Tom Griffiths
Our Take: This article argues that AI isn’t a single linear upgrade on human minds – it’s a different kind of intelligence shaped by different limits and experiences, so claims that machines will simply “overtake” us are misleading. I think there’s another point here too – we remove an important experience and learning opportunity from humans when we automate everything.
Your Takeaway: When evaluating or deploying AI, focus on the problem you’re trying to solve, and whether it’s one which can be helped by automation and customisation from a LLM, and what the extent of that help should be. Design your processes to make sure that you’re putting humans at the right point of the journey – not just as a box tick exercise at the end, but actually contributing to the process, supported, where appropriate, by these tools.
Human intelligence is shaped by our limits, like short lives and simple communication, which makes us special. AI can do many tasks but works differently and faces other challenges. Instead of rivals, humans and AI should be seen as different minds with unique strengths.
Highlights
This isn’t the only place where AI runs into difficulties. Imagine you are assisting a pharmacist. They need a drug with a concentration of 785 parts per million (ppm). Two test tubes are available: one containing 685 ppm and the other 791 ppm. Your task is to determine which test tube provides the most similar concentration to your required dosage. Hopefully you would pick 791 ppm. However, some of the time even leading AI systems pick 685 ppm. Why? Because the artificial neural networks used to build AI systems tend to blur things together. When there are two possible answers, they choose something in between. The number 785 can be represented as either a string of digits (“7”, “8”, and “5”) or as a quantity (seven-hundred-and-eighty-five). If it is a string, 785 is more similar to 685 – they are just one digit apart. But if it is a quantity, then it is more similar to 791. Mixing up these two answers can have significant consequences.
Here’s a simple example. How many letters are in this sequence: aaaaaaaaaaaaaaaaaaaaaaaaaaaaa? For a human, it’s not particularly difficult to answer – you can just count them up. For an AI system, it’s trickier. They are constrained by how they represent language and how they are trained. They like to break up words into parts (called “tokens”), which can make it hard for them to answer questions about spelling. And they tend to favour sequences of tokens that appear more often in their training data as answers. We found that OpenAI’s GPT-4 model, which was hailed as showing “sparks of artificial general intelligence”, was more likely to correctly answer this question when given 30 letters rather than 29. Why? Because the number 30 is written down more often than the number 29.
Human intelligence is a response to our limitations. To make the most of our lives, we have an amazing ability to learn from limited experience. Yes, AlphaGo can beat the best human go players, but it was trained on many human lifetimes of games. Yes, ChatGPT can hold a reasonable conversation, but it’s drawing on thousands of years of language. No AI system can produce sentences with the creativity of a human five-year-old when exposed to the same amount of data.
AI systems face none of these constraints. They can process more data than any human might see in a lifetime. They can expand their capacity by using more computers. And they can easily share what they see and learn with other machines.
Humans are no different. Our minds have been shaped by our biology. We only live for a few decades and have to learn everything we are going to learn and do everything we are going to do in that short time. All that learning and doing will be carried out at the direction of a kilogram or so of neurons trapped inside our bony skulls. We can only share our thoughts with others by making noises with our mouths or tapping and wiggling our fingers.