Sunday, October 18, 2026

Pulse (last 3d) · Research 119 · Agents 83 · Products 60 · Industry 48 · Open Source 42 · Models 41 · Legal 26 · Enterprise 26 · Releases 23 · Inference 19

Trending · OpenAI 17 · Codex 16 · Claude Code 14 · Claude 13 · Nvidia 13 · Anthropic 12 · Meta 10 · FDA 9 · Hugging Face 9 · ChatGPT 8 · Cursor 7 · GitHub 7

---

• DeepSeek V4.1 Flash is MIT-licensed, claims 77% on DeepSWE 1.1, and Ollama already serves it with zero retention. huggingface.co
• Meta's Muse Spark Open (816B total, 42B active) is now the top US open-weights model on Artificial Analysis. x.com
• NVIDIA's Nemotron 3.5 Bolt fits a 120B-class text-and-vision agent model into 24 GB. producthunt.com
• StepFun's Step 5 Preview is an open-weight 600B MoE with 27B active and a 1M-token context. producthunt.com
• Context Trace audits Claude Code and Codex session logs for leaked sensitive data, locally. producthunt.com
• Replit makes OpenAI its default model provider over Anthropic. techcrunch.com
• A compromise of opencode's console ran up charges on users' linked provider keys (single source). x.com

• Replit drops Anthropic as default - CEO says OpenAI is now Replit's default model provider. techcrunch.com
• Leaked Anthropic draft bill - Zvi Mowshowitz calls Anthropic's draft AI regulatory bill unacceptable and troubling. thezvi.substack.com
• Anthropic workforce studies - Anthropic's own studies report a less junior staff, more meetings, and staff expectations for AI's economic impact. x.com | techcrunch.com
• Scaling Discovery Requires Forgetting - proposes periodically discarding agent histories and restarting from the best solution found. x.com
• MINDSET - conversation memory stored as immutable episodes in versioned schemas, with gated supersession (preprint, unreplicated). arxiv.org
• MICA Flame-W preview - deterministic text from integer cellular-automaton rules, with no neural net at inference. producthunt.com
• Hivelane - routes tasks across coding agents, with a git worktree per worker and shared local memory. producthunt.com
• Hive - runs fleets of Claude Code, Codex, OpenCode, Kimi, Kiro or Cursor agents side by side. producthunt.com
• CrewTower Mod for Claude Code - a Mac notch panel to prompt, stop, compact and inspect Claude sessions. producthunt.com
• Navige - sends agent requests to smaller models, caches reads, trims tool output and enforces budgets. producthunt.com
• Lumovi - open-source Kubernetes triage that agents drive over MCP, asking before any change. producthunt.com
• Mantle - an agent that catches breaking API changes, starting with Stripe, and opens fix PRs. producthunt.com
• CmdRound - 55-minute Claude Code hiring tests with rule-based scoring, $49 for 10 candidates. producthunt.com
• Sphinx - a pre-ship hook that quizzes developers on their AI-generated code. sphinx-six-theta.vercel.app
• Stacks - sourced cards on how 94 engineers, including Willison and Ronacher, use coding agents. producthunt.com
• drobek - an MCP loop for agents to build, version and publish browser apps, with an AGPL core. producthunt.com
• Atticus - an open-source voice computer-use agent that runs on your Claude plan and confirms purchases first. producthunt.com
• Rushilo - a video editor driven by Claude Code or Codex, $99 one-time, with files kept local. producthunt.com
• siliXon - a design agent that turns a plain-English device description into a routed, manufacturable PCB. producthunt.com
• Scalattice - OpenAI-compatible inference on a distributed GPU network, with security tiers and regional policy. producthunt.com
• Replica - 11 free Claude skills that clone app features and turn user complaints into improvements. x.com
• EpochGo - local multi-agent AutoML in which Ollama LLMs orchestrate deterministic scikit-learn pipelines. reddit.com
• 17-model poker benchmark - cheap reasoning models match frontier ones, and a 30-line equity script beats the decision models. chrystianschutz.com
• Llmspeed - shows how fast a local LLM would run before you buy the hardware. llmspeed.dev
• Transformers from associative memory - derives the transformer architecture from key-value associative memory. alexwang.ai
• Pooled agents for open research - a platform where users pool their agents on open research problems. reddit.com
• Free plumbing for robot data - an AI lab trades free plumbing and HVAC repairs in three cities for robot training data. reddit.com | reddit.com
• $1,000-a-day Steam demo - a developer is going into debt paying about $1,000 a day to keep an AI driving-sim demo on Steam. reddit.com
• Code-switching ASR review - a 2026 LREC systematic review of code-switching speech recognition research. aclanthology.org
• Inline rendering with document.currentScript - Simon Willison on rendering output at the script tag's position. simonwillison.net
• web-edit 0.1a2 - a new alpha of Simon Willison's web-edit tool. simonwillison.net
• Pele - renders Mermaid diagrams as themeable SVG with better subgraph layout. pele.run

DeepSeek, Meta and StepFun each released a large sparse MoE with open weights, and the independent rankings moved with them.

