DailyWeights & Takes

28 Oct 2026 · small specialists, open weights

The last 24 hours, to 28 Oct 2026 09:46 ET

Edition 9 · 3 news · 1 model release · 5 repos · 5 papers · 1 thing to try

News 3

1. Axiom27 Oct

   The Classification of Finite Simple Groups is now formalized in Lean

   The classification proof runs to thousands of journal pages and is one of the largest in mathematics. Axiom's team of 31, led by Letong Hong, has turned it into 43.1M lines of Lean 4 and 2.05M declarations, drawn from 715 papers and books, with every logical step checked by Lean's kernel. It follows the first-generation proof and builds on Mathlib and the Tau Ceti project. The repository is public under Apache-2.0, and a full lake build takes more than a day, which the team plans to bring under 24 hours. The site doesn't say how much of the work was AI-written. The open question is now whether the formal statements say what the mathematicians meant, not whether the steps hold.

   226 pts69 commentsformal verification discussion

2. Anthropic27 Oct

   Anthropic reports lab results from partners using Claude in drug discovery

   The post collects partner results with numbers and linked technical reports. At BigHat, Claude was connected to the lab's data system and proposed changes to two antibody leads. About 1,010 variants were made and tested over 10 weeks, and the best bound more than 10× more tightly. At Enveda, Claude reviewed 1,061 pages of manufacturing documentation, caught all 11 issues human reviewers found, flagged 7 they missed, and cut review time by 89%. Divergence Labs generated about 35,000 antibody designs in a month. Everything is preclinical, and Anthropic is the source for every number. The pattern that works is Claude wired into a lab's own data, with experiments checking every proposal.

   264 pts256 commentsagents · science discussion

3. Artificial Analysis27 Oct

   Artificial Analysis Intelligence Index v5 swaps out saturated tests

   v5 adds FrontierCode 2.0 and Terminal-Bench-Science and upgrades HLE to HLE-Diamond, plus new versions of AA-Omniscience and CritPt. It drops GDPval-AA, SciCode and AA-LCR. On the new scale, Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55. Gemini 4 Argon (high) is sixth at 50. Kimi K3.1 Pro scores 46, but its weights aren't out yet. Among models whose weights you can download, DeepSeek V4.1 Flash 1017 and GLM-5.3 tie at 32. Scores aren't comparable with v4, so re-check any model choice you made from the old numbers.

   74 pts73 commentsleaderboards methodology discussion

Model releases 1

1. Liquid AI27 Oct

   LFM3: six open-weight ~2B models for text, vision, audio and agent decisions

   All six branch from one 1.7B base. The text model, LFM3-1.7B, has a 256K context and fits in about 1 GB at 4-bit. A 128K context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B. Liquid's average puts it at 49.37, above Qwen3.5 9B's 47.24. It wins on instruction following (IFEval 96.7 vs 92.1), tool calls (BFCL v4 71.4 vs 62.8) and web agents. It trails on GPQA Diamond (63.4 vs 80.4), SWE-bench Verified (43.4 vs 51.2) and long-context recall. On-device tests in llama.cpp at Q4_0 decoded 1.8-2.8× faster than the next model, and each model can be fine-tuned on one GPU. It's a candidate for on-device tool calling, not for coding.

   152 pts58 commentsopen weights · on-device weights discussion

GitHub repos 5

Three breakouts and two created in the window. We skipped a desktop agent with 378 stars in its first hours: its repo holds only a README and a Windows installer, with no source.

1. GitHub28 Oct

   google/elephant-goldfish-model

   A design-review plugin for Claude Code, Codex and Antigravity. An "elephant" agent with full project context drafts the spec. Then "goldfish" subagents with no prior conversation review it for hidden assumptions, and you sign off before any code is written. There is also a drift report that compares your docs with the current code. Apache-2.0, and not an officially supported Google product.

   ≈+1,750 ★ in windowwas ~53/day (breakout)

2. GitHub28 Oct

   milind-soni/OpenMausBot

   A desktop chat app where each bot is a real Claude, Codex or Grok CLI agent, with its own model and an optional cloud Linux desktop. Shell commands and file edits come up as approval cards. It also has schedules, webhooks, an MCP server and a headless CLI. API keys are stored in plaintext config.json, so set it up with that in mind. Apache-2.0, except an enterprise folder under its own license.

