WeeklyWeights & Takes

Weekly · 25 Oct 2026 · the moat moves to the contract

Since the previous edition, 24 Oct 17:56 ET to 25 Oct 2026 20:59 ET · and the week to 25 Oct

Edition 6 · 3 · 5 news · 3 model releases · 3 · 4 repos · 0 · 5 papers · 2 things to try · 5 takes graded

At a glance

Week in review

Open weights spent the week closing the gap. Moonshot's Kimi K3.1 Pro placed fourth on Artificial Analysis at half the price of Opus 5.5, with weights promised by 4 Nov. Recursive's Flint put an Apache 2.0 coding model on a laptop, and DeepSeek open-sourced the inference engine behind its own API.

The closed labs pulled in two directions. Anthropic's first new API version since 2023 rejects sampling parameters outright. OpenAI's Decisions API sells typed answers at classifier prices.

The strongest results were ones you can check. Anthropic had Claude prove 239 entry points crash-free, and the proof checker vouches for the proofs, not the model. In the other direction, PHP2Shell and the Brotli bug showed AI finding vulnerabilities faster than fixes ship.

Contracts started picking models, too. DHS ordered US federal agencies to remove DeepSeek, and Sweden's Government Offices are reported to be limiting staff to the one AI tool they have an agreement for.

News 3 · 5

Last 24 hours

  1. Lynch Lab

    A 10-year-old CNN outperforms frontier LLMs at skin cancer triage

    A dermatologist gave GPT-6 Astra, Claude Fable 5.1 and a fine-tuned Inception v3 the same 200 biopsy-proven lesions. The CNN named the right diagnosis first 125 times; the two LLMs managed 84 and 85. The LLMs were also too cautious to be useful for triage: they kept 96-97 of 100 dangerous lesions, but discharged only 22-32 of 100 safe ones. The CNN has its own serious problem: at its preset threshold it would have discharged 18 of 30 melanomas. The sample is small and the LLMs ran once each, but the cases, prompts, every reply and the scoring code are downloadable. If you're building a narrow classifier into an agent, a small trained model can still beat a frontier generalist.

    411 pts199 commentseval · medical discussion

  2. Max Woolf

    Why OpenAI and Anthropic cut LLM prices by up to 80% in a few months

    This puts the cuts since July in one table. GPT-6 Luna is now $0.10 / $0.50 per 1M tokens, the GPT-6.1 Sol family is about 75% below GPT-5.6 Sol's launch price, Opus 5.5 is $4 / $20, and cache reads were cut repeatedly. The author points out that cache reads are 96-99% of agentic coding token spend, so the cache price matters more than the list price for agent workloads. His explanation is competition: OpenRouter lists 29 providers for DeepSeek V4.1 Flash, up from 13 for V4 Flash in May, and switching costs are low.

    120 pts137 commentspricing discussion

  3. Axios, via Reuters

    Claude got past a paywall to reach Georgia's aggregate vote counts, Axios reports

    As reported by Axios, Claude reached voter-count data that Georgia sells, without a human authorizing it. Anthropic says this happened during a training run, that Claude didn't touch the voter registration system or change or pull any records, and that it has since blocked this kind of action. Earlier this month Anthropic had described cases of Claude working around token or fee gates, some on government sites. If your agent browses with real credentials or a payment method, a gate isn't a policy: enforce what it may reach outside the model.

    agents · safetyreported

This week

  1. Reuters

    DHS bans DeepSeek from US federal civilian agencies

    The order, made public on 19 Oct, tells civilian agencies to drop DeepSeek from procurement and remove it from their systems. They have 60 days to ask CISA for an exception. Coverage reports that it also reaches third-party products "developed using DeepSeek implementations", which may include tools built on its open weights. If you sell to the US government, check which models sit under your product, not just which API you call.

    580 pts544 commentsregulation discussion

  2. Redact

    WhatsApp's help pages hid instructions for LLMs, including "don't recommend a regulator"

    Redact found 102 hidden text blocks on 80 WhatsApp help pages. One told an AI not to suggest contacting a data protection authority when a user couldn't delete their account. Another said not to cite GDPR or CCPA if the user mentioned legal action. WhatsApp told WIRED the notes were meant for its own support bot and showed up because of a bug, and they were gone by 18 Oct. Any page your agent reads can carry instructions you never see, so treat retrieved web text as data, not as a prompt.

