Machines of Loving Grace — how AI might fix the world, and why the West must get scarier before that happens

We have come through a tough couple of weeks. The US elections, a sharply elevated (although perhaps overblown) threat of a nuclear event between Russia and the Western allies, extreme weather events in Spain, continuing war in the Middle East and sundry other roiling currents under which the planet seems to be submerged of late.

So, after having written a few columns about US politics and culture recently, I am returning to my more familiar beat in tech and science and I am determined to talk about optimistic things, happy things, hopeful things. Or even “... of shoes and ships and sealing wax, of cabbages and kings”, following the fine example of Lewis Carroll’s Walrus.

Which leads me to the “Machines of Loving Grace”. The phrase comes from a 1967 poem, All Watched Over by Machines of Loving Grace, by the late US poet and novelist Richard Brautigan. In it, he fondly imagines a world where humans can spend their time skipping through the daisies while benign machines take care of the drudgery of life. It’s a sweet thought, despite where we are now nearly 60 years later, surrounded by machines, without the drudgery quotient having changed much.

On 24 October this year an essay was published by Dario Amodei, co-founder of Anthropic, the company responsible for the chatbot Claude. Anthropic is perhaps the second-most-well-known of the “pioneer” AI companies and arguably the most responsible of them. Amodei’s essay is very long (26 pages, 15,000 words), more of a manifesto really, but, as occasionally happens when a respected tech titan gets into articulate future-gazing, it broke the AI internet.

It is titled (in a nod to Brautigan’s poem) Machines of Loving Grace – How AI Could Transform the World for the Better.

The reason Anthropic stands apart from OpenAI and others is their aggressive stance on safety. Amodei and his team are extremely sensitive to the dangers posed by two classes of risk – catastrophic misuse and loss-of-control over autonomous AI. They have baked in an entire series of frameworks and protocols to assess and manage safety throughout their product life cycles. Other AI companies have either followed suit or developed their own protocols, but Anthropic is generally regarded as the safety-first AI company.

Amodei’s essay is not about safety. It is about a brave new world. It is not freighted by the doomsaying and alarm of so much AI commentary. It is a praise song to this wondrous new set of tools and a roadmap to Utopia. It is, much like the original poem, a very sweet thought, as it paints an optimistic picture in which AI promises to cure cancer, eradicate poverty and clean the environment, so that the rest of us can smell the roses.

He pegs his intent at the outset:

“In this essay I try to sketch out what that upside might look like – what a world with powerful AI might look like if everything goes right.”

Amodei goes on to set the terms of the essay – he promises to avoid sci-fi scenarios like interplanetary travel, will eschew AI propaganda and avoid false prophecies of salvation. He promises to provide an evidence-led argument for where AI will take us if it is not derailed by any of the risks and threats that have so exercised his team since Anthropic was founded in 2021.

And so he gets down to the details, covering five major areas, which I will come to presently. But here we run into the rolling-eyes problem. Amodei’s timeline is five to 10 years and many of his predictions just seem, well, impossible, fantastical… magical thinking.

So please consider this: AlphaFold is a generative AI tool built to analyse one of the most difficult problems in microbiology – predicting the 3D structures of proteins. Before 2021 when AlphaFold was let loose on the problem, it would have taken thousands of scientists more than 100 years to analyse the human body’s 20,000+ proteins, using traditional tools.

It took AlphaFold a few months. And then, in 2022, it took a few more months to do the same for 200 million known proteins on Earth. That is, to put it mildly, a defining example of AI’s ability to scale quickly and exponentially.

Not just ‘sweet thoughts’