   ≈+1,416 ★ in windowwas ~59/day (breakout)

3. GitHub28 Oct

   FrigadeHQ/hob

   A desktop app for running open-weight models like Gemma and gpt-oss offline. It picks a model that fits your machine and installs it, then works through documents and fills in PDFs. Builds for macOS (Apple Silicon), Windows and Linux. MIT.

   +71 ★ in windowwas ~18/day (breakout)local

4. GitHub27 Oct

   onurozcan/omux

   A native macOS terminal for running many Claude Code sessions side by side. Each session shows whether it's working, needs you or is done, and you can answer permission prompts from the sidebar. Worktree sessions keep parallel agents on separate branches. No telemetry, but it reads your login to call an undocumented usage endpoint. Build from source, macOS 14+. MIT.

   40 ★new, created 27 Oct

5. GitHub28 Oct

   aa1143/NamiGPT

   A 15-chapter, Chinese-language course that builds GPT-2 from scratch in PyTorch, from tokenizer to attention to sampling. Each chapter has a runnable script, and the model is under 200 lines. It ends by loading OpenAI's GPT-2 weights into your own model and inspecting next-token probabilities in a Gradio demo. Code MIT, text CC BY-NC.

   87 ★new, created 28 Octlearning

Papers 5

From arXiv's 28 Oct listing and Hugging Face's daily papers for 28 Oct.

1. arXiv28 Oct

   InfraDev: coding agents on real Megatron-LM, vLLM and verl changes

   The 36 tasks each reimplement a real multi-file pull request that has to run correctly across many GPUs. The best of six frontier models resolves 38.9%. Hidden probes reject 32% of submissions that pass every official test. Given only the original PR descriptions as prompts, no model solves a single task. Agents' own tests rarely check memory, time or traffic. In infrastructure code, "the tests pass" covers far less than you'd hope.

   coding agents · benchmark

2. Hugging Face28 Oct

   h1: a plain agent loop with a fast Lean interface matches specialized provers

   h1 gives a model an efficient Lean 4 interface and keeps it going on verifier feedback, without imposing decomposition or search. The authors report state-of-the-art results on PutnamBench and FATE-X and eight settled open problems in FormalConjectures. With the same models, it matches or beats specialized workflows like AlphaProof Nexus. In ablations, the interface's speed drove cost and solve rate more than workflow design did. Before you build an elaborate harness, make the feedback loop fast.

   10 upvotestheorem proving · harness

3. Hugging Face28 Oct

   Memory Tax: many agent memory systems lose to plain RAG over raw history

   Across LongMemEval-s, PersonaMem-v2 and MINT-Eval, on three model backbones, eleven memory systems from flat to graph to trained were compared with plain retrieval over the raw history. Many came out behind. The authors call the loss "memory tax": rewriting and distilling history can hide the evidence an answer needs. Retrieving from the raw history next to memory, then reranking both, reduced it across several systems.

   68 upvotesagents · memory

4. arXiv28 Oct

   When Retrieval Is Not Enough: obsolete facts in agent memory

   The retriever found the newest version of each fact, and the model still often answered with the old one. With four pieces of evidence, swapping one unrelated statement for an outdated version of the answer dropped accuracy from 100% to 44.6% on Qwen2.5-7B and to 87.8% on Qwen3-14B. Retrieving more context made it worse. If facts in your agent's memory change, delete the old versions instead of keeping them around.

   agents · memory

5. Hugging Face28 Oct

   MoE decoding: picking expert channels needs no calibration data

   At batch size one, MoE decoding is limited by loading expert weights, so selectors read only the channels that matter. Simple top-k selection matched or beat calibrated thresholds in 40 of 43 comparisons and avoided the repetition loops thresholds caused on Gemma 4. The authors' reader keeps 98.1% of full accuracy at 17.8% of expert reads. Its Triton kernels run MoE blocks 1.6-1.9× faster than optimized vLLM at batch one.

   10 upvotesinference · MoE

Things to try 1

The take