Can we codify the innovation process, and drive humans from their last vestige of unchallenged power?
Innovation brilliance is very impressive for us humans, irrational as we are. Alas, innovation productivity is the concern for rational thinkers. To achieve high innovation productivity one must be (i) creative — yes, and also (ii) managing well one’s resources — in equal parts. As we will make the case shortly. Now, let’s argue for the time being that humans are better than machines in being ‘creative’ — they are clearly inferior to machines in resource allocation. Humans get caught up in their own irrational excitement, and are readily scared off by negative emotional experiences. AI is immunized in these aspects and pays attention to the most productive route. So it seems reasonable to conclude that AI-assisted R&D will gain superiority over human-only R&D.
Recent work in codifying the heart of the innovation process itself suggests that machines may exceed humans in the creative element too. After years of engineering practice, I have returned to academia, and to my admired professor Ephraim Kehat, with an ambitious plan to codify the innovation process. I focused in my PhD dissertation on the hard task of estimating the cost and time needed to achieve an R&D objective. I later concluded that R&D is best practiced by conducting it with a continuous measurable goal — increasing the credibility of that very estimate — cost and time to conclude the innovative mission. It sounds counter intuitive and weird to begin with — but come to think about it, the cause for poor credibility of these estimates is the degree of relevant ignorance for the project. In many years of cost engineering practice I could estimate very large and very expensive projects with a remarkable accuracy and credibility of one or two percentage points. The reason was lack of relevant unknowns. Repeat projects, even expensive and long lasting ones admit solid statistical similarity that feeds those estimates. Innovation projects, by definition, are projects that turn chunks of unknown into known. The ‘chunkiness’ of that unknown is the cause of the poor credibility of the estimates of cost-to-complete and time-to-finish. As the projects rolls ahead, areas of unknown become known — and the measure of this transformation of ignorance to knowledge is well captured by the mathematical expression of the credibility of the estimates of cost to complete and time to finish.
The implicatopns of this principle are far reaching, but here I wish to focus on how this principle suggests the prospect of effective AI-assisted R&D. A human researcher will naturally list a host of possible scenarios to advance her research. The choice among them is mostly ‘gut feeling’, less rational. One could then feed all that is known over the list of R&D attack scenarios, feed also the measure of available resources — and let the AI chart an optimal path. At least this is one course that we work on (D&G Sciences — Innovation Productivity Corporation). We stress AI, rather than an ordinary computer program because we use BiPSA — a platform that handles a myriad of heuristics.
As to autonomous AI, our expectation here are based on the codified innovation process known as InnovationSP: Innovation Solution Protocol (See my book “The Innovation Turing machine”). In brief, InnovationSP calls for efforts to solve an innovation challenge (IC) by defining it over a variety of problem-solution landscape. This ‘landscape’ identifies the terms, the assumptions, the limitations, the possibilities, and is open for creative massaging. If this direct effort fails then look for a ‘leveraged innovation challenge (LIC) such that the effort to solve it E(LIC), plus the effort to solve the original IC, given that LIC was resolved E(IC | LIC=0), is smaller than the effort to resolve IC straight on: E(IC | !LIC) > E(IC | LIC=0). This generic proposition may be further and more precisely defined and serve as a foundation for autonomous AI — that is the direction we work on, and it looks real good!