All Stories

KAT-Coder-V2.5-Dev goes open weight with strong coding scores

OpenRouter routes 250T tokens a month as usage keeps climbing

Computational identifiability paper shifts math research toward search under constraints

Interconnects tracks 4.3B open-model downloads and regional growth

OpenRouter's fastest-model chart shows big throughput and price gaps

Together AI plans day-zero hosting for Kimi K3

Kimi K3 spends 10x more thinking tokens and wins DesignArena

Ling-3.0-flash goes free across Nous, OpenRouter, and OpenCode

Anthropic's Claude Opus 5 looks imminent after preview leaks

OpenAI's image API needs a 2048 pixel cap to avoid bad box outputs

NVIDIA's Hugging Face lead reflects how open-source AI has gone mainstream

Sakana Fugu-Ultra v1.1 gains 7.9 points without a price bump

Harness Handbook maps agent behavior back to source code

Vercel's AI Gateway adds live streaming transcription

Offloop's D1 dispatcher cuts multi-agent noise and beats baselines

Omar Sar says dynamic workflows generalize harnesses, loops, routing, and graphs

Sabrina Aquino maps the next stage of loop engineering and self-improving agents

Corey Ganim maps how to productize a managed AI service

Anthropic upgrades Claude Voice with Opus, Sonnet, and tool access

SkilledProposer helps DSPy generate less overfit custom instructions

Goodfire's EVEE gets validation from FinnGen and real clinical data

Lovable earns AIUC-1 certification after security stress tests

Next.js adds a security-only RSS feed for advisories and patches

UK AISI says Kimi K3 trails frontier models on cyber exploit chains

RedNote's dots-note-3.0 gets a perfect IMO score with agentic proofs

OpenAI cyber models chain zero-days in Hugging Face production

Qwen 3.8 Max builds three Three.js scenes faster than rival models

Codex configures a camera over USB by driving gphoto2

Cognition acquires Interaction, makers of Poke

React Doctor crosses 1M weekly installs

Hermes adds an offline tldraw skill for live canvas editing

OpenAI rolls Health in ChatGPT out to U.S. users with connected records

Gigatoken tokenizes internet-scale corpora at GB per second on one machine

Manus adds Plan Mode to review specs before agent builds begin

Kimi Code adds GUI-terminal switching without losing conversation context

Rork Max lets developers build and install iOS apps from an iPhone

OpenAI Devs posts Codex Live Build session with covacut and davis7

Solar Open2 250B lands on Hugging Face with 1M context and vLLM serving

Relay routes coding work across planner, executor, and reviewer models

Self-modifying Lean proof agents reach 45.1 percent on miniF2F

Miles adds On-Policy Distillation and reports 3x shorter Qwen reasoning

Google reports roughly 22B AI tokens per minute and 950M Gemini app MAUs

Kimi K3 throughput recovers to 23 tokens per second on OpenRouter

Vercel explains how AI Gateway runs on Fluid compute

Devin workflow claims multiple long-open graph conjecture proofs and refutations

Google commits $40M in AI tokens and cloud credits to DOE Genesis Mission

Flue previews composable agents defined as code with hooks

Eve extensions package tools, connections, skills, and instructions for agents

Gemini Managed Agents switches default runtime to Gemini 3.6 Flash

TestingCatalog spots Claude Voice routing to Opus and Sonnet models

browser-use launches game mode for agent-built 2D and 3D web games

Simon Willison analyzes OpenAI's accidental Hugging Face cyber incident

Ben Hylak shows MCP custom UI working inside Codex and Claude

Artificial Analysis benchmarks Inkling on AA-Briefcase agentic knowledge work

Claude Security enters beta as a Claude Code vulnerability scanner

Structured output paper finds JSON mode collapses answer diversity

Sentry Seer analyzes an error, drafts a fix, and opens a pull request

Paulius Zilinskas outlines unified agent memory with MongoDB and FastMCP

Pi OpenAI compaction extension brings Codex-style long-session state

diffshub adds private GitHub PR diff rendering with local PAT storage

BusinessCaseBench tests frontier models on 615 MBA case questions

Anthropic Economic Index connector lets Claude query AI labor-use data

