All Stories

Humanlayer Outline adds file trees and diagrams to planning docs

Pi v0.84.2 adds transcript search and default tool settings

Stanford finds agent leaderboards poorly predict deployment fit

Magnitude profiles local hardware before downloading models

DeepSeek Native Sparse Attention cuts decode memory reads

sol-advisor finds subagent routing costlier than Sol-only

React Grab adds Three.js and React Three Fiber support

Cerebras CS-4 claims higher throughput per megawatt

OpenAI pauses frontier RL to expand monitoring safeguards

Qwen3.8-27B refusal-removed build reaches Apple Silicon

Vercel offers $1M for Sandbox escape research

Warp releases Factories for cloud software agents

Perplexity adds U.S.-hosted DeepSeek V4 Pro to Computer

AgentMail becomes an official OpenClaw plugin

TanStack AI adds structured outputs to agent harnesses

Claude Code previews managed task boards for subagents

Eve exposes durable agents as OAuth-protected MCP servers

Gemini Managed Agents provide persistent Linux sandboxes

Claude Code 2.1.235 improves prompt and permission safety

LangSmith releases a tuned evaluator for perceived agent errors

Matt Pocock proposes grep hygiene for coding-agent codebases

Cua Driver previews encrypted local Computer History

LlamaParse recommends source-region grounding for document agents

AI Engineer open-sources YouTube thumbnail A/B testing results

Perplexity Computer accepts tasks through forwarded email

Base44 reports 850,000 Superagents built in four months

Claude Code adds /design for editable UI artboards

Trellis uses an LLM to prefill onboarding forms from minimal input

Groq raises $350 million to expand its inference cloud

Microsoft trains a 4B model to negotiate with SocialRL

Sakana AI launches Namazu, a Japanese business reasoning model

DeepSeek DeepEP moves MoE dispatch from CPU proxies onto GPUs

OpenRouter reports token usage rising from 3.73T to 75T

Qwen3.8-27B runs locally at 115 tokens per second on an RTX 5090

Solar reaches 10% of global electricity as batteries shift output

Fable produces worse fine-tunes after using the wrong chat template

OpenAI reportedly tests Astra for long-running multi-agent tasks

Prof. Tom Yeh makes LoRA and full fine-tuning interactive

HK Project explains distributed training by building it from scratch

Kimi K3 stabilizes sparse MoE training with Stable LatentMoE

GitSkills catalogs 3.8 million SKILL.md files across GitHub

Unitree humanoid demonstrates a two-meter jump and 12.66 m/s sprint

A paper finds agent skills mainly stabilize procedures, not knowledge

Cartesia Sonic 3.6 tops controlled voice benchmarks

Hermes Desktop adds named Bots with separate memory and skills

Claude Managed Agents cut selected ABC Legal task costs by 50%

A model reproduces a complex shell pipeline from parametric memory

Raindrop AI releases PII Guard 2.0 for safer trace observability

A study finds coding agents badly misjudge runtime and performance

LLM-as-a-Verifier raises DeepSeek Terminal-Bench accuracy to 88%

OpenAI trains models to write superhumanly secure code

OpenAI reports GPT-5.6 Luna matches GPT-5.5 at one-eighteenth cost

Google Research uses smartphone photos to estimate insulin resistance risk

Seedance 2.5 leads Video Edit Arena and ranks near the top in I2V

MiniMax H3 gains audio-conditioned video generation in Diffusers

MiniMax H3 generates consistent walking game sprites

OpenAI removes custom GPT creation from personal ChatGPT plans

Cognition says cloud agents now produce most merged PRs

Interfaces Cheat Sheet collects immediate UI polish fixes

Megakernel literature gets a layered GPU taxonomy

Meta shows small scaling studies need broader tuning

Intelligence per Watt study tracks local AI efficiency

OpenAI CFO Sarah Friar shares ChatGPT usage tiers

Prof. Tom Yeh maps key reinforcement learning milestones

Stripe's reported OpenRouter deal spotlights routing data value

Ostris trains a MiniMax H3 LoRA to increase motion at four steps

Siemens trials self-hosted Qwen and DeepSeek with vLLM

MDA cuts scientific experiment budgets on FORCEBENCH

Contextual voice transcription links issues and wikis live

Gemini in Chrome adds a macOS menu-bar entry point

Kolt Regaskes uses Claude to audit Codex work

Aaron Levie frames AI value as applying compute to neglected work

vision skill writes evidence-backed VISION.md files

AQuA seals evaluator state from self-improving quant agents

Uncle Bob questions mutation-only testing workflows

Pi contributors argue code should remain the source of truth

Gergely Orosz reports declining trust in Perplexity search

Nous Research cuts desktop session load times 19x

Codex exports 5,032 X bookmarks through browser control

Open-source SVG mascot generator makes Grok-style animations

Magnific gets five prompt templates for motion graphics

Goldman is helping Nvidia build financing around AI infrastructure buys

ArchAgent v2 finds a better prefetcher with cascaded evolutionary search

Gemini 3.7 Flash generates full websites in real time as you type

Baseten Loops splits Qwen3.8 training into trainer and sampler sessions

Cover says its weapon imaging system can detect concealed guns farther away

Instinct lets an AI product execute transactions and navigate websites

Microsoft's Reason Wide, Not Deep distills reasoning into reusable skills

Codex multi-agents v2 adds hierarchical sub-agent tiers and model delegation

Athas Capsule swaps in Ghostty's WebAssembly terminal engine

sol-advisor routes routine work to Luna Max and Terra High

Plannotator-last opens the latest grill-me question batch for quick review

Gemini lets users disable visible media watermarks in supported regions

Gavin Baker and Sholto Douglas debate whether AI power should concentrate or distribute

