🔮 Strawberry; AI hyperpolation; digital twins; obsolete SaaS; dogs on LSD, nanotech & light hazard ++ #491

An insider's guide to AI and exponential technologies

Hi, it’s Azeem. Welcome to the Sunday edition of Exponential View. Every week, I help you ask the right questions and think more clearly about technology and the near future.

I have a handful of speaking slots left this year. Please get in touch asap.

---

Ideas of the week

The reasoning machines

I spent several hours playing with o1, OpenAI’s new model. The experience made me think that we need to reset our expectations of what problems can be solved. Most organisations have a set of implicit assumptions of what is and isn’t possible. Internal processes largely steer the organisation into the realm of the possible, and thereafter, the language of efficiency takes over. What isn’t possible is really short-hand for “things we can’t reason through, given the time and resources available.” Tools like o1 (and, more likely, its successors) will change that calculus, providing both the time and resources.

🍓 Playing with Strawberry: When machines can reason

Between, beyond and outside

Philosopher Toby Ord introduced the concept of “hyperpolation” in a new paper this week. In Ord’s words, hyperpolation “asks what lies in directions that can’t be defined in terms of the existing examples”. If interpolation is like finding a route between known cities and extrapolation guessing beyond the map’s edge, then hyperpolation is like exploring a 3D globe when given only a 2D projection. Hyperpolation may help to explain current AI’s shortfall. AI generates novelty based on remixing training data – interpolating between or extrapolating from known examples. But, it is currently unable to create or explore fundamentally new conceptual spaces and dimensions that are undefined by its existing data¹. Philosopher and computer scientist Judea Pearl captured this dissonance in his 2019 essay through the story of the ancient rivalry between Greek and Babylonian sciences: 

[T]he Babylonian astronomers were masters of black-box predictions, far surpassing their Greek rivals in accuracy and consistency of celestial observations. Yet, science favored the creative-speculative strategy of the Greek astronomers, which was wild with metaphorical imagery. [...] It was this wild modeling strategy, not Babylonian extrapolation, that jolted Eratosthenes (276–194 BC) to perform one of the most creative experiments in the ancient world and calculate the circumference of the Earth. Such an experiment would never have occurred to a Babylonian data fitter.

Pearl concluded that current AI “learning systems may get us to Babylon, but not to Athens”. Ord’s “hyperpolation” provides a term and concept to get us there. 

See also:

• A year-long study comparing AI-generated research ideas to those by expert human researchers found that LLM-generated ideas were rated as significantly more novel.

• PaperQA2 is an AI agent tool that can conduct entire scientific literature reviews autonomously, reportedly to a PhD and Postdoc-level.

• Toby and I talked about existential risks back in 2019.

It’s all me, me, me

I found this speech by Natalie Monbiot about digital twins quite intriguing. She argues that virtual dopplegangers could free us from life’s drudgery, handling preliminary job interviews or maintaining peripheral relationships. I already cede control of my rest and exercise to my fitness tracker, so why not cede other decisions? I’m not convinced that an AI avatar can capture all of a person’s essence. Even if it could, is this good for us? Philosopher Robert Nozick writes in his book Anarchy, State and Utopia that real, rather than simulated experiences, are what counts for a good life. By extension, we could never truly experience the toil of our avatars. In The Simpsons episode “Treehouse of Horror XIII”, Homer creates clones that can handle his responsibilities to avoid the commitment and grind of the day-to-day. Needless to say, it backfires. 

Storm clouds

Has AI made cloud-based SaaS obsolete? Klarna thinks so.