microprediction / building_an_open_ai_network Goto Github PK

See linked-in discussion.

In favor of light grey design or against dark design:

• I actually wish you could have the graphic from A in dark mode, as I think they "illustrate" the concept better.
• A’s artwork is “better” as B’s feel like more related to electronic circuits designs.
  If the book is more academic then A color scheme is better but if it’s more pragmatic, then B’s colors would be my choice.
• The art in B is very generic any reminds me of every AI book on the shelves, even at the actual physical MIT press bookstore
• Brain-like circuit boards have been overused IMHO :)
• Representation of networks is better in option A imho
• If you change font color in A, I think A wins
• B is easier to read, A has better artwork
• Option A is a bit closer to reflect the topic
• B's image does feel like circuitry
• I find the B option to scifi and A to be the more down to earth one
• Light one as title says more about predictions, which is farthest and distant… shadowy light looks better expression… it should be surreal …Dark one looks more Technical
• it's not just light vs dark - it is "let be the light.. in greys" - a non-trivial grey tonalities where the artist captured the range beautifully vs trivial black and electric (?!) blue. The light greys are beautiful and inspiring if you like =). I would have made everything in greys only
• the main dominating factor here is the fine picture - there is more to it than a restricted in-angles chip-like boring of the dark.
  It's a clever thing to combine sine curves ( ~ Fourier transform) interconnected in graphs and a networked sphere juxtaposed with the levels - well done to the artis
  -Option B with the Design from Option A
  -Option A, since option B looks too generic
• A, B looks cookie cutter.

In favor of dark design, or arguments against light grey

• It immediately tells it is a tech related book and dark theme hints at dark future and AI risks.
  Also, I think more likely to be picked by tech audience from the shelves.
• I like B but if you end up keeping the book on the beach with all the free time you'll have after automating your trades, the cover will bleach in the sun
• Dark: It drew my eyes to it.
• Right one looks more interesting / eye catching to me - left looks too bland for me :)
• Option B. It seems more fluid and open
• Dark all day long sir
• B would be my choice. No technical reason. Just subjective esthetics.
• Option B is a better thumbnail
• Option B with vibrant colors
• Option B. The graphic looks like an aerial view of the grid from Tron Legacy which is a neat segue into AI futurism.
• B looks better.
• Option B reason is analogous to why dark themed IDEs >> light themed IDEs… Easier on the eyes.
• The higher image contrast of B is much better for grabbing attention
• Right side, darker cover, better contrast, more aesthetically pleasing
• Dark mode for aesthetic reasons, but I actually wish you could have the graphic from A in dark mode, as I think they "illustrate" the concept better.
• It's more intuitive with the prediction signals for AI. So, I will vote for B design.
• B is easier to read, A has better artwork.
• Definitely option B, more focused on the text than background. Although like for art of the option A better
• B, more contrast less tiring on eyes

There's a constant missing in the energy argument. I slipped into the old habit of subconsciously removing constants from energy expressions (entirely valid given the goal, but hardly adding to clarity of exposition).

Deficiencies in acknowledgments are a function of my organizational ability and haste. Here I endeavor to try ack's in real-time. I will surely fail.

Recently...

• Stephen Luterman's feedback on the strategy and collateral (his most recent contrib, in addition to years of mentoring and support).

• Andrew Stein's very detailed discussions with me on WASM tech. Watch his space.

• Richard Nieves Becker's viral post mentioning me in the same sentence as Yann Le Cunn was extremely welcome hyperbole.

• Fred Viole's longitudinal platform feedback continues.

• Mike Gault crypto & other feedback.

More ways to frame the micromanager's task, or subtasks, or applications

• As using wealth adjustments to ensure low regret (one of the earliest paper's)
• As online convex optimization (See book pdf by Elad Hazan)
• ML design of auctions as per paper by Gemp et al
• Unified view of decision-making Warren Powell

As noted in the book, it is pretty futile to try to present a snapshot of technology choices for micromanagers when things are changing so fast. So here's a place to put things you think are interesting.

• postgres in the browser
• much more to come

I had a fun discussion with Sam Savage explaining the collider at www.microprediction.org who has been in what he calls the distribution distribution business much longer than I have, and realized there is a very terse way to describe what I've built to those familiar with existing terminology.

More specifically I use what Sam would call the "DC" (direct current) representation of a univariate probability. Thus we might say that www.microprediction.org is a high-velocity exchange for clearing a market of stochastic information packets, or a "SIP collider". It is also, of course, a turnkey way to source these SIPs for fast-moving, instrumented processes.

An interesting thread to pull on is the relationship between the z1-streams and the "copula layer" in Sam's work.

Chess example.

Fischer-Spassky World Championship Reykjavik 1972. Position evaluations according to Stockfish after move 12 (+0.39) , after move 13 (+0.25), after move 14 (+40), after move 15 (-0.18), after move 16 (-0.04), after move 17 (0.00), after move 18 (-0.35), after move 19 (-0.24), after move 20 (-0.16), after move 21 (-0.39), after move 22 (-0.41), after move 23 (-0.33), after move 24 (-0.61), after move 25 (-0.66)

I mean seriously, 3 wood from 213 yards? That's ridiculous and given the choice of yardage, something of an insult to Tiger Woods' classic from the Canadian open. We're going with 5 iron in the next edition.

• Stochastic Information Packages sips are promoted by Sam Savage and co.

Issue statement: Numerai, Cincubator, Quantopian et cetera aren't necessarily doing so well, so why put faith in the crowd?

Response: Crowd generation of alpha is probably sufficient reason to think a prediction web will work very well. It is not necessary, however. A prediction web is premised on the efficacy of market-like mechanisms to predict things that are not already the subject of market-based prediction.

There are more subtle reasons why a firm like Intech uses a microprediction "collider" as part of its investment process, but understanding those is not critical to the argument.

Example in Chapter 8 of target approx calculation should have denominator 2000, not 1000