Superintelligence? Not quite yet. This should have reached you this morning. I was in the middle of nowhere and forgot to send it. The AI system that runs my newsletter never asked for it, because it didn't think anything was wrong.
I started working with AI at the end of 2013, building music analysis and recommendation with a team of incredibly talented data scientists. Our model was a neural network. I asked how it actually worked. They shrugged. It just does. Fine, I said, then how do we make it better. More data. That was the answer every single time. We were told back then this could become superintelligence. It didn't.
Now I'm hearing the same tune about large language models. They're neural networks too, just enormous. An LLM is that friend with photographic memory and amnesia at the same time. Nigel Richards, the best Scrabble player you've never heard of, won the French Scrabble championship in 2015 after learning the dictionary by heart in nine weeks. He doesn't speak French. He needed a translator to thank the crowd.
Ask the engineers how the models work today and you'll get a better answer than a shrug. Ask how to make the models smarter and it's the same two words: more data. Problem is, they've already scraped the internet. Even OpenAI co-founder Ilya Sutskever said it on stage in 2024: "We have but one internet." So now they're training on data the models made up. Researchers have already shown where that leads: train a model on its own output for a few generations and it forgets the rare stuff first, then drifts into nonsense.
The companies spending trillions call that the road to superintelligence. I call it turning up the volume.
Same tune. Bigger speakers.
🎵 Daft Punk — Harder, Better, Faster, Stronger After Kraftwerk, the modern pioneers of robot music. Full playlist and archive at niccjohnson.com/drop.