All Stories

QuixiAI ports ThunderKittens kernels to RTX 3090 and Apple Metal

Meta Watermelon reportedly catches GPT-5.5 while using 10x Muse Spark compute

DeepSeek V4 Flash translates a 13.5K-token Chinese PDF at 138 tokens per second

Semantic Action RL learns robot tasks through VLA prompts

UBS raises DRAM and NAND forecasts as AI capex tightens supply

vLLM-Omni cuts Qwen3-Omni first audio latency with staged serving

Loopy adds local saving for reusable agent loops

LiveKit reports 381 ms first sentence for Gemma 4 31B voice agents

Amp cleanup workflow leaves TODO breadcrumbs for future agents

fal tutorial combines Claude Fable 5 with brand-safe video workflows

Herdr and firstmate organize multi-agent coding in tmux workspaces

emulate adds Nuxt adapter for same-origin service emulation

OpenBMB EdgeHome Harness makes MiniCPM5-1B a safe device controller

Mastra adds Apple Container sandboxes for local agent workspaces

Utopai PAI parses scripts into locked characters and keyframes

Anthropic plans to develop its own drugs after launching Claude Science

Anthropic finance team automates tax policy engines with heavy token use

PostureCheck locks your keyboard when webcam posture detection sees slouching

OpenClaw workflow gives agents a full Mac for end-to-end testing

Program-as-Weights paper proposes fuzzy functions as model weights

PagerDuty reliability concerns surface in Gergely Orosz's Uber alerting thread

Peter Gostev publishes a 60-demo Fable 5 3D prompt stress test

Midjourney seeks studio AI training records in copyright discovery

Theo and Zeeg debate when agents make code review skippable

Together says open models grew from 10% to 30% of AI tokens

Tesla's reported AI spend cap sets a $800 per month engineer budget marker

Lovable lets builders reference prior projects with @ context

Jason Zhou shares skills for cloning sites and extracting WebGL shaders

pvncher uses Codex side chats to generate steering prompts for main tasks

Matt Pocock renames /to-prd and /to-issues as spec and ticket skills

AI CLI adds terminal model metadata with context, pricing, and JSON output

NVIDIA's Bryan Catanzaro explains Nemotron and why open models matter

Base44 builds a WhatsApp and iMessage superagent skill for blind users

KernelBench reports Fable 5 wrote a fused megakernel with 18.71x speedup

Quinn Slack says banks are fast-tracking open models after frontier restrictions

Anthropic reportedly explores custom AI chip manufacturing with Samsung

BridgeMind reruns Fable 5 and finds guardrail routing crushes BridgeBench scores

Factory adds GLM 5.2 Fast to Droid with Fireworks hosting

Baseten and Cursor explore MLP compression for long-agent KV caches

openPangu-2.0-Flash specs show a 92B MoE trained on Ascend 910B

Dabit publishes Cloud Stacked Diffs skill for agent-delivered PR stacks

Matt Pocock shows claude --bg for spawning named background sessions

François Chollet says ARC-AGI-3 contenders use symbolic world modeling

Ethan Mollick uses Fable, Unity, and MCP to generate a playable FPS twist

Amazon RoPoLL paper swaps mean voting for geometric median LLM judges

ARC Prize names the first 2026 ARC-AGI-3 Milestone Prize winners

National Design Studio Rampart becomes Hugging Face's top token classifier

Avichawla explains A/B, canary, interleaved, and shadow ML testing

Kent Beck says software engineering builds trust, not just code output

EdgeBench frames agent learning as a log-sigmoid law across tasks

Scaling01 compares reward-hacking rates on DeepSWE and SWE-Marathon

Palantir says some U.S. government customers use Nvidia Nemotron open models

OpenAI reportedly floats a 5% U.S. government stake to ease pressure

Code Arena adds fullstack app generation with databases, API keys, and logins

Vik Paruchuri ships OCR safeguards that target hallucinations in workflows

Anthropic and the Pentagon fought over Claude guardrails for defense use

UK AISI finds larger token budgets stretch cyber task horizons

Will Depue sketches requirements for a trusted executive super assistant

Satya Nadella says enterprise AI moats come from private learning loops

Meta explores selling excess AI compute as a cloud business

v0 adds Nano Banana 2 Lite for image generation in the prompt bar

Meta's RLMF trains models to express uncertainty more faithfully

QuiverAI's Arrow 1.1 generates structured SVG assets through Vercel AI Gateway

Databricks tops NVIDIA's SOL-ExecBench using agentic kernel-design loops

cwolferesearch highlights practical KV cache compaction papers

Meta shows quantized reasoning models overthink correct answers

lgrammel shows AI SDK tools running inside local Docker sandboxes on Mac

MCP Server Architecture Patterns paper maps five server shapes and four anti-patterns

Pauliusztin's context layer diagram separates memory from harnesses

LangChain shows recursive language model workflows inside Deep Agents

Hermes Agent v0.18.0 ships a Judgement Release and new updater

Bloome runs Claude Code, Codex, and DeepSeek in one group chat

Hermes Agent clears every P0 and P1 issue after an all-nighter

FLARE standardizes AI flaw reports across labs and security groups

Tokenscope breaks down token spend by model, MCP calls, and skills

U.S. policy may have pushed China toward open AI ecosystems

Quinn Slack's Freedom of Intelligence rally pushes back on model gatekeeping

Perplexity CEO says power users can burn $10K a month in agent loops

Atomic.chat physics contest shows Fable 5 leads, but GLM 5.2 is far cheaper

LiteResearcher pitches scalable agentic RL for deep-research tasks

Sebastian Raschka publishes a 440-page book on reasoning models

Ramp and Revelio find heavy AI adopters grow headcount 10% in two years

Etched comes out of stealth with $800M and inference-first chip claims

Higgsfield shows Gemini Omni Flash holding up on VFX-heavy video edits

Re-llax proposes a fully non-blocking sync algorithm for recurrent attention training