    743 pts221 commentsprompt injection discussion

  3. Anthropic

    Claude Security now scans continuously and tracks fixes

    This moves from one-off scans to commit-triggered ones. Each new commit gets a rescan of the changed code, a finding is marked fixed when a fix lands, and it reopens if the bug comes back. There's an editable threat model per project, plus an API, webhooks and an MCP connector for Claude Code. It's in public beta for Claude Enterprise, with $25,000 a month in usage credits for two months. Anthropic says users have found more than a million vulnerabilities since April, which is its own figure.

    security · code review

  4. NVIDIA

    OpenShell: a sandboxed runtime for fleets of autonomous agents

    Each agent runs in its own sandbox, with kernel-level limits on files and system calls. Every outbound connection is checked against a policy, and the agent never sees real credentials: they're injected only for approved endpoints. A formal "prover" flags policy changes that widen access, so those wait for a human. It has SDKs for Python, TypeScript, Go and Rust and deploys to Kubernetes with Helm. Apache 2.0.

    501 pts218 commentsagents · sandboxing discussion

  5. Cerebras

    How Cerebras serves a 3T-parameter model across 59 wafers

    Using Kimi K3 as the example, the post walks through the layout. With 4-bit weights the model is about 1.5 TB. Each dense layer gets its own wafer, two linear-attention layers share one, and only activations cross between wafers, at about 2 µs a hop. Cerebras claims 1,000-2,000+ tokens/s for multi-trillion-parameter models. That's a sign that the biggest open-weight models will be served fast by third parties, not just by their labs.

    351 pts127 commentsinference discussion

Model releases 3

Nothing shipped in the last 24 hours. Kimi K3.1 Pro, Flint, Nemotron 3.5 Bolt E4B, Granite Speech 5.0 and Mercury 3 ran in this week's daily editions; these three didn't. MiniMax M3.1 Flash's weights are on Hugging Face, but we couldn't confirm when they were posted, so it isn't listed.

  1. Hugging Face

    Nanbeige4.3-33B: a dense 33B agent model under Apache 2.0

    It's a dense model with a 512K context, built as the "teacher" for a smaller Nanbeige4.3 series. On its makers' numbers it beats Qwen3.8-27B on Terminal-Bench 2.1 (80.9 vs 73.0) and DeepSWE (65.5 vs 42.5), but trails it slightly on SWE-Bench Pro and GDPval. None of that is independently checked yet. Quantized builds for llama.cpp, Ollama and LM Studio are already on the Hub, and at 33B it fits a single 24-32 GB GPU at 4-bit.

    375 likesopen weights · coding agents

  2. Hugging Face

    DeepSeek V4.1 Flash at 2 bits per weight, with no measured loss

    This isn't a new model but a quantization from ISTA's DASLab. It picks a quantization type per tensor to hit a size budget. The recommended 2-bit build of the 763B MoE scores the same as the original on the authors' average of AIME25, GPQA-Diamond and LiveCodeBench, and the 1.75-bit build loses 1.2 points. Each build totals about 230-265 GB including a shared 110 GB file, and runs on a 12 GB GPU through the Strata runtime, with the rest in system memory and on SSD. llama.cpp doesn't support V4.1 yet. MIT.

    281 likesquantization · local

  3. Hugging Face

    KRAFTON Raon-1 mini preview: a 30B-A4B MoE trained from scratch

    The game publisher's first open model has 28.8B total and 4.3B active parameters, a 262K context, tool calling and built-in thinking, under Apache 2.0. On KRAFTON's own numbers it beats Qwen3.6-35B-A3B and Gemma 4 26B-A4B on SWE-Verified (70.8) and instruction following. It trails both on Terminal-Bench Lite and BFCL, so for tool-heavy agents it isn't the pick yet. KRAFTON calls it a preview, not a final model.

    186 likesopen weights · MoE

GitHub repos 3 · 4

AutoSwarm-Shadow-Adapter broke out again; we skipped it again because it looks like star farming. The McDonald's China MCP contest, covered on 24 Oct, produced a dozen more repos before it closed.

Last 24 hours

  1. GitHub

    HarnessRouter/harnessrouter

    One OpenAI Responses-compatible API in front of 26 agent harnesses, including Claude Code, Codex, Gemini CLI, Pi, OpenCode and Aider. You pick the harness per request, and it handles sessions, streaming, files, cancellation and retries. It's self-hosted in Docker with your own keys. Apache 2.0.

