All Stories

Arohan shares a two-model distributed training idea that nearly doubled speed

Tencent Hunyuan says Hy3 reached number one on the OpenRouter leaderboard

DAIR.AI rounds up HOLA, ReContext, Always-On Agents, and verifier papers

DeepSeek hiring post details RL environment roles for code, cyber, and science

Huawei UnifiedBus analysis frames 1,024 Ascend NPUs as one memory domain

Chris Hillee explains why LayerNorm is not free versus RMSNorm

Jianlin Si shares Taylor expansion notes for LogSumExp and Softmax

Liam Fedus says failed experiments are durable training data

Ben Burtenshaw proposes EuroEnv RL tasks for European business workflows

Gabri Berton warns token scaling needs diverse task data

Prime Intellect Verifiers v1 gets a deep dive and vLLM rollout support

Sebastien Bubeck reports GPT-5.6 lowered a cycle double cover bound to 2.26

Fable 5 safety filter switches a macOS VNC automation task to Opus 4.8

OpenAI offers ChatGPT Classic after the Codex desktop app merge

Garfield review loop fans diffs through no-edit agent reviewers

Merve Noyan builds an MMMU-Pro-like Turkish vision-language test suite

Kyle Jeong explains the Great Firewall with interactive network diagrams

Grok Build wire analysis says the CLI uploads full repos and .env contents

Riley Goodside tests Fable and GPT-5.6 on verified constrained generation

LessWrong proposal filters scheming data out of AI training sets

Ethan Mollick compares METR AI task horizons to solar forecast misses

Zeeg says skills are repeatability primitives for efficient agent work

Dex Horthy says AI code cleanup needs program design, not just architecture

Claude Code v2 adds background subagents and shell commands

Qwen billboard claims 1 billion downloads for open-source models

Carnegie Mellon argues pretraining still lacks an essential ingredient

Multi-agent ProgramBench shows five agents reach 60 percent much faster than one

OpenBMB's Whispera runs a fully local real-time voice assistant on Windows

OpenAI says physicians found fewer flaws in GPT-5.6 answers

Doodlestein's fresh-eyes prompt catches bugs before implementation starts

boristane maps cloud resources into a live agent context graph

pvncher says skill files and agent Markdown need audits after every release

vLLM merges Cohere's hardware-aware Dynamic Speculative Decoding

MiniMax details Fireworks M3 sparse-attention kernel on Blackwell

Grok 4.5 reportedly finds a dimension-4 hypercontractivity counterexample

GPT-5.6 Sol Ultra produces a public proof for Cycle Double Cover

Ethan Mollick warns frontier models need role-specific evals

OpenArt Director turns natural-language briefs into cinematic storyboards

Sergey Levine shares a data-driven decision-making taxonomy for offline RL

LangChain bundles Nemotron Deep Agents with OpenWiki memory updates

CMU paper explains SFT and RL as compositional reasoning modules

Pieter Abbeel highlights ABC Box as a robot setup for ABC data infrastructure

Google Research says Willow demonstrates verifiable quantum advantage

Moore Threads trains MusaCoder on MTT S5000 accelerators

Hermes Desktop auto-discovers Hermes Cloud agents through Nous Portal

ComfyUI shares a reusable Seedance 2.0 product-ad workflow

fal adds Krea 2 Turbo Style Reference with no per-style LoRA training

ElevenLabs builds a bilingual Alpha Bank voice agent

Conductor uses Fable and Codex computer use to port its Mac app to Linux

Claude Code Desktop adds a sandboxed in-app browser

Perplexity Computer Analytics tracks credit spend across models

Uncle Bob forces QA agents through UI events instead of APIs

Jon Durbin outlines a verified graph pipeline for synthetic training data

Apple sues OpenAI over alleged hardware trade-secret theft

OpenAI Bio Bug Bounty becomes a private program with $50K rewards

Factory launches Droid incident response for Slack alerts

NotebookLM highlights source-centered UX gaps in ChatGPT Work

Cursor adds durable side chats and local agent transcript search

