29th February 2024
In this issue of That Space Cadet Glow I explore how reliant we might become on AI, using my own experience as someone who could easily fall into some sort of dependency. I also try and cut through some of the rubbish headlines to talk about some of the really big advances in AI that have happened in the last few weeks.
Enjoy this leap year edition of That Space Cadet Glow!
Is there such a thing as too much AI?
I have a terrible memory. If you asked me where I went on holiday a couple of years ago I would struggle to remember. It would be the same story if you asked me what gifts I received last Christmas (sorry to anyone that gave me a present!). My short-term memory, admittedly, is better, which is lucky as that’s a key facet of being a management consultant. But it’s still a challenge having to context-switch between different projects and different clients all the time. So I get a lot of help from all of the external systems I use, particularly Notion (where literally everything is recorded or written down), ToDoist, good ol’ Apple Mail and Fantastical Calendar, and now, more and more frequently, AI-based systems such as ChatGPT, Perplexity and Github Copilot. Even Notion now has it’s own LLM built in. I wanted to dig a bit deeper into this relationship between memory and AI, and see where it might be taking us (and me!).
Firstly, it’s important to understand why we forget. The New Scientist’s latest issue has some really interesting articles on neuroscience, including this one about why forgetting things is a fundamental part of the way the brain works. First, we need to distinguish between natural forgetting and pathological forgetting - the former is what most of us experience, such as the examples I used at the start of this article. The latter results from brain injuries or conditions such as Alzheimer’s. For the everyday natural forgetting, it doesn’t often help to remember every single detail, in fact it inhibits our ability to generalise - we don’t need to know every detail about a chair in order to classify it as a chair. But also, the brain is a prediction machine, where conscious experience consists of predictions about what the brain expects to perceive based on what it detects via the senses. As we experience more, we are not necessarily forgetting; we are updating our expectations based on the new information.
The parallels here with AI are particularly interesting. AI generalises on data, and an AI model that remembers too much (referred to as over-fitting) is no use for making predictions. So, should we be putting more trust, and building even more dependency on, artificial intelligence? Are we aligned enough with the machines that we can seamlessly incorporate them into our everyday lives to augment our inherent weaknesses? (BTW, I’m not going to go down the Brain Machine Interface rabbit hole here - I’ll save that juicy debate for the next edition). This article from the Financial Times earlier this month looked at how reliant on Generative AI we might become, and whether that was a good or bad thing. The conclusion, which may sound a little like fence-sitting but, as is often the case, is the correct balanced view, is that “sometimes the AI is better than you, and sometimes you are better than the AI”. This is illustrated using two examples: a quiz where you have to guess how well GPT-4 will answer a question, and a case study carried out by BCG consultants. The quiz, created by Deeepmind scientist, Nichola Carlini, shows that GPT-4 can do some things remarkably well and very fast (“Write a html page with javascript/canvas to draw the US flag that changes colour when I click on it”) but other things which most kids could do, such as playing Noughts and Crosses, it does poorly. It can compute the integral of x sin(x) from 0 to 2π, but it can’t solve much simpler maths problems. It can write poetry but it can’t play Wordle. The problem, it seems, is working out which things the AI can do well, and which it can’t. After a while you might get a feel for this, but (and here is the irony), you have to be able to remember which are the tasks done well and which aren’t.
The second example in the FT article describes an experiment carried out by the consulting firm, BCG. They gave half their consultants ChatGPT, and measured how productive each group was. As expected the AI-enabled group did much better, but they also fell into a dependency trap. One of the tasks deliberately had some confounding information in it, which could fool ChatGPT. The AI-enabled consultants fell into the trap, believing the AI’s answer more often than not, whereas even junior consultants were able to spot the error when working without the AI. So, not only do we have to remember when to use AI and when not, but also to fight the urge to want to use it all the time. The article compares this growing dependency to that of the iPhone and how quickly it has over-substituted in our lives: “We want company, but rather than meeting a friend, we fire off a tweet. We want something to read, but rather than picking up a book, we doomscroll. Instead of a good movie, TikTok”.
The confusion and dependency will only get more challenging. OpenAI are claiming to be developing Generative AI ‘agents’, where the AI carries out a series of tasks to achieve a specific objective after working out what each of these tasks needs to be. This will undoubtedly be really useful but will also obfuscate any weak points in the process. And with GPT-4 going berserk last week, many people will have trouble trusting their key tasks and activities to it.
How do I supplement my poor memory if I have to remember each time which task I should use AI for? If I do rely on it, how do I know it is doing the best thing for me? And how do I ensure that I don’t get sucked in to an AI-in-everything world and forget how to be human? Like most things in life, there is no easy solution. AI is certainly not the silver bullet that many vendors will tell you it is.
