computer vision companies are dying from long use-case → model → deploy feedback loops.
they run out of runway before they can even find the right use case.
meanwhile CV is being commoditised in real time on sn44.
every niche will be covered.
if you run a computer vision company, the miners are coming for you or are here to help depending on how you want to approach this.
“just use a foundation model” they said.
so we benchmarked 10 of them against Manako on a real forecourt damage clip.
same full 1280x720 frames, same class prompts (person, car, fuel dispenser, fuel nozzle), no crops, no zoom.
the hard class is the nozzle/hose. small, thin, occluded, and the thing that actually gets damaged.
- Grounding DINO labelled a bin as a fuel nozzle.
- Florence-2 labelled the canopy pole as one.
- OWL-ViT, YOLOE and YOLO-World got the dispenser at best, never the nozzle.
- RF-DETR doesn’t have the class.
- Gemma 4 E2B described the scene, boxed the pump, and returned no incident time.
the only open model that localised the nozzle was full SAM3. 3.45 GB, 0.2-1 FPS on a Jetson Orin Nano.
Manako (3 specialists - 29.5 MB), 3-6 FPS on the same Jetson, tracked the nozzle at 5 FPS and fired a live equipment_damage event on the hose drop at 10.3s.
117x smaller than the only model that could keep up.
open-vocab models are great at the classes the internet photographs a lot.
operations live in the long tail:
nozzles, hoses, cabinets, the small parts that break.
that’s where specialised detectors win, and it’s why we built Manako the way we did.
stuff like this is why we stuck with generative world models instead of JEPA-style ones in the end
thank god f/ai got us to talk to google just today, we'll have some fun with Argon i'm sure
In robotics, Gemini 4 Argon is speeding up real-world physical learning.
We let Argon watch a video of a human playing cornhole. It mapped the environment, calculated whole-body joint movements, and autonomously built the simulation code to practice.
When transferred to Apptronik’s Apollo humanoid, the robot sank seven straight shots in a real-world trial.
the autonomy stack that scaled in automotive is now becoming relevant for business operations.
moving from assisted detection to high autonomy requires the same kind of grounded models that made conditional and high autonomy possible in driving.
without that physical understanding, you stay stuck at alerts and supervised tasks.
that’s the gap we're going to close.
Most AI still automates tasks. The next shift is AI that can run business operations.
The progression looks a lot like autonomous driving.
Just as cars moved from driver assistance to partial, conditional, and higher autonomy, business operations will move through the same stages:
0. Manual operations → people observe, decide and act.
1. Assisted operations → AI detects events. People act. (This is where Vision Agents sit today.)
2. Partial automation → AI executes tasks. People supervise.
3. Conditional autonomy → AI runs workflows. People handle exceptions.
4. High autonomy → AI runs a defined operation, including its exceptions.
5. Autonomous operations — AI coordinates and improves end-to-end operations.
The long-term direction is clear: increasing operational autonomy.
People set the goals, the policies and the accountability. AI runs more of the operations.
Humans govern.
At Manako Labs, we’re building World Models for Business so this progression can actually happen in the physical world, starting with the cameras companies already have.
Vision Agents are the entry point.
Levels 2–5 are the ambition.
the other huge gap i keep seeing (even inside the AI community)
people who just read about AI or outsource it to their cto
vs
people who are burning tokens and pushing the frontier every day
it feels a lot like the old days when some of us couldn’t stop taking things apart just to understand how they were built
Andrej Karpathy’s point: Many people only use AI occasionally for search or writing and see little value in their daily lives. Even many professionals understand its progress mainly in theory.
People using advanced models extensively for coding and math experience huge capability gains, widening the gap in how people perceive AI.
And this impression has been confirmed here in Paris during my trip. When I talk to people, I realize they're completely unaware of how capable AI already is. They're unaware of the possibilities it offers. They haven't heard that OpenAI solved a Millennium Prize problem, or recently solved 372 math problems, some of which had remained unsolved for decades.
Currently, we're living in an ivory tower.