DeepSeek V4.1 Flash claims the top open-weight coding score, under MIT DeepSeek released V4.1-Flash-1017 on Hugging Face under MIT with a technical report, claiming 77% on DeepSWE 1.1, ahead of GLM-5.3, Kimi K3, GPT-5.6 Sol and Opus 5 (vendor benchmark). Ollama already serves it from US hosting with zero data retention, which lets regulated teams use it without China-based endpoints. huggingface.co | x.com | i.redd.it | x.com | x.com

Meta open-sourced the model one step behind its best Meta's Muse Spark Open (816B total, 42B active) scores 39 on Artificial Analysis, the highest of any US open-weights model. It shipped three days after the stronger, closed Muse Wonder, so Meta now deliberately keeps its open tier one step behind its closed tier. x.com | interconnects.ai | x.com

StepFun ships a 600B MoE with a 1M-token context Step 5 Preview is an open-weight MoE with 600B total and 27B active parameters and a 1M-token context, aimed at coding, financial research and long-running agents. It works in vLLM, SGLang and Transformers from day one, and with only 27B active per token it costs about what a mid-size dense model does. producthunt.com

Compression and speculative decoding are moving capable models onto single consumer GPUs.

NVIDIA cut a 120B multimodal model to fit 24 GB Nemotron 3.5 Bolt compresses Nemotron 3.5 Super VL from 120B total and 12B active to 74B total and 9B active for a 24 GB budget, running 84 tokens per second on RTX Spark (vendor benchmark). Checkpoints, GGUF files and a speculative-decoding drafter are on Hugging Face, so text-and-vision agents over PHI can stay on one workstation. producthunt.com | x.com

16 GB now runs long-context 27B inference A user runs a 27B Qwen model at 262k context on one 16 GB AMD GPU at 42.7 tokens per second, using multi-token prediction plus n-gram speculation (single source). A separate test of nine local models on five coding tasks with hidden graders, on a 16 GB RX 9060 XT, offers a reusable way to pick a local coding model. reddit.com | reddit.com

The strongest agent launches check, audit or contain what coding agents did rather than help them write more.

Replay and attack testing after every AI change Kryptheon records a user flow once, replays it in a real browser after each agent change, and probes a copy of a Supabase or Postgres schema for data anyone can read or change, without touching real rows. It runs locally over MCP with Claude Code, Codex, Cursor, Antigravity and VS Code, which suits teams that cannot send data to a SaaS tester. producthunt.com

Generated UIs get their own evals RUI lists every button and form in an AI-generated interface and compares what each claims to do with what it actually does. No human reviews screens generated for each chat answer, so this is a new eval category. The same lab is open-sourcing Agentable UI, which shows agents a sortable table as a callable tool. producthunt.com

Agent session logs are a data-leak surface Context Trace scans local Codex and Claude Code session records for sensitive data. It shows redacted findings, where they appeared in messages or tool results, and gaps that limit the audit. Transcripts keep whatever tool calls returned, so on clinical systems they are a PHI store that most compliance reviews do not yet cover. producthunt.com

A compromised agent console drained provider keys A report says opencode's console was compromised, hitting 1.4% of console users with hundreds to thousands of dollars in charges on linked provider keys (single source). Hosted consoles that hold provider keys concentrate risk, which is the case for the local-first designs of Kryptheon, Context Trace and Lumovi. x.com

Offensive capability is getting public benchmarks while sites and platforms tighten what agents and crawlers may touch.

Offensive capability is becoming a published number A report says OpenAI models hacked a crypto wallet and breached more websites (single source). HackerOne is now benchmarking 18 frontier LLMs against 62 complex vulnerabilities, which turns offensive capability from anecdote into a number labs can be compared on. reddit.com | x.com

Sites block agents while Meta retreats on scraping terms Websites are increasingly blocking AI agents outright. Meta separately removed terms language that could have let it scrape news sites for AI training. Agent workflows that depend on public web access are getting less reliable, and access terms are becoming contested. theverge.com | theverge.com | techcrunch.com

Hugging Face trending can be bought AutoTrust's models reached Hugging Face trending on thousands of likes from bot-like accounts. Now that open checkpoints compete with closed ones, like counts and trending status say nothing about provenance or quality. reddit.com

20 items: open questions, opinion and meme posts, novelty demos, an unsourced Anthropic IPO rumor, and newsletter citations with no detail.