Boyang "Albert" Li
@AlbertBoyangLi
Followers
970
Following
2K
Statuses
912
Nanyang Associate Prof, NRF Fellow, #NTUsg. #AI, #ML, Multimodal, Narrative Intelligence. Formerly Baidu & Disney Research. PhD Georgia Tech.
Joined February 2021
I struggled quite a bit, until someone told me to distinguish between what is psychological and what is real in math difficulties.
0
0
2
RT @nouhadziri: 📢 DeepSeek R1 still cannot solve multiplication with 100% accuracy🫠😬 Though it can achieve high scores on hard math questi…
0
153
0
Please encourage your students to apply, if they are interested in pursuing a PhD at NTU. This is a great opportunity to meet their future advisors and establish connections. We especially welcome students from "non-traditional countries", which historically did not send a lot of students to NTU.
0
4
9
Given DeepSeek R1, the next major obstacle for AI is soft reasoning, or reasoning that requires common sense and does not have definite right or wrong answers. The RL reward becomes ill defined. One needs either RLHF or curated MCQs.
0
1
4
RT @jiayi_pirate: We reproduced DeepSeek R1-Zero in the CountDown game, and it just works Through RL, the 3B base LM develops self-verifi…
0
1K
0
RT @NeelNanda5: I have fallen in love with the Google Scholar PDF Reader Extension - you can click on a reference in a paper and see title,…
0
30
0
Today I saw this formula. It's not a student's work.
0
0
1
RT @rao2z: In vs. Out of Distribution analyses are not that useful for understanding LLM reasoning capabilities #SundayHarangue IMHO, the…
0
17
0
Here is my conspiracy theory: If a self-interested person built superhuman AI, the best strategy is to hide it and let it secretly improve every business decision. However, if this person is out of money and others are catching up, the best strategy is to bluff about having built superhuman AI, and get AI regulated so much so nobody can get ahead.
0
0
0
I won't believe it until I see it. Marginal benefit diminishes. We find that, in an LLM, the training-generating-data cycle improves the model for a few iterations, but after that it begins to degrade the model. Regardless, here is a paper on using solved problems to train the next reasoning model:
Gwern thinks it's almost game over "OpenAI may have 'broken out', and have finally crossed the last threshold of criticality to takeoff - intelligence to the point of being recursively self-improving and where o4 or o5 will be able to automate AI R&D and finish off the rest." "Much of the point of a model like o1 is not to deploy it, but to generate training data for the next model. Every problem that an o1 solves is now a training data point for an o3." "I am actually mildly surprised OA has bothered to deploy o1-pro at all, instead of keeping it private and investing the compute into more bootstrapping of o3 training etc. (This is apparently what happened with Anthropic and Claude-3.6-opus - it didn’t ‘fail,’ they just chose to keep it private and distill it down into a small cheap but strangely smart Claude-3.6-sonnet.)" "If you’re wondering why OAers are suddenly weirdly, almost euphorically, optimistic on Twitter, watching the improvement from the original 4o model to o3 (and wherever it is now!) may be why. It’s like watching the AlphaGo Elo curves: it just keeps going up… and up… and up…"
0
0
1
RT @victormustar: Now that we have amazing open source TTS with fast inference, what are you building? https://t.c…
0
84
0
@ZayneSprague Maybe I misunderstood, but problems with asymmetric generation vs. verification are not scarce. Every NP-complete problem is easy to verify but hard to solve. Does CoT work on those? I haven't seen evidence either way, though my intuition is it doesn't.
1
0
1
Congratulations and welcome to NTU!
I am excited to share that I will be joining Nanyang Technological University (NTU), Singapore, as an Assistant Professor in the College of Computing and Data Science starting Summer 2025! @NTUsg @NTU_ccds I am currently recruiting PhD students for Fall 2025. If you are interested in working on robot planning and learning, you can find more information here: Thank you for helping me spread the word! #PhDOpportunity #Robotics #NTUsg
1
0
1
0
0
1
@xwang_lk Not doing a proper review is violating protocol. The question is, what mechanisms are in place to prevent people from doing that? I'm not convinced that honor is the answer. Most reviewers are PhD students who never intended to stay in academia.
0
0
5