By the Smili Phone team
On Friday, October 16, a Palo Alto startup called Elorian released its first AI model, and it has one unusual trick. The model, named Glimpse, is designed to actually look at images instead of glancing once and describing them from memory. It launched as a research preview with an API and a chat interface, and its approach to vision is different enough to be worth understanding.
Today's AI mostly describes photos from memory
Here is the problem Glimpse is trying to solve. Most multimodal AI models look at an image once, convert it into tokens, and then reason about it in words, like describing a photo from memory. That works fine for simple questions, but it falls apart on fine details.
Elorian's CEO, Andrew Dai, a former Google DeepMind researcher, found this out on game night. He photographed a board game he had just finished with friends and asked a frontier model who was winning. It could not tell. Board games each have their own visual language of icons and symbols, he told TechCrunch, and models cannot intuit what those symbols mean the way people can. His analogy: humans can look at wordless Ikea diagrams and assemble furniture, so why can AI not do the same?
The visual scratch space
Glimpse keeps going back to the source image while it thinks. Mid-reasoning, it can crop into a region for a closer look, pull specific frames from a video, or run code to measure or compare what it sees. Elorian calls this a visual scratch space. The closest everyday comparison is pinching to zoom on a phone photo to double-check what you thought you saw.
The idea is not entirely new. OpenAI added crop-and-zoom during reasoning to its o-series models in 2025. Elorian's pitch is that it built its whole model around the technique, and did it cheaply: a small team trained Glimpse on a few hundred GPUs, far less compute than the frontier systems it is benchmarked against.
What it can do
According to Elorian's own evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tests. Those are company-run benchmarks, so take them with the usual grain of salt, but the direction is clear.
The first customers Elorian is chasing are companies sitting on piles of video: robotics firms, scientific labs, media and creative companies. Longer term, the goal moves from seeing to doing, building a model of a scene and using it to choose an action, including controlling robots directly. Next year, Dai says, the aim is the world's best visual reasoning model.
Why your phone's camera is the real story
Here is the part that matters for the rest of us. Every visual AI trick, from the photo search on your phone to the apps that identify plants and translate signs, starts with a camera lens. A scratched or cracked camera lens feeds the AI blurry, flared input, and no amount of clever software fully fixes a bad picture at the source. We see it in the shop here in Escondido: photos that went soft or hazy right after a drop, even though the screen looks fine.
As AI gets better at truly seeing, the quality of the eyes matters more, not less. Keep the lens clean, use a case that guards the camera bump, and if the glass over your camera is cracked, get it fixed before it costs you the photos and the AI features that depend on them.
Related: running AI on your own Mac · Google's AI models for European researchers
