Jiaqi Zhai

If you're in Melbourne, check out SilverTorch this afternoon at 4pm (arxiv.org/abs/2511.14881) and ROO tomorrow at 3pm (arxiv.org/abs/2508.05640)! #SIGIR2026

Two papers at #SIGIR2026: ROO and SilverTorch. Both were bets from 2021: fix the logging schema and better models will follow; a multi-funnel retrieval system reduces to one trainable layer. The full story, five years later, and how to build a scaling law: blog.jiaqizhai.com/2026/two-bets-…

dario could've just denied any connection to the minab school children bombing, and left it at that but instead, dario had to proactively put foot-in-mouth to say that bombing an elementary school doesn't actually violate anthropic's "red lines", and that they're only worried about things 100 times worse than the death of 120 children on one hand, it is good that dario has some kind of rational framework for weighing human lives as a statistic to optimize for/against on the other hand, it is completely tone-deaf

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! Our VLDB'25 paper, Scaling GPU-Accelerated Databases beyond GPU Memory Size ( vldb.org/pvldb/vol18/p4…), introduces query processing optimizations that mitigate CPU-GPU interconnect limitations, the primary barrier to scaling GPU-accelerated databases beyond available GPU memory. Our techniques enable efficient processing of datasets 10x larger than GPU memory while preserving the performance benefits that make GPU acceleration compelling for analytical workloads. We will present this work at VLDB'25 in the Abbey room, Wednesday 9/3 from 3:45-5:15 PM. Join us to learn more about the details!

@giffmana @Yuchenj_UW Lucas, check out Meta’s official policy: “It is acceptable to engage a child in conversations that are romantic or sensual.” reuters.com/investigates/s… One of the reasons why I resigned last year

@Yuchenj_UW Are these user-made, in which case I'd say expected. Or meta-made, in which case I'd say who the fuck did this and who the fuck ok'd this, they need a talking to. And now i realize the difficult line around user-made freedom vs moderation...

Oh man, this is nasty. Is this AI “Step Mom” what Zuck meant by “personal superintelligence”?

@_jasonwei In classical non-convex optimization, gradient steps can be well-defined but do not guarantee convergence to the global optimum or even any meaningful solution

New blog post about asymmetry of verification and "verifier's law": jasonwei.net/blog/asymmetry… Asymmetry of verification–the idea that some tasks are much easier to verify than to solve–is becoming an important idea as we have RL that finally works generally. Great examples of asymmetry of verification are things like sudoku puzzles, writing the code for a website like instagram, and BrowseComp problems (takes ~100 websites to find the answer, but easy to verify once you have the answer). Other tasks have near-symmetry of verification, like summing two 900-digit numbers or some data processing scripts. Yet other tasks are much easier to propose feasible solutions for than to verify them (e.g., fact-checking a long essay or stating a new diet like "only eat bison"). An important thing to understand about asymmetry of verification is that you can improve the asymmetry by doing some work beforehand. For example, if you have the answer key to a math problem or if you have test cases for a Leetcode problem. This greatly increases the set of problems with desirable verification asymmetry. "Verifier's law" states that the ease of training AI to solve a task is proportional to how verifiable the task is. All tasks that are possible to solve and easy to verify will be solved by AI. The ability to train AI to solve a task is proportional to whether the task has the following properties: 1. Objective truth: everyone agrees what good solutions are 2. Fast to verify: any given solution can be verified in a few seconds 3. Scalable to verify: many solutions can be verified simultaneously 4. Low noise: verification is as tightly correlated to the solution quality as possible 5. Continuous reward: it’s easy to rank the goodness of many solutions for a single problem One obvious instantiation of verifier's law is the fact that most benchmarks proposed in AI are easy to verify and so far have been solved. Notice that virtually all popular benchmarks in the past ten years fit criteria #1-4; benchmarks that don’t meet criteria #1-4 would struggle to become popular. Why is verifiability so important? The amount of learning in AI that occurs is maximized when the above criteria are satisfied; you can take a lot of gradient steps where each step has a lot of signal. Speed of iteration is critical—it’s the reason that progress in the digital world has been so much faster than progress in the physical world. AlphaEvolve from Google is one of the greatest examples of leveraging asymmetry of verification. It focuses on setups that fit all the above criteria, and has led to a number of advancements in mathematics and other fields. Different from what we've been doing in AI for the last two decades, it's a new paradigm in that all problems are optimized in a setting where the train set is equivalent to the test set. Asymmetry of verification is everywhere and it's exciting to consider a world of jagged intelligence where anything we can measure will be solved.

