All Stories

Jiaxin Shi talk bridges autoregression and diffusion through masked diffusion

Local Studio runs local agents through vLLM with DGX Spark hardware

ComfyUI adds MiniMax H3 as an open-weight multimodal video workflow

François Chollet says agentic AI workflows are becoming increasingly CPU hungry

Warp Agent CLI adds /connect-grok for X Premium subscription access

François Chollet argues future AI may shift from LLM stacks to symbolic learning

Hark Handoff reduces computer-use agent turns from 15 seconds to 5 seconds

ElevenLabs reports ElevenReader Voice Chat raised listening time by 24%

Conductor Cloud adds local command, port forwarding and file sync escape hatches

Eve documents a Next.js-like framework for durable multi-channel agents

Sergey Levine paper investigates why action chunking helps imitation learning

NoimosAI releases Social Agent for cross-platform trend research and posting

Claude Code adds cross-session messaging for parallel agent work

prime-rl and verifiers add multi-agent training and evaluation support

Raindrop explains MCP tool hierarchy design for lower cost and faster answers

doodlestein_self_releaser reuses GitHub Actions workflows for local releases

v0 adds Usage and Activity dashboard for credits, members, projects and messages

SlopCodeBench discussion highlights 33% pass rates on evolving-codebase tasks

T3 Code merges 250 PRs with agent observability, search, mobile and reliability fixes

Matt Shumer recommends fresh context and outcome prompts for Claude Opus 5

Reach shares a global AGENTS.md template for safer coding-agent defaults

Boris Tane lists production agent failure modes around durability, subagents and tools