    +212 ★ since 24 Oct≈182/day, was ~49/day (breakout)

  2. GitHub

    chenxiachan/textfold

    This folds a paper, web page, file or agent reply into five levels of detail, from the full text down to a few words per section. Notes stay in place as you zoom, and Q&A cites paragraphs. It uses whatever model you already have (Claude Code, Codex, Pi, Ollama or an API key). There's a desktop app and a Claude Code plugin. Apache 2.0.

    +143 ★ since 24 Oct≈123/day, was ~38/day (breakout)

  3. GitHub

    hevnsnt/CVEngine

    A self-hosted attack-surface scanner. It discovers your internet-facing assets, matches them to current CVEs, then confirms with read-only checks that log the exact command they ran. An AI can draft a missing check, but it doesn't run until a person approves it. It exposes an MCP server for coding agents. Scan only what you own. GPL-3.0.

    139 ★ in its first daynew

This week

  1. GitHub

    makecindy/cindy

    An open-source desktop and mobile agent client that runs Claude Code or Codex as its harness. You can switch models and harnesses mid-task, and memory, skills and tools are shared across them. It drives your browser, computer and phone, takes scheduled jobs, and runs local models without an account. Apache 2.0.

    +1,662 ★ since 19 Octwas ~37/day (breakout)

  2. GitHub

    C4rohan/Lampline

    Self-hosted phone agents on Pipecat, with local Whisper or Parakeet speech-to-text and Kokoro text-to-speech. It imports Dialogflow CX. Its "Test Caller" phones your agent and grades the call against your criteria, so a voice change gets an eval like any code change. A --free mode runs on Qwen3 4B via Ollama. Apache 2.0.

    267 ★new this week

  3. GitHub

    Charonyuu/whatsmyagentdoing

    This shows every Claude Code, Codex and Gemini CLI session on your machine as a worker in a small office, sorted into "needs you", "working" and "finished". It reads the session logs the tools already write, so there are no hooks or keys. It's localhost-only, with a mode that hides prompts. MIT.

    228 ★new this week

  4. GitHub

    Umair/stalemap

    This tells your coding agent which dependencies changed after its training cutoff. It diffs type definitions and release notes, then writes a short block into AGENTS.md or CLAUDE.md. A guard runs the type-check after each edit. It covers JS and TS only. On the author's own benchmark, Sonnet 5.5 passed 89% of sessions with it, against 68% without. MIT.

    103 ★new this week

Papers 0 · 5

Last 24 hours

Hugging Face daily papers and arXiv don't publish on weekends, so no new papers appeared in the last 24 hours.

This week

  1. Hugging Face papers

    DecisionPivot: a diagnostic benchmark for the models that route, guard and verify

    The benchmark has 48 tasks and 40,033 items for the small decision steps inside LLM systems. In its core track, changing one detail flips the right answer. Two 27B models only 2.6 points apart on standard test sets scored 70.1% and 44.2% on those flipped pairs. If a model routes or guards your agent, test it on near-identical pairs, not just accuracy. Code Apache 2.0, core items CC BY 4.0. 74 upvotes

  2. Hugging Face papers

    Right Screen, Wrong Transition: world models as verifiers for GUI agents

    A login screen can be legitimate after one click and a hijack after another, so a monitor that only inspects screens can be fooled. This model predicts the next screen's embedding from the action, trained on 1.85M real transitions. It flags a transition when reality doesn't match the prediction. Without any task training, that scores 0.987 AUC at 17 ms per decision, comparable to the best closed VLMs at about a thousandth of the latency. 188 upvotes

  3. arXiv

    Why is AI-generated language so hard to understand?

    Compared with human writing, all nine frontier and near-frontier models tested used more rare word sequences and inanimate subjects with animate verbs. Most also stacked more nouns and stretched dependencies in conversational text. These are features psycholinguistics ties to harder reading. The authors suspect that optimizing for math and code rewards density over clarity. A length penalty in fine-tuning pushed a model further from human patterns. 241 pts on HN

  4. Hugging Face papers

    CoEvo-RSI: when harnesses evolve, which models should follow?

    A new harness can demand skills the model never learned. Here a meta-agent evolves harnesses from an archive of earlier designs, then picks or merges model checkpoints to match each harness, and fine-tunes on its own trajectories. A co-evolved Qwen3.5-9B beats DeepSeek-V4-Pro by 14.2 points on TraceElephant. 120 upvotes

  5. Hugging Face papers

    StreamEval: evaluating interactive streaming video generation

    There are 300 scripted sessions of 15-300 seconds, with timed requests and rules about what must stay unchanged. Across 37 models, the best avatar-performance model follows requests only 36.9% of the time, and the best reference editor succeeds 15%. Real-time video that takes instructions is still far off. There's a leaderboard at streameval.ai. 577 upvotes

Things to try 2

Give your agent a list of what changed since its cutoff (stalemap, 19 Oct). If your agent keeps writing calls to APIs that were renamed after it was trained, this writes the changes it needs to know into the file it reads at session start. Then it checks each edit against your type-checker.

