Wei Zhen Teoh

Independent AI research, generative models, formerly at @LyrebirdAi @descript, @LFC fan, blog:

[Wei Zhen Teoh @wez_teoh]

For trajectory forecasting models to be useful in group-dynamics domains e.g. sports analytics, they must make jointly probable predictions across all agents. CausalTraj is exactly a model built for that. [Accepted to the AI4TS workshop @ AAAI 2026]
Link: https://t.co/b56fBPiYQY

#AAAI #AI4TS

[Wei Zhen Teoh @wez_teoh]

@karpathy I find this a double-edge sword at times. It distills the mess of thoughts I have but at times also inserts opinions that I may adopt unconsciously if I’m not super aware about every line it outputs.

[Wei Zhen Teoh @wez_teoh]

Coaches often sketch basketball plays on a whiteboard and mentally simulate how teammates and defenders might react.

I explored using diffusion models to learn this kind of controllable gameplay simulation.

One intuition I had is that diffusion models repeatedly refine the entire trajectory jointly over many denoising steps, rather than committing to player movements step-by-step like autoregressive models. This makes sparse sketch-conditioned control feel surprisingly natural.

Write-up of my intuitions and implementations, code, and model available in this blog post:
https://t.co/M0CHAdrPH6

#diffusion #generativemodels #trajectory #simulation #sportsanalytics

[Wei Zhen Teoh @wez_teoh]

Tried to build a clearer mental model of the training dynamics of diffusion models. It helped to view them through the lens of Variational Autoencoder.
Wrote up my understanding here: https://t.co/Eke6FMl56P

[Wei Zhen Teoh @wez_teoh]

@Lfc_ola_ Raul Meireles

[Wei Zhen Teoh @wez_teoh]

https://t.co/X7IBVbyRoK

[The Straits Times @straits_times]

Avoidable deaths, excuses, selfish behaviour: Why Singapore’s road culture needs to change https://t.co/h8HpWpFQxP

[Wei Zhen Teoh @wez_teoh]

@wandb is your server down?

[Wei Zhen Teoh @wez_teoh]

@sedielem @M2lSchool this is a really neat introduction to the subject!

[Wei Zhen Teoh @wez_teoh]

Generative AI to simulate gameplay in team sports! I am currently presenting my work at the AI4TS workshop at #AAAI2026

📍Expo Level 2 Foyer, Slot WS176

Do drop by if you are interested about applying AI in sports analytics!

@RealAAAI

[Wei Zhen Teoh @wez_teoh]

I will present my work - Coherent Multi-Agent Trajectory Forecasting in Team Sports with CausalTraj at the AI4TS workshop, AAAI 2026 tomorrow. Feel free to drop by and speak to me if this work interests you!
#AAAI2026 #Singapore @RealAAAI

[Wei Zhen Teoh @wez_teoh]

For trajectory forecasting models to be useful in group-dynamics domains e.g. sports analytics, they must make jointly probable predictions across all agents. CausalTraj is exactly a model built for that. [Accepted to the AI4TS workshop @ AAAI 2026]
Link: https://t.co/b56fBPiYQY

#AAAI #AI4TS

[Wei Zhen Teoh @wez_teoh]

Thanks for reading. CausalTraj is a step towards better joint multi-agent trajectory forecasting for team sports. Check out the paper, code, and demos:

paper: https://t.co/DTaofdook8
code: https://t.co/cyP78O7Dv6
project page: https://t.co/Fu5PZZ6lkY

[Wei Zhen Teoh @wez_teoh]

For trajectory forecasting models to be useful in group-dynamics domains e.g. sports analytics, they must make jointly probable predictions across all agents. CausalTraj is exactly a model built for that. [Accepted to the AI4TS workshop @ AAAI 2026]
Link: https://t.co/b56fBPiYQY

#AAAI #AI4TS

[Wei Zhen Teoh @wez_teoh]

Qualitatively, CausalTraj produces more coherent gameplay scenario samples than previous SoTA works: players move in strategic formations, adjust directions to track opponents or teammates; the ball tends to follow realistic straight-line passes rather than unrealistic arcs.

[Wei Zhen Teoh @wez_teoh]

as an independent researcher I am grateful to @vast_ai for their affordable compute in asian region at the moment

[Wei Zhen Teoh @wez_teoh]

@FabrizioRomano And Liverpool sold Ben Doak to Bournemouth to replace Ouattara, now it all makes sense