Learn AI by Building
From your first dataset to production agents — deep-dive series, hands-on notebooks, and experiments you can rerun yourself.
Paper of the Week
All Issues →Paper of the Week #9 — The Best Agent Passed 46 of 168 Physics Research Runs
This week's roundup. In FrontierPhysics, 42 physicists turned their own research into 56 tasks, and the best agent configuration passed 46 of its 168 runs (27.4%); the failures were mostly loose reasoning, not missing physics. IdeaTaste finds that agent-written research proposals overlap with the work researchers actually published at about 18% of the rate human papers overlap with each other. Also: Route Anything, where a perfect per-question model picker would gain 5.12 points but trained pickers gained 0.30; CoworkOS, a harness that ran five rounds of fine-tuning on another model; and LoCo, which cuts a 7B reasoning model's answers by about 44% on math.
Premium Series
Our Products
Courses and code built from what we measure here
Video Courses
13 hands-on courses — quantization, diffusion, RAG, on-device AI, decision models. $199 lifetime bundle, first 3 lectures of every course free
LLM Quantization and Compression Hands-On
GPTQ, AWQ, GGUF, QLoRA — fit LLMs into the memory you have. The course behind our KV-cache measurements
Free Benchmark Code
The scripts and raw results behind the measurement posts, starting with the KV cache harness. Free with an account
Premium Series
140 deep-dive posts across 21 series, bilingual KO/EN, with production-ready code and notebooks
Latest Posts
View All →
One Sentence Through GPT-2, Qwen3 and a Mamba Hybrid: Where Does the Name Come From?
We wrote the forward pass of GPT-2, Qwen3-0.6B-Base and Granite-4.0-H-350M by hand and followed the last token of one sentence through every layer. All three answer Alice. In the hybrid, switching off the four attention layers changes nothing at 18 tokens, but at 1,410 tokens the model without them falls to 0.002, while the full model stays at 0.990.
28 Attention Designs, Each With a GIF and a Short Function: From Bahdanau to Latent Sparse Attention
Additive attention, MHA, GQA, MLA, sinks, BigBird, DSA, MoBA, Reformer, ClusterKV, LSA, Gated DeltaNet and more, grouped by what each one changes. Each has a GIF, a short PyTorch function and what it improved, why and at what cost, with the papers' own numbers. The functions are teaching versions checked on CPU: we did not train them or measure their speed.
Paper of the Week #9 — The Best Agent Passed 46 of 168 Physics Research Runs
This week's roundup. In FrontierPhysics, 42 physicists turned their own research into 56 tasks, and the best agent configuration passed 46 of its 168 runs (27.4%); the failures were mostly loose reasoning, not missing physics. IdeaTaste finds that agent-written research proposals overlap with the work researchers actually published at about 18% of the rate human papers overlap with each other. Also: Route Anything, where a perfect per-question model picker would gain 5.12 points but trained pickers gained 0.30; CoworkOS, a harness that ran five rounds of fine-tuning on another model; and LoCo, which cuts a 7B reasoning model's answers by about 44% on math.
Latent Sparse Attention: 72x in One Layer, 9x for the Whole Model at 10M Tokens
The paper reports 72.4x prefill in one Qwen3.8-27B layer and 9.21x to first token for the whole model at 10M tokens, with almost no first-token gain at 512K. We follow real queries through a small LSA index on Qwen3-8B to show how it did, what its limits are and what it missed. We did not run the LSA server.
We Forced Qwen3.8 27B Into 8GB. It Failed the Same Task 9 Times.
Qwen3.8 27B with every layer on the GPU stays inside a 7,664 MiB budget at about 50 t/s. On four Python tasks it scored 0.462 against 0.781 for Qwen3.5-9B on a 16K f16 cache, and one task got 0 in all nine runs across three settings. Measured on an A100.
What Phone Harness Sends by Default: One Agent Script, Captured on an Android Emulator
We ran Phone Harness 0.3.3 on an Android emulator against a local stand-in server. Default settings produced one event holding the script, the typed text and the printed screen text, and a failed tap sent unprinted screen text too. Each of the three off switches stopped it.