The best demonstration of understanding is action. MolmoWeb extends Molmo2’s visual understanding → taking actions in the browser. What started as a small 2 person effort grew over a year—alongside my kiddo, who recently turned 1 Grateful and proud of the team at Ai2 that made this possible and of course, my better half for being my guide through it all ❤️

Come help us build: hark.com/careers

Just wrapped up my first week at Hark! The team, the speed, the conviction, the product vision — been loving it. I'll be working with @adcock_brett and the team to shape our computer-use (CUA) efforts. I am passionate about the future of personal computing — how we interact with our computers and what we expect of them. And that's exactly what Hark is building: personalized intelligence that integrates seamlessly into your life and brings frontier-level task planning and execution to your everyday goals, big and small. Excited to have found a home at Hark and to learn and grow with the team. And if this resonates with you, we are hiring! Link in the thread.

Ending the day with my wife with delicious pizza, pasta, and a generous serving of “oh shit…this is real now”

End of an era! Thank you, Ai2 (@allen_ai), for being my home these last six years. My time here has been professionally transformative but also marked with monumental life events — I met my partner, got married, bought our first home, had a kid. Those moments are etched in my memory right alongside the work. My last day was overwhelming in the best way: saying goodbye to wonderful people, sitting with fond memories, making my last latte. I feel lucky to have been sent off with so much warmth and grace. I want to wish everyone at Ai2 — the senior leaders, but especially the younger champions — all the best in shaping the organization and its mission in the years ahead. ❤️ (More on my next chapter soon)

@zeddotdev now for my wishlist - I have 3 main requirements that aren’t quite well supported yet - 1. Allow searching by hostname when opening remote projects - i often need to ssh to different machines and “open remote project”. This works but the UI shows this large list of servers with no way to search and it’s not even sorted by hostnames. Scrolling through to find the server i want to connect to gets annoying. I know i can do zed ssh://host/path but would be nice to be able to do this in the UI as well. Also, the cli opens the remote project in a different window instead of existing one even with -e 2. Attaching to a running docker container on a remote machine - some of my workflows involve connecting to a remote server where i run code inside a container. I would like to be able to attach Zed directly to the code inside the container. If we can get the agents panel running inside the attached session (meaning claude code sees files inside the container) that would be a plus! 3. Markdown - latex support and better rendering of front matter. Add live preview to that and you have replaced Obsidian for me!

I switched to Zed (@zeddotdev) in late 2024. Some useful features that I had come to expect from VS Code were missing in Zed back then (eg. git and agent support was nascent). But there was an elegant minimalism and snappiness to the Zed UI that made it hard for me to stop using it even though sometimes I still reached for VS Code for certain things. In the last 6 months I have completely switched to Zed and haven’t missed VS Code. In fact, I have worn my coworkers’ ears off yapping about Zed over the past few months. With parallel agents and other recent updates and fixes, I just can’t imagine going back to VS Code. I still have a wishlist for features I would like in Zed (see thread) but really happy for @zeddotdev team for v1.0 launch. Congratulations on making a beautiful product and keep cooking! 🚀

Computer/web-use agents will redefine how everyone—from our grandparents to our children—interface with computing devices. MolmoWeb is a step towards building this technology in the open. Read the blog for a behind-the-scenes look into why we built it and where it might be going…with a sprinkling of enthusiastic quotes from your truly!

MolmoWeb extends the same visual grounding idea into the browser. Given an instruction & a screenshot, it predicts the next action—clicking, typing, & navigating from the visual interface alone. ↓ youtube.com/watch?v=rzkBE8J7…

When we released Molmo, it was a bet that open vision-language models could compete with closed systems. Since then, we’ve expanded Molmo into a family of open visual AI building blocks for pointing, web interaction, 3D perception, & robotics. 🧵

You can now train, adapt, and eval web agents on your own tasks. We're releasing the full MolmoWeb codebase—the training code, eval harness, annotation tooling, synthetic data pipeline, & client-side code for our demo. 🧵

72hrs after the release, looking at the community’s excitement around MolmoWeb, I have been reflecting on what leading this project throughout the past year was actually like. It didn’t feel like winning. It felt like a constant uphill battle. Making the case that this is worth building. Building a team around the project from the ground up. Working through compute constraints and org-wide competing priorities. Showing early demos that didn’t quite land. And so on. But reading people’s comments, it is clear that builders wanted an open web agent they could run locally. They wanted MolmoWeb. For me, it is a powerful reminder that sometimes you must go against the grain. Sometimes you must work in silence until your results can speak for themselves. If you are wrong, you will learn. If you are right, you might just give the world what it needs.