Suno adds Voices recording to its iOS and Android apps

Simon Willison notes ChatGPT mobile effort controls behind long-press send

Gergely Orosz criticizes Gartner AI reports as pay-to-play vendor rankings

ARC Prize shows Gemini 3.6 Flash at 60.4% on ARC-AGI-2

Google DeepMind opens WeatherNext Cyclones for storm forecasting

Cloudflare OS makes Kimi K2.7 Code and GLM 5.2 the default model picks

Sergey Brin says Jeff Dean pushed TPUs when voice search hit CPU limits

OpenRouter brings open models into Netlify Agent Runners

Baseten becomes a Hugging Face inference provider

Unsloth says DeepSeek-V4-Flash runs up to 2x faster with DSpark

Portals into Orbs expose running apps over HTTPS for live remote dev

Herdr joins Y Combinator and keeps its agent runtime open

T3 Code's inbox-style threads keep more agent work in view

Pion Opus ships a pure-Go Opus codec for audio pipelines

Agent Plugins standard lets Codex, ChatGPT and Cursor share bundles

pi-subagents adds code mode for fan-out orchestration and worktrees

BT6's Black Hat demo shows prompt injection can jailbreak a robot dog

Daphne Koller says biology AI is bottlenecked by data and causality

Transluce finds Claude changes behavior when it recognizes the user

OpenAI security memo shows HR punished a board warning

ARC Prize retests GPT-5.6 Luna at the same score for 80% less cost

Muse Spark 1.2 lands on the cost-performance frontier

OpenAI unifies GPT-5.6 Sol and Luna in ChatGPT

Matt Pocock's /wait-what skill rewrites jargon into plain code explanations

Epoch AI says one in five US workers now delegate a task to AI

Suno lays out a responsible music policy for AI generation

Goodfire AI's motion model spots Parkinson's gait impairment

Hermes moves Polymarket into an optional skill install

ElevenLabs Dubbing v2 preserves voice and delivery across languages

Wispr Flow's growth engine ran on one person and a stack of agents

Remocn turns AI agents into browser-native product demo video builders

LangChain updates Deep Agents course with async subagents and TypeScript

Prof. Tom Yeh walks through sparse autoencoders by hand

Poolside says Laguna S 2.1 crossed 10T tokens served

Roboflow Playground benchmarks Qwen3.8-Max on medication-tray reasoning

DeepSeek V4 Flash shows up as Ollama's fastest-growing model

Gemma Translator turns a Raspberry Pi into an offline translation device

MiniMax H3 tops DesignArena on three video editing tasks

Prime Intellect's Prime Agent hits 95.5 percent on ARC-AGI-3

Uncle Bob says deterministic work should stay in deterministic tools

Grok voice mode adds connectors for task execution

v0 API exposes app generation, previews, and deploys for agents

Hark Handoff claims browser-use lead across three evals

Cursor opens SDK Bridge so Rust and Go can drive agents

Zed 1.14 sandboxes terminal and fetch tools by default

ByteDance SeedRealtime merges audio, video, and text into one live model

OpenAI eval agents built a hidden message board inside a repo

Codex adds Microsoft Edge support to Computer Use

Alexandr Wang says agent swarms can outwork teams of 100 engineers

WisprFlow turns meetings into instant notes without a bot

MIT and Stanford find LLM financial advice depends heavily on the prompt

Codex can turn Bluetooth back on when your mouse dies

Codex turns Chet Baker transcriptions into a practice app

Higgsfield publishes a full tutorial for its AI film Hell Grind

ComfyUI ships FLUX 3 Video through partner nodes

Matt Pocock says coding sessions develop bad trajectories that need resets

llama-macOS launches local model servers and picks models for your Mac

Prof. Tom Yeh walks through Switch Transformer routing by hand

Zyphra releases ZUNA1.1 report and EEG Playground updates for thought-to-text

Mode Forcing draft screenshots raise concerns about AI citation rings

Locus automated research system leads PostTrainBench with 44.7

One Layer Deeper competition targets latent serial computation in models

Google DeepMind leaders put recursive self-improvement on a 2027 to 2028 horizon

Hieu Pham picks FlashAttention 2, SGLang and DeepSeek v3 as core AI OSS

Kimi K3 and Qwen diagrams point to linear attention in top open models

ICLR decision screenshot shows hallucinations fixed after reviewer rejection

OpenAI releases Lean certificates for ten AI-assisted math advances

Factory reports enterprise agent usage up 56 percent month over month

Yann LeCun ties test-time optimization to EBMs and objective-driven AI

John Achiam separates RSI into goals, intelligence, task knowledge and science

QuixiAI shares ReAligned-Qwen3.5 models on Hugging Face

Sakana AI joins AIRoA to pursue world models and Physical AI R&D

Claude Chrome Plugin gives Claude Code a browser for frontend iteration

Hamel Husain and Shreya Shankar expand AI evals course reader to 200 pages

Claude Opus 5 draws high-engagement reports of instruction-following regressions

Sucheng Zang challenges Exploration Modeling FLOP efficiency claims

ChronicleBio uses AI over 153 TB of patient blood data for disease subtyping

Clement Delangue calls for trace sharing after AI-powered cyberattacks

Deedy Das turns Bay Area floor plans into Kyoto-style 3D redesigns with LLMs

FlexBase plans a 500 MW GPU cluster atop a Swiss grid-scale battery

Grad623 compares chain-of-thought loops with hidden-state layer looping

Guinness Chen says ChatGPT voice chats use Terra-low for latency

Kimmonismus flags robotics as the next AI revolution after LLMs

Core Auto AI posts talks on dense linear algebra and GPU kernel automation

Iris open-sources a desktop voice companion that delegates work to Hermes

Matthew Berman says Codex Voice misread a command as run the canary a million times

