This new meet-up is focused on Generative AI such as LLMs and diffusion models.

Expect to see product demos, papers discussed, emerging business models analysed, best practices, infrastructure, tooling etc. These will generally be discussed along technical, social, economical, philosophical themes. The format is simple: some talks, chats, pizza, refreshments.

How to train an LLM for free

Fionnán Alt — Founder, MainsailAI

Large Language Models (LLMs) present a step-change in technology capability as well as profound and wide-ranging societal, legal, ethical, philosophical and economic impacts. In addition, due to their size and complexity, building or training an LLM from scratch is often viewed as being prohibitively expensive outside of large tech companies or well-funded AI labs.While state-of-the-art models such as OpenAI's GPT-4 are estimated to cost $100m+ for a single training run, recent advances have reduced the overhead required for running and training LLMs.

In this talk, Fionnán demonstrates how to train a capable LLM from scratch using free resources kindly provided by the TPU Research Cloud. We will explore using modern ML frameworks and technologies such as JAX, Ray, and Tensor Processing Units (TPUs) to efficiently scale out neural net training tasks for models with billions of parameters.

Uncovering granular insights about your LLMs

Morgan McGuire — Head of Growth ML, Weights & Biases.

Weights & Biases new LLM debugging tool, with Trace Timeline and Trace Table, is a natural extension of its scalable experiment tracking, designed to support ML practitioners working on prompt engineering for LLMs. It makes it easy review past results, identify and debug errors, gather insights about model behavior, and share learnings with colleagues.

In this talk, Morgan will demonstrate how to introduce a streamlined debugging workflow into your process, with some practical examples to get you started.

Name classification: How does ChatGPT compare to machine learning language models?

Pedro Tabacof — Staff ML Scientist, Intercom.

In this presentation, Pedro shows how to use machine learning models and ChatGPT to classify names into person or company (e.g. "Google" is a company while "Barack Obama" is a person). The task of name classification is just an illustrative example of text classification. The target audience includes data scientists interested in applied natural language processing (NLP) and/or how to use ChatGPT in practice.

First, we present the Kaggle datasets used and how we can start with a simple baseline using scikit-learn logistic regression. Then, we show how to use two more complex models for the same task: FastAI LSTM and Hugging Face DistilBERT pre-trained language models. Finally, we show how we can leverage ChatGPT for its predictive capabilities: This only requires calling its API, a fallback method when it refuses to classify a name, and some textual post-processing to extract the answers. As simple as it sounds, this achieves higher accuracy than the previous models without any training whatsoever. The whole presentation can be reproduced from a publicly available Jupyter notebook.

Intercom  2nd Floor, Stephen Court, 18-21 St Stephen's Green, Dublin 2, D02 N960