
After five incredible years on the faculty at @harvardhbs @hseas — including a year spent at @GoogleAI — I'm leaving to build something new. Back to startups. More at noon PT 🚀

cofounder @elorianai prev: @harvard asst professor, rs @googleai gemini, @welligence cofounder

After five incredible years on the faculty at @harvardhbs @hseas — including a year spent at @GoogleAI — I'm leaving to build something new. Back to startups. More at noon PT 🚀

New work from @ElorianAI let by @dzyan01 on grounding, a core element of visual thinking! Whether grounding helps is determined by whether the reference can be captured by text. This exposes a key weakness of existing benchmarks for visual reasoning -- the relevant visual information can be concisely described in text. To address this we introduce a new benchmark BrickVision.

Excited to have co-led ScienceVision and contributed to DecoyCount! Both explore where visual understanding and reasoning still break down—from interpreting scientific 3D structures to counting everyday objects. Huge thanks to our teammates and collaborators at @UniPat_AI and @mercor! ScienceVision Blog: elorian.ai/blog/sciencevisio… Paper: arxiv.org/abs/2611.04087 DecoyCount GitHub: github.com/elorian-ai/decoyc… Paper: arxiv.org/abs/2610.34855
Tired of seeing benchmarks saying we are at AGI but models still can't understand a 3D protein view or a floor plan? When a visual model gets something wrong, is it failing to see or failing to reason about what it sees? We've built two benchmarks that answers these questions. Explore DecoyCount and ScienceVision here: github.com/elorian-ai/decoyc… and huggingface.co/datasets/Unip… And thank you @mercor and @UniPat_AI for your partnership in building these!

Proud to have contributed to this work with @ElorianAI. Separating what models fail to see from what they fail to reason about makes benchmarks more useful. Thank you to @AndrewDai and everyone who helped bring this research to life.
Tired of seeing benchmarks saying we are at AGI but models still can't understand a 3D protein view or a floor plan? When a visual model gets something wrong, is it failing to see or failing to reason about what it sees? We've built two benchmarks that answers these questions. Explore DecoyCount and ScienceVision here: github.com/elorian-ai/decoyc… and huggingface.co/datasets/Unip… And thank you @mercor and @UniPat_AI for your partnership in building these!

Pushing the frontier in visual thinking requires evaluations that truly test whether models can understand 2D & 3D scenes and reason over them... We're excited to release DecoyCount that measures visual misalignment in counting tasks with @mercor and ScienceVision to benchmark understanding of 3D scientific data with @UniPat_AI
Tired of seeing benchmarks saying we are at AGI but models still can't understand a 3D protein view or a floor plan? When a visual model gets something wrong, is it failing to see or failing to reason about what it sees? We've built two benchmarks that answers these questions. Explore DecoyCount and ScienceVision here: github.com/elorian-ai/decoyc… and huggingface.co/datasets/Unip… And thank you @mercor and @UniPat_AI for your partnership in building these!

can I claim a Chanel bag as a business expense?

glimpse is a big fan of the Divide album
ChatGPT looked at a concert, saw the lights, and decided they were all Yellow from @coldplay. Our new visual thinking AI model, Glimpse, knows the Shape of You, @edsheeran. It got it right on the first try. Try it out!


Coach Glimpse 🏌️🫡
ChatGPT got me into AI via Ghibli style. Golf got me hooked on Glimpse 🏌️♂️ Going from text prompts to frame-by-frame video analysis is next level. Try the demo: elorian.ai/blog/introducing-… #Glimpse #ElorianAI


LLMs are either not capable or not affordable for visual reasoning tasks at scale. The Elorian team - one of Jeff Dean’s best reports who led Gemini, an Apple LLM lead, xAI lead, and a Harvard professor - just launched an Astra-level visual reasoning model, Glimpse, for 17x cheaper. Watch it break down the flying Tesla video physics here. Extremely critical for video inputs in manufacturing, robotics, gaming, content moderation and more.
Spent years watching models ace a single image and completely whiff on anything that moves or changes angle. Today we’re fixing that. Try Glimpse now, our first visual thinking model with competitive frontier performance compared to Astra on key public visual reasoning benchmarks.

My mom somehow also thinks she came up with the name Glimpse, but I think Dustin has receipts here. What should the next model be called??
The name Glimpse was inspired by one of my favorite documentaries: John Berger’s Ways of Seeing (1972). Berger shows how photography and television changed our relationship with centuries of art: a painting could leave the gallery, enter our homes, and take on new meanings through cropping, commentary, or music. Today, AI is changing how we create images and films. I find myself returning to his questions about seeing, authorship, and interpretation. Highly recommend watching. Episode 1 below.

according to Elorian's evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks. spr.ly/6015BHzLYR

Working with Le has been one of the many highlights of building glimpse. He is one of the most talented researchers in multimodal today!
Introducing Glimpse, our first model launch — a multimodal Visual Thinking Model 🚀🚀🚀 Had a lot of fun leading the training effort on this one with such an amazing team! Give it a try — we'd love your feedback. Things are speeding up, more to come!

if @elonmusk could stop building literal flying cars for just 1 day so ppl could focus on our model launch that would be great tysm

Thanks Jeff!! 🫡🫡
Have a glimpse at this! Nice work, @AndrewDai and Elorian folks!