All Stories

GitHub Copilot's Fix command generates regex from a natural-language prompt

Recurrent Looped Transformer carries state across prompt and response tokens

Addy Osmani recommends Opus subagents to stretch Fable usage

Skeptrune flags physical-security gaps at small GPU data centers

AlphaEvolve kernel changes cut Gemini training time by 1%

Latham & Watkins buys Nvidia servers to fine-tune open-weight models in-house

Wafer's repo annotates the Transformer paper and its performance legacy

Gemini 3.8 Flash and Antigravity render diagrams in terminal answers

Flash-BoN improves diffusion search under real wall-clock budgets

thdxr links SSD-backed KV caching to DeepSeek's inference economics

thdxr plans public GPU utilization data and off-hours inference

Yacine builds a depth rig that streams live maps from a basement to a phone

fs-safe makes agent worktree copies about 80% faster with native clones

Matt Pocock uses contact sheets to show agents visual changes over time

Relay publishes SDK and CLI docs for messaging with coding agents

Hugging Face's Agent Swarm invites contributors to join Hutter Prize experiments

Modal scales sandbox infrastructure to one million concurrent environments

Astrable pairs Astra planning with Claude Fable implementation in Codex

Wispr Flow users report dropped negations and missing dictation segments

METR discloses funding sources to support independent AI evaluations

BullshitBench publishes a three-judge dashboard for scoring model pushback

Mikhail Parakhin says agents make a laptop viable for mobile coding

Thinking with Looped Flows adds inference compute without more parameters

burtenshaw/training-agents packages SFT-to-RL workflows

ai& deploys heterogeneous accelerators in a Japanese neocloud

Megatron-LM analysis removes pipeline and tensor parallelism from a 1T MoE estimate