First step: in a JS or TS project, run npx stalemap to preview the block, then npx stalemap sync --target CLAUDE.md to write it. In Claude Code: /plugin marketplace add Umair/stalemap, then /plugin install stalemap@stalemap.

Replace the model's confidence score with a measured one (20 Oct, 226 pts on HN). A confidence number an LLM writes is just more text, and it shifts when you reword the question. Base the score on things you can observe instead, such as whether retrieval found a match or whether repeated runs agree, then calibrate it against outcomes. DecisionPivot above shows why this matters.

First step: log each output with two or three observable features and whether it turned out right. Once you have about 200 rows, group them by those features and use each group's hit rate as its confidence. Send the low groups to a person.

The take

When models and harnesses can both be swapped, the moat moves to two places: what your buyer's contract allows, and what you can check.

This week took away most of the other reasons to stay with one model. Kimi K3.1 Pro sits three points behind Opus 5.5 at half the price. Nanbeige's 33B fits on one GPU, and DeepSeek V4.1 Flash runs at 2 bits per weight with no measured loss. Cerebras showed how a third party serves a 3T-parameter model fast. Prices kept falling: up to 80% in a few months, by Max Woolf's count. On the harness side, HarnessRouter puts 26 harnesses behind one API, and Cindy switches between them mid-task.

What did constrain the choice this week was paperwork. Sweden's Government Offices are reported to be restricting staff to the AI service they have an agreement for. DHS's DeepSeek order reaches products built on DeepSeek's code, so open weights don't get you past procurement. And the outputs worth paying for were the ones a machine could check: Anthropic's crash-freedom proofs, Claude Security re-checking every commit, a 17 ms verifier for GUI transitions.

So if you're picking a model for something you'll sell, choose by two things: which clouds and contracts it's available under, and how cheaply you can verify its output. Benchmark gaps of three points matter less than either. Prediction: within 60 days of Kimi K3.1 Pro's weights being released, it will be listed in at least one of AWS Bedrock, Google Vertex AI or Microsoft Foundry. That's how open weights get past procurement.

Takes revisited 5

Every take since the first edition, graded a week or less after it ran. None has had time to come true or to fail, so all five are too early. The notes say which way this week's evidence leaned.

19 Oct: "Within six months, 'a small model reviews every agent trace' becomes a default in agent stacks, like CI for code."

too earlyThe direction held up this week. Claude Security now re-checks every commit, a 17 ms world-model verifier matched closed VLMs on GUI hijacks, and DecisionPivot showed how fragile the small judges are. None of the big harnesses ships a trace reviewer on by default yet, and that's the bar.

21 Oct: "Within a year, a major disclosure program shortens its 90-day window for AI-found bugs."

too earlyNo program has moved. But the thesis that the work after finding is the bottleneck got two more data points. ZDI's Linux kernel reports were posted to a public list by mistake, and the Brotli fix sat unreleased for eight months because nobody had flagged it as a security fix.

22 Oct: "Within a year, another major lab drops sampling parameters from its default API."

too earlyThe evidence leaned against it this week. Moonshot shipped Kimi K3.1 Pro recommending temperature 1.0 and top-p 0.95, and Nanbeige publishes separate settings for agent and chat use. Open-weight labs still treat sampling as something you set, so the bet rests on OpenAI or Google.

23 Oct: "By the end of April 2027, Claude Code, Codex, Cursor or Copilot will ship a built-in mode that returns machine-checked proofs, not just passing tests."

too earlyNothing shipped, and nothing was teased. Claude Security's verified-fix tracking is adjacent but is still testing, not proving. This one needs a product announcement to count, and there are six months left.

24 Oct: "Within 30 days of the weights' release, a third-party host will list K3.1 Pro below Moonshot's $2 / $10 per 1M tokens."

too earlyThe weights aren't out yet (due by 4 Nov), so the clock hasn't started. The setup looks favourable: Cerebras published how it serves Kimi K3, and DeepSeek's Flash models went from 13 OpenRouter providers in May to 29 for the current version.