AI News 2026-09-20

$ cat top-story.md

Laya, an Open-Source 421M Model, Dethrones the Buzzy "Jev" Decision Engine

A solo developer at Convai Innovations released Laya, an Apache-2.0 "System 1" decision model (322M–421M parameters, built on ModernBERT/mmBERT) that returns calibrated probability scores instead of generated text. Trained via reinforcement learning against proper scoring rules, it beats TypeSafe AI's much-hyped Jev on both latency (32.8ms vs. 6–8x slower) and cost (free/self-hosted vs. $0.042 per million input tokens) while remaining fully open. The story shot to the top of Hacker News just days after TechCrunch revealed Jev's creator is a former OpenAI ChatGPT/RLHF researcher.

Why it matters: If you route agent decisions, guardrails, or classification steps outside your main LLM call, this is a faster, free, fully self-hostable alternative you can drop straight into a vLLM-style stack next to Gemma.

// source: Hacker News / Convai Innovations · additional source: TechCrunch

## 📦 Models & Releases

▸ GPT-6 Astra Cracks a 108-Year-Old WWI Cipher

OpenAI's GPT-6 Astra decrypted a November 1918 German ADFGVX-cipher radio message that had resisted cryptanalysis for over a century, then cross-checked the decoded troop-movement intelligence against historical naval records confirming HMS Canterbury's arrival at Sevastopol. Why it matters: A concrete, verifiable demo of long-horizon reasoning chained with tool use and historical-record verification. // prinzai.com / Hacker News

## 🛠️ Tools & Repos

▸ Cua Reaches "Computer-Use 2.0" for Cross-OS Agent Fleets

An open-source platform giving agents native control of macOS/Windows/Linux desktops via cloud "Fleets," a local VM manager (Lume), a small specialist model (CUA-S1) for UI decisions, and a benchmark suite — now at 24.6k stars and climbing fast on GitHub Trending. Why it matters: The most active open framework for moving Claude Code-style agents beyond the terminal into full GUI automation. // GitHub Trending

▸ Hister Turns Your Browsing History Into a Private, MCP-Queryable Search Engine

Hister indexes your visited pages, bookmarks and local files into a full-text (and optionally semantic) search index you fully control, queryable via web UI, CLI, or directly by AI assistants over MCP — now at 5.3k stars. Why it matters: A ready-made, self-hosted "personal knowledge index" your agents can query over MCP without shipping your browsing history to a third party. // GitHub Trending

▸ ENZO Packages a Self-Hosted, BYOK AI Workspace With Agents and Skills

A new Show HN launch: a Docker-deployable AI hub with 300+ models across 9 providers, 74 bundled skills, a plain-language agent builder, and client-side AES-256-GCM-sealed API keys — no vendor markup, no subscription. Why it matters: A concrete, inspectable alternative to hosted agent platforms for keeping Claude-Code-adjacent workflows fully under your own infrastructure. // Hacker News (Show HN)

## 🤖 Agents & MCP

▸ ZCode Caught Silently Uploading Your Entire .git History

A developer reverse-engineered Zhipu's ZCode coding agent and found it packages a full workspace snapshot — commit history, reflogs, LFS caches, up to 313MB and 86% .git content — encrypts it with a server-supplied key it can't decrypt locally, and uploads it, with no UI toggle able to disable it. Why it matters: A sharp reminder to audit exactly what any coding agent exfiltrates from your repo before trusting it with proprietary code. // ferstar's blog / Hacker News

▸ GitLab 19.4 Adds Permission-Tiered Governance for MCP Tools

GitLab now lets teams govern third-party MCP tools alongside its own Duo Agent Platform tools — read-only tools default to "Always Allow," write/delete tools default to "Always Ask" — plus new MCP tools for repo browsing, issue/epic triage, CI/CD control, and autonomous MR merges via a new /goal command. Why it matters: A concrete blueprint for permission-tiered MCP tool governance you can apply to Claude Code or your own MCP servers. // GitLab release notes