fofrAI publishes an Omni API skill for video editing workflows

OSWorld2 benchmarks computer-use agents on long-horizon desktop tasks

Spellbook launches autonomous contract management across the full deal lifecycle

Anthropic frames managed agents as a stable layer over changing harnesses

Acti turns the keyboard into an action layer for agent workflows

Google Omni Flash keeps the speaker's voice in video translation demos

Higgsfield launches Seed Audio 1.0 for voice changes and dubbing

MCP Server Architecture Patterns paper maps five recurring server shapes

Booster Robotics launches an IDE for simulation, debugging, and deployment

Paper adds tokens, pen tool, cursor chat, folders, and Figma paste

CyberBench tests whether models can find and patch real open-source vulnerabilities

Jason Zhou gives a coding agent a repo map and richer grep hooks

Zenith harness pushes open models above frontier baselines on hard tasks

Gemini Enterprise adds an Inbox for review, in progress, and done work

Warp adds custom routers for task-based model selection

Google says full-stack TPU ownership is what makes AI reliable

thdxr slices bare metal servers into per-engineer VMs for agent work

Meta's Brain2Qwerty v2 decodes sentences from non-invasive brain signals

vLLM wires in DSpark speculative decoding and pushes throughput higher

ComfyUI's Comfy MCP exposes saved workflows to agents

Cursor launches an iOS app for cloud agents and remote control

Hermes Agent moves session DB writes off the event loop

IRIS packages Gemini Live plus Hermes Agent into a public repo

Hamel Husain says hard-to-eval AI products are usually a design smell

NVIDIA's HORIZON turns hardware design into repository-level agent work

Dexhorthy says to layer small agent loops onto existing systems

LangChain shows how to turn a LangGraph agent into a voice agent

JetBrains Air brings Codex, Claude, Gemini, and Junie to Windows

H company opens its computer-use agent API in beta

T3 Code mobile lets one app steer work across multiple computers

Google's Paper Assistant Tool automates scientific review

Qwen maps the verification horizon for coding-agent rewards

Candidly reads support chats live and steers resolution in real time

LangChain starts rolling out Trace Judge for cheaper agent trace checks

Pauliusztin's raw, wiki, outputs folders bootstrap a second brain

DiffusionBench argues ImageNet alone misses real DiT progress

Paper adds selected text layers as prompt presets for image generation

Gergely Orosz says hosted open models vary by provider and quantization

Nvidia's AI chips keep adding RAM every generation, now up to Rubin

BIS warns AI capex and debt could become the next credit shock

PlanBench-XL tests agents across 1,665 tools and hidden dependencies

BrowserBC turns browser sessions into reusable skills and lifts web scores

Towards an Agent-First Web adds identity, purpose, and agents.txt to requests

Instructor 1.15.4 ships a maintainer sweep of parsing and doc fixes

LangChain ties DeepAgents, sandboxes, evals, and model integration together

Fleet agents can be shared directly in Slack and Teams channels

Anthropic says the real breakthrough is cloud AI wired into internal systems

OpenAI tests a Codex effort slider and reworks real-time voice support

OpenCode signs up for Telnyx by following agent-specific onboarding steps

ChatGPT's plugin menu exposes Agent mode, Netlify, Supabase, and Slack

Think orchestrates Claude Code in Cloudflare sandboxes with zero API keys

Codex desktop shows all remote sessions on mobile and other devices

Are They Ready For An Agent-Native Memory System? maps four memory modules

Mark Chen says OpenAI models are nearing self-sustained research

Techhalla's Blender MCP workflow turns viewport scenes into stylized video

OpenRouter's GPT-5.5 performance page shows 41 tok/s and latency charts

Micron says the global RAM crunch should ease by 2028, slowly

NVIDIA's SOLAR automates speed-of-light bounds from PyTorch and JAX code

Opus 4.8 forgets names in long Claude Code sessions, hinting at KV limits

TriAttention compacts vLLM KV cache and frees memory from paged blocks

OpenAI plans GPT-5.6 Sol on Cerebras at up to 750 tokens per second

Codex helps reproduce a touch-sensing robot paper in a few hours

PaddleFormers 1.2 adds DeepSeek-V4 training support and 128K long contexts

Sebastian Raschka shows how to run local coding agents with open-weight models

Seed Audio 1.0 generates dialogue, voices, and scene-level audio together

Nous Research adds cron-backed dashboard scripts with sandboxed paths

The writing-agent repo lays out a Python agent stack from configs to observability

OpenTrons OT-2 and Claude Code power a garage Alzheimer's screening workflow

Corey Ganim says AI should filter email and draft replies, not send them

NUS survey maps World Action Models from prediction to action

Josef Chen shows routing and voting hit a co-failure ceiling across 67 models

AISI's cyber horizon chart says frontier cyber capability is moving fast

cwolferesearch shows how realistic cyber evals are built

Gergely Orosz says bogus copyright claims can erase articles from Google

Ford rehiring 350 specialists shows factories still need human failure memory

Hamel Husain says binary LLM judges beat Likert scales for actionable evals

Cursor helps build an Electron app that fixes Samsung webcam flicker

willmcgugan's writing app targets distraction-free drafting with typography controls

OpenKnowledgeAI turns papers into visual explainers and markdown notes

Bank of Korea finds AI cuts task time 3.8 percent but not output

hf-claude lets Claude Code use over 100 open models

Anthropic predicts autonomous self-improvement could arrive by the end of 2028