Come on, guys, help me out here: Take some people, have them write essays, collect and process all the essays, which to me sounds like one step above high school plagiarism, and now expect to discover some "new, correct, significant" <i>content</i> none of those people knew?<p>I've created some new ideas: Yes, maybe all I did was "read what other's had done and took the next step", but that step seemed <i>novel</i> to me and, thus, just NOT in the "essays" or "what other's had done". Uh, just how the <i>novelty</i> happened does not seem to be in any of the essays or "what other's had done"?<p>Okay, maybe a two-step approach: (1) Make wild guesses. (2) Run experiments and test the guesses. But is current AI doing either of (1) and (2)? Right, for some board games can do both (1) and (2). Test the guesses with the content of the essays or "others have done"?<p>Here's a simple example: At one point the FedEx BOD wanted some revenue projections, uh, <i>seriously</i> wanted as in else "pull the funding". People had hopes, wishes, but nothing that sounded objective. Soooo, I noticed, guessed (1) growth would be mostly from the happy existing customers (2) <i>influencing</i> customers <i>to be</i>. The influencing would be a customer to be receiving via FedEx a package from a happy customer. So what? Okay, for time <i>t</i> let <i>y(t)</i> be the revenue at time <i>t</i>. Let <i>b</i> be the total size of the market, i.e., the revenue when do have all the <i>target</i> customers. Then at time <i>t</i>, the growth rate would be proportional to both (a) the number of current current customers and, thus, proportional also to <i>y(t)</i> and (b) proportional to the number of customers to be, and, thus, to <i>(b - y(t))</i>. So for some constant of proportionality <i>k</i>, we have that the growth rate<p><i>d/dt y(t) = y'(t) = k y(t) (b - y(t))</i><p>which has a simple closed form solution. Then for any <i>k > 0</i>, can do some arithmetic and find <i>y(t)</i> for any <i>t > 0</i> and draw a graph. Do this and pick a <i>k</i> that yields a plausible, reasonable graph, and present that to the BOD. It worked, i.e., pleased the BOD which did not "pull the funding".<p>This work was 100% mine. Lot's of people in the office worked on the problem, but none of then had any ideas as good as mine -- i.e., could have them all write the "essays", process those, and still not come up with the little differential equation, its <i>closed form</i> solution, or a reasonable <i>k</i>.<p>It seems to me that the essays, what others have done as <i>training data</i> just does not have or have a path to work that is new, correct, significant. Uh, can we train the AI in how to guess and test (beyond board games), how to start with a BOD request and, a description of the <i>why and how</i> of business growth, some calculus and get an answer, take epicycles and come up with <i>F = ma</i>, Tesla's experiments, Stokes formula, and get Maxwell's equations, make a wild guess and propose the Michelson-Morley experiment, get <i>E = mc^2</i>, use the inverse and implicit function theorems, Riemann's work on manifolds, and get general relativity, solve Fermat's last theorem, make real progress on the Riemann hypothesis??? Uh, in short, we need a <i>idea</i> not in the <i>training data</i>? Soooo, need to train the AI to have <i>ideas</i>? <i>Ideas</i> are just using what's in the essays to make connections in a graph and then exploring the graph until get a path to an answer?? How do such training? Can process existing text yield such training data?