Episode 4 - Large Language Models


Welcome to episode 4: LLMS

In this episode we talk about LLMs, the predicted impact they will have on jobs, the kind of work they can help with and how you can go about using them.

We start by talking about the $135bn market value of AI and the predicts of potential doom from Goldman Sachs as well as a take from McKinsey. We attribute the success of LLMs to OpenAI before moving on to the kinds of use cases for these new tools - essentially the possibility to eliminate boring, repetitive tasks that many of us face today (an no I haven’t got around to setting up an LLM to produce these summaries…).

We discuss code completion using tools like AWS CodeWhisperer and Github CoPilot as well as document outlines using Notion AI and Khanmigo for that personalised tutor experience. We then discuss some of the issues that folks face, including hallucination, non-determinism and that these tools are not general AI, they’re just good at predicting the next best word (ish), we finish with the disappointing test of Khanmigo.

We shrug off the negative and move to disabuse folks of the idea that AI, General AI and Generative AI are all the same thing. There’s more to AI and it’s not just Transforms, although they really are very good!

We finish out by discussing how you can use these kinds of tools, first suggesting that most companies aren’t Bloomberg and don’t have the experience or money to build these things in house. This leave two options: APIs or Fine-tuning but things are changing rapidly, with OpenLLM from Bentoml providing a way to run models, smaller high quality datasets looking to yield good performance and a massive $1.3bn spend as Databricks strikes a deal for MosaicML. GO OPEN SOURCE!!!

The AI train is only getting faster! join us on a high level, mostly positive conversation about LLMs, machine learning and how we probably aren’t going to be working for our AI overlords before 2024.

Additional Sources Sal Khan - How AI Could Save (Not Destroy) Education
Falcon LLM
Artificial Intelligence: A Modern Approach
Dan’s Thoughts on AI eating Senior Engineers
Simon Willison on Benchmarking with GPT

Understanding LLMS

Jay Alammar
Visual Intro to Neural Networks

This is the Everything Delivery podcast

In this podcast we will discuss anything and everything that helps deliver value to clients while maintaining a healthy and happy Organization.

You can find this podcast on: Spotify: https://open.spotify.com/show/0hDIMNruStpK9aE9eHC7NZ
Apple podcast: https://podcasts.apple.com/it/podcast/everything-delivery/id1680879029?l=en