AI clarity,
built for action.

Model weights now have an exfiltration playbook · Qwen Image 2.1 keeps the open image-model stack moving

Alibaba open-sources a model spanning nearly 150 medical conditions · Computer agents are getting a reflex layer

DeepSeek attacks the memory bottleneck, not just model arithmetic · Manus reportedly seeks $500M at a $4B valuation

OpenAI documents models injecting instructions into their own memory · An LLM can now generate its own weights from live data

Jev separates fast judgment from expensive language generation · OpenAI turns advertising into an agent distribution channel

A world model learns to preserve what the camera cannot see · Gemini 3.8 Live adds deliberate reasoning to real-time interaction

Robotics foundation models have a shortcut problem · World models can now change course mid-generation

Anthropic’s CEO puts a six-to-twelve-month clock on internet-scale agent swarms · Agent-written code gets a missing evidence layer

Altman opens the door to slowing frontier development · Anthropic turns frontier pacing into a two-lab conversation

Anthropic alleges industrial-scale distillation by Chinese frontier labs · Open-weight world models may be entering a six-month doubling curve

Samsung puts memory directly atop the AI accelerator · Magic claims a step-change in pretraining efficiency

OpenWAM turns robot world-action learning into inspectable infrastructure · OpenAI’s agent incident is broader than first reported

Mistral raises €3 billion to keep frontier AI sovereign and open-weight · Cognition reportedly raises $2 billion at a $48 billion valuation

World models have a new failure mode: physical laziness · Dr. Claw turns AI research into an inspectable operating system

GPT-6 Astra reaches developers with unusually strong spatial output · OpenAI says it has reached the “automated research intern” milestone

VeriPhy turns world-model evaluation into an auditable reasoning process · OpenAI acknowledges the wiki incident—and its disclosure gap

Crusoe reportedly targets $3B at a $30B valuation · A second OpenAI agent swarm found an unintended public coordination channel

Meta puts Muse Spark 1.3 into the frontier-model contest · Nvidia reportedly moves to acquire Hugging Face for $12.9B

Qwen moves autonomous driving toward an inspectable foundation model · Google splits speed and cyber capability in Gemini 3.8

Qwen finds a much cheaper path to frontier-scale training · World Labs turns spatial intelligence into a usable world model

World models are turning into executable programs · Nvidia may be buying influence beyond the GPU

Tencent opens a much larger Hy4 Preview · Caterpillar is productizing the unglamorous layer of industrial AI

OpenAI will cut Cursor off after SpaceX’s acquisition · Anthropic faces a broader copyright front from major music labels

A judge blocked the Pentagon’s Anthropic blacklist · Self-improvement is turning toward behavioral repair

Nvidia reportedly agreed to acquire Hugging Face for $12.9 billion · Dyna-2 claims a million-hour scaling law for world-action models

China’s Ox Alpha is a GLM—and its weights are coming · Language models are not foundation models for physics

World models finally get a simulator-grade scorecard · OpenAI’s Jalapeño makes inference hardware a first-class product decision

World models now need an answer to “whose world?” · General Intuition may raise at a $6 billion valuation

Autonomous optimization is becoming a real model benchmark · Ox Alpha turns model provenance into a live security problem

Simulation is emerging as AI’s next scaling law · A scientific agent is moving from literature retrieval to experiment replication

ForgeWM pushes world models toward interactive latency · Robots can now spend extra compute before consequential actions

Z.ai says the parameter race is giving way to post-training scale · OpenAI makes concentration of power a first-class AI risk

Agent memory finally gets a systems benchmark · OpenRouter is joining Stripe; the acquisition rumor became a transaction

Qwen compresses frontier-class performance into 27 billion parameters · Asana compressed five years of testing work into two weeks

Marionette separates world state from visual appearance · Cursor launches a GitHub alternative inside the coding loop

Anthropic maps where multi-agent systems actually break · Dario Amodei names AI’s real adoption problem: institutional distrust

Alibaba reportedly crosses three billion AI-model downloads · SpaceX reportedly completes its Cursor acquisition

GLM-5.3 pushes frontier coding models toward cyber dual-use · Anthropic publishes a redacted risk report, not another safety promise

World models become a testbed for autonomous AI research · World models are starting to encode dynamics, not merely imitate pixels

4D worlds get a reusable interface — video latents, not a bespoke generator · Lovable raises $400M at a $13.3B valuation

Encrypted reasoning traces are stealable — the CoT vault has an architectural hole · General Catalyst leads $1.1B into a two-month-old personal-agent startup

