AI is a Bubble

Like dot-com it'll take decades, not years, to see the real value

Context - I love AI and it’s a huge part of my life

I did my PhD in AI. I’ve published several AI research papers. I spent four years as co-founder of an AI startup (which failed). I work as an AI consultant and I’m starting another AI company. And, most importantly, I pay for AI products I use every day:

• Midjourney helps me make personalised birthday cards for my friends

• ChatGPT helps me revise my written work

• Github CoPilot helps me write code

Sadly, I’m now certain that AI is a bubble. Far too much funding is pouring into the space given the expected value AI will create. It’s also clear that many people are investing their time & money in AI because they see others doing it, rather than because they understand what it’s capable of.

This may seem contradictory, given I’m claiming to use AI daily. But the key part of each use case is that AI helps me. There’s no way I’d let any of these AI products write code, or edit my work, without checking its output. This puts a sharp limit on how much value is created. I don’t think most businesses spending money on AI, or investors pouring in funding, recognise how sharp this limit is.

First a step back: why is software so powerful?

Computers can reliably complete incredibly complex tasks. But under the hood, all software is built from super simple functions. They take some input and transform it into an output. This output is passed to another function. Input, transform, output, repeat (not as catchy as eat, sleep, rave, repeat … I know) 

Long chains of simple functions create huge value

The software we use is just long chains of millions of these functions. They work because each function in the chain is 100% reliable. If each new function you add is 100% reliable, you can keep making the chain longer and it will still work reliably. Over time, these chains have kept getting longer and software more powerful.

But what if functions weren’t 100% reliable? What if they were 99% reliable instead? Suddenly, you can’t build these long chains any more. Forget millions of functions: 1000 functions of 99% reliability each would only work 0.004% of the time. 

Chains with a broken link have little value

AI is not 100% reliable. Not even close. It never has been - boring old machine learning, which generative AI comes from, wasn’t reliable either. So you can’t just treat AI as another function to stick somewhere in a long chain. The fact that “it could” open up new capabilities doesn’t mean that “it will”. Unfortunately, “it could” is equivalent to “it won’t” in most software use cases. 

But you’re already using AI every day?

I’m not arguing that AI won’t create any value - it already has. But the use cases are much narrower than most people think. Both the following conditions must be satisfied for generative AI to provide value: 

1. The AI must be at the end of the chain of functions so that nothing downstream gets messed up

2. There must be a human checking the output in some form. 

This means the bar for reliability can be lower than 100%, in some cases much lower. 

Let’s take an image generation example: how “reliable” is Midjourney at making friends’ birthday cards for me? On average, I generate 50 images before landing on one that seems good enough. Arguably, its reliability is ~2%. In this case, 2% is ok because I’m always checking the output. I would never in a million years just let Midjourney do its thing and send my friend their birthday card.

Midjourney therefore isn’t a function that can be slotted into another product. A card delivery company, like Moonpig, can’t just add Midjourney to its tech stack and offer its users a new feature like “make a personalised card with a simple prompt”. The feature would have to be “make personalised cards with 50 prompts and several hours of your time”. This is far, far less valuable.

I want to flog this horse well beyond its death, so here’s another example. Take GitHub Copilot, which I use to help me write code. When I ask Copilot to write a function, it’s correct about 50% of the time. However, I always have to check what Copilot has written, because it’s not that reliable and, while it will improve, probably won’t be reliable any time soon. While it’s a productivity improvement, it’s not a huge one because I still always have to check (and make changes when it’s wrong).

It’s not good enough for Copilot to be “nearly correct”. If it writes a 50-line function, but one line has a mistake, the whole function is useless. These “nearly correct” cases are particularly dangerous. If Copilot writes code that looks correct but is actually wrong, I’m less likely to spot the errors.  

The people who are most excited about Copilot - who think it will replace software engineers - are the people who’ve never used it. Often they’re the people who can’t code at all.

Who’s inflating the bubble?

The AI bubble is a perfect storm of:

1. People who don’t understand AI, but are pouring money into it

2. Experts who do understand AI, but are incentivised to keep the money flowing

People who don’t understand AI

People pouring money in can be divided into i) companies and ii) investors. 

Let’s start with companies. A big company CEO uses ChatGPT and thinks wow this is amazing, it just wrote me a rap battle between Stalin and Hitler.  Despite being “amazing” ChatGPT has created nothing of real economic value here. Anyway, their amazement quickly turns to fear of missing out (FOMO), so they instruct their CTO to create an AI strategy and start building features.  

Now the CEO isn’t unreasonable, so he allocates a budget for the CTO to start moving things forward. My calls with companies always go something like this:

Company: We’re so excited to start working with you. We’ve already got budget approval for some big AI projects!

AI Consultant: Great! So what would you like to do with AI?

Company: Well…we were hoping you could tell us that.

Clearly, AI is the answer. Companies just don’t know what the question is yet.

What about investors? They’re encouraged that companies are already spending lots of money on AI, so it can’t be another bubble. Investor FOMO takes the form of lots of time spent on LinkedIn keeping up to speed with all the latest demo videos. They get particularly excited about the possibilities of linking new AIs together, like Google’s MultiModal model. This combines text, image and video AI to create something even more powerful!!! 

Sadly, most investors never take the time to use AI extensively themselves. They never bother to deeply interrogate whether AI is adding as much value as people claim. After all, if they waste time doing their due diligence, they may miss the next demo video.

AI Experts

What about the people that do understand? Some big tech companies, like Google, certainly know about the limits of AI. Unfortunately, they too are incentivized to continue the illusion. Look at the billions of venture capital that’s gone to AI startups with no customers. Look at big tech share prices since ChatGPT. Remember that Google MultiModal AI demo video? Turns out they faked it - they want the AI bubble to keep inflating too.

Almost no one has an incentive to burst the bubble. I too felt conflicted writing this because I’m founding another AI startup - I’d probably sign up more customers and raise more money if I just played along. But I’m too much of a zealot to keep my mouth shut. Hopefully, my girlfriend will forgive me when we next argue about whether we can pay our rent with the truth.

Where do we go from here?

I don’t want to come across too negative. There’s genuine value being created with AI right now, and I use several AI products daily. However, the funding piling into AI is completely disproportionate to its value, not only its value today but its potential any time soon. Unreliable functions just don’t fit well with the rest of software.

Instead, AI will follow the same pattern as the internet. In the ‘90s, some genuine value was created from businesses moving online. But the internet wasn’t ready yet for most businesses. The value created wasn’t close to justifying the funding piling in and, eventually, there was a big market correction. The dot-com bubble burst. 

Decades later, most businesses are online and the internet has transformed society. The same will probably happen for AI but it will take decades, not years, to figure out how AI fits in with the reliable functions that make up the rest of software. This is what a top AI engineer does - they break a high-level problem into the correct components for AI to tackle. For the foreseeable, this will be on a case-by-case basis. 

I’m staying in AI and I’m excited for the future. Really, I am. In the meantime, I’ll refrain from having my mind blown each time a new demo video, based on cherry-picked examples, goes viral. It’s well known that 1,000 monkeys with 1,000 typewriters can make something pretty impressive.