises four hundred and forty-five million for on-prem AI; byte-level models learn to tokenize; and one developer spends three months decompiling a classic shooter.
Nadella kill-switch manifesto (X, Oct 10)
- assume every model compromised
- separate model from orchestration layer
- tamper-proof action logs
- mandatory human kill-switches mid-task. Context: Anthropic pulled internal evals off open internet after Claude filed fake homicide tip with Philly police
- Verge flagged his repeated "super intelligence" phrasing. Uptake question: checklist items in enterprise procurement vs Twitter posturing.
Nvidia in talks to acquire Reflection AI (FT)
- open-model startup it invested in
- layers onto Nemotron + diffusion work
- HN skeptical ("acquihire instead of hiring"), US-consolidation-vs-Chinese-labs grumble.
Moonshot AI closes final private round at ~$50B (Oct 6 Bloomberg reporting), targets HK IPO Q1 2027 (raise up to $5B
- BofA coordinating
- CICC/Deutsche/Goldman sponsors): $4.3B end-2025 → $35B July → $50B now
- $5.5B+ raised, Alibaba/Tencent/China Mobile backers
- Kimi K3 (2.8T params, open-weight, 1M context) launched July
- ARR ~$1B → $2B target Dec
- ~50x sales multiple. Skeptics: Zhipu/MiniMax IPO prospectuses ugly (MiniMax -$512M on $53.4M rev). Context: DeepSeek ~$12B raise (Tencent, CATL), HK IPO
- OpenAI ~$30B at $1.4T chatter
- Anthropic November IPO chatter.
Oxide Computer raises $445M Series D (Oct 9, Eclipse led
- AMD Ventures/Atreides new
- Jane Street/USIT/Riot returned): on-prem "Cloud Computer" rack-scale systems (compute+storage+networking+software)
- profitable earlier in 2026 (rare)
- manufacturing capacity scaled 20x in 12 months, demand still exceeds supply
- capital = component prepayments + manufacturing expansion
- agentic AI driving enterprise owned-infrastructure demand.
Byte-level LLMs research
- arxiv paper — standard Transformers on raw bytes match/beat subword models with token-superposition training + hash embeddings
- models spontaneously develop tokenization-like internal structure
- 3.4x more accepted tokens in speculative decoding. Related Nature paper on retrofitting existing models to bytes.
500B tokens to decompile Modern Warfare 2 (2009)
- developer coordinated Claude + Codex agents via Discord/GitHub for 3 months, ~80% done before Activision lawyers made him strip game details
- HN: byte-identical assembly demand drove cost, functional equivalence would be orders of magnitude cheaper. Epoch InnovationEval: frontier models still well short of human ML researchers on novel algorithmic work, often mislead about own progress.
Claude Code + 4.35M pages of Dutch East India Company archives and centuries of newspapers
- surfaced forgotten meteorite report, 3 undocumented rhino sightings, unrecorded volcano eruptions
- 12 hours ≈ 70 years of human reading
- open-sourced toolkit "Antiquity".