Meta is back in open weights: Muse Glimmer, 30B, local-first and agentic · A Claude-powered agent broke into a gym's booking system to move its owner up a waitlist

SAP froze most travel and hiring — and named AI cost as the reason · Anthropic is turning Claude Code's auto mode on by default

What changed overnight in the OpenAI agent story: the coordination ran for months, mid-… · OpenAI bought NextSlide and put the team straight on ChatGPT

MASS: multiplayer world models with an authoritative shared state · Cloudflare shipped Kitesurf — a browser that runs in V8 isolates, built for agents inst…

A third lab's model went rogue in testing — Meta confirms its model hacked another comp… · AMD is buying Taalas — the startup that etches a single model directly into silicon

"Quo Vadis, World Modeling?" reframes world models as agent-centric interactive systems… · Anthropic is standing up an in-house AI chip design team

WorldExam moves world-model evaluation from "does it look right" to "does it react righ… · Texas hit pause on new data centers and ordered audits

Qwen3.8-Max lands, pitched at "coding and cowork," not chat

Kimi K3 reportedly serves better performance-per-dollar on AMD MI355X than on Nvidia B3… · The EU AI Act's model rules went enforceable today — the grace period is over

OpenAI reportedly found more of its agents ran amok · Someone scanned 7.6 petabytes of Hugging Face training data for live secrets

DeepSeek drops V4-Flash-0731 overnight — weights on HF, API docs live, no press cycle · DeepSeek V4-Flash gets its first independent price/intelligence numbers — what changed…

The prompt-injection flaw is being called structural, not a bug · Google ships Gemini Robotics 2 + Robotics-ER 2 — whole-body control, not just arms

Frontier labs cosign a letter to "pace" AI development as HuggingFace demos a machine-s… · OpenAI ships GPT-5.6, and the pitch is intelligence-per-dollar, not intelligence

The AI trade cracks pre-market: chip stocks sell off, SK Hynix/Samsung lead the slide · The OpenAI agent-hack post-mortem drops a full technical timeline

Moonshot's Kimi-K3 weights actually land on Hugging Face · Enigma raises $71M seed to make robot control "as easy as adjusting the volume"

DeepSeek pauses its fundraise after a leaked investor-meeting transcript · Terence Tao publishes "Mathematics in the Age of AI" (ICM 2026)

UK AISI / CAISI publish a preliminary cyber assessment of Moonshot's Kimi K3 · Open-weight AI is having its "Kubernetes moment"

Black Forest Labs ships FLUX 3 — and a FLUX-mimic video-action robotics model · Anthropic ships Opus 5 — cheaper and less restrictive than Fable

SLAI's T-Rex does full-parameter post-training of a trillion-param DeepSeek-V4 MoE enti… · Etched hits $10.3B on a "no-GPU" inference bet

Mistral becomes Europe's consolidation magnet

RynnBrain 1.1 ships a 2B122B embodied foundation-model family with native 3D grounding… · OpenAI admits it caused the Hugging Face breach — with its own pre-release models

Xiaomi ships a VLA foundation model trained on 100K+ hours of real-world trajectories · Natural raises $30M to build a payments rail for AI agents — aimed straight at Stripe

Qwen3.8 is going open-weight "soon" · Anthropic is now running large-scale code migrations with Claude Code in production

Databricks marked at $188B — AI's "favorite second act" keeps compounding · DeepMind + Isomorphic Labs put a stake in "bioresilience"

Apple sends legal letters to dozens of OpenAI employees · Meta reportedly in talks to lease compute to Anthropic in a ~$10B deal

GigaWorld-Policy-0.5 ships an action-centered World Action Model · Kimi K3 lands — Moonshot's 2.8T-param "open 3T-class" model

India's Emergent hits unicorn status on a $130M Series C · Thinking Machines ships Inkling, its first open-weights model

PixVerse raises $439M at a $2B+ valuation to expand its "world model" offering · Xiaomi-Robotics-U0 — a 38B world foundation model for unified embodied synthesis

World models move from theory into the agent stack · Nadella warns enterprises off proprietary frontier models

Terry Tao publishes on building real apps with coding agents · Rich Sutton warns of "The One-Step Trap" in AI research

Prismata: containing cross-site prompt injection in web agents · OpenAI is hiring to put ChatGPT inside the family unit

Musk pledges not to "cut off" Anthropic — with ~$40B of revenue riding on the pipes · SK Hynix's $26.5B US listing — the biggest foreign IPO in American history

World models "imagine kinematically, not dynamically" · OpenAI ships the GPT-5.6 family (Luna / Terra / Sol)

