Lightnews — Scholar-powered news

Kuzman Ganchev

@ganchev.bsky.social

1.4K followers 33 following 3 posts

Research Scientist at GoogleDeepMind (formerly at Google Research). UPenn graduate.

Posts Media Videos Starter Packs

Reposted by Kuzman Ganchev

Ethan Mollick @emollick.bsky.social · May 4

Study in Nature: “Across 30 out of 32 evaluation axes from the specialist physician perspective & 25 out of 26 evaluation axes from the patient-actor perspective, AMIE [Google Medical LLM] was rated superior to PCPs [primary care docs] while being non-inferior on the rest.”

(& AIME is an older LLM)

4 16 70

Reposted by Kuzman Ganchev

Gus @gusthema.bsky.social · Apr 30

Gemma 3 explained: Longer context, image support, and a new 1B model. → goo.gle/4lV8iaw

Other key enhancements:
🔸 Best model that fits in a single consumer GPU or TPU host
🔸 KV-cache memory reduction with 5-to-1 interleaved attention
🔸 And more!

Read the blog for the full details on Gemma 3.

Gemma explained: What’s new in Gemma 3- Google Developers Blog

Google's Gemma 3 model includes vision-language support and architectural changes for resource-friendly multimodal language models.

goo.gle

1 8 22

Kuzman Ganchev @ganchev.bsky.social · Dec 17

There's a link to a really nice interactive viewer for a sample of the data (will only make sense after you read the post). There's some examples that I would have expected (where something is implied but not directly stated) but also a surprising number of kind of topical things.

Tyler Chang @tylerachang.bsky.social · Dec 13

We scaled training data attribution (TDA) methods ~1000x to find influential pretraining examples for thousands of queries in an 8B-parameter LLM over the entire 160B-token C4 corpus!
medium.com/people-ai-re...

1 3

Reposted by Kuzman Ganchev

Andreas Steiner @andreaspsteiner.bsky.social · Dec 5

Want to get started using PaliGemma 2?

🎤 developers.googleblog.com/en/introduci...
🤗 huggingface.co/blog/paligem...
💾 kaggle.com/models/googl...
🔧 github.com/google-resea...

7/7

1 7

Kuzman Ganchev @ganchev.bsky.social · Nov 11

Wanted to share that Varun Godbole recently released a prompting playbook. The title says prompt tuning, but this is text prompts, not soft prompts.

github.com/varungodbole...

GitHub - varungodbole/prompt-tuning-playbook: A playbook for effectively prompting post-trained LLMs

A playbook for effectively prompting post-trained LLMs - varungodbole/prompt-tuning-playbook

github.com

7 14

Reposted by Kuzman Ganchev

Jacob Eisenstein is at CoLM @jacobeisenstein.bsky.social · Oct 24

I’m pretty excited about this one!

ALTA is A Language for Transformer Analysis.

Because ALTA programs can be compiled to transformer weights, it provides constructive proofs of transformer expressivity. It also offers new analytic tools for *learnability*.

arxiv.org/abs/2410.18077

ALTA: Compiler-Based Analysis of Transformers

We propose a new programming language called ALTA and a compiler that can map ALTA programs to Transformer weights. ALTA is inspired by RASP, a language proposed by Weiss et al. (2021), and Tracr (Lin...

arxiv.org

2 16 53

Kuzman Ganchev @ganchev.bsky.social · Oct 25

Not news, but I recently saw the zed.dev demo and it looks amazing. Has anyone used it or something similar?

Zed - The editor for what's next

Zed is a high-performance, multiplayer code editor from the creators of Atom and Tree-sitter.

zed.dev