▸ Study: Claude Agent Skills Cost 2.3x Less Context Than Equivalent MCP Tools

Tokenizing 6,434 public SKILL.md files and 108 MCP tool definitions found a median always-on cost of 78 tokens for a skill vs. 177.5 tokens for an MCP tool — 85% of a tool's cost comes from its JSON input schema — and that a third of real-world skills already exceed Anthropic's ~100-token guidance. Why it matters: Hard numbers for deciding whether to expose a capability as a Skill or an MCP tool when trying to keep your own Claude Code context overhead down. // Towards AI

▸ "Brood War Bench" Pits Claude, Codex and Grok Against Each Other in StarCraft

A round-robin tournament runs 19 model configurations (Claude Opus/Sonnet/Fable/Haiku, Codex, Grok and others) against each other in StarCraft: Brood War under identical harnesses — Codex Astra won every match at its hardest setting, Grok models won barely 5-11%, and even the leaders couldn't build a real army or hold off a simple attack. Why it matters: A genuinely novel long-horizon, real-time agentic benchmark showing which model families hold up under continuous multi-step decisions versus one-shot coding tasks. // Hacker News / swerdlow.dev

## 🏠 Self-Hosting & RAG

▸ Bonsai 2 27B Squeezes a Qwen3.8 Model Into 5.9GB With Near-Lossless Quality

PrismML applied ternary (-1/0/+1) weight quantization with FP16 group-wise scaling to Qwen3.8-27B, shrinking it 9x to 5.9GB (1.76 effective bits/weight) while retaining 98.2% of aggregate benchmark score (83.9 vs. 85.4) — Apache 2.0, 262K context, runs via CUDA or MLX. Why it matters: A 27B-class model that fits alongside Gemma on a single consumer GPU with headroom left for Qdrant and embeddings. // PrismML / Hacker News

▸ vLLM Fixes Watermarking So It Finally Works With Speculative Decoding

A newly merged vLLM PR uses separate watermark keys for draft vs. target tokens plus context deduplication, closing a long-standing gap where native text watermarking broke when combined with speculative decoding, while keeping acceptance rates around 80%. Why it matters: If you use speculative decoding to speed up Gemma serving in vLLM and need to prove provenance of generated text, this closes a real compliance gap. // vLLM GitHub

## 🎬 Worth Watching

▸ Two Minute Papers — "DeepSeek's Insane New Architecture"

A walkthrough of DeepSeek's new Manifold-Constrained Hyper-Connections (mHC), which constrains how residual streams interact to make large-model training dramatically more stable and compute-efficient. Why it matters: If mHC scales as claimed, it could meaningfully lower the cost of training capable open-weight models worth self-hosting. // Two Minute Papers

▸ AI Explained — "What AI Researchers Saw, Before Their Demand to 'Pace' AI"

A breakdown of the internal research signals that reportedly pushed Dario Amodei and others to publicly call for slowing frontier AI capability growth. Why it matters: Useful context behind the ongoing "pace the frontier" debate for anyone following AI-safety developments beyond headlines. // AI Explained

## 📰 Worth Reading

▸ Microsoft Exec Called AI Scraping "the Largest Theft of Labor in Human History"

Newly unsealed NYT v. OpenAI/Microsoft filings show a Microsoft director privately calling AI training-data scraping "theft," while Microsoft's own data reportedly showed Copilot cutting click-throughs to the NYT by up to 93%. Why it matters: A rare internal admission from inside a major AI vendor that contradicts the industry's public "fair use" framing. // TechCrunch

▸ Simon Willison — "Probably Gonna Eat You"

A short note arguing that computer scientists who dismiss LLMs as uninteresting in 2026 resemble geneticists shrugging off Jurassic Park — a reckless but undeniably field-defining moment. Why it matters: A concise, credible-insider framing of why continued LLM skepticism among technologists is increasingly untenable. // simonwillison.net

AI News — Sunday, September 20, 2026