Tahmin Mahmud

building something that's cool
doctoral researcher @PurdueECE
Alum @WSUVoiland @BRACUniversity
Ask me questions:

[Tahmin Mahmud @tmhd99]

San Francisco stole my heart— amazing Golden Gate Bridge, China Town’s fried rice and beef steak, and coffee shops that scream ‘stay forever.’

[Archer @flyarcher]

Anduril Thunder for defense. Archer Halo for commercial.

Same clean-sheet autonomous, hybrid-electric VTOL platform, built with @anduriltech to serve two different domains. 

Halo is designed for an array of uncrewed commercial missions across search & rescue, logistics, offshore energy, humanitarian & medical cargo, maritime operations and more.

Marubeni Aerospace Corporation is one of our first strategic launch partners for Halo eyeing cargo and logistics use cases, as well as search & rescue. More partners expected to be announced this year.

https://t.co/Fcovh8X01I

[djcows @djcows]

BREAKING: Haaland has joined Anthropic as Member of Technical Staff

[Steven Brunton @eigensteve]

Optimization Bootcamp now live on YouTube!!!
https://t.co/tx6TLLvfUD

Videos will release every few days, so keep tuned!

Book available on Amazon July 16
Free PDF of book soon!

[Tahmin Mahmud @tmhd99]

🇺🇸🦾🚀

[Tahmin Mahmud @tmhd99]

@tomas_salvo22 lfgg

[Tahmin Mahmud @tmhd99]

@NielsRogge @huggingface W

[Tahmin Mahmud @tmhd99]

@MartinShkreli 😂😂😂

[Tahmin Mahmud @tmhd99]

@sashaprzybylski Can we get mozart in our lab over this summer?

[Tahmin Mahmud @tmhd99]

@ti_morse @SecretaryWright @ScottNolan @generalmatter Great conversation! 👍🏼

[Tahmin Mahmud @tmhd99]

@luanalopeslara Happy Birthday, TWIN 🎉

[Alec Helbling @alec_helbling]

In flow matching, a coupling determines how noise and data samples are paired during training.

The choice of coupling is important because it influences the geometry of trajectories at inference time.

The simplest choice is the independent coupling, where noise and data points are paired arbitrarily. This can lead to curved trajectories as the model averages over many conflicting pairings.

However, if we use optimal transport on batches of pairs, this leads to fewer ambiguous intersections that the model must resolve, leading to straighter trajectories at inference time.

[Tahmin Mahmud @tmhd99]

@LubaYudasina @3blue1brown Great pod!

[Tahmin Mahmud @tmhd99]

@Ravenismeee

[Julia Turc @juliarturc]

Finally found a paper that clarifies the difference between a world model and a general-purpose video generation model:

Framing:
- NVIDIA's Cosmos-Predict = x2world model
- x can be text, image, video, or even robot action
- world = video (in this particular case)

Compared to a general x2video model:
- training data is highly curated (e.g. remove cartoons, focus on autonomous driving, etc)
- swap architectural components with specialized ones (e.g. replace T5 text encoder with a VLM that describes motion in more detail)
- a slew of fine-tunes that either adapt the API for special inputs (e.g. robot action, segmentation map, video from physics engine)

Generally, a very good starting point for making sense of the fuzzy world models landscape. Much more eye-opening than LeCun's work, which is more narrow and opinionated.

[Keller Cliffton @Keller]

Sci-fi Western

[Tahmin Mahmud @tmhd99]

@ashebytes @minafahmi @sandbar This is the coolest bg for a tech pod! 👌🏽

[Tahmin Mahmud @tmhd99]

W

[Laura Ruis @LauraRuis]

Noticed guerrilla posters at iclr; hanging posters on empty spots and presenting like its an accepted paper until the actual assignee to that spot showed up and kicks you off. Definitely a move

[Tahmin Mahmud @tmhd99]

@alexkehr Yes

[Tahmin Mahmud @tmhd99]

Glad to be part of the IEEE IDCD 2026 this spring. I'll have two presentations lined up. Looking forward to good discussions on intelligent drive systems.