Anton Slesarev

Cameras are the eyes of AI. Building Agnify to turn video into intelligence. Former co-founder & CTO of Avride. Previously Yandex

[Anton Slesarev @omgsdgomg]

@alifarhat79 First day outside the training distribution. https://t.co/5jEOh12UD1

[Anton Slesarev @omgsdgomg]

BREAKING NEWS: regular computer vision modesl can't detect flying cars

[Anton Slesarev @omgsdgomg]

@elonmusk Please give the training datasets time to catch up. https://t.co/5jEOh12mNt

[Anton Slesarev @omgsdgomg]

BREAKING NEWS: regular computer vision modesl can't detect flying cars

[Anton Slesarev @omgsdgomg]

BREAKING NEWS: regular computer vision modesl can't detect flying cars

[Anton Slesarev @omgsdgomg]

@n0mad_0 And video ai is video generation

[Anton Slesarev @omgsdgomg]

@Szy_Paula Real floors in sim is the right call. We do the same for live ops: build a 3D model of the plant or warehouse from its existing cameras, then track people and products moving through it on real shifts.

[Anton Slesarev @omgsdgomg]

Breakdown, models behind each step, and CLI templates to run it on your own footage: https://t.co/FM2wYBmK5v

[Anton Slesarev @omgsdgomg]

One ceiling camera over a hand-fed paper-bag press. Same feed answers four questions: who's in shot, how many bags, is the press running, is anyone at the station.

Every overlay here is model output. 🧵

[Anton Slesarev @omgsdgomg]

"Is the operator still there?" is a different question from "is the press running". That one goes to a VLM reading short clips: 6.5 s from walking away to the alert.

People are masked before any other model sees the frame.

[Anton Slesarev @omgsdgomg]

@SoniaBaschez Did you ask your AI husband to build AI wife?

[Anton Slesarev @omgsdgomg]

https://t.co/oHY0f2yGgd

[Anton Slesarev @omgsdgomg]

@yaroslavvb Super cool!

[Anton Slesarev @omgsdgomg]

@hr0nix @Plinz is the best! Loved getting to see his talk in person this summer.

[Anton Slesarev @omgsdgomg]

@lucataco @ArgoSaakyan check it out!

[Anton Slesarev @omgsdgomg]

@polynoamial @OpenAI We saw this in our video tests. Flash was much cheaper than Astra and closer in tracking accuracy than I expected given the price gap.

[Anton Slesarev @omgsdgomg]

@ArtificialAnlys We ran into a different issue on video: Sol and Astra sometimes changed the name of an object between frames, so our tracker treated it as a new object. Giving each object a number helped.

[Anton Slesarev @omgsdgomg]

@sama Sol was faster than Astra in our video tests too, with very little difference in tracking accuracy.

[Anton Slesarev @omgsdgomg]

@OpenAI We saw something similar in our video tracking tests. Sol came close to Astra and cost much less to run. https://t.co/Pt3SwlKZJ8

[Anton Slesarev @omgsdgomg]

Updated the VLM tracker. GPT-6.1 Sol is close to Astra on video tracking and much cheaper.

Closed models still lead on accuracy. Qwen and Nemotron cost less to run, but aren't there yet. Curious to see what @Alibaba_Qwen does next!