[3/3] This work would not be possible w/o Linjian Ma, Xing Liu and Carole-Jean Wu initiating & driving this effort, as well as many contributors across Meta, NVIDIA, AMD, and Stanford (full list in MLPerf post).
Hope to see your results, and please reach out if you are interested in collaborations!
[2/3] The MLPerf RecSys benchmark is uniquely challenging: it needs to represent production workloads while maximally preserving user privacy. Our synthetic data approaches have enabled 20x model size increase, 6,500x FLOPs increase -- while keeping everything open-source.
[1/3] Excited to share that HSTU/Generative Recommenders has been selected w/ DeepSeek-R1, GPT-OSS, etc as part of MLPerf!
Since our initial results in 2024, we’ve seen widespread adoption of related techniques, from LinkedIn, Shopify, Twitter to RedNote, Meituan, Yandex, and Alibaba.
Scaling GPU-Accelerated Databases beyond GPU Memory Size
For the first time, we show that GPU-accelerated database systems can be both faster AND cheaper than their CPU counterparts, with a proof-of-concept on Microsoft SQL Server in Azure running TPC-H 1TB with a single A100/H100! Check out our VLDB'25 paper (www.vldb.org/pvldb/vol18/...) for details!
At #TheWebConf2025 today? Join our talk on next-gen retrieval paradigm at 10:30am in C3.3! Also moderating Responsible Web session (2:30-4pm, C3.4) on misinformation demonetization, cookie compliance & factchecking - crucial topics for building a better internet. See you there!
To some, AGI may be just consolidation of intellectual capital under existing power structures
If bigco knows an image is likely bad, they would not have allowed it to be generated in the first place. So the attacks are most likely coming from (finetuned) oss models anyway.
🧵(4/4) Learn more:
Paper: arxiv.org/abs/2407.15462
Code: github.com/bailuding/rails
Interested in pushing the boundaries of foundational technologies together? Find @bailuding.bsky.social or I at WWW 2025 in Sydney or reach out!
#TheWebConf25 #WWW2025 #RAG #RecSys #LLMs #NeuralRetrieval
🧵(3/4) MoL achieves dense retrieval-level speed on GPUs, while delivering 20-30% better Hit Rate@50-400 on 100M+ item DBs (arxiv.org/abs/2306.04039, arxiv.org/abs/2407.13218). These gains make a compelling case for migrating web-scale vector databases to Retrieval with Learned Similarities (RAILS).
🧵(2/4) Our paper introduces Mixture-of-Logits (MoL) - a universal approximator of similarity functions unifying sparse/dense retrieval, multi-embeddings, and generative approaches. Beyond theory, MoL achieves SotA from sequential retrieval on Transformers/HSTU to finetuning LMs for RAG/QA.
🧵(1/4) Interested in the next generation retrieval paradigm or making your recommendations, search, or RAG/LLMs applications 20-30% better? Check out Retrieval with Learned Similarities, a collaboration with Microsoft Research, accepted as an oral presentation (155 out of 2062 papers) at WWW 2025!
I don’t really have the energy for politics right now. So I will observe without comment:
Executive Order 14110 was revoked (Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence)
Issue is, we've already had a historic opportunity to make social media fully user-controlled and unbiased, yet our lawmakers chose to optimize for tech oligarchs' net worth instead...
The flood of papers in turn pushes flashy marketing and rushed publishing over substance. A self-reinforcing cycle that we should break.
Recent controversies with BlueSky data being scraped for training GenAI models demonstrates that no one has thought through how decentralized social networks are supposed to work at all. Treating all internet engagements as public still seems like a safe default option.
Source email is under Exhibit 13 for anyone else wondering www.courtlistener.com/docket/69013...
Do you happen to have links to the actual emails?
Strange how clarity strikes when flying above the attention economy's reach. The ground is nothing but a notification prison now.
