Blog
Agents, context, and the harness.
Essays and notes on operable AI — the stack around the model, not prompt tricks.
AI Report — October 20, 2026
DeepSeek DeepInfer + US ban, llm-d 1.0, TorchTPU, NVIDIA Cosmos3-H, FairStress, ALTK-Evolve, Mistral voice.
AI Report — October 14, 2026
LiteLLM gateway security wave, NVIDIA UK and EU city agents, Allen AI HiRO-ACEv2, Meta Muse, open-weight news.
AI Report — October 13, 2026
Meta Muse, Anthropic self-replication research, AI infra capital risk, LiquidAI on-device agents, NVIDIA city agents, MCP/LiteLLM security.
AI Report — October 12, 2026
NVIDIA Vera Rubin at OCP, OpenAI math backlash, Qwen3.8 and Mistral ML4, decision-model benchmarks, Reflection deal talk.
AI Report — October 11, 2026
Decision-1 + OpenAI Decisions API, NVIDIA–Reflection deal talk, Anthropic police-tip incident, PoeLLM, Ecosia.
AI Report — October 10, 2026
Microsoft Decision 1 (Qwen), open decision scorers, Mistral benchmark follow-up, Qwen image model, weight-theft policy.
AI Report — October 9, 2026
OpenAI math backlash and GPT-6, Claude Haiku 5.5, Ecosia leaves Mistral, AllenAI GPU scheduling, NVIDIA Omniverse agents.
AI Report — October 8, 2026
RTX Spark + Windows agents, Mistral Large 4 follow-through, Ecosia model switch, PoeLLM and Korea bank attacks.
AI Report — October 7, 2026
Mistral Large 4 and Le Chonk, Nemotron olympiad fine-tunes, LMCache RCE, open-weight geopolitics.
AI Report — October 1, 2026
Innermost Loop + NVIDIA: agentic AI with CoreWeave, AI factory ROI, fellowship cycle.
Future workflows: routines without cron
Background jobs in tmux pane 2 that die with the session — and post results back to channel.
The execution loop: dispatch, work, reply
How JARVIS keeps chat, agent pane, and host aligned for every run.
context.md as the agent contract
Why scope rules live beside the stack, not inside the system prompt.
Case study: JARVIS v4 in production
The harness we run on ourselves — open source under MIT.