💡Can we trust synthetic data for statistical inference?
We show that synthetic data (e.g., LLM simulations) can significantly improve the performance of inference tasks. The key intuition lies in the interactions between the moment residuals of synthetic data and those of real data
We’ll know he’s really made it when Alcatraz gets autocorrected to Alcaraz.
Small idiosyncratic language errors in otherwise clear technical papers are now a plus; assurance that I'm not wasting my time trying to read GPT spam. #neurips2025
Spent the whole day reading unlearning papers. If any of the methods worked, I might hope to recover from the experience.
Final 5m of NBA games solved. If you foul:
• opposing team gets 1 FT
• & keeps possession w fresh 24
Next up, Israel/Palestine.
Google miles behind Open/Anth til overnight boom 2.5.
Mosaic models underwrite $1B until boom mothballed.
Deepseek turns world upside down, but only for 5min.
The speed at which the winds turn here is truly wild.
What problem were airplane designers hoping to solve when they proposed this monstrosity?
50M human hours have been spent adjusting airplane seats into upright position before takeoff, saving a grand total of 0 human lives
Unlike OpenAI, ostensibly controlled by a powerful holding entity, here have effectively 0 board power. This saga's shaping up like the Asimov's Spacer novels, but instead of each book exploring a different ratio of humans to robotic servants, it's a different level of corporate oversight.
OpenAI folks getting creative about corporate governance. I've seen this movie before...
www.theinformation.com/articles/thi...
We even found that sometimes with deep nets imputing 0 is better than running a fancy imputation algorithm because it’s more obvious which values are imputed so there’s less destruction of information.
Yeah, it makes sense, it gives you all the info you have. A powerful enough model can learn any reasonable imputation strategy implicitly.
Taking me back to 2016 baby. proceedings.mlr.press/v56/Lipton16...
In some clinical tasks the indicators are so useful, they’re almost as informative as the features themselves.
On my way to Singapore for #iclr2025. Excited to pop my head out of AI product development to get inspired, connect w the next generation of researchers, and also for food. hmu
Thus the “controversiality”. Point is, hard to imagine any non-political reason to get any other compact EV than a tesla. Seems like nobody is even close, which i found surprising given the growing market.
Controversial opinion. It’s impossible to lease an EV in the US that’s even half as good a deal as @Tesla model 3/Y.
Plugin that automatically clicks thru "reject" to select "only strictly necessary" on the cookie pop-ups. 💵
CTO’s the most slippery title in the world. Depending on the company, you’re either
1. A technical founder who never wanted to manage, now an IC engineer
2. A VP of Engineering
3. Leading all of EPD (+AI)
4. An influencer with no internal responsibilities
what sort of model was involved in the retrieval step?
Just as the jaws of tetragenarianism are closing in, stole this W from the jaws of impending awardless decrepitude.
...you’re thinking about infrastructure, user experience, orchestration, governance, and automation at scale. That’s a different game altogether."
per ChatGPT:
"AI is no longer just “machine learning at scale.” It’s a full-stack engineering and economic transformation, where ML is just one piece of the puzzle. If you’re building modern AI systems, you’re thinking beyond ML—...
Phase 1 of my research career: "ML" = rebranding of the successful subset of AI.
Phase 2: "AI" = rebranding of the successful subset of ML.
Phase 3: AI & ML truly starting to diverge.
Can't find any person or AI that can explain Luka → Lakers. Only satisfying theory I can conjure: behind-the-scenes feud → "calf strain" masking a relationship strain → trade demand. But according to ChatGPT's research, Luka was blindsided by the trade.
Just another boring week in the AI capabilities landscape.
How to **efficiently** build personalized language models **without** textual info on user preferences?
Our Personalized-RLHF work:
- light-weight user model
- personalize all *PO alignment algorithms
- strong performance on the largest personalized preference dataset
arxiv.org/abs/2402.05133
First #runconference @neuripsconf.bsky.social #NeurIPS2024 was great! Will share tomorrow's deets later today, join us!
@zacharylipton.bsky.social @adamjelley.bsky.social @random-steve.bsky.social
optimizing for inclusivity, pace depends on crowd, never crazy, maybe as fast as 8:00 miles, as slow as 10:00, depending on the crowd
