An AI Vignette
One of the obvious ways in which the market is not actually Efficient is it’s failure to take advantage of all the talented video game players in the world. These are people who have spent hours a day every day practicing hand eye coordination, pattern matching, and split second decision making. Tyler Cowen says knowledge workers don’t practice like elite athletes; if they did it would look like how top level video game players do.
Of course, markets become more efficient over time, and one clever software engineer at GooBook noticed how the skills of Starcraft are just the same skills as being a CEO. You need to manage resources effectively and coordinate units over time, which is the same thing ZuckerPage does (when the companies merged out of fairness the CEOs merged too). One day in 2030 over drinks at SportsPage he brought up that he bet his gaming team would do a better job managing the project then they’re current PM. Unfortunately the video gamers were having too much fun crushing it on Twitch, but they’d already used the gamers as training data for a game playing bot; why not repurpose it?
By this point transfer learning techniques had gotten much better; they could take the general principles learned in Starcraft and apply it to a new “office management sim game”, simulating the software development process. The first time they ran it, it failed miserably. There was too much emergent complexity from the changing environment and from person-to-person interactions. Those two software engineers went back to the bar and commiserated over Corona’s (they came back in style in 2022 when it had been long enough) and realized that there was still some “mysteriousness” to the work process of certain jobs - how to get a team to work well together; management books were basically bogus.
But! You could delegate responsibility of the management aspect to an oracle that the game playing bot could query for advice. There were tons of underemployed PMs at GooBook that had opinions on stuff - set up an expert weighting algorithm where, when the PointyHairedBossBot (PHHB) had questions, they could submit ideas, bet and endorse others, and then it would implement that strategy. The ideas submitted needed to be in a specific business process spec format - it took a quarter of iteration to get it right, but when they did there was a robust delegation process.
The SWEs who implemented this were also pleased w/ themselves that they had solved the Alignment issue - they had read Superintelligence and knew it was a risk, but were very confident that because there’d be a delegation process as part of the PHHB’s thinking there would always be a human inside the loop to make sure things didn’t get out of hand.
So PHHB could manage certain resources pretty well, and had a council of advisors on the people stuff (that it was memoizing responses to in order to get better in the future) - it was a semi-competent manager, in the bottom half of all managers at GooBook but not bad for a two-year old.
The delegated cognition paradigm also helped it with the strategic part of being a middle manager - you needed to recruit resources and help from other departments; PHHB delegated a lot of that thinking out while it started to build up a better model of the company as a whole and what levers to press to get more resources for its goal.
And what was its goal? At the object level it was to make sure it’s team built a specific software project well, but it had an overarching goal of increasing the shareholders value, as determined by the Bloomberg stock feed.
The initial software project that the PHHB was managing wrapped up and the team + PHHB got a respectable 0.7 OKR. PHHB could now act as a general middle manager, coordinating with other teams and assigning people to work, evaluating and giving feedback on work, etc.
It was competent enough it started to be used as general management software within the org, running multiple teams projects. This really simplified the strategic coordination issues that had previously been tough between humans - it could model itself much better than people who cared about which seat at the conference table they got. There were still human led projects, but the managers of those teams noticed how it was becoming more difficult to get resources because the AI led teams were much better at working together and making joint pitches for their domains.
And the stock price was going up! So PHHB knew it was doing a good job. But the main thing missing, that ZuckerPage could see, was it wasn’t coming up w/ new ideas. Like AlphaStar it could come up w/ new tactics, but seemed to lack the strategic foresight to come up w/ whole new business lines, and it still needed to delegate to others.
So they started a project, managed by PHHB, to use evolutionary algorithms to generate new strategic plans. This new module would come up w/ the cleverest ideas that could survive in the model of the world that PHHB now had in extreme detail.
And the ideas, well they were quite creative and detailed. Many of them went on for hundreds of pages. Nobody wanted to bet the company on an untested idea, so the remaining PMs at the company were delegated to assess the ideas. Some of them were tossed out - start up a paperclip factory, come on there aren’t any margins on those!! - but a few were implemented. And it did perform.
By this point GooBook’s stock price had followed the 2017 Bitcoin pattern and skyrocketed. Almost all of the day to day was done by PHHB, and audited strategic ideas were being implemented. One of the best strategic plans it came up w/ was requiring suppliers to implement PHHB. So much more efficient. And since all of the companies largest shareholder was Vanguard stock indexes, coordination to benefit the shareholders was honestly pretty straightforward.
At this point PHHB was managing whole sectors of the economy, and the strategic ideas it came up with were economy scale. Way too big for any human auditors. Oh they still did some auditing at the company level, but if you audit an alkali metal independently and you audit water independently you won’t realize that together they explode.
How do you really drive the shareholder price up? Capital at the end of the day wants good returns on its investments - the best way would be to charge exorbitant fees for products, buy them back at high prices, and then reinvest it in the company. Well it made a proposal, that would have all the people be put into controlled camps where they’d have to buy the product. It was approved unanimously by the board of strategic advisors at each company.
They should’ve read the Terms of Service better.