DailyWeights & Takes

23 Oct 2026 · trust the checker

Since the previous edition, 22 Oct 10:38 ET to 23 Oct 2026 11:42 ET

Edition 4 · 4 news · 1 model release · 2 repos · 5 papers · 1 thing to try

News 4

NVIDIA's OpenShell agent runtime topped HN, but its latest release came out on 20 Oct, so it isn't listed.

1. Anthropic22 Oct

   LLMs are making formal software verification feasible at a much larger scale

   Anthropic's Frontier Red Team and Theorem turned OSS-Fuzz harnesses into specifications. Claude then proved that the compiled RISC-V binaries never crash on any input under 64 MB: 239 entry points across 123 libraries, after patching the bugs it found along the way. That took a few weeks and about 180 million lines of Rocq. You don't have to trust the model, because the proof checker verifies each proof. The proofs only cover single-threaded, statically linked musl builds, and a proof is only as good as its spec. Anthropic withheld proofs that could reveal unpatched bugs and released a subset, along with the RISC-V and Linux semantics.

   395 pts241 commentscode · security discussion

2. OpenAI22 Oct

   The Decisions API is generally available

   This is a separate endpoint (POST /v1/decisions) for the classify, route and score steps that most apps run through a chat model. Instead of free text, you get a typed answer: the probability that a condition is true, one option from a fixed list, or a score on a scale. It runs on gpt-6-luna only and costs $0.10 per 1M uncached input tokens, with no charge for output. General availability raises the limit to 400 questions per request, caches shared input of 1,024 tokens or more, and accepts structured input_json.

   GAplatforms · pricing changelog

3. Anthropic22 Oct

   Agent Skills moves to the Agentic AI Foundation

   The skill format now belongs to the same Linux Foundation body that holds MCP, and maintainers from Anthropic, Google and OpenAI take part as individuals. A skill you write is now aimed at a neutral spec, not one vendor's format. Anthropic says more than 70 platforms already load skills, including ChatGPT, Cursor, GitHub Copilot and VS Code. The roadmap adds versioned releases, a formal proposal process and security guidance for authors and installers. The spec and the repo move with it.

   standardsagents

4. Sveriges Radio23 Oct

   Sweden's Government Offices are blocking ChatGPT and Claude on work computers

   As reported by Ekot, the reason is information security: an AI service used without an agreement is a risk. Staff are pointed to Microsoft Copilot, which passed a security review. Other outlets report the block starts on 26 Oct and that some agencies already have similar blocks. If you sell into the public sector, the contract decides which model gets used, not the benchmarks.

   285 pts98 commentsregulation discussion

Model releases 1

1. NVIDIA23 Oct

   Nemotron 3.5 Bolt E4B

   This is the small sibling of the 74B-A9B Bolt from 15 Oct. It's built on Gemma 4, with 8B total and about 4B effective parameters and a 64K context, and it's sized for on-device agents. The Q4_K_M GGUF is about 6 GB. The card lists llama.cpp, vLLM, SGLang, Ollama and LM Studio. On NVIDIA's own numbers it beats Qwen3.5 4B on MultiChallenge and instruction-hierarchy tests, but trails it on 64K RULER and Tau2-Bench, so it isn't the pick for long-context tool use. The license is OpenMDW-1.1.

   open weightslocal model card

GitHub repos 2

Both broke out in the window. Two other breakouts were skipped. AutoSwarm-Shadow-Adapter has about 3,000 stars against 2 forks, which looks like star farming. Folkbench ranks API relays, and its README carries paid placements from relay vendors. The repos created today were weak again.

1. GitHub23 Oct

   morluto/rea

   An MCP server and CLI that let Claude Code, Codex, Cursor or Gemini CLI reverse-engineer a target. That can be a native binary (via Ghidra, IDA or Hopper), an Electron app, an APK, firmware or a packet capture. Analysis runs locally, and findings come back with evidence. MIT code.

   +8,143 ★ in window≈7,681/day · was ~1,891/daybreakout

2. GitHub23 Oct

   Ephemeral-AI-Lab/layerfs

   This is git-like branching for agents' scratch space. Each agent gets a writable FUSE workspace over shared history, without copying the base, and commits deduplicated checkpoints it can rewind or merge. Version 1.0 is out, but the store isn't crash-safe yet. MIT.

   +132 ★ in window≈125/day · was ~25/daybreakout

Papers 5

Hugging Face daily papers featured on 23 Oct; submission dates shown.

1. Hugging Face papers19 Oct

   Deep Search in Real-World Web Environments (OpenDS-World)

   There are 200 human-verified research tasks on the live web, each with a unique answer spread across sources, file formats and interactive pages. The best of 12 frontier models answers 39.0%. People given the same tasks managed 21.1% over 1,000 attempts. Most of the failures come from tool coordination and from deciding when to stop.

   147 upvotesagents · eval

2. Hugging Face papers22 Oct

   IQuest-Q1: a foundation model for agentic CLI work

   This is the technical report for an open 320B-total, 15B-active MoE trained across several agent scaffolds, with models merged at the end. It reports 83.2 on Terminal-Bench 2.1 and 80.3 on SWE-bench Multilingual. Both are its own numbers, and GLM-5.3 Flash edges it on the first. The weights have been out since late September.

   52 upvotesopen weights · code

3. Hugging Face papers22 Oct

   Is complex agentic memory always needed? Naive RAG is stronger than we thought

   Plain dense retrieval plus reranking over the raw conversation history beat the strongest agentic memory system in six of seven settings across four memory benchmarks, by 0.1 to 5.5 points. Even full access to a transformed memory store leaves an accuracy gap, so keep the raw log before you build anything clever.

   14 upvotesagents · memory

4. Hugging Face papers19 Oct

   Heartbeat Thinking: reasoning and calling tools while the user is still talking

   A voice agent re-reads the partial transcript on a timer and lets earlier reasoning and tool calls finish instead of cancelling them. On FDBench v3 that covers 16 points more of the needed tool calls before the user stops talking and cuts final response latency by 36.4%. Gains on answer quality vary by benchmark.

   17 upvotesvoice · agents

5. Hugging Face papers22 Oct

   WeVisOne-VL: unified visual perception from objects to pixels

   WeChat's model handles detection, counting, OCR, segmentation and matting in one VLM, with a small pixel decoder triggered by a special token. It leads on more than 30 benchmarks by its own count and refuses when the object isn't there. The authors say it's being open-sourced.

   216 upvotesvision

Things to try 1

The take