Glimpse: a visual thinking model
The Hook
“I'm Andrew Dai, co-founder and CEO of Elorian AI, and previously I was at Google DeepMind and originally Google Brain for the past 12 years.”
Product Reveal
“Today, we're very proud to announce our first visual thinking model, Glimpse. Glimpse is the first in a new category of vision-specialized models, systems that are able to reason about the physical world from what they see.”
Feature Teaser
“These models need to follow objects as they move and through changing viewpoints. They need to understand how parts fit together and connect an observation with what happens next.”
Feature Teaser
“One of the reasons Glimpse is able to do this is by employing a visual scratch space. It conditions reasoning steps on transformed images, feeding crops, selected frames, and new visual views back into its reasoning.”
Feature Teaser
“Glimpse delivers frontier-level performance, competitive with Astra across benchmarks spanning counting, scientific chart reasoning, and video and robotic captioning. And it does this at just one-sixteenth the cost of tokens compared to Astra.”
Feature Teaser
“Glimpse was built with a small, talented team using only a few hundred GPUs. And we believe it's an early step towards models that understand the physical world for robotics, science, engineering, and more.”
Call to Action
“(No spoken dialogue — ambient warm synth chord resolve)”
Synthesizable in Remotion · key components of the deep-tech-launch-engine template.
Dynamic 2D bounding boxes snapping around objects with smooth interpolation and attached color tags.
Use Remotion AbsoluteFill with SVG rect elements mapped to frame-interpolated coordinate matrices.
Dot-matrix wireframe grid projecting onto camera background with telemetry readouts.
Render Three.js PointCloud in Remotion with ThreeCanvas, animating point coordinates along a parabolic trajectory.
Layered technical blueprint splitting into neon glowing components along the Z axis.
Stack layered SVG paths with CSS drop-shadow filters, interpolating translateY and opacity via spring curves.
Minimal segmented vertical bar comparison revealing pricing differences with numeral counters.
Map bar height via interpolate(frame, [0, 20], [0, targetHeight]) accompanied by a formatted spring-based numeral counter.
Purpose. Build trust through executive pedigree and research heritage.
Execution. Warm, cinema-lit founder talking head with clean typography.
Purpose. Demonstrate tangible spatial intelligence beyond basic LLMs.
Execution. Robotic manipulation and spatial trajectory prediction HUDs.
Purpose. Explain mechanism of action via visual scratchpad conditioning.
Execution. Blueprint morphs and exploded-view car rendering.
Purpose. Prove parity with market leaders at an order of magnitude lower cost.
Execution. Side-by-side benchmark matrix and clear cost comparison charts.
Purpose. Showcase complex reasoning workflow and close with brand identity.
Execution. Interactive optical physics prompt resolving into clean logo animation.
Objective
Launch of a spatial vision model for autonomous surgical robotics.How the recipe adapts
Replace basketball trajectory and car blueprints with surgical field tracking and 3D organ telemetry, keeping the benchmark split-screens and cost efficiency comparison.Scene durations (7 scenes, 90s)
Key Remotion components
Introducing Griffin, the first AI to pass the video Turing test
The most emotive and controllable TTS model
We just released QVAC TranscribePsy, a speech-to-text model small enough to run inside a phone app. It covers 25 European languages on a laptop and 8 on a phone, and on an iPhone 17 Pro Max's Ne