Yoshinari Fujinuma

I played around with the mlx and 4-bit version of Qwen3-30B-A3B locally on an Apple M4 Max chip, with Japanese and English. This is amazing. It seems feasible to run it locally for tasks other than long-horizon or complex ones.

Oh yeah, I can easily imagine that'll be outputted by the model 🙂

To be clear, the recipe to replicate o1 style models is not new techniques, but applying them in a new way.
This shouldn't be surprising.

I've just played around with DeepSeek-R1 and wow, such a long thoughts for a simple question "What is the square root of 16?" 😀

My LinkedIn feed is full of AWS re:Invent posts (since I work at AWS, and many colleagues share about it), Twitter/X is a mixture of everything, and Bluesky posts are mostly academic. Welcome to the filter bubbles!

I was taking a stab at the responding to author discussions for ARR for the last few days, but some common issues I see is that submitted drafts are pretty exaggerating how good the results are.

Now Hear This: World’s Most Flexible Sound Machine Debuts Fugatto generates or transforms any mix of music, voices and sounds described with prompts using any combination of text and audio files.

Where is the open-weight model that we can try it out? 😀
blogs.nvidia.com/blog/fugatto...

Look who's here
@pnas.org ✔️
@science.org ✔️
@naturecellbiology.bsky.social ✔️
@natrevgenet.bsky.social ✔️
@naturebiotech.bsky.social ✔️
@naturemicrobiol.bsky.social ✔️
@naturechemistry.bsky.social ✔️
@genesdev.bsky.social ✔️
@cellchembiol.bsky.social ✔️
@genomeresearch.bsky.social ✔️
@jcellbiol.bsky.social ✔️

I like typeset.io/pdf-to-video 's pdf-to-video feature for getting a quick overview of the paper. Looking forward to having more fine-grained video (or even customizable controlled generation of pdf video summary) version of it 😀

Just a heads up to everyone: @deep-mind.bsky.social is unfortunately a fake account and has been reported. Please do not follow it nor repost anything from it.

Here's the starter pack for AI/ML/NLP conferences that I was able to find as of now. I couldn't remove myself from the starter pack so feel free to unfollow me after hitting the "follow all" button 🙂 go.bsky.app/9QQXJ1u

AI Bluesky Join the conversation

Great AI people starter pack from @chris.bsky.social!

go.bsky.app/KRsy8pF

📣 I am sure we have reached only a small fraction of New York's ML community in bsky. Please repost 🔁 this if you think you may have interested people close to you in the social graph.

Someone should really treat me some coffee for asking me to assign & finish the emergency review within a day 😀

1. Find your friends! I've found most of mine with:

- Starter packs blueskydirectory.com/starter-pack...
- the Chrome extension 'Sky Follower Bridge' www.sky-follower-bridge.dev
- @theo.io's Follow Finder, which lists people who are followed by lots of people you follow bsky-follow-finder.theo.io

Does your LLM truly unlearn? An embarrassingly simple approach to recover unlearned knowledge Large language models (LLMs) have shown remarkable proficiency in generating text, benefiting from extensive training on vast textual corpora. However, LLMs may also acquire unwanted behaviors from th...

Today's paper reading: Interesting, quantization can reverse unlearning up to 83% arxiv.org/abs/2410.16454

I did a starter pack of people in New York (City) working on ML/AI. Please distribute and feel free to self nominate!

go.bsky.app/BoEtagz

@ramon-astudillo.bsky.social self-nominating myself :)

On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes Knowledge distillation (KD) is widely used for compressing a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, current KD methods for auto-...

Today's paper reading: arxiv.org/abs/2306.13649

Too many people... @Shibuya station, Tokyo, Japan

Wow, more than 2000 papers were accepted in total for EMNLP

Looking forward to catching up with old friends and meeting new friends :)

References:
[1] arxiv.org/pdf/2305.112...
[2] arxiv.org/pdf/2310.163...
[3] aclanthology.org/2023.conll-1...

[3] Bonus. 12/7 1:45pm Though I'm not the author, I'll be helping out presenting the poster at #CoNLL co-authored by my colleague titled "Cross-Document Event Coreference Resolution: Instruct Humans or Instruct GPT?"
(my first attempt and let's see how this turns out :) )

Heading to #EMNLP ! Co-authored papers 👇
[1] 12/9 11am in-person poster by Sharon Levy title "Comparing Biases and the Impact of Multilingual Training across Multiple Languages"

[2] 12/8 2pm virtual poster titled "A Multi-Modal Multilingual Benchmark for Document Image Classification"

Resolving Latex errors for uploading to arxiv...

GPT-4V just fixed my circuit breaker (where I had been struggling for 10+ mins at midnight)

Done finishing up the EMNLP findings camera ready

I finally read the attention sink paper [1] and the HF blog article [2]. Seems like another interesting data point that the models we usually interact with strongly attend to the first few tokens...
[1] arxiv.org/abs/2309.17453
[2] huggingface.co/blog/tomaars...

New conference alert! COLM (“collum”) seeks a broad range of work on language modeling. 9 pages due Mar 8: colmweb.org

Preparing an EMNLP camera ready version of our accepted paper ✍️✍️✍️