One way of looking at the world, and the vast interconnected economy and agents of capital, is that it is the product of accumulated variance. Humans are quirky, there is a monetary feedback function, particularly for large scale and late stage businesses, but it is slow, often indirect, and incredibly inefficient in who actually gets rewarded. This accumulation tends to hide the sheer unlikeliness of many outcomes in our world.
When thinking about the so-called “drop in remote worker”. What I keep coming back to, as a heavy user of the models, is variance. It is incredibly hard to really push the models to do something random. There are various terms for this Model collapse or Entropy collapse, sometimes called narrowness. Every 1 shot vibe coded website has a “smell”. Leaving the engineer to hammer away at these distributions, prodding them with the right image inspiration, the right nudge, the right scaffold of .md files.
There’s an obsession on the frontier, particularly among researchers with this replacement of humans. Beyond being an own goal that has calcified the general population against the labs and AI infra at large, is even the right approach? Why do I need a drop in worker when I have codex? Codex performs the tasks I ask of it, leaving product decisions to me - just the way I want it. It is an amplifier. Codex is faster per useful task (if I choose the right model) than it was a 1/3/6 months ago and is more capable of understanding what I want. Yet it is still quite rare that codex builds me something I didn’t ask for that I actually want. Perhaps skill issue and perhaps the domain I operate in (web dev) is simply yearning for better reinforcement learning (RL) environments.
Something distinctly human is missing, a layer of consistent noise that is effectively my taste profile and my accumulated variance. My sense of how I want to build applications, what I want things to look like, how I am approaching a new idea. Is there a “correct” way of creating a new map for something? Or a correct way to display a specific type of data in a chart or table - not really. I’d go even further to say most of the successful work I’ve done has been distinctly out of distribution.
A different example of this sort of accumulation is the S&P 500, 500 different companies, with different cultures, backstories, people, CEOs. Ultimately only ~4% of these 500 companies make up the net overperformance of the index vs T-bills [1]. New companies come in with novel technology, recycling out the stale and receding past. The RL pilled will point to the notion that given over a century of data we could go back and cultivate a RL environment keyed on those 4% of companies at/before the moments they overperformed. Such a task feels a bit like simulating the universe, simply too many variables and unknowns. Let’s suspend reality though and entertain such a dataset, a reward hacking agent built on this dataset is still collapsed. In this reality models would push users towards a much smaller scope of ideas, expanding the opportunity set for the proverbial ideas guy, who is decidedly uncollapsed.