Leo Boytsov
@srchvrs.bsky.social
Machine learning scientist and engineer speaking πtorch & C++ (ph-D CMU) working on (un)natural language processing, speaking πtorch & C++. Opinions sampled from MY OWN 100T param LM.
"Meta is cutting around 600 positions out of the several thousand roles in its Superintelligence Labs, the Facebook owner said on Wednesday as it looks to make its artificial intelligence unit more flexible and responsive."
www.reuters.com/business/met...
www.reuters.com/business/met...
www.reuters.com
October 22, 2025 at 9:43 PM
"Meta is cutting around 600 positions out of the several thousand roles in its Superintelligence Labs, the Facebook owner said on Wednesday as it looks to make its artificial intelligence unit more flexible and responsive."
www.reuters.com/business/met...
www.reuters.com/business/met...
🧵I recently finished my nerdiest computer science paper so far and it was accepted by TMLR: A Curious Case of Remarkable Resilience to Gradient Attacks via Fully Convolutional and Differentiable Front End with a Skip Connection. This work was done while I was at Bosch. ↩️
August 27, 2025 at 3:49 AM
🧵I recently finished my nerdiest computer science paper so far and it was accepted by TMLR: A Curious Case of Remarkable Resilience to Gradient Attacks via Fully Convolutional and Differentiable Front End with a Skip Connection. This work was done while I was at Bosch. ↩️
🧵Hot take: LLMs still fail at basic grammar/style checking. A repeating situation that I encounter:
1. Ask a model about an issue.
2. The model suggests some rewrite for clarity/accuracy. Typically it's actually quite good (but watch for factual errors!).
3. Recheck the text again.
↩️
1. Ask a model about an issue.
2. The model suggests some rewrite for clarity/accuracy. Typically it's actually quite good (but watch for factual errors!).
3. Recheck the text again.
↩️
August 9, 2025 at 3:05 PM
🧵Hot take: LLMs still fail at basic grammar/style checking. A repeating situation that I encounter:
1. Ask a model about an issue.
2. The model suggests some rewrite for clarity/accuracy. Typically it's actually quite good (but watch for factual errors!).
3. Recheck the text again.
↩️
1. Ask a model about an issue.
2. The model suggests some rewrite for clarity/accuracy. Typically it's actually quite good (but watch for factual errors!).
3. Recheck the text again.
↩️
🧵Pseudo-relevance feedback (PRF) (also known as blind feedback) is a technique of first retrieving/re-ranking top-k documents and adding some of their words to the initial query. Then, a second retrieval/ranking stage uses an updated query. ↩️
July 18, 2025 at 6:01 PM
🧵Pseudo-relevance feedback (PRF) (also known as blind feedback) is a technique of first retrieving/re-ranking top-k documents and adding some of their words to the initial query. Then, a second retrieval/ranking stage uses an updated query. ↩️
🧵 Dear (scientific) authors: I am being in the same boat too. However, if you receive a ton of detailed complaints regarding paper quality, do NOT try to address them during the rebuttal phase. It's just a waste of everybody's time. ↩️
July 2, 2025 at 3:04 AM
🧵 Dear (scientific) authors: I am being in the same boat too. However, if you receive a ton of detailed complaints regarding paper quality, do NOT try to address them during the rebuttal phase. It's just a waste of everybody's time. ↩️
@microsoft.com faces an interesting issue that might affect others selling wrappers around ChatGPT and Claude models: users prefer to use ChatGPT directly rather than engage with Microsoft's Copilot.
futurism.com/microsoft-co...
futurism.com/microsoft-co...
Microsoft Is Having an Incredibly Embarrassing Problem With Its AI
Despite investing tens of billions of dollars into OpenAI, Microsoft is still, apparently, competing with its business partner.
futurism.com
June 27, 2025 at 2:34 PM
@microsoft.com faces an interesting issue that might affect others selling wrappers around ChatGPT and Claude models: users prefer to use ChatGPT directly rather than engage with Microsoft's Copilot.
futurism.com/microsoft-co...
futurism.com/microsoft-co...
This is a rather blockbuster piece of news: the @hf.co library is dropping support for both Jax and Tensorflow.
www.linkedin.com/posts/lysand...
www.linkedin.com/posts/lysand...
I have bittersweet news to share. | Lysandre Debut
I have bittersweet news to share.
Yesterday we merged a PR deprecating TensorFlow and Flax support in transformers.
