May 2nd, 2019. Five hundred years ago today, Leonardo da Vinci died. Some say he was the last man to master frontier knowledge in all scientific fields.
The AI arms race appears destined to follow an unavoidable concentration pattern. Almost as the natural consequence of the fact that corporations leading the AI development embed the winner-takes-all economy we live in. It is hard to avoid top companies hoarding AI scientists. Not to mention the prize at the end of the rainbow – singularity – that would potentially be the ultimate step when the first in closes the gates for others. One AI to rule them all, so goes the omen.
For the last few hundred years, scientific progress has been a key drive for productivity and economic value creation. And this very science – built by Newton, Bayes, Godel, Turing, Bohr, to name a few – is the giant shoulder AI stands on. If we look back, despite the glorious contribution those geniuses individually made, the nature of the scientific progress in unquestionable a collaborative one.
Not that all science is to be replaced by AI, of course. For the time being, at least. Some science is, though. Additionally, a big part of scientific production now relies back on AI. It is hard to imagine theoretical physics, chemistry, or genetics nowadays without AI. This feedback loop would place scientific production into an analogous winner takes all path. Competition, not collaboration, would be the way to go.
Now: is it so? Let me dare to propose not:
I do not want to get to the endless debate if machines are able to get human general intelligence capabilities, what intelligence is to begin with, if human brain is a biological machine, nor if the singularity would come when quantum time entanglement is able to reenact the awakening of the Budha. All tempting chats, but to my limited point: What are the contingencies involved?
I will restrict this argument to predictive power. A very narrow slice of knowledge indeed but a significant share of what AI investment seeks. Ruling out psychics and fortunetellers, prediction power requires material resources. Among others such as skill, any prediction requires a mix of time, energy, and experience. And these are all limited resources. Even if time itself is not, predictive power application of time is.
I argue that such limited property is in the heart of collaboration of modern science. No single scientist has the time to reinvent all the underneath science needed, nor the resources to get the latest in all fronts. Even if that was possible, objective revision would be unfeasible. As books, universities, method and the enlightenment curiosity played their part, collaboration sprung science productivity.
Are those resources in any matter more available today? Indeed; digital data brings along tons of experience, energy is increasingly abundant, and computers make calculations in fraction of a second. Are they unlimited, though? No, they aren’t. That’s our queue.
How can we share time? In the form of computing power, there are initiatives that optimize excess computing power. Harnessing the unused computing resources globally could power a shared AI enterprise. Energy sharing has an analogous case.
What about data? Some argue we should own our data, some say data belongs to all. Anyway, data collection is a private business. Not focusing on the capitalist ‘private’ here, the fact remains that collecting and organizing data requires a considerable effort.
I hope finding a way out of this data bottleneck could free the genius of AI collaboration. Cryptography perhaps can help us there. The idea is that data collectors would be able to post encrypted data – AI engines would be able to train and produce learning material on that. Data collectors would not see expropriated their efforts. On the contrary, those same data collectors would benefit from AI engines learning on other data collections. Ai engines would similarly develop their own parameters in-house, but make them available cryptographically for the community to lever on. Competition of most efficient AI, most rich data collections would still push for new applications.
This system (yes a very quick sketch but as said time is limited, I promise I’ll elaborate soon) could in theory be more dynamic than a single centralized organization working solo. If modern science analogy holds, productivity dividends would make collaboration unstoppable, not the reverse.
Thomaz Teixeira
Any cryptographers out there willing to collaborate? Lets discuss how – send me an email to thomazbt@yahoo.com