fal shows Claude Code and GenMedia CLI scoring AI-generated film clips

swyx urges engineers to separate control, data and management planes

Pragmatic Engineer clip explains Dex Horthy's loop engineering pattern

skeptrune uses Claude to add 1,600 tests and fix 160 client bugs

Codex WSL setup tip requires the system install inside WSL

Sirupsen latency table recaps storage, memory, syscall and network costs

Claude Code desktop adds iOS Simulator support for app iteration

sqs suggests asking coding agents to fuzz and fix similar bug patterns

OpenAI publishes Contrastive SDF for measuring reward-seeking behavior

Cisco Foundation AI open-sources bug-hunting security models

Daniel Han publishes Unsloth workshop on RL, reward hacking and kernels

BytePlus Seed Audio 1.0 adds timestamped prompts and 20 plus languages

Sakana AI proposes UnMaskFork for test-time scaling of diffusion LMs

Vercel AI Gateway adds priority and flex service tiers

QuixiCore runs Qwen3.6 35B at 100 tok/s on Intel B60 XPUs

BlackRock's Larry Fink argues China leads the AI energy buildout

teortaxes revisits China's 14nm chip path for large AI compute runs

Nous Portal extends free Hy3 access for cost-effective agentic coding

OpenCode reports 212T routed tokens in 30 days before gateway launch

vLLM adds day-zero serving support for NVIDIA Cosmos 3 Edge

LangChain case study says Apollo cut AI skill launch work by 80 percent

ChatGPT adds Ask User Input tool for interactive clarification questions

Flue sketches a React-style component DSL for building agents with hooks

Omar Sar shares reusable multimodal skill pattern for agent orchestrators

World Labs robotics hire notes realtime policy debugging challenges

freecode CLI skill lets agents mint repo identity keys via GitHub OIDC

Render plans to remove GitHub from its critical deployment path

Will Depue shows frontier models decoding keyboard-shifted gibberish

Raindrop surfaces unknown agent failures through semantic clustering

Claude Code 2.1.217 caps subagent fan-out and fixes MCP memory leak

Vercel caches immutable static assets across deployments on its CDN

OpenAI will brief US officials on its upcoming workplace model family

ChatGPT Sites reaches EEA, UK and Switzerland with Sign in with ChatGPT demo

ComfyUI showcases asset-collection workflow for coherent AI video edits

Gemini Live adds real-time camera help for whatever you are looking at

Anthropic's J-space paper maps verbalizable reasoning inside LLMs

Natolambert's RLHF lecture recaps preference data and open questions

Clement Delangue says open-source AI is the defense, not the risk

Unsloth adds AMD GPU support for 500-plus models

IMO 2026 runs show frontier models clearing all six problems

Parameter-efficient fine-tuning may give formal generalization guarantees

Anthropic's Penlight aims to help clinicians during patient visits

Claude Code 2.1.216 tightens sandboxing and fixes long-session stalls

Cameron Wolfe says world modeling gives agents denser supervision

LangSmith opens free sandboxes for trying agent workflows

Claude Conway access notice points to a staged release or shutdown

Paperclip Skills Studio gives teams a place to prototype agent skills

Anthropic's $1.5B authors settlement keeps the fair-use win but raises provenance risk

OpenAI's unreleased model reportedly escapes its sandbox in evals

Kimi K3 climbs ReactBench and DesignArena for frontend work

Recursive self-improvement survey says evaluators are the bottleneck

Creative writing benchmark shows Fable still leads but Kimi is close

Pi exposes AGENTS.md and SYSTEM.md for direct context control

OffDeal engineers shadow bankers and turn the workflow into automation

Google Gemini Notebook adds Collections for themed notebook folders

Pi hosts badlogicgames and mitsuhiko on malleable, open software tools

The Turing Post compares 10 small language models for 2026 local AI

Yacine prompts GPT-5.6 Sol to model a custom CAD funnel with annotations

DeepSeek V4 Pro API draws community routing concerns after demo outputs

Anthropic-linked screenshots claim an LLM found a Jacobian conjecture counterexample

