It’s the most important question: What really matters? We could journey far back in history, when the ancient Greeks discovered the phalanx and the hoplite, thereby democratizing power structure. A few ordinary citizens could now halt a nobleman in his chariot from the mythical Trojan age and that in essence is how democracy could be born – and with it explosion of human ingenuity and science and art. Similarly, making access to AI easy truly matters. We don’t want a scenario where a single superpower monopolizes control over this extremely crucial tool. Just like we wouldn’t have modern day world if not for Linux, imagine world where there is just one company that controls operating system for all the cloud. It’s essential that a vast array of organizations can create their own versions. There should be thousands of ChatGPTs, not just a few and definitely not just one. Every rebel who opposes the establishment should have the opportunity to produce their own version, every university or research agency should be able to create their own. Training must become much cheaper and way faster—costs in data and computational power need to be normalized by at least an order of magnitude.

It’s fundamental question, what really is important for future of AI. Speed is all we need. Imagine if chat-gpt would be 10 times faster, so it can talk with you in realtime, generate actions (set calendar events, make notes, send messages) while converting speech to text and validating every prompt against hallucinations and other mistakes and comparing it with some reliable data sources. It all can be done right now, but it would take computational power that isn’t affordable for great part of society, even if we speak about quite wealthy western society. Imagine if such a chain of chat-gpts could generate code – run it – debug it – write unit tests – validate them – do code review to clean up code – write documentation – finally publish them. It’s all quite possible, but it’s so computationally expensive. We need speed, affordable speed. We need order of magnitude better solutions in terms of speed/buck.

What we really need is affordability and abundance. And we cannot rely solely on hardware; it’s neither wise nor efficient. Over the course of an entire lifetime, the human brain processes merely a minuscule fraction, essentially a rounding error, of the data that large language models are fed during just a few months of training, yet it outperforms them in many, many ways.