Predictive Analytics

"Bayes theorem predicts that Bayesians will win." Nate Silver

 Lingering emotions about free will and volition disappear with any growing understanding of predictive analytics. As more and more of the behaviors (and proxies or precursors for behaviors) appear as continuously collected personal data, the ability to quantify the self (the I in "I believe") gets off-loaded to cognitive means in the environment. As more and more data is indirect (moving to direct) metrics for modes of action:
1) It becomes harder and harder to surprise 'the predictor'
2) it becomes more and more costly to miss collectiing the best data.
3) it becomes easier and easier to indirectly and directly demand/suggest next steps.

The question: is the range of response of a Tachikoma predictable within error bars and if not, why not? Could the input stream (percept) run on a lab bench AI (an AI-head in a vat) produce like response? The simulation is highly dependent on initial conditions. But in a robust multi-player, massively parallel simulation what surprises are likely? That is, would the final act of the Tachikoma horde bringing down the satellite which has the download/synch occur in some, none or most forward looking simulations.
Also, in Dan Dennett's modified, limited free will -- does the Tachikoma horde have this kind of limited free will? That is, can I share my rendition of cause/effect with others and multiplex the results (recommendations) to synthesize a collective will to (strictly limited, highly constrained) power?

We cannot predict all outcomes but can we predict the behaviors that result from group action and hence the narrow range of plausible outcomes?