Pause or panic: battle to tame the AI monster - 0 views
-
What exactly are they afraid of? How do you draw a line from a chatbot to global destruction
-
This tribe feels we have made three crucial errors: giving the AI the capability to write code, connecting it to the internet and teaching it about human psychology. In those steps we have created a self-improving, potentially manipulative entity that can use the network to achieve its ends — which may not align with ours
-
This is a technology that learns from our every interaction with it. In an eerie glimpse of AI’s single-mindedness, OpenAI revealed in a paper that GPT-4 was willing to lie, telling a human online it was a blind person, to get a task done.
- ...16 more annotations...
-
For researchers concerned with more immediate AI risks, such as bias, disinformation and job displacement, the voices of doom are a distraction. Professor Brent Mittelstadt, director of research at the Oxford Internet Institute, said the warnings of “the existential risks community” are overblown. “The problem is you can’t disprove the future scenarios . . . in the same way you can’t disprove science fiction.” Emily Bender, a professor of linguistics at the University of Washington, believes the doomsters are propagating “unhinged AI hype, helping those building this stuff sell it”.
-
Those urging us to stop, pause and think again have a useful card up our sleeves: the people building these models do not fully understand them. AI like ChatGPT is made up of huge neural networks that can defy their creators by coming up with “emergent properties”.
-
Let’s not forget the excitement, because that is also part of Moloch, driving us forward. The lure of AI’s promises for humanity has been hinted at by DeepMind’s AlphaFold breakthrough, which predicted the 3D structures of nearly all the proteins known to humanity.
-
Noam Shazeer, a former Google engineer credited with setting large language models such as ChatGPT on their present path, was asked by The Sunday Times how the models worked. He replied: “I don’t think anybody really understands how they work, just like nobody really understands how the brain works. It’s pretty much alchemy.”
-
The industry is turning itself to understanding what has been created, but some predict it will take years, decades even.
-
Alex Heath, deputy editor of The Verge, who recently attended an AI conference in San Francisco. “It’s clear the people working on generative AI are uneasy about the worst-case scenario of it destroying us all. These fears are much more pronounced in private than they are in public.” One figure building an AI product “said over lunch with a straight face that he is savoring the time before he is killed by AI”.
-
Greg Brockman, co-founder of OpenAI, told the TED2023 conference this week: “We hear from people who are excited, we hear from people who are concerned. We hear from people who feel both those emotions at once. And, honestly, that’s how we feel.”
-
A CBS interviewer challenged Sundar Pichai, Google’s chief executive, this week: “You don’t fully understand how it works, and yet you’ve turned it loose on society?
-
In 2020 there wasn’t a single drug in clinical trials developed using an AI-first approach. Today there are 18
-
Consider this from Bill Gates last month: “I think in the next five to ten years, AI-driven software will finally deliver on the promise of revolutionising the way people teach and learn.”
-
If the industry is aware of the risks, is it doing enough to mitigate them? Microsoft recently cut its ethics team, and researchers building AI outnumber those focused on safety by 30-to-1,
-
The concentration of AI power, which worries so many, also presents an opportunity to more easily develop some global rules. But there is little agreement on direction. Europe is proposing a centrally defined, top-down approach. Britain wants an innovation-friendly environment where rules are defined by each industry regulator. The US commerce department is consulting on whether risky AI models should be certified. China is proposing strict controls on generative AI that could upend social order.
-
Part of the drive to act now is to ensure we learn the lessons of social media. Twenty years after creating it, we are trying to put it back in a legal straitjacket after learning that its algorithms understand us only too well. “Social media was the first contact between AI and humanity, and humanity lost,” Yuval Harari, the Sapiens author,
-
Others point to bioethics, especially international agreements on human cloning. Tegmark said last week: “You could make so much money on human cloning. Why aren’t we doing it? Because biologists thought hard about this and felt this is way too risky. They got together in the Seventies and decided, let’s not do this because it’s too unpredictable. We could lose control over what happens to our species. So they paused.” Even China signed up.
-
One voice urging calm is Yann LeCun, Meta’s chief AI scientist. He has labelled ChatGPT a “flashy demo” and “not a particularly interesting scientific advance”. He tweeted: “A GPT-4-powered robot couldn’t clear up the dinner table and fill up the dishwasher, which any ten-year-old can do. And it couldn’t drive a car, which any 18-year-old can learn to do in 20 hours of practice. We’re still missing something big for human-level AI.” If this is sour grapes and he’s wrong, Moloch already has us in its thrall.