Pavan Kalyan

Founder & CEO @ TensorTest (YC W26). Prev Microsoft Research; IIT Bombay; NeurIPS/ACL/Interspeech.

[Pavan Kalyan @tpavankalyan23]

@GuninGupta907
@NirmitArora26

[Pavan Kalyan @tpavankalyan23]

1/5

We read the last message from 896 AI video-editing runs.

93% said the job was done. 14% passed.

"Done" tells you nothing.

[Pavan Kalyan @tpavankalyan23]

5/5

Training a video or computer-use agent? We'll run it privately and show you what it claims next to what it ships: https://t.co/PoDs1QSCGe

[Pavan Kalyan @tpavankalyan23]

4/5

For anyone training agents: self-report can't be the reward or the stop signal. The verifier has to look at the cut.

TimelineBench: 56 real editing jobs, scored against 43 professional editors.
Paper: https://t.co/w6zqAJ3Kt0
Leaderboard: https://t.co/GjtI2SQv4r

[Pavan Kalyan @tpavankalyan23]

@CSProfKGD Give LLMs a reference and they’re exceptionally good at copying. We saw this while building TimelineBench. But reproducing a reference isn’t the same as solving a real-world task from a brief. That’s where they still struggle
https://t.co/w6zqAJ3Kt0

[Pavan Kalyan @tpavankalyan23]

@rezoundous everyone these days!

[Pavan Kalyan @tpavankalyan23]

@nitin_builds SI

[Pavan Kalyan @tpavankalyan23]

@ashoKumar89 Resign before getting fired

[Pavan Kalyan @tpavankalyan23]

@AbhiCodes15 Learn to learn from every ai coding chat.

[Pavan Kalyan @tpavankalyan23]

@moiiikaaa Codeveloper

[Pavan Kalyan @tpavankalyan23]

@himanshustwts This is so truuu

[Pavan Kalyan @tpavankalyan23]

Am i the only one not looking at linkedin messages anymore? New linkedin is needed :/

[Pavan Kalyan @tpavankalyan23]

The performance review: Opus 5.5 in Claude Code resolved 18 of 56 real video-editing jobs.
Leaderboard: https://t.co/GjtI2SQv4r
Paper: https://t.co/w6zqAJ3Kt0
@GuninGupta907 @NirmitArora26

[Pavan Kalyan @tpavankalyan23]

Opus 5.5 made an ad for its own performance review.

Bold career move. 📷

Considering it only scored 32.1% on TimelineBench.

[Pavan Kalyan @tpavankalyan23]

@saenxz @carrynointerest Was planning to lurk, but that’s a strategically important “me.” Mind if I DM?

[Pavan Kalyan @tpavankalyan23]

@lennysan @tryramp Spot on! we need more evals to check an agent’s “done” actually means done. we observed the same while developing TimelineBench, agents claimed success on 93% of failed edits.
Paper - https://t.co/w6zqAJ3Kt0

[Gunin Gupta @GuninGupta907]

Introducing TimelineBench: Evaluating Agents on Realistic Video-Editing Tasks, from Raw Footage to Final Cut

We tested 16 agents on 56 real video-editing jobs. The best passed just 26.8%

https://t.co/RmCUx8aroL

[Pavan Kalyan @tpavankalyan23]

I'll be at Sarvam Epoch: India's Frontier AI Conference. July 30-31, Bengaluru. https://t.co/gttkHzpqPJ