This New York Times article focuses on the impact of AI on work, but its sentiment applies to my questions too: we need to believe in the potential of humans just as much as we do in the machines. It’s clear we all need to be acutely aware of AI’s strengths and weaknesses, but we need to flip the debate. The focus should not be on how AI can threaten us, but what are the unique strengths and abilities of humans that we can promote and develop, whilst using AI as tool to help us get there. The article quotes Minouche Shafik, who is now the president of Columbia University, who said: “In the past, jobs were about muscles. Now they’re about brains, but in the future, they’ll be about the heart.”
So, I should really stop worrying about forgetting stuff (it’s normal), make sure I use the right tools just enough (so I can do my job better) but not too much (so that I don’t become overly-dependent), all of which means I should hopefully be able to focus on doing the things that I’m actually pretty good at.
The Wrong Story
The product releases, especially around Generative AI, are relentless at the moment. But what actually makes the headlines are often the follow-on stories, most of which don’t deserve the column inches they attract. Last week was a prime example: Google had just released an updated (and renamed) version of its LLM, Gemini, but within days there were posts across social media telling of racially-inappropriate images that the model had generated, including racially-diverse Nazis and Black vikings. The headlines shouted about how the model had gone ‘too woke’, as though being woke was a bad thing. But they were all shouting about nothing much at all. Google had simply over-compensated some of the guardrails following the (rightfully) stinging criticism of previous models that were clearly exhibiting racist and misogynistic views. This time the images were just historically wrong, and now Google are looking to fix it, but not before their share price dropping by over $70 billion, the CEO reportedly issuing an internal memo and Elon Musk mocking his rival in the worst possible taste - all because of the narrative portrayed in the headlines. This article by Tess Buckley for WeAndAI highlights the role that headlines can play in people’s perceptions of AI, leading to a lot of the irrational fear that deafens people to understanding the real benefits and the real risks of AI.
Meantime, the really interesting story was lost in the mist: Google’s most advanced LLM, Gemini 1.5, may, at last, be the one that dethrones GPT-4. Key to this is its ability to ingest up to 1 million tokens in each prompt. That equates to around 10 novels. Compare this to GPT-4’s 128,000 tokens and you quickly realise this is a huge step forward. For those of us building LLM applications this removes one of the biggest headaches - that of having to chunk up large texts and then retrieve the bits we need to answer the question, which, as well as building in technical debt, often leads to incomplete or unsatisfactory answers. With its huge context window, Gemini 1.5 can read everything all at once and provide very precise answers. I’ve always said that the real value from LLMs come from when you point them at a specific corpus of data, and now Gemini 1.5 will allow you do that that much, much more easily and with much better results. I can’t wait to get my hands on it.
And while we are talking about big steps forward, OpenAI released Sora, a video generation foundation model. Just like Gemini 1.5, the advance over previous models, such as Runway, is massive. We have gone from the ability to make slightly weird videos that are a few seconds long to realistic one-minute videos. The results aren’t perfect (just watch the feet of the lady walking through Tokyo) but, jeez, they are impressive. We are now getting into the realm of AI-generated advertisements, and, before long, short TV programs.
Thee Sacred Souls - Tiny Desk Concert
A few issues ago (#130) I recommended you listen to Thee Sacred Souls because of their beautiful soul sound and meaningful lyrics. Since then I have come across their Tiny Desk concert, which is simply a joy to watch. I could write loads more about why they are so good, but it’s easiest if you just watch and listen to the 5 tracks they perform here:
I wanted to tell you about a new paid subscription level for That Space Cadet Glow that is available. For the price of a round of drinks (£40) you get access for a whole year to a special podcast version of this newsletter which includes bonus content and music snippets, plus emailed extracts from my updated book and extra interim content via the Substack app. I’ve made the December audio version (which includes additional content) available to everyone so you can get a feel for whether you prefer listening or reading (or both!). Just click here to access it for free. If you like it, all you need to do is go to this link here to sign up to access all future episodes and the bonus content.
Finally, I wanted to mention the number of paywall links that I’ve used in this issue. Normally I try to find free links to the same subject, but increasingly I’ve found that the best content, the best writing and the best perspectives are through subscriptions. I really don’t have an issue with this, as I’m glad to pay for quality content, and I would suggest that this is the most moral way to get the information you need.
The next edition of That Space Cadet Glow will dive into the world of Brain Machine Interfaces, an AI-enabled technology which is making big advances at the moment, but also causing quite a bit of concern to ethicists. In the meantime, look out for the audio version of Issue #133 in the next week or so.
Thank you for reading.