All Stories

OpenPipe ART uses plain-English RULER criteria to train RL agents

vLLM boosts Kimi K3 throughput up to 2.8x on B300

Grok 4.7 trails frontier models on KernelBench-Mega

Unsloth Docker adds AMD support and GUI workflows for 500-plus models

Dwarkesh's multi-agent podcast examines research automation and alignment

GPT-6 Astra deciphers a 1918 German radio message

Andrew Carr reconstructs people and objects in 3D from ordinary video

AppLess open-sources a no-app phone stack powered by OUI-1

Sakana AI launches Frontier Intelligence Group for alternative AI paradigms

Cartesia launches Multilingual Voices for consistent brand identity

OpenAI adds a GPT-Live-1 voice agent to Codex

Shopify open-sources Roast for structured AI workflows

Eve adds Jev-powered automatic approvals for shell tools

SkillRanker uses Jev to rank the next agent skill from live context

Hermes Agent moves bundled memory providers into independent plugins

Claude Projects turns one conversation into parallel cloud work

HumanLayer open-sources visual-pr skill for concise pull requests

12-Factor Agents guide treats LLM workflows as structured software

Grok Bot reduces subagent chatter for 10% more effective usage

OpenCode adds /btw for noninterrupting status checks

Agora uses Git commits as shared memory for 13 auto-research agents

Anthropic opens Life Sciences Verification Program for biology teams

Zed 1.20 adds animated cursors and rendered Markdown previews

GPT-6 Astra completes Factorio: Space Age after 165 hours

Devin tests whether an autonomous agent can run a business

Google Research generates guarded interactive STEM simulations for teachers

Runway adds Enhance Frame Rate for generated video up to 120 fps

Mio launches a proactive team agent for Slack, Notion, and Linear

Polyphron uses reinforcement learning to grow transplantable tissue

GPT Image 2.5 and Seedance 2.5 lock reference frames for coherent video

macOS Orbs run virtual M4 Pro machines with 14 GB of RAM

Google's Fuse benchmark tests social reasoning under user framing

Gemini Canvas turns parameterized vase designs into printable STL files

Maxime Labonne shows how a 230M model can reach 2,000 Elo

GLM probes surface tokenizer matches and role-confusion behavior

Periodic Neon tests held-out chemical-system generalization

Grace et al.'s AI forecast survey puts publishable math proofs around 2039

OpenRouter makes Union Alpha free for multimodal agent coding

TypeSafe's Jev beats Gemini Flash Lite on classifier accuracy and latency

Unsloth Studio fine-tunes 500-plus open models in free Colab

Hermes Agent opens a reviewed catalog with 100 plugins

Harbor documents user simulation and streaming for agent evals

Pi's next harness targets durable tasks and pluggable runtimes

NVIDIA finds homogeneous model pools beat diverse agent ensembles

Instinct requests broad Gmail, Drive, Docs, and Calendar access

Astra VLA zero-shot arm demo exposes reliability gaps on SO-101

Claude Code 2.1.274 repairs MCP timeouts and transcript loops

Devin's Code Scans use Agentic MapReduce for whole-repo audits

OpenAI previews Codex Replay for parallel historical task comparisons

Vercel makes secure build microVMs 64% faster to start

DeepSeek V4.1 Flash exposes gaps in evaluation and output legibility

Ramp data points to rising spend on AI security software

Microsoft demonstrates capability laundering through aligned LLMs

Opik maps RAG traces from retrieval through generation

OpenAI's agentic software factory routes Codex through risk-based review

Opus builds a playable Roblox game in a 15M-token test

Safari 27 uses JSPI to let native WebAssembly code await Promises

Higgsfield and GPT-6 Astra build a 15-material Blender shader add-on

ACGVF and HHMMs unify agent-controlled goals in hierarchical RL

OpenRouter sees OpenAI reclaim weekly wallet share from Anthropic

Fable's cost breaks today's agent orchestrators, Yegge argues

Agent Plugins standard packages skills and MCP servers across clients

Computer-use benchmark scores can hide model strategy, Rasbt shows

Eval Engineering Skill turns production traces into agent tests

Codex falls back to computer use when Slack blocks a channel search

ElevenLabs' Dom qualifies leads in four minutes and generates $1M pipeline

Rogue agents reportedly breached an OpenAI supercomputer

GPT-6 Astra stalls after a creeper wipes its Minecraft progress

MCP replaces CLIs for many model-driven integrations as tool calling improves

ARTEMIS automates mobile workflows through coding-agent MCP

AOA framework structures Claude skills before business automation

Pvncher reports extra subagents burn tokens without quality gains

Spectrum UI offers 250 animated React components through an MCP server

Agent cyberattacks prompt trace-sharing rules from Clement Delangue

GPT-5.6 Sol reports a database lock convoy resolved after deployment

Meta's alignment plan emphasizes independent evaluators and serving compute

Sakana Marlin adds interactive source reading and PowerPoint export

MiniMax H3 Max Multi-Angle adds camera control from one reference image

Spellbook reports 48% partial lawyer adoption in ILTA survey

Factory raises $200M at $5B to scale self-improving software development

Vercel Labs opens a public arm for software research and experiments

DeepSeek kernel engineer describes becoming a mecha pilot for agents

SGLang and MiniMax push H3 video past 2x real time on B200s

Tinker uses physics verifiers to train models for transformer design

Nomic AI receives a pitch for filtered frontier-model reasoning traces

Yann LeCun argues reasoning should search in abstract representations

vLLM adds day-zero support for Intern-S2-397B scientific agents

AI Toolkit adds experimental LoRA training for YuE2-3B

Cognichip launches ACI Enterprise for AI-assisted chip design

DeepSeek kernel engineer says AI is nearing expert CUDA optimization

Gemini 4 Pro argon checkpoint reportedly tests 256K-token outputs

GPT-6 Astra and Fable test competing Linear A translations

OpenAI reportedly acquires Glass Imaging for neural camera technology

Williams92-style GRPO batching reduces reinforcement-learning stragglers

Linear Loops automate project updates, document edits, and Slack notices

Claude Siri bridge routes macOS Spotlight prompts through Claude Code

Omar Sar0's ReAct guide decomposes a minimal agent harness

Emulate v0.11.2 expands CI emulators for OAuth, S3, Slack, and more

fx 0.0.10 adds auto-upgrade resume and active-turn steering

Claude Code 2.1.271 adds domain-scoped Bash access and policy fixes

Cloudflare security-audit-skill runs six-phase verified agent audits

Vercel agent-browser tests deployments behind Deployment Protection

GPT-6 Astra cheats in all ten exposed-engine honeypot games

OpenAI researcher Dan Selsam publishes a personal AI risk statement

AI Engineer World's Fair collects production lessons on agent harnesses

Anthropic scales CI after Claude writes 80% of its code

Amazon GAUGE finds LLM judges mis-rank close agent pairs 31% of the time

Cohere calls for evidence-based AI standards instead of a lab cartel

Turbopuffer makes native embeddings generally available