In this article:
Jonny, ready for the rise of the machines?
Haha, this is not the Terminator.
But is it coming for all our jobs?
AI is a big theme, but there's a lot of misconceptions in the press and we should understand what it is and what it isn't. There's advanced general intelligence, the big kahuna, but we’re a long way from that. Within that, there's machine learning and then there is deep learning, where computers behave in a humanlike manner using neural networks. Deep learning is where some pretty amazing things are happening.
Where does it go from here?
There's a lot of work going on and what’s amazing is that people are finding that bigger models are yielding better results. When you think about complex systems, normally if you make a bigger model, you have more crud, more connections, and diminishing returns. But actually what we’ve found recently is that for large language models, where GPT 2 has gone from 1.5 billion parameters to over 1 billion in GPT 3, it has produced amazing new results for Google, Facebook, Microsoft, and other people. So these guys all have a new tool and a new weapon to use against each other.
But it's not really an edge for any of them against each other.
We can split this up into two parts. First there are algorithmic advances in terms of formulas and approaches, and there's data which you need to feed the beast and train these models up. A lot of the algorithm advances have been open sourced - OpenAI is one of the leaders here - which is unusual. So everybody can access the same formula.
Now second, does everyone have the access to the same data to train these models on? That's where incumbency has an advantage, both in terms of data to train and the financial resources to fund the training, because these models can cost hundreds of millions of dollars to train.
I think also in terms of routes to markets, the large companies’ apps are already piped straight onto your smartphone and effectively straight into your brain because you spend your whole day on the smartphone. So this favours the large incumbents against disruptive new entrants.
But this isn't really something that's going to take us all the way to AI doing all of the tasks humans would rather not do.
Deep learning systems use neural networks and they are amazing at two things. One thing is they can find a needle in a haystack. So, for example, show it 10,000 pictures and it will find the one with the cats better than any human can do. More importantly, it can also do that procedure in reverse. So it's really good at generating amazing pictures of cats. ChatGPT does amazingly lifelike speech; and stable diffusion produces amazingly lifelike images. All of which is this cat generation function on steroids.
The problem is the computer doesn't know it is a cat. It just knows it's generating a picture relating to previous statistical items it's seen, it doesn't understand. And that means it's very good at generating a lifelike response, but it has no idea if it's the correct response.
Now, if you have roles where what matters is generating something very lifelike with, say, 95 per cent accuracy, like generating the background to a movie set or generating a cartoon or a picture, great. These machines can do that all day long, better and cheaper than a human. But if you have a task where it needs to be correct, like landing a plane on a runway with a fly-by-wire... bad idea. Because if you're wrong one per cent of the time, bad things happen.
Those margins matter and so we know we still need humans to do those things.
Yes. So there are creative industries where what matters is being realistic and not correct. And that's what these models are great for. There are other industries where being correct matters and this is a bad product-market fit.
Also one of the big plays has been moderation companies out of India. Is this a big step forward for them?
These amazing bots can be used as a tool to enable humans to do more value-added things. The AI is great at filtering out the noise, but you still need human moderation at the end to figure out, you know, the one per cent that was or was not right. What happens here is the human uses a tool and it doesn't take away his job. It just changes the job and lets humans do higher level tasks.
Who do you think will benefit commercially?
There will be people who gain who make the software, make the services. But it's so hard to tell at this point. The analogy I always use is that 4G was amazing technology. 4G created Uber, 4G created Airbnb, 4G created Instagram. All came about because you had a mobile and you were connected at all times. But it was very hard to forecast ahead of time.
What you could predict was that 4G was going to happen and Apple was going to sell iPhones. And similarly, if you think an AI revolution is going to happen, then Nvidia and others are going to sell chips. And therefore, you know, the easy way to play is via the arms dealers. Because it is so hard to pick the winners higher up the stack.