Doomsday AI?

Table of Contents

We’re living a very special moment in Generative Artificial Intelligence. Frontier models, including soon-to-be open-weight ones, are becoming very capable and effective at sustaining days-long tasks (burning several million tokens along the way).

AI-inflated expectations are not new. The history clearly marks at least two major “AI Winter” periods where Artificial Intelligence as a practice field retracted due to limited real-value creation — or simply very high inflated expectations.

With the current advancements, what is happening is a bit different: Instead of skeptics, we’re seeing the rise of Doomsday Prophets of GenAI. Those think current GenAI systems are the ultimate and final stage of technology evolution. Those also think there will be one lab/model to rule them all. And they also believe the only possible outcome is unparalleled unemployment, poverty, and uncontrolled cybercrime. In essence, they don’t believe humans will triumph over a technology humans had created. Spoiler: that never happened before in the history of our civilization. Humans dominated Earth due to reasons, including our biological sense of survival combined with our collective wandering and inventiveness.

Be warned about those folks: they’re either being driven by bad intentions (i.e. spreading fear, uncertainty, and doubt), being plain stupid (Dunning-Kruger effect is endemic on the Self-Proclaimed GenAI Experts Wonderland), or trying to over-sell something to you. Most of the time, they’re a very dangerous mix of all 3-in-1.

Yes, it’s concerning to see what the latest Anthropic and OpenAI models can do around “cyber offense capabilities”. And assuming frontier Chinese labs are not there too (or even further) is just naive. But, all those risks do not outweigh the potential benefits of this same technology being used safely and responsibly. We need to be vigilant and do our part to make sure the current and next-generation of AI systems will benefit society more than harming it. Formal methods, defense-in-depth, stringent sandbox boundaries, tamper-resistant hardware attestation procedures, and a long etcetera existed well before GenAI. And they’re as important and current as ever.

Embargo, export controls, and other physical-world mitigation options may help in the very short time, but they will not age well (they never did). Actually, there are already early signals that the current silicon embargo does not help keeping “bad actors” from pushing the frontier forward. Scarcity, as in other times of tech history, is the mother of invention.

I’m choosing to be ahealthy paranoid-optimist about the future and advocate for a thoughtful, sometimes contentious, debate around AI regulations. Not driven by government and public politics. But by acredible, well-known, well-managed, and well-funded body(https://en.wikipedia.org/wiki/Self-regulatory_organization).

The Perfect is the Enemy of Good

I really hate this phrase most of the time. I love striving for perfection whenever I can. But with this unprecedented pace of innovation on Large Language Models (and on what’s next to them), if we wait for the perfect solution to land, it will be born already obsolete.

Government and parliaments around the world are trying to encode AI laws they’re not prepared nor willing to really debate, in good faith, the pros and cons of this generation of AI systems. Weak countries, institutionally speaking, will take the wrong decisions –– and be wiped out from the potential benefits of this transformative technology.

Responsible Self-Governance

The securities industry in the US is a great example. It’s not infallible, but they don’t pretend to be perfect either. US Securities Markets are not only the biggest in the world. They’re one of the most credible and highly supervised markets on Earth. Meet FINRA.

And read more from Demis Hassabis from DeepMind.