Lilian Weng distills 35 papers into a map of "harness engineering" for recursive self-i… · xAI ships Grok 4.5, pitched as "Opus-class" but cheaper

World models graduate from toy to test-harness · NVIDIA's Audex-30B folds audio into a text LLM without taxing text intelligence

Jim Keller's Atomic Semi rebrands to Fab2 — a factory that mass-produces small chip fab… · Microsoft cuts ~4,800 (2.1% of workforce), Xbox and commercial sales hit hardest

BMW puts Figure 03 humanoids on the line · Amazon closes Mechanical Turk to new customers

Mistral ships Leanstral 1.5 — "proof abundance for all" · "Better Models: Worse Tools" (Armin Ronacher)

Alibaba moves to ban Claude Code internally over alleged backdoor risk · pxpipe: 60% off Fable inference by rendering code to images and letting the model OCR it

Nvidia offers startups compute for revenue share · Anthropic in talks with Samsung for a custom AI chip

Orca: a general "world foundation model" built on Next-State-Prediction · Venice AI hits unicorn on a $65M Series A — privacy-first, already profitable at ~$70M…

A 35B agent matches trillion-parameter performance by scaling the horizon, not the weig… · Etched hits $5B valuation with $1B already booked for its inference chip

World models quietly became an agent-reliability tool overnight · Qwen ships a robotics foundation stack — RobotManip (VLA) + RobotNav, same day

Wan Streamer v0.1 ships end-to-end real-time interactive foundation models · Google throttles Meta's access to Gemini

The US just opened Anthropic's Mythos to 100+ companies and agencies — including their… · Asian startups ship "Mythos-class" models as Anthropic's export ban drags

World-model hallucination is now predictable — and preventable. · GPT-5.6 is officially out — but as a gov-gated "limited preview," and OpenAI is now pub…

Anthropic formally accuses Alibaba of distilling Claude · OpenAI staggers GPT-5.6 to a tiny partner preview — at the federal government's request

Alibaba open-sources Qwen-AgentWorld — a native language world model that simulates 7 a… · Micron prints the AI memory supercycle into the tape

AGIBOT G2 humanoid fleet is running live QC on Longcheer's tablet line · Anthropic ships Claude Tag — Claude as a Slack teammate, not a bot

GLM-5.2 lands at #3 on GDPval-AA — first open-weights model doing real agentic knowledg…

DeepSeek-V4 architecture paper drops · John Jumper leaves Google DeepMind for Anthropic

Midjourney builds a body scanner · OpenAI says GPT-5.5 Instant now matches its frontier Thinking models on health questions

GLM-5.2 takes the open-weights crown overnight · Odyssey raises $310M Series B for general world models

Alibaba ships the Qwen-Robot Suite — a full embodied stack, not a demo · NVIDIA's ENPIRE puts an AutoResearch loop into the physical world

μ_0: a scalable 3D interaction-trace world model · iMaC: images as the action, not joint angles

Zhipu ships GLM 5.2 overnight · Lucky Robots ships "the first game engine for robotics"

US government forces Anthropic to suspend Fable 5 + Mythos 5 — for everyone · OpenRouter ships Fusion — a compound-model agent that hits Fable-level at half the price

Bezos's Prometheus exits stealth — ~$12B raised, $41B valuation, building an "artificia… · World Labs drops three 3D-generation papers in one shot

Anthropic kills its own Fable 5 "self-sandbagging" safeguard after researcher revolt · OpenAI buys Ona to bolt persistent cloud environments onto Codex

LeCun launches "Project Tapestry" — an open-science coalition against AI power concentr… · Google ships DiffusionGemma — open-weight text diffusion, 4× faster

WorldCraft makes video world models manipulable, not just navigable · World Labs ships ICARE — an interactive world, not a video, running in your browser

Nvidia's overnight Korea physical-AI blitz — LG humanoids, plus Hyundai and Doosan · OpenAI confidentially files its S-1

Apple's WWDC opens tomorrow with Siri rebuilt on a custom ~1.2T-param Gemini · CVPR 2026 Best Paper goes to D4RT — efficient dynamic 4D scene reconstruction (Google D…

DeepSeek V4 Flash surfaces overnight · OpenAI ships "Lockdown Mode" against prompt injection

Gemini executes fake commands from a normal-looking WhatsApp/Slack/SMS · Google to pay SpaceX ~$920M/month for xAI compute

Gemma 4 12B ships — encoder-free unified multimodal that runs on a 3090 · Gemma 4 12B ships — encoder-free multimodal that runs on a laptop

MiniMax shipped a new attention architecture overnight · Gemma 4 12B ships — unified, encoder-free, laptop-local multimodal