King Charles recently sat down with tech overlords to talk about AI. When a hyper-privileged old dude is freaking out about the implications of AI it’s a clear sign we have reached the peak of inflated expectations. But Charles isn’t alone in his fears. The leaders of the European Union, France and India have all been doing the same. The South Korean president has been getting involved and former UK Prime Minister Rishi Sunak is all over AI. Heck even political podcaster Rory Stewart has recently taken to explaining AI to his audience. And at Sentient Consulting we’ve been asking ourselves WTF does Rory Stewart know about AI…
Over at the Diary of the CEO podcast the analysts have discovered that inviting gurus to warn that humanity has six months to run is really good business. Quick, book another crank our Ad revenues are through the roof…
The general line is that artificial intelligence is advancing at an exponential pace. The internet is flooded with sages and seers foretelling that today’s models will soon morph into hyper-intelligent digital gods.
Meanwhile Sam Altman and Elon Musk and their chums are pleading with government to regulate them and slow them down. But we suspect their motivation is less likely to be genuine fear about the future and more likely that they have realised progress is about to stall anyway, whether they get regulated or not.
Because without a fundamental change in approach—essentially throwing current models in the trash and starting over—we are already staring at the ceiling of large language model intelligence. First we have to understand how AI got so impressive in the first place.
Modern AI models didn't get smarter because researchers unlocked the deep, magical secrets of human consciousness. They didn’t get smarter because the AI started training itself and writing its own code though god knows the folk at Anthropic would like you to believe that’s true. AI models got smarter because engineers threw unprecedented amounts of compute, data, and electricity at them. It’s called Scaling.
If a model with 10 billion parameters was impressive, a model with 100 billion was astounding! So, naturally, the industry decided that the answer to everything was simple: Just keep scaling.
Imagine sitting in front of a chess board and placing a dollar on the first square and then doubling to two dollars on square two. Scaling 2X to the second square only cost you one dollar. But doubling again will cost you four dollars and the next iteration will cost eight. Pretty quickly, long before you get even half way up the chess board, the sums involved in scaling every square get astronomical. And that’s at 2X never mind the 10X scaling that AI has been through over the last few years. But the next scaling round will cost trillions.
That is the simple wet blanket that means that AI progress will be slower in the future until someone comes up with a radically different architecture. And that we suspect is why Tech CEOs are calling for a slowdown.
AI improvement follows what scientists call neural scaling laws. In plain English: to get a small, incremental improvement in how smart an AI behaves, you don't just add a little more data or power—you have to multiply it exponentially.
Imagine you are studying for a high school exam:
- 1 hour of studying gets you a D.
- 10 hours of studying gets you a B.
- 100 hours of studying gets you an A.
- 1,000 hours of studying gets you an A+.
Notice how the amount of time you invest grows exponentially but the marginal improvement shrinks with each investment. AI models face this problem, but on a catastrophic financial scale. Each iteration needs 10x more investment for smaller and smaller improvements.
And as for the data needed to train the new models? Well they have already fed AI pretty much the entire internet. Short of training AI on its own AI-generated slop the world is running out of new training data.
Farmers in the UK already tried that idea in the 1980s when they started feeding their cattle with the remains of dead cattle and the result was Mad Cow Disease.
When the difference between a "good" model and a "slightly better" model is the cost of a large country’s GDP, the math simply stops working.
At its core, a Large Language Model doesn’t actually think. It is just an impossibly sophisticated version of the auto-complete feature on your smartphone. When you type "I'd like a cup of...", it guesses that the next word is likely "coffee" based on billions of sentence examples, and what it knows about you.
By scaling up compute and data, they made that auto-complete guesser mind-bogglingly accurate. It can mimic a lawyer, summarize a legal brief, or write a poem in the style of Shakespeare. But it is still, fundamentally, guessing the next token.
It has no internal world model, no genuine power of reasoning, and no understanding of cause and effect. Adding another trillion words to its training data won't give it common sense—it won’t even extend its vocabulary. AI has already absorbed every word we know and broken out the relationships between words from all the examples it was trained on. AI already knows that “coffee” has a strong relationship with “cup” and a weak relationship with “umbrella”.
Does this mean AI is dead? Not at all! It just means this specific method has peaked.
To achieve super-human intelligence (or even just an AI that doesn't occasionally hallucinate places that don’t exist), researchers will likely have to start over from scratch. They will need fundamentally different architectures.
Until that technological breakthrough happens, get used to the AI tools you have right now. They'll get a bit faster, and they are going to get a lot more integrated into your apps and devices and that’s about it.
All of which means that the skills you learn today, will still be relevant five years from now. There has never been a better time to take a course, experiment with some use cases and figure out how to make today’s AI tools work for you. Because today’s AI tools are going to be around a lot longer than people want us to believe.