Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
Jev introduces a new shape of LLM—System One, aka Decision Models
Simon Willison reviews TypeSafe AI’s Jev: a cheap, fast decision model that returns calibrated floats for yes/no, choice, and score questions—and why black-box bias and evals matter more than for chat LLMs.
3 min · 796 words
Halo: Frontier-Lab Training for Everyone
White Circle open-sources Halo, a Hugging Face–native training framework claiming up to ~2.8× faster post-training than stock TRL with lower memory use—from single GPU to multi-node, keeping checkpoints in native HF format.
18 min · 4,098 words
Reducing Image Generation cost with AMD and the Luminal Compiler
Luminal engineers show Flux.2 Klein 9B image generation costs cut by up to 47% on AMD MI300X versus an Nvidia H200, using the Luminal compiler.
15 min · 3,489 words
What Is RLCD? The Secret Behind Jev
Di Zhang explains RLCD (schema-conditioned Plackett–Luce reward modeling) and how Jev turns calibrated multiway decisions into a product—making the reward model the model rather than hiding it behind a generator.
10 min · 2,324 words
How GPT-6 Astra ascended NetHack: setup, agent loop, tool use, failure modes, and what beating a famously hard roguelike says about LLM agents in open-ended environments.
11 min · 2,477 words
Amit Shekhar walks through how LLM design moved from RNNs to attention, Transformers, scaling laws, Mixture of Experts, and the open problems still ahead.
25 min · 5,729 words
Dan McKinley argues that obsessing over prompt text misses the point: build interlocking evaluation and optimization pipelines, and treat LLMs as non-conscious systems whose 'meaning' is mostly our projection.
12 min · 2,673 words
Multimodal Agents: From Perception to Action
Illustrated notes from Berkeley’s LLM Agents lecture 7: OSWorld outcome checks, AgentTrek trajectories, TACO tools, and Aguvis grounding—why reading a screen is not the same as finishing the task.
8 min · 1,915 words
A technical walkthrough of what Jev-style decision models appear to be doing—single-token decisions, structured outputs, and why the architecture matters for agent tooling.
20 min · 4,510 words
The new season of Survivor airs next week. My favorite thing about watching Survivor is arguing about the strategy of the show: who should have won the season? Is it better to play under the radar? Or does making big moves set you up to win? Who is the greatest player of all time: Cirie? Tony? Boston Rob? I thought it’d be fun to build a machine learning model to help answer these questions.
13 min · 3,012 words
Pirate Face: The permanence layer for sovereign AI
Pirate Face mirrors open AI models from Hugging Face as checksum-verified torrents—magnet links meant to keep LLMs, image/audio models, and datasets alive peer-to-peer without a single owner or takedown point.
5 min · 1,214 words
Duality in Optimization: A Visual Tutorial
Mohini Bariya’s arXiv tutorial builds geometric intuition for Lagrangian duality in optimization—bridging solver techniques and solution interpretation with visual explanations of the dual.
1 min · 114 words
I built non-autoregressive decision models with RL a year ago
Convai Innovations’ Nandakishor recounts building Laya—a ~33ms multilingual non-autoregressive decision engine with calibrated probabilities—via RLCD a year before frontier labs framed similar System One models as breakthroughs.
8 min · 1,852 words
USRA Contributes Planetary Science Expertise to NASA-IBM Lunar Foundation Model
USRA describes an open-source NASA–IBM lunar foundation model that fuses diverse Moon datasets to support scientific analysis, including ice-prospectivity patterns near the poles.
3 min · 792 words
The Right Answer Is Not a Proof: Put Verification Inside the Reasoning Loop
Cognaptus explains PRoSFI: a 7B model emits small machine-checkable reasoning steps that Lean/Z3 can verify, raising measured soundness far more than final-answer accuracy alone on ProverQA-Hard.
6 min · 1,480 words
An Empirical Study of Harness Design for Coding Agents
Fan et al. ablate planning, action space, and context management in a fixed coding-agent loop across 176 SWE-Bench/Terminal-Bench settings, finding when context management, planning, and predefined tools help—and when bash-only is enough.
1 min · 291 words
Scaling Discovery through Test-Time Communication
Research paper showing that test-time communication among identical agents sharing discoveries can beat independent parallel search on ARC-AGI-3 and transfer to research tasks like polyomino packing and MNIST compression.
54 min · 12,394 words
Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint
PrismML introduces Bonsai 2 27B, a near-lossless compression of a 27B-class multimodal model into roughly a 9× smaller footprint aimed at efficient on-device and local inference.
4 min · 931 words
I had Gemini train its own replacement for $9
Gemini 3.1 Pro labeled 4,290 Reddit comments for $9; a fine-tuned GLiNER model now tags brands, models and materials locally at 0.83 F1 — including the tensor-mask bug that wiped five of ten runs.
7 min · 1,532 wordsagent-assisted
My AI Predictions: What Did I Get Right So Far?
Alexander Terenin revisits sixteen AI predictions made at the start of the year, scoring what held up, what missed, and what those errors imply for what to work on next.
18 min · 4,181 words