Xiangyu Qi
@xiangyuqi_pton
Followers
2K
Following
4K
Media
27
Statuses
956
Research @openai | PhD @Princeton | Prev @GoogleAI @GoogleDeepMind
Joined December 2019
I was laid off by Meta today. As a Research Scientist, my work was just cited by the legendary @johnschulman2 and Nicholas Carlini yesterday. Iโm actively looking for new opportunities โ please reach out if you have any openings!
282
384
5K
https://t.co/kECiVya8Ii Itโs remarkable that what Dr. Yang observed back in 1988 still holds true in 2025 when comparing the East and the West.
0
0
1
Fine-tuning APIs are becoming more powerful and widespread, but they're harder to safeguard against misuse than fixed-weight sampling APIs. Excited to share a new paper: Detecting Adversarial Fine-tuning with Auditing Agents ( https://t.co/NqMeGSCQIF). Auditing agents search
arxiv.org
Large Language Model (LLM) providers expose fine-tuning APIs that let end users fine-tune their frontier LLMs. Unfortunately, it has been shown that an adversary with fine-tuning access to an LLM...
10
50
462
BTW, Weโre growing a safer foundation model research stack at @tiktok_us โsafety pretraining, RLHF/RLAIF, evals. ๐จIntern + FTE roles. ๐จDMs open.
@ICCVConference vibe kicking in with @liang_weixin and the very shy @xiangyuqi_pton. Batch-norm the smiles, dropout the shyness๐
0
1
20
ChatGPT Atlas is here! Our new browser has ChatGPT built in so it can help you across the web and, if you want, remember what youโve done online and use that context for future requests. More of my thoughts on why we built this here: https://t.co/8Fnx6Stszc
232
259
3K
Yesterday Seb shared our recent work (thanks, @SebastienBubeck) on @OpenAI GPT-5 and its ability to solve simple mathematical conjectures. Since there has been a lot of discussions, let me clarify a few points. We ask whether large language models can handle new but very simple
0
9
92
This started as a fun personal project. The sample is small, so nothing definitive, but a few patterns emerged when we put GPT-5 to the test. โข When the path was clear, it did great: nearly correct proofs in 3/5 problems. โข On Problem 2, it surprised us with a new
It's becoming increasingly clear that gpt5 can solve MINOR open math problems, those that would require a day/few days of a good PhD student. Ofc it's not a 100% guarantee, eg below gpt5 solves 3/5 optimization conjectures. Imo full impact of this has yet to be internalized...
6
11
143
Announcing strategic partnership with @nvidia for millions of GPUs โ about as much compute as they've shipped in 2025 in total โย and an investment up to $100B as these GPUs are deployed:
openai.com
OpenAI and NVIDIA announce a strategic partnership to deploy 10 gigawatts of AI datacenters powered by NVIDIA systems, with the first phase launching in 2026.
121
230
2K
Congrats to the team on another ๐ฅโwith a perfect score! A fitting way to close a chapter where intellectual competitions defined the frontier. Today, new horizons beckon. I'm glad our โจexperimental reasoning modelโจ (same one from IMO/IOI) got one last golden run!
1/n Iโm really excited to share that our @OpenAI reasoning system got a perfect score of 12/12 during the 2025 ICPC World Finals, the premier collegiate programming competition where top university teams from around the world solve complex algorithmic problems. This would have
12
17
364
1/n Iโm really excited to share that our @OpenAI reasoning system got a perfect score of 12/12 during the 2025 ICPC World Finals, the premier collegiate programming competition where top university teams from around the world solve complex algorithmic problems. This would have
140
449
3K
Some quick thoughts on the recent copyright litigation developments: "Anthropic Settles Its Copyright Litigationโand Why That Was the Right Move" ๐๐
1
1
5
What happens when AI is guided by law-like principles? Can we design some computational tools to "debug" rules? Check out our new work ๐ ๐๐ฅ๐๐ฅ๐ฆ๐ฅ๐ ๐ฃ๐ช โ๐ ๐๐ค๐ฅ๐ฃ๐ฆ๐๐ฅ๐๐ ๐ ๐๐๐ ๐๐๐ฅ๐๐ฃ๐ก๐ฃ๐๐ฅ๐๐ฅ๐๐ ๐ ๐๐ ๐ฃ ๐ธ๐๐งโโ๏ธ to find out more! ๐งต(1/10)
1
26
88
New research explains why LLMs hallucinate, through a connection between supervised and self-supervised learning. We also describe a key obstacle that can be removed to reduce them. ๐งต https://t.co/6Lb6xlg0SZ
openai.com
OpenAIโs new research explains why language models hallucinate. The findings show how improved evaluations can enhance AI reliability, honesty, and safety.
105
344
1K
๐กNew on the CITP Blog: "Statutory Construction & Interpretation for AI" > What if an LLM concludes a user's behavior is โegregiously immoral" -- & contacts authorities? CITP researchers with Prof @PeterHndrsn's POLARIS Lab provide a possible explanation.๐๐@PrincetonCS
1
5
7
Meanwhile, ICMLโs submission deadline always seems to land right around Chinese New Year ๐
ACL 2026, a top tier NLP conference, being over the July 4th weekend is so bizarre (4th 2026 is a Saturday). I'll be turning down any more talk invites in lieu of touching grass. People should probably consider submitting elsewhere.
0
0
15
Wonder why Claude decided to report users to the authorities? It might be because its constitution says Claude should choose responses in the long-term interest of humanity! But what if we could leverage computational and legal tools to "debug" or "lint" AI rules/laws for
3
9
27
๐๐๐ I'm excited to be on the faculty job market this fall. I updated my website with my CV. https://t.co/4Ddv6tN0jq
stephencasper.com
Visit the post for more.
8
22
172
I'm starting to get emails about PhDs for next year. I'm always looking for great people to join! For next year, I'm looking for people with a strong reinforcement learning, game theory, or strategic decision-making background. (As well as positive energy, intellectual
2
31
245