MolmoWeb picked my night out from comedy shows in seattle. even my weekends run on open-source now. see y'all at Moore.

Excited to share my hopes for AI research in 2026 in the New Year issue of The Batch, alongside a fantastic lineup of researchers and thought leaders! TL;DR: Models that predict are not the same as systems that act. We need more of the latter. Also, great to see alignment with @AndrewYNg’s proposal of Turing-AGI Test as a measure of progress for increasingly intelligent systems. Big thanks to @tedgreenwald (editor of The Batch) for the opportunity and valuable feedback!

Great analysis below to support the hypothesis that “visual” web agents would generalize better than DOM / Axtree with concrete examples of when and how DOM fails the agent! Intuitively, I find “visual” web agents more appealing for 3 reasons: 1. Humans have developed a universal visual design language for web pages. We generally know where to expect navigation bar, search boxes, social media links, T&Cs etc on any web page. We know that clicking on hamburger icons reveals more content, calendar icon in a form lets you choose a date using a grid of numbers, forms usually have a search or submit button nearby, and so on. 2. Market forces enforce visual homogeneity within a website category. For eg, a new shopping website is likely to look visually similar to other shopping websites so that it’s intuitive to use for online shoppers worldwide. 3. Code behind these visually similar web pages on the other hand could look very different depending on web frameworks used, programming choices, and quirks of the developer. And to weigh in on the discussion about re-designing web exclusively for agents — to me it seems more likely that web would evolve to support “co-habitation” of people and agents. Re-designing the web solely for agents is the digital equivalent of terraforming Earth for robots instead of building humanoid robots that learn to exist alongside and assist humans in a world designed for humans.

As academic researchers releasing a scientific body of work, they didn’t have to include this cinematic shot of the NY subway Or, this poetic introduction - “The world sees us as we see it. It streams through us…and then sometimes something stands out - a shape lighter than air, a sudden flash of color…” But they did 🤩

Doctoral Consortium starting soon at 11am in 301 A B! If you are presenting a poster you can come by and set it up earlier. Mentors see you all shortly.

Had the surreal experience of telling a room full of computer vision researchers at the ICCV25 AC workshop why “computer vision researcher” won’t be a thing in 5 years 🌶️ Of course, this was an extreme stance to keep things lively in a fun debate setting but it echoed some of my own internal monologue over the past few years as someone who has identified as a computer vision researcher for the last decade. The argument went as follows: ⚡️a research community needs a set of core problems and methods that are specific to that community ⚡️the vision community used to have these 10-15yrs ago but today’s general purpose multimodal architectures assume very little about the input/output modality and are likely to subsume more tasks and modalities over time ⚡️time and again we have had to swallow the bitter pill — methods that bake human intuition into learning algorithm might show gains in the short term but are eventually surpassed by more general methods that utilize more data and compute - llms, vlms, sora, genie etc ⚡️gains in vision systems over the last many years have come from things that have nothing to do with vision or images but general advances in deep learning - optimizers, normalization layers, attention, residual connections, quantization, parallelization methods, larger models etc. Computer vision ends at tokenization and then deep learning and distributed systems engineering take over. ⚡️so not only would “vision researcher” be obsolete, we must actively fight the urge to play a “computer vision researcher” to avoid our biases from creeping into our AI systems ⚡️in short, there is nothing uniquely vision in today’s computer vision research and there is too much overlap with other specialized communities like robotics, graphics etc Thanks again @ICCVConference PCs for hosting the debate and @anikembhavi for inviting me to participate! It was incredibly awesome for everyone at the AC workshop to take this discussion in a fun spirit 🙌 Arguing for the motion with me were @sarameghanbeery and @RoozbehMottaghi In our opposition were @HildeKuehne @aagrawalAA and @bluevincent If you are a vision researcher, share your thoughts whether you agree or not!

If you are a near-graduation PhD student in computer vision, consider applying to the ICCV 2025 Doctoral Consortium (DC). It is a chance to be mentored by an experienced researcher in the vision community to help you transition to your post-PhD career in academia or industry. It is also an opportunity to share your work with other senior researchers and peers. First time hearing about DC? 🤔 You still have time to apply! A lot can be done in a single day and you have 10 DAYS. Start now and get your application in before Aug 8 🚀 @anna_kukleva_ and I are honored to be your ICCV DC chairs this year and we look forward to your applications! #ICCV2025