Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
Handwired: turn your regular JavaScript code into a neural net
Cheng Lou's Handwired transforms ordinary JavaScript control flow into exact, fully differentiable neural nets that run in the browser without a GPU—including demos like a mini desktop, trainable gallery, and parallel JSON search.
18 min · 4,245 words
Everything Is a Leaky Abstraction (in machine learning)
Elijah Potter, building on Andrej Karpathy's case for understanding backprop by hand, argues that unlike sealed abstractions such as TCP/IP, TLS or PostgreSQL, every part of a machine learning system leaks into the cost function, so practitioners should understand the whole model and training regimen before they start.
2 min · 402 words
We fine-tuned a 235B model on four GPUs. Then we didn't need it.
Vertexcover's Savitha Suresh recounts fine-tuning Qwen3-235B-A22B-Thinking for document entity classification on four GPUs: Axolotl, Unsloth, FSDP+QLoRA and LlamaFactory setups failed before DeepSpeed ZeRO-3 with CPU offload worked, lifting F1 by 7.6 points, before newer open models made it unnecessary.
7 min · 1,541 words
Introducing Glimpse: Visual thinking that stays grounded
Elorian introduces Glimpse, a research-preview visual thinking model that reasons directly with visual evidence: it zooms into regions, rewatches video intervals, tracks objects through motion and uses multimodal tools and subagents. Elorian reports frontier-level multimodal benchmark results at $0.60/$3.00 per million input/output tokens.
12 min · 2,751 words
Reflection releases the weights of Beam, its first open-weight model, a 501B-parameter sparse Mixture-of-Experts with 23B active parameters, on Hugging Face under Apache 2.0, with a model card and safety technical report. The post covers early developer demos, BF16/FP8/NVFP4 builds, the open-source Mirror CLI harness, fine-tuning and RL recipes, and more than 40 launch partners.
6 min · 1,393 words
Welcome Beam: a 501B open model you can own, customize, and scale
context window of 1M tokens, and is released under the Apache 2.0 license. Beam is focused on coding and agentic use cases. It is for teams that want a frontier-scale model on their own hardware, infrastructure, or cloud, and want to customize it to their own use cases. Beam ships with its own dedicated open source Rust based harness that delivers the best performance from the model.
18 min · 4,113 words
Neural Networks Are Already CleanA working hypothesis about neurons, generalization, and the ecology of circuits
Nick Cammarata of Goodfire argues neural networks are already 'clean': polysemanticity and messy circuits may be less fundamental than assumed, and studying neurons and circuit ecology points toward more interpretable structure than the superposition story alone suggests.
139 min · 32,079 words
Frontier Pre-training at Microsoft AI
Microsoft AI details pre-training MAI-Base-2, a 2.8-trillion-parameter MoE foundation model (80B active) trained on 16k GB200 GPUs over 38 days, which it says beats the open-weight models it tested on held-out evaluations at 4-12x better pre-training FLOPs efficiency, covering scaling ladders with source scaling, hyperparameter scaling laws, MoE and Muon optimizations, activation gradient clipping and native vision data.
12 min · 2,874 words
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B-parameter open weights model for text-to-image generation, editing and multi-reference generation with up to 10 references, with native 4K output and bounding-box control. It ships base and step-distilled checkpoints plus FP8, INT8 and NVFP4 builds under the FLUX Kommunity License, and is available through the BFL API.
5 min · 1,130 words
Brian Potter of Construction Physics looks at whether robotics can repeat AI's scaling playbook, explaining what robot training data is, how it is gathered through teleoperation, simulation, and other sources, and how large today's robot datasets are compared with those behind language models.
16 min · 3,630 words
Introducing Timestral: Mistral's time-series foundation models
Mistral releases Timestral, its first family of time-series foundation models in four sizes from 5M to 209M parameters. The largest ranks first on GIFT-Eval, TIME and fev-bench, delivers probabilistic multivariate forecasts with past and known-future covariates, runs efficiently on CPU, and ships under a modified MIT licence.
6 min · 1,385 words
Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
Hugging Face's guide to Meta's Muse Spark Open, an 816B-parameter (42B active) open-weight multimodal mixture-of-experts model with a 1M-token context: architecture, perception and audio encoders, and how to run it with Transformers, llama.cpp, SGLang, vLLM and MLX, deploy it, fine-tune it with TRL, plus demos.
19 min · 4,307 words
You could have invented transformers from memory
Alex Wang builds intuition for linear maps, MLPs and self-attention as different kinds of memory stores, using the interference-to-signal ratio to estimate each one's capacity. With 64-dimensional keys, capacity grows from about 64 memories for a linear layer to about 2,700 for an MLP and over a million for self-attention.
21 min · 4,738 words
Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio
Apple Machine Learning Research explains how MLX plus RDMA lets large open-weight models run locally and across networked Mac Studios—covering distributed inference setup aimed at the Muse Spark Open and similar workloads.
12 min · 2,785 words
Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date
Meta Superintelligence Labs releases the weights of Muse Spark Open, a natively multimodal 816B-parameter (42B active) model with a 1M-token context window built for coding and agentic work. It runs on two Apple M5 Ultra Mac Studios via MLX, is offered through hosted inference partners and major model hubs, ships with Helion GPU kernels for SGLang, and is released under the Muse Model License after safety evaluation.
3 min · 644 words
Ultramarine 1 Flash: Scaling Non-Verifiable RL
Paper Instruments presents Ultramarine 1 Flash, an agentic model for professional knowledge work trained with large-scale reinforcement learning on GLM-5.3 Flash. It improves averages across six knowledge-work benchmarks while generating fewer tokens, and costs about $0.13 per task on their internal suite.
32 min · 7,459 words
The Emperor Pelican Has No Clothes
Aman Rusia shows that frontier LLMs keep drawing nearly identical SVG pelicans on a bicycle—and contrasts that collapse with diverse human-drawn alternatives, calling it an LLM diversity collapse problem.
1 min · 87 words
LFM2.5-24B-A2B: Built for Fast On-Device Agents
Liquid AI releases LFM2.5-24B-A2B, an open-weight 24B-parameter reasoning model with 2B active parameters aimed at fast, private on-device agents. The post covers changes since LFM2-24B-A2B, agentic and tool-use benchmarks against Qwen3.6-35B-A3B, inference speed on Apple and NVIDIA hardware, and how to get started.
6 min · 1,330 words
Lost in Implicit Routing: How LLM Agents Fail to Load Relevant Skills at Scale
arXiv paper showing skill routing collapse: as irrelevant skills are added, LLM agents fail to load the right reusable skills on SkillsBench even when golden skills stay fixed.
2 min · 399 words
Google's Gemini 4 Argon post for robotics: physical reasoning that leads ERQA, with representative prompts for zero-shot control, Real2Sim practice, visual-coordinate harnesses, and agentic verification workflows across the robotics stack.
13 min · 3,033 words