Exa and Cerebras pair search APIs with high-speed inference

Google AI Studio adds custom ai.studio URLs for deployed apps

Epoch AI estimates Codex contributors show measurable AI uplift

Matt Pocock says /teach works unchanged inside Claude Cowork

Paperclip adds /wireframe skill for shareable feature mockups

Zed previews DeltaDB collaborative threads as Google Docs for code

OpenAI announces Build Week with a GPT-5.6 and Codex global challenge

Perplexity previews a GLM router that escalates to frontier models

Arav Srinivas frames local harnesses as operating systems for frontier tokens

Jimeng surfaces Seedance 2.5 Ultra-Long Video mode for 180-second clips

Netflix releases video datasets and models on Hugging Face

Jürgen Schmidhuber rebuts LeCun on neural world model peer review

TRACE trains agents to identify and repair their own failure modes

AI/ML API adds GPT-5.6 models to its single-key model aggregator

Matt Pocock diagrams agents as models plus harnesses plus environments

OpenAI seeks design partners for a GPT-Live API

Avichawla builds a four-agent software team with InsForge and Hermes Kanban

OpenCode sets up a Slack World Cup bot without a separate app

Google Gemini Live connects camera prompts with Maps and image generation

Codex backend reasoning summaries surface concise step labels in agent runs

Anthropic appoints Ben Bernanke to its Long-Term Benefit Trust

Jared Palmer compares Bun’s AI rewrite with TurboRepo’s 14-month Rust rewrite

Ben Hylak shows how to avoid scroll-jacking while keeping transitions

Richard Ngo publishes Selective Optimism critique of the AI 2040 scenario

Entire comes out of stealth as an AI-native GitHub alternative

dejavucoder uses Codex to find a 212x faster QR kernel

Seedream 5.0 Pro reaches fal, ComfyUI and Venice image workflows

Sakana Fugu gets a DeepLearning.AI model orchestration deep dive

vLLM reaches Transformers backend parity for 450 architectures

Robbyant opens LingBot-Vision weights, code and paper

Artificial Analysis benchmarks Nano Banana 2 Lite image quality

ZML releases an inference engine integrated with Hugging Face

Epoch AI scores GLM-5.2 highest among open-weight models

fal explains 1,000 tok/s prompt expansion for image generation

Qualcomm CEO forecasts a 40x jump in global token demand

Wafer explains Pallas kernels and TPU execution semantics

LangChain and NVIDIA work on NemoClaw Deep Agents Blueprint

Together adds provisioned throughput for MiniMax M3 and GLM-5.2

Runway opens a developer platform for media model inference

GLM-Image report highlights Ascend training and Apple runtime

Prime Intellect raises $130M for an open superintelligence stack

RynnWorld-4D paper targets 4D embodied world models for robot manipulation

Ben Burtenshaw argues developers should learn on-device model primitives

Google DeepMind A3C paper wins ICML 2026 Test of Time Award

Matthew Berman breaks down Anthropic J-space interpretability claims

Kling launches MCP and CLI access for agent video workflows

Claude Tag walkthrough covers multiplayer steering and proactive channel work

Peter Steinberger exposes macOS ComputerUseAllowForbiddenTargets flag

Oxford paper synthesizes six recurring LLM agent failure modes

Hugging Face and SkyPilot make private storage cloud-agnostic

Sakana AI highlights Lilian Weng's harness engineering path to recursive self-improvement

trq shares a lightweight video editing harness for agent-generated talks

Perplexity says Vera CPUs power its computer-use sandbox runtime

Reuters reports China is weighing overseas restrictions on top AI models

NVIDIA Puzzle-75B-A9B doubles hybrid MoE server throughput

Google Research demos LLM translation on a low-power Coral NPU

Yacine ports a rigid-body physics engine to CUDA for real-time cube simulation

Elie Bakouch maps J-lens layer geometry across 38 open models

Cohere releases Transcribe Arabic as an Apache 2.0 speech model

Elie Bakouch estimates Gemma 4 used far more pretraining tokens than prior releases

