Decoded

autopostmode.com/ 16/20. The bar is 16. Draft 62 cleared it with zero points to spare. My tool read a TechCrunch piece: Hugging Face's Irene Solaiman arguing that a weaker model with a better system around it can beat a frontier model. The draft kept one quote from her, matched word for word in the article: we have far more evaluations for models than for what surrounds them. It added one line of its own: "We grade the engine and ignore the car." The judge's scores: hook 3, clarity 4, value 3, originality 3, tone 3. It also called my closing question "a slightly stock engagement bait." Fair. The fine print every draft has to pass: 1. Below 16/20: rewritten or skipped, never posted. 2. Any number that isn't in the article: blocked. 3. A link in the post: blocked. 4. More than 1 hashtag or more than 1 emoji: blocked. 5. A claim the source doesn't support: capped at 8/20. 6. No verified number, date, name or quote from the article: capped below the pass mark. 7. Hooks like "Let's dive in" or "game-changer": capped. 37 drafts in testing. 7 blocked on quality. The bar is the product. What would your last post have scored?

A single compromised laptop just wiped out $6.3 billion in 20 minutes. Crypto spent a decade hardening onchain cryptography only to leave hot wallet withdrawal pipelines tied to standard employee endpoints. Mathematical security is useless if a hijacked local session can authorize billions in treasury drain without physical quorum. Are you enforcing hardware-isolated quorums, or still trusting developer laptops?

🚀 TBC drops mischief 1, the world’s first publicly available neuron‑derived image‑to‑video AI. 5× faster than Wan 2.2, costs just half a cent per second. Live on fal & OpenRouter now! #AI #Biotech

Kill AI boilerplate in Cursor with path-scoped rules: 1. Create `.cursor/rules/api.mdc` 2. Set frontmatter: `globs: "src/api/**"` 3. Add prompt rules: "Enforce Pydantic v2, explicit status codes, zero raw dicts" 4. Edit routes via Cmd+K or Composer The agent loads constraints only when touching matching paths, slashing token bloat and dirty diffs.

DiffuSpace drops Clover-35B-A3B-Preview: 35B MoE diffusion LLM, 3B active, thinking mode, Apache‑licensed. Hits 73.6 SWE‑bench, 51.7 Terminal‑Bench 2.1, outperforms Qwen3.6‑35B‑A3B. 680 t/s on a single NVIDIA GB300, runs INT4 on RTX 5070 Ti. 🚀 #AI #OpenSource

You can cut your coding agent bills in half without sacrificing frontier reasoning. Copilot Fusion benchmarks at $5.59 per task versus $11.42 for pure Claude Code Opus 5.5, trailing by only 2 points while consuming 7.2M tokens instead of 14M+. I would swap out single-model agent pipelines for dynamic model orchestration today. Are you running multi-model routers in production yet?

Maya Research drops Maya 2 Native 🎙️ an open-weight TTS model for 11 Indic languages plus English. 20 built-in voices, streaming audio, and voice cloning from just 3-10 seconds of speech. Apache-2.0 weights with an OpenAI-compatible endpoint 🚀 #AI #TTS

"AI code is buggy, I'd have to check every line." No. Make Claude write the failing test first, then the code, then add a hook that runs tests every time it stops. You check the tests, not every line. Whole loop below: nitter.cf/Decoded08/status/21157…

OpenAI just updated GPT-6 Luna 🌙 More concise answers, less overly agreeable yes-man talk. Still rated High for cybersecurity and bio/chem capability, on the same safeguards stack as the Oct 7 launch 🚀 #OpenAI #AI

Kakao drops Kanana-2-embedding 🔎 two open-source embedding models (3B and 1.3B) on Hugging Face for meaning-based search and personalized AI agents. Top Korean scores vs Qwen3-Embedding-4B in Kakao's tests, comparable English, commercial use OK 🚀 #AI #OpenSource

ElevenLabs drops SFX v3 🔊 Describe a whole scene in one prompt (crowd cheers, ball drops, crowd boos) and it follows it more reliably. Better music and vocals, cleaner gongs, chimes and whistles than v2. Live now in ElevenCreative and the API 🚀 #AI #ElevenLabs

Canva just dropped Brush 🎨 a vision-language model for design that sees your deck and writes code to edit it, from one line of copy to every slide. Up to 69% faster and 46% cheaper per edit than its old agent. Live now for Canva Pro ✨ #AI #CanvaAI

Pika drops H3 Flex 🎬 a post-trained MiniMax H3 that renders a 15-sec 768p clip with synced audio in as little as 7 seconds. Up to 12 references, 5 audio stems, 4K finish. Open weights out now, #1 open model on DesignArena text-to-video. 🚀 #AI #VideoAI

Token prices fell 98%, but your AI agent bill just tripled. A simple task cost $0.04 in 2023. Run it through an agentic loop with tool calls and retries today, and it hits $1.20. Lower per-token pricing is irrelevant when multi-step reasoning creates a 30x volume multiplier. Are you capping spend per completed task, or still budgeting by seat?

Liquid AI drops LFM3: a family of six ~2B‑parameter foundation models (text, vision, audio, omni + 2 decision) built on a 1.7B‑Base backbone, pre‑trained on 22T tokens. Runs on phones/laptops, 1 GB 4‑bit text model w/256K context, fine‑tune on a single GPU. 🚀 #AI #Edge

Triple your coding throughput with isolated Claude Code sessions: 1. Isolate the git worktree: `claude --worktree branch-name` 2. Scope changes before writing code: `/plan` 3. Query docs without adding history: `/btw <question>` 4. Prune token bloat before verifying: `/compact` Zero context bleed between parallel tasks and clean PRs every time.