damo

PhD in CS

[damo @tugarin62]

@teortaxesTex @xlr8harder There's a good bench for russian called RuQualBench. Measures mistakes and hallucinations on generating russian text.

[damo @tugarin62]

@EFCxGoat Now count UCL titles

[damo @tugarin62]

@techwith_ram turboquant is not SOTA quant method. so anything built on top of it is by design outdated. please do not fall for X hype.

[damo @tugarin62]

@teortaxesTex That paper is a joke. They ran TWO models for the whole methodology and got accepted to TMLR (most reputable ML journal) as is. There's no way this could have been accepted in purely blind submission process.

[damo @tugarin62]

@My__Regis @gaoj0017 I agree. I feel like a 10 *might be* a proper rank for this paper. BUT if you rank paper 10 though it has cons you should not try to make the paper sound worse than it is post factum. You either rank lower OR you agree that the paper is a 10 despite these problems with baseline

[damo @tugarin62]

@teortaxesTex Just a jealous guy trying to ride the hype train 🤣

[damo @tugarin62]

@colorful_kiki @gaoj0017 Moreover, I consider both the OPs work and the TurboQuant the ordinary good quant papers, but not breakthroughs. Just the latter got overhyped due to X shitposters.

[damo @tugarin62]

@My__Regis @gaoj0017 Obvious things. https://t.co/VcWoXKN2Rc

The reviewer WFrV is the only mentioning same issues with RaBitQ. I am pretty damn sure it is either the OP, or his coauthors.

Anyway, it is always miserable to try join the hype train and raising CoNcErNs after you rank the paper 10.

[damo @tugarin62]

@colorful_kiki @gaoj0017 Neither

[damo @tugarin62]

@s_pect_re @gaoj0017 Am i wrong?

[damo @tugarin62]

@predict_addict Very easy to be fair. At least let's hope this test is not for maths majors.

[damo @tugarin62]

@TheTuringPost @christoph_wertz @GoogleResearch I'd say because infomaniacs from X overhyped the claims of the original paper. Just look at the barplot, the X scale starts from 48 🤣

What a shame.

[damo @tugarin62]

@Zai_org let's gooooo

[damo @tugarin62]

@askalphaxiv A year old paper with MODERATE gains yet X goes crazy for the whole week. What is wrong with you people???

[damo @tugarin62]

@_vmlops this is a useless vibe-coded ai slop. nobody will ask these questions at a real interview

[damo @tugarin62]

@a_weers would be nice to add two sorting ways for these RL methods: one order will be chronological, another one in terms of performance (unsure about performance as in literature the newer methods are *usually* better than old ones)

[damo @tugarin62]

@nalinrajput23 just use 90-99% isopropyl alcohol and you will be fine

[damo @tugarin62]

@Iam_No_One____ @nalinrajput23 that is actually on overkill, you can buy just methanol and wipe the keyboard with methanol. it will remove the fat from the fingertips.

[damo @tugarin62]

@NTFabiano what about pullups?

[Jack Morris @jxmnop]

"there's nothing interesting on arxiv these days!"
- the words of an uncurious mind

i have personally been blown away by the volume of interesting papers posted over the last few months, and eagerly following daily digests

here are some papers i enjoyed the most:

- Pre-training under infinite compute (September 2025, https://t.co/3Q838oO6ei)
- Fresh in memory: Training-order recency is linearly encoded in language model activations (September 2025, https://t.co/V9qCttiFPJ)
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data (July 2025, https://t.co/eJrGChfq1d)
- Memory Limitations of Prompt Tuning in Transformers (September 2025, https://t.co/AJR17dkVUx)
- Behavioral Fingerprinting of Large Language Models (September 2025, https://t.co/ZdHMlIdcYP)
- Language Self-Play For Data-Free Training (September 2025, https://t.co/9kLvY8dNbe)
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs (September 2025, https://t.co/X7bwtKE8xe)
- Do Natural Language Descriptions of Model Activations Convey Privileged Information? (September 2025, https://t.co/4qjWhFJVUG)
- Beyond the Leaderboard: Understanding Performance Disparities in Large Language Models via Model Diffing (September 2025, https://t.co/2ejyGDCSVF)
- Stochastic activations (September 2025, https://t.co/1xoXmLeIiF)
- PonderLM-2: Pretraining LLM with Latent Thoughts in Continuous Space (September 2025, https://t.co/gZW50tvCIK)
- Words That Make Language Models Perceive (October 2025, https://t.co/IDQEXdeAGv)
- Language Models Do Not Embed Numbers Continuously (October 2025, https://t.co/g8Cw3yNcoV)
- Learning Facts at Scale with Active Reading (August 2025, https://t.co/aw3fE8dKiJ)
- OverFill: Two-Stage Models for Efficient Language Model Decoding (August 2025, https://t.co/Wku5FXbGEz)
- Retrieval Capabilities of Large Language Models Scale with Pretraining FLOPs (August 2025, https://t.co/TWgqTCHjuZ)
- Reasoning-Intensive Regression (August 2025, https://t.co/2G8Lxn323A)
- Watch the Weights: Unsupervised monitoring and control of fine-tuned LLMs (August 2025, https://t.co/im0qdNorNQ)
- On the Theoretical Limitations of Embedding-Based Retrieval (August 2025, https://t.co/7haVnfNpTp)