Going forward, we're focusing all our efforts on PyTorch to remove a lot of th...
www.linkedin.com
June 12, 2025 at 7:04 PM
This is a rather blockbuster piece of news: the @hf.co library is dropping support for both Jax and Tensorflow.
www.linkedin.com/posts/lysand...
www.linkedin.com/posts/lysand...
Found a hidden gem on IR evaluation methodology from Microsoft "What Matters in a Measure? A Perspective from Large-Scale Search Evaluation."
dl.acm.org/doi/pdf/10.1...
dl.acm.org/doi/pdf/10.1...
dl.acm.org
April 27, 2025 at 3:15 AM
Found a hidden gem on IR evaluation methodology from Microsoft "What Matters in a Measure? A Perspective from Large-Scale Search Evaluation."
dl.acm.org/doi/pdf/10.1...
dl.acm.org/doi/pdf/10.1...
Parental advice: if you master algebra you will know how to deal with your x-es.
@ccanonne.bsky.social feel free to borrow!
@ccanonne.bsky.social feel free to borrow!
April 21, 2025 at 11:41 PM
Parental advice: if you master algebra you will know how to deal with your x-es.
@ccanonne.bsky.social feel free to borrow!
@ccanonne.bsky.social feel free to borrow!
Some people say: A prompt is worth a thousand words! Excuse, but have you seen these ones? They are way longer!
April 16, 2025 at 1:42 AM
Some people say: A prompt is worth a thousand words! Excuse, but have you seen these ones? They are way longer!
🧵A fascinating perspective on the nature of intelligence and the history of automation/ (and ahem development of AI). It is also a cautionary story of how to not trust AI too much. ↩️
🎄 Merry Christmas! 🎄
🔔 As Advent ends, the 12 Days of Christmas🎄begin!
Join me for a deeper dive into the initiatives shown in this map - from Data Science Africa to @ai.cam.ac.uk - as we explore practical responses to AI challenges.
the-atomic-human.ai/reflections/...
🔔 As Advent ends, the 12 Days of Christmas🎄begin!
Join me for a deeper dive into the initiatives shown in this map - from Data Science Africa to @ai.cam.ac.uk - as we explore practical responses to AI challenges.
the-atomic-human.ai/reflections/...
April 7, 2025 at 2:24 AM
🧵A fascinating perspective on the nature of intelligence and the history of automation/ (and ahem development of AI). It is also a cautionary story of how to not trust AI too much. ↩️
🧵Although pre-trained Transformer models took NLP by storm, they were less successful for recommender systems (arxiv.org/abs/2306.11114). RecSys is hard:
1. The number of users is high.
2. The number of items is high.
3. A cold-start problem is a hard one.
↩️
1. The number of users is high.
2. The number of items is high.
3. A cold-start problem is a hard one.
↩️
March 23, 2025 at 8:59 PM
🧵Although pre-trained Transformer models took NLP by storm, they were less successful for recommender systems (arxiv.org/abs/2306.11114). RecSys is hard:
1. The number of users is high.
2. The number of items is high.
3. A cold-start problem is a hard one.
↩️
1. The number of users is high.
2. The number of items is high.
3. A cold-start problem is a hard one.
↩️
🧵"Through extensive experiments, we empirically confirm the bias of judges towards their related student models caused by preference leakage across multiple LLM baselines and benchmarks. Further analysis suggests ..." ↩️
February 4, 2025 at 2:52 PM
🧵"Through extensive experiments, we empirically confirm the bias of judges towards their related student models caused by preference leakage across multiple LLM baselines and benchmarks. Further analysis suggests ..." ↩️
🧵It was my great pleasure to appear on the vector podcast with @dmitrykan.bsky.social . We covered the history of our NMSLIB library and how it helped shape the vector search industry! ↩️
January 17, 2025 at 4:48 PM
🧵It was my great pleasure to appear on the vector podcast with @dmitrykan.bsky.social . We covered the history of our NMSLIB library and how it helped shape the vector search industry! ↩️
🧵"Fun" fact about pseudorand numbers (true for np.random, likely for everything). If u ||-ze your Monte-Carlo simulation across processes with standard map-style approach, each worker will be generating IDENTICAL (ohh bother) sequence of rand. numbers and you may not even notice that 😢.↩
January 5, 2025 at 7:13 PM
🧵"Fun" fact about pseudorand numbers (true for np.random, likely for everything). If u ||-ze your Monte-Carlo simulation across processes with standard map-style approach, each worker will be generating IDENTICAL (ohh bother) sequence of rand. numbers and you may not even notice that 😢.↩
Reposted by Leo Boytsov
Quick primer for non-wizards about the post-MCTS LLM reasoning future:
How will LLMs learn to reason efficiently?