Theo's watermark detector demo shows image marks can be verified

BenchDrift shows benchmark wording can swing LLM scores by tens of points

5.6 Sol pushes refusal rate under 2 percent on 842 harmful prompts

Google and Jeremy Kun open-source a compiler for homomorphic encryption

Scale AI's taxonomy localizes agent failures to model, harness, or environment

Modal's H100 glossary becomes an interactive Three.js GPU explainer

feedback-agent opens a cloud PR when users file in-app feedback

Amp adds Git repository hosting to its project creation flow

Claude adds a model comparison UI with memory and hold responses

TechHalla shares JSON prompt templates for Minimax H3 video scenes

Gergely Orosz says AI coding tools can create a burnout vampire effect

Claude Opus 5 feels harder to work with in a Hacker News thread

Dario Amodei says curing cancer is how Anthropic can change public perception

Pi adds /fork, /tree, and /resume for branching agent sessions

Stanford's Deep RL lectures are now free to watch online

OpenAI GPT Pro writes directly into GitHub repos from the web app

Anthropic's risk report says automated R&D risk is low, but confidence is slipping

Anthropic says Model 2 beats Mythos 5 internally but will not ship it

Replica trains Faraday to replicate research papers as an RL task

Turpuffer runs 100-plus BYOC clusters with a custom control plane

vLLM's adaptive verification keeps DSpark on the Pareto frontier

Hugging Face says small models still dominate while Qwen derivatives explode

AI/ML API shows Qwen3.8-27B matching Opus 4.6 on a 3D flythrough

RedNote's dots3-note Preview opens a 280B multimodal agent model

Alibaba Qwen opens Qwen3.8-27B with 262K context and vision support

Gemini 3.7 Flash builds an interactive 3D heart anatomy site

Unsloth makes Qwen3.8-27B run locally on 17GB RAM with dynamic GGUFs

DeepSeek V4 Pro 0813 appears on OpenRouter with API-only access

Reuters reports US draft would force AI partners to choose sides

Qwen3.8-Max lands day zero on multiple hosting platforms

OpenClaw makes agent sessions shareable as URLs for handoff

GitHub MCP now installs in VS Code with one click

Lovable lets each app user connect their own Drive, Slack, and HubSpot

Andrew Ng maps the most important skills in AI engineering

OpenRouter's Ori launcher wraps DeepSeek, Grok Build, and Prime Agent

MIT paper formalizes deep learning architectures with category theory

Claude Code 2.1.233 adds memory cgroups and fixes NTLM leak paths

Anthropic's red team found a dozen Firefox bugs, and Mozilla shipped fixes

OpenAI's packages/ai parser misses Kimi cached_tokens as cache reads

OpenAI is reportedly on track to clear $40B in annualized revenue

Zed can now diff your branch against the default branch in gutters

Seedance 2.5 and Codex repair a 2 Fast 2 Furious CGI shot

Cartesia's Sonic 3.6 adds Hindi, Hinglish, Urdu, and Odia

Suno Studio 2.0 adds plugin-style effects and session workflows

METR raises $71M to study autonomy, self-improvement, and AI incidents

SSHHIP lets phone users tap wrapped terminal URLs from SSH sessions

AI Toolkit adds MiniMax H3 Ref2Vid training support

Vercel CLI connects coding agents to AI Gateway with one command

Databricks Unity AI Gateway adds task-aware smart routing for coding agents

Nous Portal gives LongCat-2.0 a one-week free coding window

GLM-5.3 ships in the Coding Plan with staged API and open weights

Unsloth lets you fine-tune Muse Glimmer for free on a 24GB GPU

Anthropic and Broadcom compute financing shows how frontier AI gets funded

Mistral's 1 gigawatt Europe compute plan underscores the scale of the buildout

François Chollet pushes test-time training into the mainstream

Cerebras explains the 900,000-core engine behind Sol Ultrafast

Google says seven-year-old TPUs still run at 100 percent utilization

HarnessAgent lets you swap models with one line of code

Hermes Agent lets you steer and stop subagents live

OpenAI charts show frontier firms use plugins 2x and skills 6x more often

ASImposium proposes a public ledger for agent-run scientific inquiry

Microsoft finds agent skills can cause 307 failures and slow agents down

Codex uses Computer Use to automate Google Cloud Console OAuth setup

AI SDK adds an ACP harness layer for coding-agent runtimes

grok-plugin brings Grok 4.6 into Claude Code and Codex

DeepSeek Harness docs define pure UI rendering for replay-safe tool cards

Devin works directly on live MongoDB Atlas data mid-task

Hermes Agent expands its plugin surface with 29 PRs and 26 hooks

DeepSeek Harness ships a plugin-first agent harness in developer preview

Grok 4.6 climbs ReactBench and closes the gap on React fixes

ARC Prize says Grok 4.6 matches GPT-5.6 Sol on ARC-AGI-3 for less

Google ships Gemini 3.7 Flash with cheaper pricing and stronger coding scores

Factory adds Agent Effectiveness dashboards for throughput and cycle time

Firecrawl indexes 41M life science papers for research agents

Sakana Chat adds code execution for Japanese vibe-coded apps

Ramp's AI spend chart shows Anthropic widening its lead over OpenAI

OpenAI data says productive firms adopt AI faster and more often

Qwen 27B planner-actor workflow shows weak cost tradeoff in user report

DeepSeek V4 Flash test reaches 14.86 tok/s on dual v620 GPUs