Yannick Versley

Computational linguist & curious person

[Yannick Versley @glowingfreezer]

I created an overview on common techniques used by today's open-weight models like DeepSeek V4.1, GLM 5.3 or Nemotron 3 Ultra, automatically extracted from respective code and tech reports.
https://t.co/F0YT1YVqtQ

[Yannick Versley @glowingfreezer]

@Real123Here @pmddomingos If you look closely at History, you'll find a lot of parallel discoveries and people refining each other's concepts. Today's transformer or LSTM are collages of dozens of refinements of an original idea, rather than independent and unique.

[Yannick Versley @glowingfreezer]

@NandoDF Important to note that regulation can guide the companies towards sane behavior. E.g. Germany requires 50% of salary for the duration of post-contract non-competes, and as a result companies only do it when necessary and don't treat it as a freebie

[Yannick Versley @glowingfreezer]

@daniel_271828 @McclaneDet @hendrycks I'd rather say, alignment research, if successful, puts the decision of what the system does in a central place - which may be divorced from what users want. Centralizing helps to keep users from designing Super Mario mousepads or writing haikus on political figures.

[Yannick Versley @glowingfreezer]

@ilyasut yet at a certain time DL practitioners fully expected you to believe in tanh nonlinearities and deep networks without residual connections

[Michael Saxon @m2saxon]

What an amazing opportunity for the most well-resourced private school students and children of researchers :D

[Yannick Versley @glowingfreezer]

@yoavgo How about going to "HLT" (human language technology), no geographic indicator? *ducks*

[Joris Baan @jsbaan]

Interested in uncertainty in NLP and NLG systems?

We are organising the first workshop on uncertainty-aware NLP, co-located with #EACL2024 in Malta!

Consider submitting and/or attending!

More info (like awesome speakers and a broad range of topics) at https://t.co/XiurnFqWgi.

[David Bamman @dbamman]

This might be flying under the radar (so please RT!), but the US Copyright Office is soliciting comments for its decisions on training ML/NLP/AI systems on copyrighted material (even *non*-GenAI). Researchers, please comment! Deadline Oct 30. https://t.co/4WUhBOz6Dv

[Taylor Belrose @norabelrose]

I'm opposed to any AI regulation based on absolute capability thresholds, as opposed to indexing to some fraction of state-of-the-art capabilities.
The Center for AI Policy is proposing thresholds which already include open source Llama 2 (7B). This is ridiculous.

[elvis @omarsar0]

🎓 Probabilistic Machine Learning

This is genuinely a one-of-a-kind resource for students looking to get well-versed in machine learning.

The trilogy book includes:

- Book 0: Machine Learning: A Probabilistic Perspective (2012)

- Book 1: Probabilistic Machine Learning: An Introduction (2022)

- Book 2: Probabilistic Machine Learning: Advanced Topics (2023)

These books also have accompanying notebooks/codes.

An impressive book series by @sirbayes 👏

books: https://t.co/HfrZNpki5t
notebooks: https://t.co/xgzrH2TpCK

[UKP Lab @UKPLab]

PhD students wanted! 🎓 The European Laboratory for Learning & Intelligent Systems @ELLISforEurope is a pan-European AI network of excellence. The 2023 Call for Applications of the ELLIS PhD Program is now online & can be found here ➡️ https://t.co/ZMYCYUsvJn #JoinELLISforEurope

[Yannick Versley @glowingfreezer]

@gchrupala @srchvrs @cohenrap Agree - with ARR chasing the reviewers was always a chore from AE perspective. Did this change when ARR explicitly asked reviewers to opt-in to each cycle instead of assuming everyone has constant capacity?

[Yannick Versley @glowingfreezer]

@nlopitz @AnetteMFrank Congratulations!

[Jim Pickard 🐋 @PickardJE]

lefty voices criticising “greedflation” now include none other than - squints at notebook - the IMF

[Yannick Versley @glowingfreezer]

@jochenleidner @stanfordnlp Multidisciplinary as in, "only applied maths"? I think that is making explicit the impoverishment that AI has undergone on the way from trying to understand things to just building big transformer models. Not disputing that knowing maths (+ being a polymath) gives you an edge.

[Andrej Karpathy @karpathy]

Very simple, minimal implementations for LLM inference at the edge with a lot of momentum, and a number of developing extensions across GPU support, quantization++, training/finetuning, etc.
👏 looking forward!

+"Inference at the edge" manifesto good read:
https://t.co/v8HaHALY7f

[Joshua Goodfield @joshuagoodfield]

💀💀 a Toronto restaurant renamed its menu items as office supplies so people could expense their meals. I’ve seen it all.

[Yannick Versley @glowingfreezer]

Kind of like rule-based coreference systems of the 2010s being "unsupervised". And many "zero-shot" papers of today do the same thing and put 90s heuristics on top of LLMs. 2/2

[Yannick Versley @glowingfreezer]

I feel like our field has become a bit dumber in the last few years - where people previously had "domain adaptation" or "rule-based" or "unsupervised learning" papers now only say "zero-shot" and want you to be impressed instead of telling you which one it is. 1/2