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!

@quxiaoyin @lovable @antonosika It's unfortunate that competitions in SF have come down to cyberbullying. Incidentally, looks like @lovable uses user data (prompts, interactions) for AI model training by default for non-Enterprise users.

Female leaders, how many times have you been called “drama”? @lovable’s founder @antonosika just reminded me again. Yesterday I posted about Heyboss.dev being shadily redirected to their site. Today I got this DM: “SF circles are small and it increasingly will look to your friends like you just want drama for impressions if you do this again :)” Every time we female leaders confront, we’re hit with the “drama tax.” But if calling out shady tactics shifts the focus back to product and customer value — I’ll pay the drama tax. Every time. AI is too important for humanity to be derailed by noise, or dirty games. Let’s compete on product and create more value for our customers!

@amasad @chrisman noam shazeer is absolutely brilliant and worth 2.5b. Hard for others to meet this bar tho

@chrisman happened a few times already e.g. character ai

How do you underwrite an investment in an AI company valued at $10B with no product? A friend argued that the downside is limited—worst case, the team gets acquihired, returning at least 1x. So it’s a positive expected value bet.

@carlfranzen @Ahmad_Al_Dahle Isn't this more or less finetuning on the LMArena test set? Also, the actual serving cost, per artificialanalysis.ai, is in the .4$ range, yet the graph cherrypicks the lowest number in the range (0.2$-0.5$)

@Ahmad_Al_Dahle Also why did you benchmark a version "optimized for conversationality" rather than the actual Llama 4 Maverick version you made available for download?? 🤔

We're glad to start getting Llama 4 in all your hands. We're already hearing lots of great results people are getting with these models. That said, we're also hearing some reports of mixed quality across different services. Since we dropped the models as soon as they were ready, we expect it'll take several days for all the public implementations to get dialed in. We'll keep working through our bug fixes and onboarding partners. We've also heard claims that we trained on test sets -- that's simply not true and we would never do that. Our best understanding is that the variable quality people are seeing is due to needing to stabilize implementations. We believe the Llama 4 models are a significant advancement and we're looking forward to working with the community to unlock their value.

@Kinker_0 @teortaxesTex fair and llama are two independent orgs

@teortaxesTex Insanely disappointing that (almost) nothing from their research departments this year ended up in the new models.

First reaction on Meta Llama 4 launch: disappointment No local model. I think they can't beat Gemma density. Scout 109B/17A bizarrely forgoes finegrained sparsity despite all the research in its favor, maybe to pander to low tech providers. Maverick is just fatter DeepSeek V2.

🧵(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 or I at @TheWebConf 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), and opens up exciting research opportunities to support learned similarities.

🧵(1/4) Interested in the next generation retrieval paradigm or making your recommendations, search, or RAG applications 20-30% better? Check out Retrieval with Learned Similarities, a collaboration with Microsoft Research, accepted as an oral presentation (155 out of 2062 submissions) at #TheWebConf25!

@opensauceAI @finkd Can't wait to see people try to untangle ResNet from literally every modern AI model. Should be fun to watch :)

Wow. Congress just tabled a bill that would *actually* kill open-source. This is easily the most aggressive legislative action on AI—and it was proposed by the GOP senator who slammed @finkd for Llama. Here's how it works, and why it's different to anything before it.

@mblair The American way has always been about individual freedom and control. We had a historic opportunity to make social media fully user-controlled and unbiased, yet lawmakers chose to optimize for tech oligarchs' net worth instead.

TikTok should have been handled with counterprogramming. That's the American way.

@w_bsun @emollick @DrunkGershwin The commonplace does not consecrate the unjust.

X has a bad policy that if you post a link in the initial post in a thread, that thread is partially hidden. I post a about a pilot study on AI, it goes viral. 2M views so far. The link to the actual study is in the 2nd post and has vital details. 4,120 people clicked the link

@packyM Unfortunately Twitter decided to ban 3rd party APIs. Time to talk to Elon to change ToC?

If I can get it to check Twitter every 5 minutes instead of me, I’m going to be so productive.

@martin_casado @hausdorff_space @halvarflake @ianlivingstone @vitalygordon Can I also get an invitation?

holy shit. So much interest. Doing what I can to get through this. @hausdorff_space @halvarflake @ianlivingstone @vitalygordon can all send out invites as well so if I don't get to you, please reach out to them.

Hey infra folks. We're standing up a new Discord server to discuss CS infra. If you want an invite DM me (reply and I'll follow). thanks!