LLM Workshop - Introduction

Overview

• History of AI
• Compute and AI
• Ethics and concerns
• Introducing the workshop hardware

History of AI

• How has AI developed over time?

AI as a term is coined

We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. [...] An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.

• Darthmout Summer Research Project

Early NLP – 1956-1966

• Rule based, lots of manual effort
• Lots of LISP
• Used for:
  • Information retrieval
  • Basic chat-bots e.g. Eliza
  • Limited translation systems

AI Winter I – 1974-1980

• NLP Winter started even earlier

Statistical NLP – 1980s

• Using statistics of the corpus
• Bag-of-words, N-grams

AI Winter II – 1990s, early 2000s

• The word "AI" is not a buzzword
• Research continues under other names

Deep Learning – 2012-

• 2012: AlexNet has less than 25% error on ImageNet challenge
• 2017: Transformer architecture (Attention Is All You Need, retrospectic)
• 2019: GPT-2 released. Surprisingly good.
• 2020: GPT-3 released. Surprisingly still improving.
• 2022: ChatGPT (GPT-3.5) released. General public starts to take notice.

Deep Learning – Transformer architecture

No recurrent connections, thus more parallelizable.

Deep Learning – Attention mechanism

• Scaled Dot-Product Attention $$ \mathrm{Attention}(V, K, Q) = \mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_K}}\right)V $$
• Cross-attention: \(Q = X_\mathrm{dec} W_Q\), \(K = X_\mathrm{enc} W_K\) and \(V = X_\mathrm{enc} W_V\)
• Self-attention: Same \(X\) used for all matrices
• In decoder, self-attention masks future tokens
• Autoregressive unimodal LLMs usually decoder only

Compute and AI

• What has changed?
• The bitter lesson by Richard Sutton, 2019

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.

Compute use over time

• Compute, but also data, architecture and algorithms

Compute and performance

What was new with ChatGPT?

• Base models are pure language models
• Chat models are:
  • Instruct tuned (supervised)
  • Reinforcement Learning with Human Feedback

RLHF

• Enables RL when no clear scoring function available
• Relatively little human input needed

Further scaling

Ethics and issues

• Societal concerns
• Misuse concerns
• Misalignment concerns

Societal concerns

• Perpetuated bias
• Confident falsehood and sycophancy
• Copyright & IP issues
• Distribution of wealth and the job market
  • E.g. GDPval benchmark
• Climate footprint
  • Word for word comparisson, AI is cheaper
  • But, AI can generate a lot more text (rel. Jevons paradox

Misuse concerns

• Mass spear phishing, disinformation campaigns, ...
• Cyberattacks
  • Finding and exploiting vulnerabilities
• Enabling bad actors
  • Dual use of artificial-intelligence-powered drug discovery

Misalignment

• RLHF is only a step in the right direction
• Goodhart's law
• Misaligned leadership
  • What future are they aiming for?
• Superintelligence

Introduction to the hardware for this workshop

• Main reference: Alvis introduction material
• Compute clusters
• GPUs as compute accelerator
• Multi-GPU
• Containers
• Batch queue system

Compute clusters

The compute node

• Speed-up by parallelization
• Feeding data to GPU memory (VRAM) often bottleneck

Software

• Default software environment intentionally sparse
• Use modules or containers to run software
• (Follow our recommendations when installing Python packages)
• We will use containers in this course

Software – containers

• Apptainer containers
• A single file for your software and all dependencies
• (Building containers)
• Running software in a container

  apptainer exec your_container.sif python your_code.py
  

SLURM

• Batch queueing system
• Allocates resources in a fair and effective manner
• Resources are finite so expect queue times occasionally

SLURM workflow – Preparing job

SLURM workflow – Submitting job to queue

SLURM workflow – Job starts

Partial command overview

• View queued and running jobs squeue [--me]
• View previous jobs sacct
• Submit jobs sbatch <JOBSCRIPT>
• Cancel queued or running jobs scancel <JOBID>
• More complete overview at SLURM documentation and Alvis intro

Summary of Introduction

• Compute a key component to the success of LLMs
• Use and development of AI is not without its issues
• The hardware you can access
• Containers for accessing software
• SLURM batch queue system for running things on the cluster

Excercise

1. Finish prerequisites
2. Navigate to your instance of LLM-workshop
3. Do a git pull to get the latest changes
4. Launch an interactive session through https://alvis.c3se.chalmers.se/
5. Modify and then submit LLM-workshop/excercises/introduction/hello-llms.sh TODO