The title is from the perspective of optimization theory. I'm trying to generalize the idea in this note. A [decision variable](https://www.baeldung.com/cs/optimization-terminology) is the variable(s) in the optimization equation that are being optimized, whereas the constraint variable(s) are just the constraints. The point of this note is that if we, as humans, treat a constraint variable as if it were a decision variable then we extrapolate wildly incorrect predictions. [The AI boom: lessons from history - The Economis](https://www.economist.com/finance-and-economics/2023/02/02/the-ai-boom-lessons-from-history?utm_content=ed-picks-article-link-6&etear=nl_weekly_6&utm_campaign=a.the-economist-this-week&utm_medium=email.internal-newsletter.np&utm_source=salesforce-marketing-cloud&utm_term=2/2/2023&utm_id=1472299) > AI might well augment the productivity of workers of all different skill levels, even writers. Yet what that means for an occupation as a whole depends on whether improved productivity and lower costs lead to a big jump in demand or only a minor one. When the assembly line—a process innovation with gpt-like characteristics—allowed Henry Ford to cut the cost of making cars, demand surged and workers benefited. If AI boosts productivity and lowers costs in medicine, for example, that might lead to much higher demand for medical services and professionals. This paragraph in particular stands out to me because it reminds me of the counter-intuitive phenomenon of cities investing in better road infrastructure to reduce traffic. (I need a source for this but) it just so happens that when a city invests in infrastructure to reduce traffic, rather than reducing traffic, it pretty much stays the same over the long run. In the short term the traffic gets better, but over time people move further away from the city, and this increases traffic. So the traffic in a city is not a function of the road infrastructure, it's a function of our (humanity's) tolerance for traffic. We, as a whole, have some tolerance for traffic, and we will maximize other aspects of our life (e.g. living in a more favorable area) within the constraints of our tolerance for things like traffic. So the optimization equation of "lets minimize traffic" turns out to be misplaced because humanity doesn't operate with traffic as an optimization; instead humanity operators with traffic as a constraint. Coming back to the quote above, the follow sentence stands out: "If AI boosts productivity and lowers costs in medicine, for example, that might lead to much higher demand for medical services and professionals.". So what some people are predicting is that AI will replace jobs, which is certainly true. But in the long term, the overall "human happiness" optimization will shift dramatically in a different direction because a constraint has been dramatically shifted. In the long term we won't lose jobs, they will just be different. In my head I'm picking this problem as an optimization space with some fixed form based on the set of constraints in the space. Then when a constraint changes, the form of the space shifts with it. The followup question to this idea is: Given that NLMs like GPT will massively alleviate a constraint, how does humanity's overall optimization equation change? What is the new limiting constraint? This idea of decision variables vs. constraint variables, and the malleability of the optimization space as a function of its constraints, may become a [[published/The Foundations|foundation]].