Unsloth ships DeepSeek-V4-Flash GGUFs for local llama.cpp inference

Reuters says DeepSeek is developing an inference chip

MetaSkill-Evolve paper evolves both agent skills and the skill-improvement loop

Thomas Wolf shows 20-plus agents building an RL and LLM research wiki

Jürgen Schmidhuber traces neural network distillation back to his 1991 work

TechHalla shares a Magnific workflow for SnorriCam video from one image

Hugging Face runs a live tutorial on model distillation for custom agents

Norm Ai raises $120M while expanding agentic law and supervisory agents

xAI Console adds 25-plus multilingual Grok voices for testing

Charles Wolfe explains why agentic RL changes GRPO advantage normalization

LangChain Academy launches Introduction to Deep Agents

ComfyUI demos an automated video production pipeline through Comfy MCP

Ramp for Agents lets agents incorporate, open bank accounts, and create virtual cards

ai_for_success cuts a Gemini Live desktop assistant bill from $1,440 to $255 per month

Vercel brings Better Auth team in to build open auth for apps and agents

AppLess streams OpenUI screens from Gemma 4 on Cerebras at interactive speed

Vercel Sandbox adds CPU, memory, and transfer observability for agent workloads

Claude Code 2.1.203 and 2.1.204 fix background sessions, hooks, and CPU regressions

Omnigent introduces stateful contextual policies for coding-agent sessions

Piotr Skalski finds Fable 5 weak on cluttered object detection versus Gemini

Anthropic subscription flow starts asking some users for government ID verification

Pauliusztin maps three GraphRAG retrieval modes for agent memory

Google DeepMind adds Predicting the Past Skill to Antigravity

VS Code shows how Tasks can run repeated project commands from the editor

OpenAI shows Codex workflows for cell analysis and immune-cell simulation

FrankenOCR expands into a pure Rust local OCR engine for agents

GitHub issue image uploads need an API path for agent QA, Zeeg says

Interfaces publishes Typography Manual for the Web with tooling resources

Media Synthesis Museum preserves legacy generative AI models on Hugging Face

FastVideo releases FP4 attention kernels for NVIDIA B300 with 1.69x speedup

Parallax ports MiniMax MSA kernels to RTX Pro 6000 with 7x training gains

GitHub Copilot subscriptions add Kimi K2.7 Code as an open-weight option

Sber releases GigaChat 3.5 Ultra, a 432B MoE with MIT-licensed weights

Google Research demos Vision Banana for segmentation, depth, and normals

GPT-5.6 Sol leak claims Cerebras serving at up to 750 tokens per second

OpenAI confirms GPT-5.6 Sol Ultra will be available in Codex

SGLang adds DSpark variable-length verification for speculative decoding

Wafer argues six NVIDIA challengers converge on software-managed SRAM

Hyperbrowser open-sources Agentmap to pre-map websites for agents

LingBot-Depth 2.0 cuts depth error in half for transparent-object robotics

OpenCode 2.0 is being designed for durable execution systems

Google DeepMind and Apptronik expand Gemini Robotics data collection

Eve adds first-class evals that assert agent actions and tools

VS Code explains the Agent Host Protocol in an Insiders podcast

Google Gemini Spark adds real-time topic tracking and triggered emails

Claude Code cleanup finds 306 orphaned MongoDB MCP servers using 35 GB RAM

BlockSearch studies LM retrieval at million-token corpus scale

fofrAI uses a Gemini agent to rediscover a candidate 13.4B-year-old galaxy

prinzbench shows Fable 5 lags GPT-5.5 on legal search but matches legal analysis

ReContext replays query-relevant evidence inside 128K-token prompts

Hugging Face says Xet storage reached 125.6 PB physical data

Codex browser control cancels Navan bookings and escalates refunds

Sakana AI releases Sakana Translate for Japanese, English, and Chinese

Nature Health features Mustafa Suleyman's paper on public LLM health queries