AI Code and Coffee Warsaw

31/01/2025

Wes Roth gives an overview of Berkleys reproduction of DeepSeek R1 core technologies, in the CountDown game

https://www.youtube.com/watch?v=E_h8xt0X1Kg

Links to original sources:

GitHub:

https://github.com/Jiayi-Pan/TinyZero

(the author):

https://x.com/jiayi_pirate/status/1882839370505621655

Berkeley Researchers Replicate DeepSeek R1's Core Tech for Just $30: A Small Model RL Revolution

https://xyzlabs.substack.com/p/berkeley-researchers-replicate-deepseek

A Berkeley AI Research team led by PhD candidate Jiayi Pan has achieved what many thought impossible: reproducing DeepSeek R1-Zero's key technologies for less than the cost of a dinner for two.

05/12/2024

Thanks to everyone who came to yesterday's AI code and coffee

We talked about AI in areospace, LLMs, Karpathy's great advice on training models

https://karpathy.github.io/2019/04/25/recipe/

and 3blue1brown's amazing visualizations on AI

https://www.youtube.com/c/3blue1brown

Many thanks to IDEAS for resharing the event! We had lots of amazing people turn up

05/05/2024

Google Deep Learning Tuning Playbook

"there is an astonishing amount of toil and guesswork involved in actually getting deep neural networks to work well in practice"

A playbook for systematically maximizing the performance of deep learning models. - google-research/tuning_playbook

18/04/2024

Thanks to everyone who came to AI code and coffee yesterday!

We talked about Model Collapse as a Change of Scaling Laws (https://arxiv.org/pdf/2402.07043.pdf),

Luk described tricks and techniques for building an AI startup on stable diffusion

Besher talked about Kubernetes, and operating it for ML at scale. Some great discussions, about how compression represents a form of intelligence.

21/03/2024

3 months ago there was a wild rumor regarding a new powerful technique at OpenAI called Q-Star. Turns out it was short for Quiet-Star - an 'inner monologue' and it massively improved its performance

Scientists trained an AI system to think before speaking with a technique called QuietSTaR. The inner monologue improved common sense reasoning and doubled math performance.

19/03/2024

Trying to keep up with AI? Building an AI startup?

Come to our AI Code and Coffee Warsaw meetup tomorrow, Wednesday 20th March, where we will be discussing quickly building an AI startup using, python, FastAPI and Async Django.

Location: Bella Kafeteria, Freta 30/32 34, 00-227 Warszawa,

Time: 18:00-20:00

Date: 2024-03-20

We will have a short presentation (20min), abstract can be found

https://pycon.lt/2024/talks/YCMZQ8

Django's async capabilities and batteries-included tooling make it an ideal framework for quickly building MVPs and iterating. This talk demonstrates building a document search MVP with Django templates, ChromaDB, and self hosted large language models. It then shows how to refactor and scale it using Elasticsearch, Celery/RabbitMQ workers, React, self-hosted vLLM, and auth. With Django async, you can rapidly build, constantly improve, and deploy the latest AI models in your product.

Afterwards coding and networking

Link to meetup:

https://www.facebook.com/events/403202712319205/

https://www.meetup.com/code-coffee-warsaw/events/299655475/

AI Code and Coffee 𝘄𝗮𝗿𝗺𝗹𝘆 𝗶𝗻𝘃𝗶𝘁𝗲𝘀 𝘆𝗼𝘂 𝘁𝗼 𝗼𝘂𝗿 𝗰𝗼𝗱𝗲 𝗮𝗻𝗱 𝗰𝗼𝗳𝗳𝗲𝗲 𝗲𝘃𝗲𝗻𝘁! Stuck on some code? Working on something exciting? Want to learn

14/03/2024

"There's only two types of companies in this world, those great at AI and everyone else. If you don't know AI, you are going to fail. Period. End of story" (Mark Cuban). Agree or disagree?