Yu-Xiang Wang
@yuxiangw_cs
Followers
3K
Following
973
Media
171
Statuses
873
Faculty @hdsiucsd, director of S2ML lab. Visitor @awscloud. Prev @ucsbcs @SCSatCMU. Researcher in #machinelearning, #reinforcementlearning, #differentialprivacy
Joined August 2021
Open educational materials can often go a long way. I benefitted from countless such materials myself (e.g., those from @GilStrangMIT, Steve Boyd & @AndrewYNg). Now it's time to contribute back. Here we go -- 52 videos from three grad-level courses I taught at UCSB. Enjoy.
4
62
314
Can't wait to interrogate them a bit. .
Our open models are here. Both of them.
0
0
3
0
1
3
RT @lyang36: šØ Olympiad math + AI:. We ran Googleās Gemini 2.5 Pro on the fresh IMO 2025 problems. With careful prompting and pipeline desiā¦.
0
118
0
Okay. So the proposed safe AGI are like the āTrisolaransā from the Three-Body problem and human beingās last stand is our ability to plot and scheme? š¤£ .(Sorry for the spoilers to those who havenāt read the series yetā¦ ).
A simple AGI safety technique: AIās thoughts are in plain English, just read them. We know it works, with OK (not perfect) transparency!. The risk is fragility: RL training, new architectures, etc threaten transparency. Experts from many orgs agree we should try to preserve it:
2
1
12
To clarify a bit: Weak AI + Weaker Human > Strong AI + Same Human. Two partners must take turns without communication. Of course, weak human cannot understand strong AI's "brilliant" moves to play along, and strong AI does not know how to dumb up its game.
0
0
1
Wonder how online learning helps statistical estimation? Sunil Madhow and @_dheeraj_b will present AKORN at West Exhibition Hall B2-B3 W-1006. On a high level, it provides an appropriate **attention map** that optimally balances bias-variance locally. Come see us now!
1
0
9
Stronger AI is always better? Weaker chess AI that learns to blunder like their weaker (human) partner easily beats stronger AI in human-collaboration mode.
1
0
7
I am offering a course on *Safety in Generative AI* at UCSD next fall. I'd like to focus on the most important problems today. Mighty #AcademicTwitter: what are some topics that you think I must cover?.
5
3
29
RT @jxmnop: new paper from our work at Meta!. **GPT-style language models memorize 3.6 bits per param**. we compute capacity by measuring tā¦.
0
383
0
RT @xuandongzhao: š Excited to share the most inspiring work Iāve been part of this year:. "Learning to Reason without External Rewards"ā¦.
0
512
0
RT @rui_xin31: Think PII scrubbing ensures privacy? š¤Think againā¼ļø In our paper, for the first time on unstructured text, we show that youā¦.
0
21
0
An upcoming workshop on "Inference Optimization for Generative AI" organized by a few awesome people I worked with at AWS. If you are going to #KDD2025, don't miss the workshop! Link to CfP:
2
2
11
Some winning attacks are surprisingly simple, such as shifting the watermarked image by 7-8 pixelsš This turns out to be very effective for Stegastamp. See more details about this NeurIPS '24 competition here
0
0
2
Okay. Time to reveal. The winning attacks for the black-box track involve humans eyeballing the images by their unique artifacts (small blurry patterns), guessing the associated watermark, and applying the known effective attacks from the white-box track.
The secret is revealed! The six ways images were watermarked in the blackbox track of the #NeurIPS2024 competition for watermark removal attacks. Guess what the most successful attacks were?
2
3
14
The secret is revealed! The six ways images were watermarked in the blackbox track of the #NeurIPS2024 competition for watermark removal attacks. Guess what the most successful attacks were?
0
0
9
Itās very fitting that the watermarking workshop at #ICLR starts with Scott Aaronson. Interesting program for the rest of the day too.
0
0
11
Curious about the state of watermarking and its role in AI safety? I will be at #ICLR Poster 237 (Hall 3) from 3pm this afternoon. Come and have a chat!
0
4
26
It turns out that ARE had an epic comeback in the next talk by Babak Hassibi that generalizes LQR to Bregman Divergence losses ā the best part: it satisfies the separation principle as in LQG (ie one can plug in with Kalman Filter like state estimates).
0
0
3
So the answer is no, but we got a lot of interesting stuffs about ābow tiesā and āhourglassesā. John made a strong case that these things are universal.
1
0
0