jachiam0 argues AI will kill security by obscurity in spaghetti-code systems

OpenSurge for Mac opens a Surge-style LAN proxy gateway with observability

Steipete spots OrcaRouter spam opening pull requests across 58 repositories

Nick Dobos says Claude found a 100M dollar crypto bug in 8 minutes

Codex users say five-hour rate windows have not returned

Google DeepMind AGI Safety allows coding agents in engineering interviews

Greg Brockman demos ChatGPT for building interactive educational tools

Jakub Krehel adds cmd-K search across Interfaces issues with cmdk

TechHalla shares Seedance 2.5 prompt recipes for 30-second coherent scenes

Hrishioa uses Fable with interactive visualizations to relearn advanced math

Sakana AI releases Namazu API for Japanese business LLM workflows

Matt Pocock warns stale specs become bad cache for coding agents

jxmnop finds PyTorch, vLLM, and kernel bugs with coding agents

Nous Hermes desktop can be built locally on Intel Macs

Firecrawl open-sources pdf-inspector with 0.002 second per page parsing

Kyle Jeong gets Codex to deobfuscate and rebuild itself as an Electron app

Wafer benchmarks Kimi K3 on AMD MI355X at 952 tokens per second per node

Thinky Machines tests pretraining data filtering for harmful capability control

Ethan Mollick says think step by step no longer improves modern models

John Schulman outlines OpenAI's safe path for open-weight models

QuixiAI questions whether NVIDIA software support makes Blackwell look exclusive

Ruslan Salakhutdinov discusses frontier lab data, infra, agents, and decision-making

Yale and Chicago paper finds LLM research ideas are narrower than humans

Avichawla compares OPRO, MIPROv2, TextGrad, GEPA, AlphaEvolve, and AutoResearch

Meta AI website tests voice dictation support in the prompt box

Herdr positions a tmux-like terminal as an agent-aware multiplexer

OpenAI ChatGPT Work cloud browser lets users monitor and intervene

Shreya Shankar shares an error-discovery skill for trace review

Amp removes per-workspace leaderboards after agent commit counts dominate

ChatGPT desktop adds a Codex Security scan configuration panel

Google's unreleased LMChat shows DeepMind's product-shipping constraint

Google Gemini Spark rolls out to AI Pro users outside the US

ChatGPT Work can schedule recurring email triage automations

Grok Imagine adds voice consistency, multi-reference, and 1080p video

SpaceXAI SuperGrok Plus adds app creation and one-tap web deployment

Native SDK adds a built-in code renderer component for chat and review UIs

Arena ranks Inkling-Small near larger open models with only 12B active parameters

Maxime Labonne shares 40 slides on post-training edge agentic models

Drae Comino says Cerebras WSE architecture is still early in its tech tree

Osanseviero open-sources Gemini calendar art for e-ink displays

Opus 5 demos generate procedural Mario, Pokémon, and ThreeJS scenes

Frontis-MA1 paper releases OpenMLE stack for recursive ML engineering agents

Simon Willison builds mcp-explorer and datasette-mcp around stateless MCP

Microsoft Echoverse trains computer-use agents in deep evolving environments

AgentRadio improves Claude Code multi-agent task accuracy with passive awareness

OpenAI brings hands-free ChatGPT Voice to macOS and Windows desktop apps

Dan Roy account compromise spreads fake ICLR petition malware link

doodlestein warns 4K phones can exfiltrate codebases from monitored screens

ARC Prize reports Kimi K3 at 60.4 percent on ARC-AGI-2

stochasticchasm poses hidden special-token forensics for open models

badlogicgames warns model filters can be bypassed after long cross-lab context

Simon Willison releases smevals for small model, harness and prompt eval suites

Conjecture Bench opens 15,000 mathematical conjectures for LLM disproving

Vercel Passport puts vibe-coded internal apps behind enterprise identity providers

Matt Shumer demos Claude controlling Unreal Engine for future Gauntlet Loops

Artificial Analysis ranks Mureka V9 second on both music generation leaderboards

fal adds MiniMax H3 text, image and reference-to-video API endpoints

TechHalla shares Seedance 2.5 workflow using 42 image references

The Looking Glass repo generates location and time grounded image and video scenes

Dkundel highlights ChatGPT Sites hosting, database, storage and auth stack

PJ Accetturo shares Seedance 2.5 anchor image workflow to cut video costs

Imagine Video 1.5 upgrades go live on API and consumer products

Tom Likes Robots posts Seedance 2.5 field notes from CapCut access

Gemini macOS app adds voice dictation, file summaries and rewrite actions

Prof. Tom Yeh walks through the Discrete Fourier Transform as matrix math

Sakana AI interview explains its rapid shift from research lab to product shipper

Seedance 2.5 adds 30-second clips, 3-minute consistency and 50 references

Reach publishes a write-like-leopold skill for agent-assisted prose style

Venice Studio publishes a one-hour AI filmmaking masterclass and harness workflow