I was excited to read Rebooting AI (website), to find inspiration and tools for doing things better. Here is the book in one great quote:
For now, we are in a kind of interregnum: narrow but networked intelligences with autonomy, but too little genuine intelligence to be able to reason about the consequences of that power.
There is a lot to like. Marcus & Davis clearly map out the history and landscape of AI challenges, plus plausible elements of future solutions. They provide useful tools for thinking about problems with partial solutions to intelligence, such as the “fundamental over attribution error” and the “illusory progress gap”. They show how current ML solutions based on big data can be opaque and brittle. They demonstrate how key attributes of human intelligence instead allow the development of rich cognitive models – such as how language and the real world work – and how solutions incorporating such models would address current shortcomings, enabling AI to tackle open-ended tasks. This is great material for a general reader.
Where I felt the book fell short was that it didn’t build many bridges between our current “narrow but networked intelligences” and the authors’ posited future state capabilities. The future state reads like Artificial General Intelligence (AGI) by another name, fleshed out by scenarios that are short on implementation detail. Though sometimes mundane, from our current perspective, Arthur C Clarke might describe as them “indistinguishable from magic” and hence Rodney Brooks would say they are “no longer falsifiable”. We know there’s a massive chasm between current ML solutions and AGI, but I didn’t find much to close or bridge the chasm in Rebooting AI.
Some of these future capabilities are illustrated by domain-specific modelling techniques – like formal logic – that would be familiar to many computer science students. But I found this a little incongruous because these techniques have also failed to deliver on promises of realising intelligence, and not done any more to squash the “long tail of edge cases” than other narrow intelligences. Given the diverse facets of intelligence, maybe the paradigm of “narrow but networked intelligences” is the best way to achieve or approximate intelligence, or maybe it’s ultimately illusory progress, but these illustrations didn’t help me resolve that.
There is undeniable value in the current generation of ML solutions. How do we build on these? A detailed analysis of a number of key avenues of short to medium term progress was lacking. For instance, starting with current ML solutions, the authors could have explored:
- various designs of hybrid human-machine decision-making systems that augment human abilities while remaining resilient to new scenarios that stump machines;
- transfer learning, few-shot learning and sophisticated representation learning like transformers, that have potential to increase the representative and reasoning power of solutions;
- the role of ecosystem design and governance, including ongoing monitoring and data curation to correct issues (for instance bias testing, CD4ML, etc).
Instead, ML was stereotyped as fully automated, tabula rasa, E2E.
Finally, to know things are getting better, we need the right baseline and measures. While the language examples clearly demonstrated superficial artificial understanding, and self-driving vehicles have a ways to go, some issues raised were not assessed against incumbent human capabilities on narrow tasks in a like-for-like comparison, but rather against posited capabilities of a future AGI system. I would agree that humans can individually reflect and introspect to recognise their mistakes, but it is still the case that, in operational scenarios, humans make mistakes like artificial systems do. These operational mistakes are moderated by the wider ecosystem in which humans operate, in the same way as predictive inference mode is moderated by a wider human-machine ecosystem. I felt the core issue in some instances was structurally unsound or concentrated decision-making without proper governance, rather than whether or not mistakes were made, and this confounded the analysis. I would have liked to have seen these factors teased out so comparisons could be made in a way that would help to measure progress.
Marcus & Davis do lay out a helpful framework for building trust in AI systems, including stress testing, understanding costs of failures, building in modularity and maintainability, etc. This is good guidance but it would be really helpful to see more detail or case studies under these headlines, to the specificity of other works like Weapons of Math Destruction and Made by Humans.
So, maybe I was hoping for “Refactoring AI” rather than “Rebooting AI”. The book certainly clearly describes problems with the current state, and desirable characteristics of the future state. On balance, the technical arguments may indeed be sounder than my concerns. If you’re curious, I would encourage you to read it and draw your own conclusions. Ultimately, however, I’m disappointed because I didn’t leave inspired and equipped with new insight and new tools for improving AI today, tomorrow, and the day after.