Tuan-Duy Hien Nguyen (@duynht) header
Tuan-Duy Hien Nguyen (@duynht) profile picture

Research Potato in CS @NUSingapore

Ho Chi Minh, Vietnam
Joined March 2019
Photo by Tuan-Duy Hien Nguyen (@duynht)
I’m happy and grateful to share that I’ve started my PhD @NUSingapore under the @AISingapore PhD Fellowship!
2
9
714
Tuan-Duy Hien Nguyen retweeted
First up a hackathon: 20+ researchers · two weeks · @OxfordStats × @NUSComputing × @NTUsg and tokens on tap by @AnthropicAI . Outcome: Five open‑source AI agents solving common workflow frustrations across the whole research lifecycle. 🧵
6
14
2
22
5,943
Tuan-Duy Hien Nguyen retweeted
🎉Excited to share our new work: "StochasTok: Improving Fine-Grained Subword Understanding in LLMs"!🎉 We rethink tokenization and allow LLMs to ‘see inside’ individual tokens by stochastically splitting them into equivalent pairs. See👇! Thanks to my incredible co-authors!🥰
🚀Introducing “StochasTok: Improving Fine-Grained Subword Understanding in LLMs”!🚀 LLMs are incredible but still struggle disproportionately with subword tasks, e.g., for character counts, wordplay, multi-digit numbers, fixing typos… Enter StochasTok, led by @anyaasims! [1/]
1
10
26
3,440
Tuan-Duy Hien Nguyen retweeted
🚨My awesome labmate Yiran @yiran_zhao924 is presenting our #ACL2024NLP poster: “Prompt Optimization via Adversarial In-context Learning”. 🚀 Want to chat about optimising LLM hard prompts by Adv. Learning? 🌟Please check out his presentation on Tue. 5-6pm, Poster Sec. 5!
1
3
15
986
Tuan-Duy Hien Nguyen retweeted
NUS CS Research Week researchweek.comp.nus.edu.sg… is in full swing with Hamed Hassani @HamedSHassani now giving a talk on robust machine learning.
3
53
3,782
Tuan-Duy Hien Nguyen retweeted
Very excited about hosting many distinguished researchers for our upcoming @NUSComputing CS Research Week, Jan 10-12, 2024. researchweek.comp.nus.edu.sg… The event is public and needs no formal registration.
13
83
10,475
Tuan-Duy Hien Nguyen retweeted
My DGX Station is the most gorgeous piece of technology that I’ve ever had.
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
113
35
12
1,131
496,654
Tuan-Duy Hien Nguyen retweeted
“Remote work was a mistake, we need to return to the office.”
30
442
36
5,309
428,433
Tuan-Duy Hien Nguyen retweeted
The rise of LLMs has been the main push for the adoption of vector databases. Turning text (or even multimodal) data into an embedding vector can significantly improve performance of any AI application. However, it is quite possible that we can get even better performance if we first build that basis workhorse of the Internet era - a search engine. Read more in the following blog post: “Build a search engine, not a vector DB” blog.elicit.com/search-vs-ve…
15
36
11
509
97,582
Tuan-Duy Hien Nguyen retweeted
📢 My book on Fairness and ML with @s010n and Moritz Hardt has been published by @mitpress! mitpress.mit.edu/97802620486… The draft (fairmlbook.org) has already been used in many courses. We've also released a set of exercises and discussion prompts: docs.google.com/document/d/1…
13
205
5
941
82,417
Tuan-Duy Hien Nguyen retweeted
We have been working on fairness problems in federated learning. We showed that FL may propagate bias from the most biased parties to the less biased parties. Check our paper here arxiv.org/abs/2309.02160 and a blog post here montrealethics.ai/bias-propa….
2
14
1,327
Tuan-Duy Hien Nguyen retweeted
realtime autoregressive poster generating
12
83
11
1,027
106,263
Tuan-Duy Hien Nguyen retweeted
bridgit mendler is THE ONLY person who can rival that korean american Navy Seal Doctor Astronaut guy, and it's not close she's only 30 so she has time... Harvard JD MIT PhD and Masters USC BA Space Tech Startup Disney Channel Lead Star @bridgitmendler
CAN WE TALK ABOUT HOW BRIDGIT MENDLER FROM GOOD LUCK CHARLIE USC undergrad, MIT masters and PhD, HARVARD LAW??? Oh And she’s working on a Stealth startup here in palo alto
14
45
7
419
125,412
Tuan-Duy Hien Nguyen retweeted
I'm happy to announce the start of a new free and open online course on neuroscience for people with a machine learning or similar background, co-developed by @MarcusGhosh. YouTube Videos and Jupyter-based exercises will be released weekly. There is a Discord for discussions.
22
375
23
1,664
274,867
Tuan-Duy Hien Nguyen retweeted
Replying to @brickroad7
Hey! I'm so glad you found this exciting -- I've been really excited about how it might help and delighted to be able to share it. With that said, I do want to caution that a lot of work remains to be done. teapawt.org/ch402/status/171046145…
Replying to @ch402
Just to be clear: - I'm very optimistic, but I don't think superposition is "fully solved". - Superposition was the biggest bottleneck on mech interp in my view, but not the only one. - Mech interp may help us tell if models are safe, but it doesn't solve alignment in itself.
2
30
5,335
Why does this read to me as "Just grab VC money without building usable and scalable products"? Maybe I am missing something.
Replying to @thedanigrant
It's a myth that VCs need you to have a technical co-founder You just need an answer to "How are you going to build your product?" “We'll prototype with no-code tools then hire an engineer” is 100% OK. VCs want to see you can build a team, not write code
1
1
293
Tuan-Duy Hien Nguyen retweeted
New paper! We ask: How important is VQ for neural discrete representation learning? 🤔 I always felt it should be possible to “absorb” the VQ on top of a deep net into the net and use a simple grid-based quantization instead, without sacrificing much expressivity. Summary 🧵👇
7
56
3
336
66,266
Tuan-Duy Hien Nguyen retweeted
I ask DALLE-3 to generate a Pepe but each time I tell it to make it "more rare."
191
971
220
11,049
1,765,161
Tuan-Duy Hien Nguyen retweeted
also, the video we made for dalle 3 is SO CUTE:
876
3,385
1,301
27,340
4,996,752
Tuan-Duy Hien Nguyen retweeted
Here's a tutorial that I wrote about training and fine-tuning LLMs. I try to make the motivation and details of RLHF clearer and discuss a new method called direct preference optimization. borealisai.com/research-blog… Hope this is helpful. Feedback welcome.
3
65
1
316
46,528
A response to a student who was surprised that I still bother working on "old" (non-AI/ML) topics in graphics.
16
94
7
824
107,416