86. Andy Jones - AI Safety and the Scaling Hypothesis
Towards Data Science
06/02/21
•85m
About
Comments
Featured In
When OpenAI announced the release of their GPT-3 API last year, the tech world was shocked. Here was a language model, trained only to perform a simple autocomplete task, which turned out to be capable of language translation, coding, essay writing, question answering and many other tasks that previously would each have required purpose-built systems.
What accounted for GPT-3’s ability to solve these problems? How did it beat state-of-the-art AIs that were purpose-built to solve tasks it was never explicitly trained for? Was it a brilliant new algorithm? Something deeper than deep learning?
Well... no. As algorithms go, GPT-3 was relatively simple, and was built using a by-then fairly standard transformer architecture. Instead of a fancy algorithm, the real difference between GPT-3 and everything that came before was size: GPT-3 is a simple-but-massive, 175B-parameter model, about 10X bigger than the next largest AI system.
GPT-3 is only the latest in a long line of results that now show that scaling up simple AI techniques can give rise to new behavior, and far greater capabilities. Together, these results have motivated a push toward AI scaling: the pursuit of ever larger AIs, trained with more compute on bigger datasets. But scaling is expensive: by some estimates, GPT-3 cost as much as $5M to train. As a result, only well-resources companies like Google, OpenAI and Microsoft have been able to experiment with scaled models.
That’s a problem for independent AI safety researchers, who want to better understand how advanced AI systems work, and what their most dangerous behaviors might be, but who can’t afford a $5M compute budget. That’s why a recent paper by Andy Jones, an independent researcher specialized in AI scaling, is so promising: Andy’s paper shows that, at least in some contexts, the capabilities of large AI systems can be predicted from those of smaller ones. If the result generalizes, it could give independent researchers the ability to run cheap experiments on small systems, which nonetheless generalize to expensive, scaled AIs like GPT-3. Andy was kind enough to join me for this episode of the podcast.
Previous Episode
85. Brian Christian - The Alignment Problem
May 26, 2021
•66m
In 2016, OpenAI published a blog describing the results of one of their AI safety experiments. In it, they describe how an AI that was trained to maximize its score in a boat racing game ended up discovering a strange hack: rather than completing the race circuit as fast as it could, the AI learned that it could rack up an essentially unlimited number of bonus points by looping around a series of targets, in a process that required it to ram into obstacles, and even travel in the wrong direction through parts of the circuit.
This is a great example of the alignment problem: if we’re not extremely careful, we risk training AIs that find dangerously creative ways to optimize whatever thing we tell them to optimize for. So building safe AIs — AIs that are aligned with our values — involves finding ways to very clearly and correctly quantify what we want our AIs to do. That may sound like a simple task, but it isn’t: humans have struggled for centuries to define “good” metrics for things like economic health or human flourishing, with very little success.
Today’s episode of the podcast features Brian Christian — the bestselling author of several books related to the connection between humanity and computer science & AI. His most recent book, The Alignment Problem, explores the history of alignment research, and the technical and philosophical questions that we’ll have to answer if we’re ever going to safely outsource our reasoning to machines. Brian’s perspective on the alignment problem links together many of the themes we’ve explored on the podcast so far, from AI bias and ethics to existential risk from AI.
Next Episode
87. Evan Hubinger - The Inner Alignment Problem
June 9, 2021
•69m
How can you know that a super-intelligent AI is trying to do what you asked it to do?
The answer, it turns out, is: not easily. And unfortunately, an increasing number of AI safety researchers are warning that this is a problem we’re going to have to solve sooner rather than later, if we want to avoid bad outcomes — which may include a species-level catastrophe.
The type of failure mode whereby AIs optimize for things other than those we ask them to is known as an inner alignment failure in the context of AI safety. It’s distinct from outer alignment failure, which is what happens when you ask your AI to do something that turns out to be dangerous, and it was only recognized by AI safety researchers as its own category of risk in 2019. And the researcher who led that effort is my guest for this episode of the podcast, Evan Hubinger.
Evan is an AI safety veteran who’s done research at leading AI labs like OpenAI, and whose experience also includes stints at Google, Ripple and Yelp. He currently works at the Machine Intelligence Research Institute (MIRI) as a Research Fellow, and joined me to talk about his views on AI safety, the alignment problem, and whether humanity is likely to survive the advent of superintelligent AI.
If you like this episode you’ll love
Promoted