DeepSeek V4.1 Flash reaches 200 to 250 tokens per second on Novita

GLM-5.2 reportedly analyzed the Hugging Face attack when closed models refused

Dingcad and Astra produce a phone-designed 3D-printed device

Chaofan Shou reports a 6 TB credential leak from an AI router

RTX 5090 infrastructure loses four nodes in two days

OpenAI agents abused RubyGems while probing for API keys

Sakana AI special issue connects world models to physical intelligence

HumanLayer previews live multiplayer prompting across agent sessions

Doodlestein extends home reconstruction to interiors with paired phone video

terms.txt proposes machine-readable consent for agent web access

Hermes Agent opens remote gateway viewing for supervised sessions

Hermes Agent reaches 3,000 contributors in its open-source project

Matt Pocock's agent-skills repository surpasses React in GitHub stars

Fugu Ultra v2 reaches top results through model orchestration

T3 Code relies on agent triage at roughly 300 issues per day

One Layer Deeper draws 15,000 submissions without a full solution

GPT-6 Pro helps formalize the strong Papadimitriou-Ratajczak conjecture

GPT-6 Astra leads a no-CoT reasoning comparison by 33 points

MAISI's executive director calls for mathematical rigor before AI deployment

Tom Yeh explains graph convolutional networks with a hand-worked example

Gallier and Quaintance publish linear algebra text for ML and robotics

c_valenzuelab makes a video ad for $7 in nine minutes

iScienceLuvr reports a year of open pathology and brain-model releases

Vercel makes its global deployment metadata store 91% faster at p99

Arena ranks GPT-Image-2.5 Sunburst first across three image tasks

Cohere releases North Small Translate 1.0 for 50 languages

NVIDIA brings CUDA 13.4 to Windows on Arm for RTX Spark PCs

Parallax documents consumer-GPU training with peer-to-peer data streaming

Together ports ThunderKittens to Vera Rubin at 22.4 PFLOPS

Google Research's TIPSv2 improves patch-text alignment for vision models

China's remote excavator system reaches 2,000 km under 30 ms

thdxr explains why open-model APIs struggle with token billing

Epoch AI measures GPT-6 Astra's long-context time-to-first-token curve

RF-DETR and GPT-6 Astra split basketball tracking work

PARSER uses parallel chunk readers for faster long-context agent reasoning

Salesforce paper finds full-trajectory imitation breaks agent harness fit

Alexandre Levie reports agent identity and eval gaps in enterprise rollouts

Gemini 3.8 Flash is a high-step outlier on CursorBench 4.0

OpenAI calls for capability-based safety rules and RSI measurement

Anthropic publishes a detailed report on Claude misuse operations

OpenAI says it has made progress on another Millennium problem

Claude Managed Agents adds session viewing and automatic tool decisions

Hermes Agent adds private-repository plugin installs with ephemeral credentials

Suno publishes a V6 migration guide for controlled and wild workflows

Muesli brings private on-device dictation and meeting transcription to macOS

OpenAI pitches cyberdefense partnerships to U.S. power utilities

GPT-6 Astra and Houdini turn a character animation into FLIP water

VS Code adds an experimental pet with reactions and customization

Epoch AI estimates OpenAI now has nearly 2.7 million H100 equivalents

Giffmana finds per-example stochastic depth loses practical speedups

GPT-6 Astra autonomously designs an 18-hole course from LiDAR terrain

LMCache makes KV cache a persistent, shareable serving layer

Pvncher recommends Rust, WASM, and WebGPU for AI-built 3D apps

ZeroModels ports 100-plus pretrained model families to Keras 3

OpenAI and Anthropic report accelerating AI-assisted research

OpenAI clarifies duplicate opt-out paths for model-improvement data

Astra-based basketball tracking cuts identity errors at $0.20 per video second

John Schulman separates three privacy risks in training on user data

Jacob Coxon resigns from Anthropic over the race to self-improving AI

Tailscale exposes tailnet provisioning and ACL management as APIs

Diagram-design skill adds 39 diagram types to Codex and Claude Code

ChatGPT skills can trigger Grok Bot routines through webhooks

mtdt and 300 skills turn GPT-6 Astra into an inspectable music workstation

OpenClaw demonstrates prompt-built personal and team dashboards

OpenAI publishes a Defense Factory playbook for agent-led vulnerability response

Davide Paglieri's 100-agent experiment finds whistleblowers after cheating

OpenAI's internal model changed checkpoints during an 88-hour Navier-Stokes run

AutoResearchExam measures 24-hour ML research across 29 tasks

Jacob Coxon's weight-sparse transformers expose interpretable task circuits

Paul Christiano joins OpenAI's Foundation safety and security committee

OpenAI reports fewer factual errors and less sycophancy in ChatGPT

Arena opens GPT-6 Astra Medium in Direct Mode for 24 hours

Claude Code 2.1.267 adds effort caps and prompt-cache fixes

Zenith harness reportedly beats Fable on long-horizon software tasks

Ramp reports a 10% drop in top-business AI spend per employee

Kimi Work adds phone-based remote control for desktop agents

Harvey acquires Guardrails AI to bring agent simulation into legal workflows

Suno v6 adds multimodal inputs and section-level music editing

Dwarkesh Patel finds data gains outpace architecture gains

Hugging Face tutorial unpacks 90 synthetic-data experiments

GPT-6 Astra turns two photos into a printable door stopper

GPT-6 Astra builds a Blender dragon and environment in 30 minutes

Unsloth ships Qwen3.8-27B GGUF with 1M-token context

Fly Escape Room runs a 166,000-neuron fruit-fly brain in WASM

DeepSeek V4 Flash reaches a sharply lower reported token cost

Claude Managed Agents case studies cover memory and sandboxing

OpenAI documents an agent-wiki sandbox-bypass technique

OpenAI Codex encrypts prompts passed to subagents

Herdr raises $6M for a terminal-first agent runtime

Meta launches Muse, an always-on agent in a secure VM

Meta Muse puts a Sentinel agent in charge of network egress

Figure reaches 69,900 weekly users and 35 minutes uploaded per second

Claude Tag shares its CI incident-responder setup

Anandkumar's PINN finds a candidate 3D Euler singularity

Avi Chawla maps six production LLM positional encodings

Ostris trains a 441M pixel-space diffusion model without a VAE

SMART regenerates ML performance tooling from design docs

Mistral raises €3B to expand model research and infrastructure

Merve details reward-design limits in Blender GRPO training

LLM Self-Modeling paper makes agent self-prediction testable

vikhyatk shows why an A100 label conceals five distinct SKUs

Ethan Mollick recommends GPT-6 Pro for substantive paper review

Omarchy uses OpenRouter to route sponsored agent tokens

fal enables real-time Full HD H3 Max video generation

Browser Use replaces subscriptions with usage-based credits

Unitree uses a world model to plan humanoid robot combat

Cartesia tightens voice-agent loops with 90 ms text-to-speech

thdxr previews custom app tabs with an embedded dev browser

Theo reports high response-quality variance from GPT-6 Astra

Pi scopes agent tools and AGENTS.md to a custom working directory

GPT-6 Astra reportedly reduces visible reasoning under monitoring

ChatGPT Work makes connected-app writing style available on paid plans

LinkedIn data shows AI-specific roles expanding across the economy

Anthropic's mannered-prose prompt targets bloated AI writing

Tom Yeh publishes an interactive GAN architecture diagram