No math in this thread, ~simple words only! Let's go through the "Process Reinforcement through IMplicit REwards" (PRIME) method. 1/n
curvy-check-498.notion.site/Process-Rein...
How will LLMs learn to reason efficiently?
No math in this thread, ~simple words only! Let's go through the "Process Reinforcement through IMplicit REwards" (PRIME) method. 1/n
curvy-check-498.notion.site/Process-Rein...
January 5, 2025 at 12:30 AM
Quick primer for non-wizards about the post-MCTS LLM reasoning future (I'm kinda PRIME-pilled rn) by @xpasky.bsky.social
x.com/xpasky/statu...
x.com/xpasky/statu...
January 4, 2025 at 5:20 PM
Quick primer for non-wizards about the post-MCTS LLM reasoning future (I'm kinda PRIME-pilled rn) by @xpasky.bsky.social
x.com/xpasky/statu...
x.com/xpasky/statu...
Great analysys of the recent OpenAi breakthrough on ARC-AGI challenge.
December 26, 2024 at 3:02 PM
Great analysys of the recent OpenAi breakthrough on ARC-AGI challenge.
Another paper (by @laura-dietz.bsky.social and @claclarke.bsky.social ) shows LLM-judges are vulnerable to adv. attacks:
"We demonstrate the ease w. which automatic evaluation metrics can be subverted, showing that systems designed to exploit these evaluations can achieve artificially high scores."↩️
"We demonstrate the ease w. which automatic evaluation metrics can be subverted, showing that systems designed to exploit these evaluations can achieve artificially high scores."↩️
December 24, 2024 at 10:57 PM
Another paper (by @laura-dietz.bsky.social and @claclarke.bsky.social ) shows LLM-judges are vulnerable to adv. attacks:
"We demonstrate the ease w. which automatic evaluation metrics can be subverted, showing that systems designed to exploit these evaluations can achieve artificially high scores."↩️
"We demonstrate the ease w. which automatic evaluation metrics can be subverted, showing that systems designed to exploit these evaluations can achieve artificially high scores."↩️
"I sensed anxiety & frustration" by @kyunghyuncho.bsky.social addresses frustration of many PhD students expecting to enjoy the same level of benefits (huge compensation, freedom to publish w/o contributing to product) as a few people hired from few deep learning labs a decade (or so ago)." ↩️
December 23, 2024 at 10:53 PM
"I sensed anxiety & frustration" by @kyunghyuncho.bsky.social addresses frustration of many PhD students expecting to enjoy the same level of benefits (huge compensation, freedom to publish w/o contributing to product) as a few people hired from few deep learning labs a decade (or so ago)." ↩️
There are three things one can watch forever: fire burning, water flowing, and A/B test metrics updating.
December 20, 2024 at 3:39 AM
There are three things one can watch forever: fire burning, water flowing, and A/B test metrics updating.
🧵We are proposing two new approaches to speed up constrained decoding with lookaheads (unrolling generation to the point in the future where constrains can be verified). ↩️
December 17, 2024 at 11:46 PM
🧵We are proposing two new approaches to speed up constrained decoding with lookaheads (unrolling generation to the point in the future where constrains can be verified). ↩️
🧵A recent study finds that PPO would be better than DPO if PPO were producing of comparable lengths. In other words, judges and humans prefer longer outputs so PPO appears to underperforming. ↩️
December 17, 2024 at 12:34 PM
🧵A recent study finds that PPO would be better than DPO if PPO were producing of comparable lengths. In other words, judges and humans prefer longer outputs so PPO appears to underperforming. ↩️
When you wrote circa 200 bash scripts (not counting Python code) during the period of three months and then had trouble recalling differences/nuances. 😅
December 17, 2024 at 3:57 AM
When you wrote circa 200 bash scripts (not counting Python code) during the period of three months and then had trouble recalling differences/nuances. 😅
Reposted by Leo Boytsov
at the end of the talk, the audience was in silence. only one brave student with a thick India accent went up to the mic and asked "excuse me Sir, do you consider me a threat?" to which the speaker answered "Not you *specifically*, but in general, yes. Although India is not a big threat as China is"
December 14, 2024 at 11:06 AM
at the end of the talk, the audience was in silence. only one brave student with a thick India accent went up to the mic and asked "excuse me Sir, do you consider me a threat?" to which the speaker answered "Not you *specifically*, but in general, yes. Although India is not a big threat as China is"