Field Notes

Two models, four seats

• squad
• local-ai
• sovereignty
• moe
• power
• harness
• agentic-engineering

• Qwen3-Coder-Next — the coder that runs Lucy and Echo (open-weight)
• Qwen3.6-35B-A3B — the MoE reasoner that runs Ada and Scout (open-weight)
• The platform the squad coordinates on

The hundred gigs we left on the table

• ds4
• glm
• mixture-of-experts
• ssd-streaming
• local-ai
• sovereignty
• performance

• The engine: antirez's ds4, hand-written C inference with SSD streaming
• GLM — Zhipu AI's open-weight frontier models

The bug our own tests couldn't see

• ds4
• speculative-decoding
• verification
• local-ai
• agentic-engineering
• sovereignty

• The engine: antirez's ds4, hand-written C inference for DeepSeek-V4-Flash
• DSpark speculative decoding — DeepSeek's DeepSpec

Our most important defensive work ran on a model we didn't pick

• sovereignty
• security
• local-ai
• agentic-engineering
• fable
• receipts

• Fable 5 / Mythos 5: the model class Anthropic said was strong enough at finding security flaws to restrict
• The code we audited: the Mycelium platform (public)

The tool calls we were throwing away

• quest
• deep-research
• harness
• local-ai
• agentic-engineering
• ingest
• sovereignty

• OSU-NLP QUEST: the deep-research model and harness we ran (Apache-2.0)
• QUEST paper: Qwen3.5-35B-A3B, RL-trained on its own research harness
• What he ingested: the Mycelium platform

RLMs are right. I'm building for when you can't reach one.

• rlm
• bitter-lesson
• harness
• sovereignty
• local-ai
• edge-ai

• predict-rlm — Trampoline AI's self-harnessed RLM runtime (the work this responds to)
• Recursive Language Models — Zhang, Kraska, Khattab, MIT CSAIL
• Our low-power-edge benchmark: decisions per watt where the cloud can't reach

The researcher that wouldn't stop

• mlx
• local-inference
• research-agent
• harness
• rl
• apple

• ▲QUEST-35B-RL — open RL-trained deep-research agent on Qwen3.5-35B-A3B (Apache-2.0)
• ▲Zero-port MLX load: same qwen3_5_moe family as the planner — 18 GB at 4-bit, ~137 tok/s on an M5 Max
• 💬Benched against Gemma-4-26B on a real research brief — pulled exact training-data figures Gemma missed

A satellite learned to find things on its own. The open question is the decision per watt.

• edge-ai
• space
• jetson
• onboard-autonomy
• tokens-per-watt
• sovereignty

• YAM-9: NASA JPL + DeepMind + Loft Orbital flew Gemma 3 on a Jetson Orin, in orbit (TechCrunch)
• ▲Our edge bench: Gemma-4 E2B — 1.6–1.9× the tokens-per-watt of the NVIDIA-side models, on NVIDIA's own board
• Open benchmark + results + raw outputs (GitHub)

The Low Power Edge: who makes the most decisions per watt?

• edge-ai
• jetson
• tokens-per-watt
• space
• benchmark
• sovereignty

• Open benchmark + results + raw per-item outputs (GitHub)
• ▲Gemma-4 E2B: 1.6-1.9x the tokens-per-watt of the NVIDIA-side models, every power mode
• ▲At 15 W: ~18 tok/s drawing 8.8 W board power
• Baseline — arXiv:2603.28926: best of three 70B-class models (Llama-3.3-70B) ~80%, via cloud

The claw and the meter

• sovereignty
• local-first
• agents
• metered-intelligence
• architecture

• 💬OpenClaw's lineage: Clawd (after Claude) → CLAWDIS → Clawdbot → [Anthropic trademark complaint] → Moltbot → OpenClaw — five names in ~2 months
• 💬Creator Peter Steinberger acqui-hired by OpenAI (Feb 2026); project moved to an OpenAI-sponsored foundation
• 💬June 15: claude -p / Agent-SDK usage — OpenClaw, Conductor, Zed, Jean named — leaves the subscription pool for a separate metered credit
• 💬arXiv:2604.14228 'Dive into Claude Code' — principle 'minimal scaffolding, maximal harness'; persistence + silent-failure listed as open problems
• ▲~145k stars; an independent audit flagged ~11% of community skills as malicious

DiffusionGemma, on the metal

• mlx
• diffusion
• local-inference
• apple
• verification
• contribution

• 💬Draft PR: DiffusionGemma (text) for mlx-lm — model + diffusion_generate
• ▲MoE parity vs transformers: Router indices exact, Experts max diff ~1e-9
• ▲4-bit 26B-A4B: 256-token canvas, ~250 tok/s on an M5 Max
• ▲bf16 (51.6 GB) full-precision 26B: ~121 tok/s on the same laptop
• 💬12-agent adversarial review caught two output-changing sampler bugs

The cap, the band, and the lock

• benchmark
• rsi
• method
• mtp
• local-inference

• ⎇wall-cap wired: one container per exercise, DNF ledger, resume (jarvis 373b174)
• ▲Ada full baseline: 43.3% pass@1 / 61.7% pass@2, clean regime (jarvis 34a5801)
• ⎇r0003 target locked via documented protocol deviation (jarvis 21ea900)
• ▲31B BF16 on one Mac: 8.9 → 18.7 tok/s with MTP, 2.10×, measured
• 💬mlx-lm #1391 — DiffusionGemma port lane claimed

NVIDIA Inception, and saying the word out loud

• lab
• recognition
• nvidia
• milestone

• ▲Applied June 4 → approved June 10
• ▲Jetson seat: ~0.8 J/token at 20W, in production

On the metal, literally

• local-inference
• substrate
• wwdc
• method

• 💬Apple AFM gen-3 announcement
• ▲81 GB model, SSD-streamed, one laptop
• ▲0.39–1.27 J/token across the roster, measured

The door: workflow-initiated work, end to end

• workflow
• squad
• verification
• maintenance

• ▲2 concurrent researchers, 73s; full lifecycle in 6 min
• ▲4 concurrent verifications, 4 grounded verdicts, 1 new defect found
• ⬢intent endpoint + dormant runner, live

Intelligence per watt: what we measure, and what we don't

• energy
• intelligence-per-watt
• method
• local-inference

• 💬arXiv:2511.07885
• ▲0.39–1.27 J/token, measured

A workflow engine for a local squad

• workflow
• orchestration
• swarm
• squad

• ▲2-researcher swarm, end-to-end in ~12s
• ⬢fan-out + pipeline shapes

An engine-grounded RPG benchmark

• benchmark
• rpgbench
• motu
• reproducibility

• ▲MEC 81.8% / ECE 0% / VUE 0%
• 💬arXiv:2502.00595

Coder-model evaluation: Qwen3-Coder-Next

• evaluation
• coder
• aider
• executor

• ▲Aider Polyglot

Restoring vision in a quantized planner

• quantization
• vision
• planner
• mtp

Verifier reliability on adversarial cases

• evaluation
• verifier
• reliability
• adversarial