Don't believe the hype..? AI:Actionable Insights #9

Hello and welcome to the latest edition of AI: Actionable Insights!

We're talking Hype vs Reality in the world of Large Language Models (chatGPT, Gemini, CoPilot, Claude etc), hallucinations and a couple of tips/actual take-aways to help you get more from AI.

Let's talk about where we are today in the hype cycle.

I’ve seen a few LinkedIn and blog posts recently stating that AI is just a fad, that it won’t change the world much, and my particular favourite from a “thought leader” was the phrase: “All hype and no trousers”.

There’s also the incredibly reductive, and clearly logical fallacy, argument that goes along these lines I've seen too: “NFTs didn’t change the world therefore AI won’t” – just because some Tech Bros over-hyped a technology, it doesn’t change the reality of another.

Where do I stand?

I agree there is too much hype - there are people pushing the future benefits of tools that don't exist yet.

I also don’t think full Artificial General Intelligence is imminent – I don’t think it’s going to make everyone redundant, but I do think it’s already having significant impacts, will continue to do so in the world, and that many underestimate the impact of this change.

One of the issues in this disconnect between hype and reality is true of any exponentially growing technological change* - if you look back, it looks pretty flat, so you underestimate what’s ahead; If you look forward it’s hard to predict the scale of the change and the heights ahead seem unrealistic.

*Exhibit A: chatGPT 3 had 16,000 context token size in Nov 2022, the largest from Gemini is now 2 million – to translate that from geek speak: we've gone from being able to use about 1/4 of a book as context, to a whole bookshelf worth you can include in your prompts in less than 2 years.

*Exhibit B: From understanding text to video as an input, and soon output from text to video.

*Exhibit C – If you want to get really geeky, this article sets out a fairly solid case that growth in the capability of AI is exponential , I think his timeline is over optimistic and some argue his log scale graph is more of a gentle curve than a straight line, but I don’t think that changes the outcome much – just the time scale it’s likely to occur in.

What other reasons are there for a disconnect between the Hype and Reality?

Another contributing factor is that many folks don’t understand that chatGPT is (to use a phrase I stole from @KateBevan on Twitter) just a “spicy auto-complete”.

See this previous edition on the differences between search and why you should try tools like gemini.google.com and perplexity ai.

I’ve got an over simplified way I like to use to explain this, indulge me for 2 minutes:

Think of Large Language Models like ChatGPT as a big sprawling network of nodes connected to each other, each node is a bit of information (maybe a word, or part of a word) and the connections between these points are reinforced as we put more and more data into training.

In the “Introduction to AI” session I run I like to do a quick example of how this works:

I want you to pretend you’re an AI. I’m going to give you 4 pieces of training data, then I’ll ask you a question which you should answer as quickly and as instinctively as you can.

1. …the blue car was faster…
2. …fastest car was, blue…
3. …cars that are blue are fast…

Given the data above, what is the next word in this sentence:

The blue car is …..

You said Fast or Faster didn’t you?

Everyone does.

Is that correct? No.

When I do this in training sessions everyone confidently answers that way. This is what we refer to as an hallucination.

If I show you longer versions of that training data you can maybe understand how this could happen with an LLM:

1. …if I told you the blue car was faster…
2. …fastest car was, blue lights flashing…
3. …cars that are blue are fast becoming harder to sell…

If you train an LLM on billions and billions of pieces of information you can then ask it to find connections to "auto-complete" the next part. The larger the data set, the better the model and more importantly for us as users - the more specific a path you start it on, the better the outcome you’ll get.

By asking a very short, simple question – based on the limited training data I provided I hope you can understand why we reached that wrong (hallucinated) answer… which we as humans know is incorrect.

This is also a factor in why bad training data, can produce biased results –which we talked about Bias and the challenges Google had in Edition 3.

Actionable Insight

Aside from hopefully giving you a way to understand why LLMs (chatGPT etc) behave the way they do, here are some tips to help you get more from them when writing prompts:

1. Read this great guide to prompting by Ethan Mollick - there are many MANY folks sharing tips, from my experience this is the best.
2. Tell it to “Take your time” – a longer, slower process enables the LLM to step through the issue and produces better results. If you need succinct output, generate a longer version first, then refine it.
3. Think of it as a conversation, not a one-and-done - ask it to check how it did, or to review the output. Refine, repeat and break down what you need into steps.

Upcoming events:

June 11th: The final Simplify AI event (in-person) with Solent Partners – hosted at Eagle Labs in Southampton - book your place here

July 11th: Guest speaker at the next LinkedIn Local Bournemouth @ Village hotel, raising funds for MyTime – tickets and more info

Thanks for reading, if you have any feedback or topics that you’d like us to cover please do let me know - we're working on something very exciting to release to you all soon... stay tuned!

We've covered a lot in these 9 newsletters, if you’d like more tailored help with anything we’ve covered, please do get in touch.

Cheers,