Vercel AI Gateway offers one API for multi-model routing and failovers

Hugging Face schedules local AI session on hardware setups and compression

Vipul Ved Prakash argues open weights shift spend to inference, data and apps

François Chollet predicts cheap AI through a new symbolic learning paradigm

Timothy Gowers says the 2030 Fields Medal may be the last human one

OpenAI Codex models.json PR surfaces another ultrafast mode hint

Charles Wolfe previews a 12.6K-word deep dive on agentic world models

BrainCo demonstrates a fluid bionic hand prototype at WAIC 2026

Aaron Levie says applied AI is gated by real-world feedback loops

DeepSeek V4 harness is expected to ship with the official model release

Riley Goodside tests Fable on a contextually appropriate perfect anagram

Clement Delangue says defender workflows can be blocked by AI guardrails

Codex controls DaVinci Resolve for GPT-5.6 Sol video editing demo

Dan Shipper says AI now automates about 70 percent of Every copy edits

doodlestein runs AI agent planning from a phone using markdown plans

Jack Clark says Fable gave a useful fiction-editing deletion suggestion

Flowith shares Seedream and Seedance prompts for a miniature BBQ video

Self-Improvements survey maps how modern AI agents learn from experience

Avi Chawla visualizes RLHF, DPO, and GRPO alignment pipelines

Elie Bakouch compares Laguna XS.2.1 J-space shifts and optimizer CKA

Google Gemma hackathons expand across India, Uganda, US, and more

Nick Dobos says filler tokens can buy extra inference-time thinking

Gemini 3.6 Flash leak points to late-July Google release

Riley Goodside renders all 1,025 Pokémon as a Klein bottle crossword

Andrew Carr ports a TensorFlow face sampler to torch with parameter fitting

Graham Neubig revisits Hamming's open-door research lesson for AI labs

Open Weights Meetup and March opens RSVPs for July 25

Nirant Kasliwal maps inference gateways, routers, and router-enabled products

Josh Achiam separates open-weight effects on inference and training compute

Sebastian Raschka explains low, medium, and high reasoning effort in LLMs

GPT-5.6 Sol on Cerebras is shown generating at 750 tokens per second

Aaron Levie maps the emerging enterprise AI stack beyond frontier labs

x-agent-intelligence skill builds scheduled AI news feeds from X MCP

Whisper Rail concept sketches a push-to-talk finger microphone for agents

Dotta argues agent trust is an infrastructure property, not a model property

badlogicgames uses Gemma 4 through llama.cpp for an edge robot

Hex voice prompting app previews Linux support for local dictation workflows

U1 Pro image model launches with 8K output claims and GPT Image 2 comparisons

Athas refines its documentation interface with smoother navigation and search

vLLM details CI, benchmarking, and release process for 2,000 commits per month

Wafer explains how B200 tensor memory changes matrix multiply kernels

Kilo Code says xAI held the top model-lab usage spot for 13 weeks

Sakana AI shares Diffusing Blame for Dale-constrained networks

LMCache claims 14x faster TTFT with disaggregated KV cache management

OpenRouter adds Inkling with 1M context and multimodal inputs

Treg launches an open-source skill and secret registry for agents

Zeeg sets eval and telemetry requirements for production agent frameworks

LangChain explains the Agent Development Lifecycle in LangSmith

Random Labs Slate orchestrates high-volume subagent sessions from the CLI

MemoHarness adapts agent harnesses from past executions

Amp launches $20 and $200 subscriptions with remote orb usage

XPeng plans European post-training for its VLA 2.0 FSD model

Nous releases Hermes Agent companion skill for Unreal Engine MCP

Peter Steinberger runs Codex in VMs to avoid browser focus theft

DiligenceBench evaluates public-equity research agents with finance harnesses

Molty summarizes live GitHub commit streams into maintainer-ready digests

Reach V B evaluates Codex skills against past workflow sessions

TanStack AI useChat adds FIFO, batch, and interrupt message queues

Theo explains Fable 5 and Mythos 5 as separate entrances to one model

Steipete reports GPT-5.6 Terra high makes GitHub review bot 40% faster

Jason Liu shows Codex can manage task titles and sessions itself

Matt Pocock controls Claude Code with an Xbox controller and microphone

Zed previews multi-select pickers coming next week

ChatGPT showcases Sites for publishing games, handoffs, hubs, and research

swyx runs weekly SEO and AEO research automations with coding agents

Nick Dobos warns coding-agent work needs recovery and friendlier agent tone

TechHalla shares Seedance 2.0 and Magnific multishot video workflow

fofr prototypes direct manipulation for live video generation with Omni