2025, I made hard decisions. Turned down acq offers to solve what's 10x. Turned down millions of funding. Became a solo founder. Cut team 6 → 3; rebuilt to 7. One key responsibility of a leader is to make hard decisions, quickly. The other is to empower. I trusted someone with no experience in growth or startup to lead our UGC growth from 0 to 6M views/mo in just two months. I hired someone in early 20s to lead our new mobile app dev, uiux, paywall from 0 to 1, almost unsupervised. I guided our amazing sr. eng into a powerful CTO that backs the whole team firmly. I never felt this much of responsibility to lift others' career while being supported back wholeheartedly. What I gave out now started to empower me. An exciting new year ahead. Be brave.

Our new #NeurIPS2026 paper explores how learned priors can override visual evidence in VLMs, and how a few in-context examples can help overcome this bias. Check out the thread from our fantastic project lead @KunWang0129 for more details! 😃

Today, I’m excited to share we’ve raised an additional $12.8M at a $500M valuation. Muse, Instinct, Grokbot, Dots, etc are great for consumers, but businesses need more if they want to delegate real responsibilities to AI. @TaskletAI is built with collaboration, automation, security, and control at its core, and our model-agnostic harness & cloud infrastructure gives you the best value for every AI dollar you spend. Investors in the round include @usv, @359Capital, @salil, @RakeshAgrawal, $RVI, @patrickc and others.

Today, we are announcing our Series A and a new product: @raindrop_ai Simulations. We've now raised $50m from @CRV and @lightspeedvp to protect the world from agent failures, big and small.

For AI to work with us, it needs to understand us Today, we're introducing Persimmon, the first large-scale model designed to realistically simulate how people talk and interact

🚨 Call for live demo / paper at CoRL 2026 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗼𝗯𝗼𝘁𝗶𝗰𝘀 𝗪𝗼𝗿𝗸𝘀𝗵𝗼𝗽! No paper required for demo track. Due Sep 27. 🔮 The real magical power of agent is live interactive demo with the audience, which is why we provide all-out support: - Robot, compute, and API are ready for you in the conference room - Custom robot / sim demo also welcomed - Just submit a video proof - friendly to industrial participants agentic-robotics-workshop.gi… Paper submission also welcomed! We also have an amazing speaker lineup from CMU, UC Berkeley, DeepMind, NVIDIA, and Tencent.

🔥 Recursive self improvement is all the heat right now. But before an AI agent can learn from its own failures, there is a more basic question: Can LLMs really understand their own failures? Introducing Who&When Pro, a large scale benchmark for failure attribution in AI agents. It contains over 12k failed trajectories across text, image, and video, covering 26 benchmarks and 15 agent frameworks. The benchmark spans both single agent and multi agent systems, from reasoning and coding agents to web agents, GUI agents, and multimodal agents. And the answer is not as reassuring as we might hope. Our results show that current LLMs can sometimes locate where a failure happened, while identifying the underlying failure type remains much more difficult. This leaves a major gap for agents that are expected to diagnose, learn from, and improve on their own mistakes.

Today, we're introducing @shotwellst. We're solving the hardest problem in robotics: understanding your data There’s no point collecting robotics data if you don't understand whats happening in it, but your options are limited. If you need dense, accurate labels with precise subtask boundaries, VLMs don't cut it, an in-house annotation army is a huge operational sink, and today's vendors charge an arm and a leg for low quality work Shotwell combines the best of both worlds by training our own annotation models and letting humans deal with the ugly edge cases. Its the ONLY way to consistently guarantee 100% quality in a scalable way. Humans are inconsistent, and models are inaccurate.

We are a hiring a Head of Design Engineering at @Tasklet. This is a key leadership role. TC can exceed $1 million / yr. We're looking for an ex-founder who's built zero-to-one who's both a builder AND a designer. You'll get a chance to invent how people and teams work with AI. Full job posting here: tasklet.ai/careers/head-of-d…

being using @FishAudio at @MathosAI since its launch. best, ultra fast, vivid, expressive audio. cheaper but better, built by the best founders @rissa_cao @LengyueMatrix

A billion people produce the most valuable dataset in the world every day – and delete it every night. We’re recording it. Introducing @AttentionInc

We are starting Intent Lab, building an autonomous team we call "fleet" that turns intent into production software. Today we are sharing some early results: the fastest GLM5.2 inference engine, one shot database creation, and a fully verified agent filesystem. intentlab.ai/blog/turn-your-…

If you are interested in Who&When Pro but couldn't join #ICML2026 in person, we built a live paper QA on discord for you! Scan QR code below or join via discord.gg/6adJVU7wNJ to learn more!

So honored to receive the outstanding paper honorable mention for our motion attribution for video generation paper at #ICML2026 This wouldn’t happen without my amazing mentors and advisors! @paschalidoud_1 @JunGao33210520 Antonio Torralba @lealtaixe @orussakovsky @FidlerSanja @jonLorraine9 and many other friends and colleagues, this work would not have been possible without all of you and the huge support from @NVIDIAAI @PrincetonCS @MIT_CSAIL @UMichCSE @UofTCompSci @VectorInst 💪

Autoresearch just left the sandbox and entered the embodied world. We are excited to introduce 𝐄𝐍𝐏𝐈𝐑𝐄: a system that drops frontier coding agents onto a fleet of real robots and hands them the entire loop: reset the environment → search the literature → implement ideas and build the infra → train and deploy → self-verify → analyze the logs and rewrite the code → repeat, until the policy is reliable in the real world. No human in the loop. Guided only by the robot's self-proposed, heuristic-based success signal, the agents hill-climb to 99% on dexterous real-world tasks: organizing pins into a box, seating GPUs, tying zip-ties. We envision the bottleneck in robotics shifting — from building smarter algorithms to building the closed physical feedback loops an agent can finally turn on its own. 🔗 research.nvidia.com/labs/gea… From @NVIDIA @CMU_Robotics @Berkeley_AI 🧵

amazing: soon we can dictate to robots like we delegate tasks to llm agents today

New #NVIDIA Paper We introduce Motive, a motion-centric, gradient-based data attribution method that traces which training videos help or hurt video generation. By isolating temporal dynamics from static appearance, Motive identifies which training videos shape motion in video generation. 🔗 research.nvidia.com/labs/sil… 1/10

Claude Cowork proves the interest in AI coworkers. Orion (meetorion.app) takes the autonomous working experience to another level: • Works with and enhances your experience with your favorite apps (messages, docs, emails, sheets, files, calendars…) and AI tools (ChatGPT, Claude Code, Gemini, Grok…) • Supervises workflows when you are away, from vibe coding to media creation • Coordinates AI agents across all your devices (macOS, windows, VM), each operating one device independently Your personal AI superpower. New surprising ways of vibe working every day - I just tried using Orion to add subtitles to the video and it worked! Watch it handle multiple tasks autonomously from a single cmd+k.

My new blog to wrap up the year: haoruxue.github.io/taas/ Training-time data scaling makes robotics work. Deploy-time data scaling makes robotics business work. My prediction for 2026? A new word to add to robotics startup ideas: teleoperation-as-a-service (TaaS)