From the Lab

Thinking Out Loud

Research notes and perspectives from the Primate Intelligence team.

Engineering

What Scene Understanding Looks Like in a Real Security Camera Pipeline

A concrete walkthrough of what changes — and what stays the same — when you add Primate Vision to a security camera CV pipeline.

MMMatt Miesnieks··7 min read
Research

Primate Vision vs YOLO v8: A Head-to-Head on Real Security Camera Footage

We ran both models on the same real-world VIRAT dataset clips. Here's what the results actually look like.

MNMehdi Nikkhah··8 min read
Perspective

The False Tradeoff in Computer Vision

For five years, CV teams have been forced to choose between reliable (but rigid) detection and flexible (but unreliable) VLMs. That tradeoff is an artefact of architecture, not an inherent property of the problem.

MMMatt Miesnieks··5 min read
Engineering

Why 'Deterministic' Matters More Than Accuracy in Production CV

Production CV systems don't just need models that are accurate. They need models that are predictable. Here's why that's a harder problem.

MNMehdi Nikkhah··6 min read
Perspective

The $6 Billion Bet on World Models — What It Means for CV Developers

AMI, World Labs, General Intuition — over $6B has flowed into world model companies. Here's what CV developers should actually take away from it.

MMMatt Miesnieks··5 min read
Research

How JEPA Learns to Predict Scenes Without Reconstructing Pixels

A technical explainer of Joint-Embedding Predictive Architectures and why they're fundamentally different from generative video models — and why the difference isn't just architectural.

MNMehdi Nikkhah··11 min read
Announcement

Introducing Primate Vision

Today we're launching Primate Vision — the first JEPA-based scene understanding API for CV developers.

MMMatt Miesnieks··5 min read
Research

Darwin: Video JEPA model that outperforms SOTA models while running on edge CPU

Language models had their GPT moment. Vision, especially video, hasn’t yet!

MNMehdi Nikkhah··5 min read

Subscribe via RSS. Every post is also available as raw markdown — append .md to any post URL.