James Bond villains and Anthropic’s CEO say the same things. That they have built something that can end the world. The difference is that Dario Amodei appears to feel bad about it. He reminds me of a Marxist who flies business class and probably collapses onto its flat bed in deep tragic empathy for those in coach.
Why do Big AI CEOs reveal that their products are dangerous? In these admissions, there is a bit of bragging about how advanced they are. For instance, Meta’s AI lived in the disgrace of not being able to hack anyone until a few days ago, when it too made the cut. Also, they want to say they are good guys.
But a more likely reason is they want to keep doing what they love without ordinary people restraining their pursuits through their government.
Scientists are artists. They work on beautiful ideas for the sake of the ideas, but must appear to make the world a better place or the world won’t leave them in peace.
So, the humans of AI, in their desperation to keep working on their tech, have a reason to pre-empt regulation by creating it themselves.
In July, Demis Hassabis, Google DeepMind’s then CEO and now chairman, published A Framework for Frontier AI, proposing a US-led independent body modelled on financial self-regulatory organizations to evaluate advanced AI models before they go public.
The industry is mistaken in thinking self-regulation will save it from governments. In any case, appeasing governments may do little to assuage the anxieties of regular folks.
So, I have an idea. Big AI should stop pandering to all and just focus on a majority. It can create a giant body funded with billions of dollars from AI firms to fight an evil that is not perceived as evil— gatekeeping, which is a tyranny of elite opinions in many fields.
In any endeavour where tens of thousands apply or create something, there is no human way to fairly analyse all of them, and the inevitable subjective basis of judgement is often intellectually corrupt and opaque.
The internet did diminish some kinds of gatekeeping, especially in the arts, but still the fraud of prestige has largely survived.
I feel that AI can completely destroy human gatekeeping. Consider one way in which science is allowed to become science. If Albert Einstein had sent his revolutionary ideas today to a reputed science journal, they would not even be read because why must an editor read the thought experiments of a patent clerk with no credentials? He was lucky to belong to a time when the world of science was very small.
God knows what gems are buried in the slush pile of scientific journals. There are just too many ideas today because of creeping equality, and the gatekeepers of prestige and credibility can’t do justice to all.
The anti-gatekeeping foundation, let’s call it ‘Humans of AI’ (or ‘No-gates Foundation’?) would be the repository of all human knowledge, opinion and art. It will ‘read’ every word of all scientific papers sent to it in any language. If a paper has merit, Humans of AI would approve, with the reputation of its tech tied to the quality of the papers it identifies as real science.
Thus, the awarding of prestige would move from a few to all scientists on record who ever lived. We may discover gems from the self-taught and geniuses from obscure nations.
Also, the foundation can transform college admissions. Ivy leagues that do not have an objective selection process claim to consider thousands of applications fairly. This is humanly not possible. They filter them using social and other biases. Humans of AI can take over the filter, or the college can admit it does not wish to have a transparent process.
It can transform arts. There is a famous story of how Matt Damon and Ben Affleck, in their debut script for Good Will Hunting, which they sent to many studio executives, added a random gay-sex scene in the middle of a conversation. It was to check if anyone was reading. Only one executive noticed.
Stories are among the most oversupplied things in the world. So, gatekeepers read very little from unknown writers. Humans of AI can be trained to read millions of works and compare them to everything that has ever been written, and judge them by their potential to be popular or even touch our souls.
Awards in subjective fields are bizarre. I myself have been a beneficiary, and a part of a literary jury—for a Commonwealth prize. Unknown hands filtered out hundreds of novels and sent me 50 or 60. I couldn’t read more than a few pages of most. In any case, how does my appreciation matter to people who are not me?
Humans of AI can be trained by actual readers, not just writers and humanities professors. Millions of interviews with readers, when they are sitting alone, watched by no one, telling AI what it is that they like in a book, play or movie, will assist AI to identify what is ‘good.’
This will not work as a mass recommendation, no matter how good the tech is. Every person would need his own list of recommendations. AI would do better than the current ‘you may also like’ farce that takes genres and awards too seriously, as though they are real categories of stories.
There is, of course, a risk in transferring gatekeeping to AI. Creators who are ignored by the mainstream, who feel belittled, at least have the dignity of assuming human bias. It can be devastating to know that all of humanity, too, probably rejected him. And AI could be wrong, too.
But this is still better than the overt conceit of the acclaim industry where a handful of people decide for the rest what is good, or what “can work.”
The author is a journalist, novelist and screenwriter. His latest book is ‘Why the Poor Don’t Kill Us.’