What Would Change Your Mind?
In my free time, I enjoy reading about history, and one of the things that fascinates me is how systems of belief form around periods of significant change. In retrospect, history tends to look much cleaner than it actually was. We know which technologies mattered, which companies survived, which predictions were right and which ideas disappeared. The people living through those periods had none of those advantages. They were trying to understand something new while it was happening, usually with incomplete information and often with substantial incentives to believe one version of the future over another.
I think about this frequently when I hear predictions about technology. Markets manufacture narratives around genuinely uncertain technological change, and to some extent they have to. Investors need some understanding of what the future might look like before allocating capital. Founders need a reason to spend years building something that may not work. Companies need to decide which technologies deserve investment, and workers need to decide which skills will remain valuable. But we can’t see the future, so we must make decisions with the evidence available to us and act accordingly.
The complication is that there is often a great deal of money riding on those predictions being right. Founders need to raise capital, investors need returns, companies need growth, consultants need transformations to sell, and entire sectors of the economy can become attached to a particular understanding of where a trend is going. None of this requires anyone to be dishonest. But the natural tendency to filter a vision of the future through biases anchored in the past can cloud an already murky picture. Human beings are remarkably good at constructing narratives that reinforce existing biases, connecting new facts to old conclusions, and changing the explanation without changing the underlying belief.
I first noticed this clearly during the crypto boom. Gun.io did a fair amount of business in the industry. We worked with companies building blockchain products, saw substantial demand for Solidity developers, and watched enormous amounts of capital enter the space. There was significant technology being built and significant economic activity happening around it. Alongside that came much larger predictions about what the technology would eventually mean: cryptocurrencies would displace fiat currencies, traditional financial institutions would be disintermediated, and blockchain would reorganize significant portions of the economy.
What fascinated me was how easily changes in the evidence could be incorporated into the same underlying narrative. When cryptocurrency prices rose, the market was recognizing the future and adoption was accelerating. When prices fell, the assets were temporarily cheap and it was time to buy the dip. Either explanation could be reasonable on its own. What was more difficult was identifying the set of circumstances under which someone deeply committed to the thesis would actually conclude that the thesis itself needed to change.
That distinction matters because there is a difference between updating your explanation for why something happened and updating your belief about what is going to happen. When presented with new information, does it actually change the underlying prediction, or does it merely produce a new explanation that preserves it? If the explanation can change indefinitely while the conclusion remains fixed, the original prediction becomes increasingly difficult to either prove or disprove. At some point, a hypothesis that can accommodate any possible outcome stops functioning as an empirical claim and becomes a matter of belief.
I see some of the same dynamics today around artificial intelligence, including among people who were deeply skeptical of the crypto boom. Gun.io is doing a great deal of business in this boom too, across different parts of the industry. We work with companies building AI products, companies incorporating AI into existing products, and companies trying to use it internally to change how they operate. We see demand for engineers building the underlying systems, engineers applying the models, and teams trying to turn increasingly capable technology into something commercially useful. So my skepticism about some of the predictions is not skepticism about whether there is already real economic activity. I can see the activity directly.
If anything, being this close to the activity makes the uncertainty more interesting. There is clearly something consequential happening, but observing that something consequential is happening is not the same as knowing where it leads. “AI will replace software engineers, destroy BPOs and service firms, automate enormous categories of knowledge work, and fundamentally change the economics of software,” or so some of the more ambitious predictions go. Some of these may turn out to be correct, and some are already correct in narrower forms. The capabilities of the technology are extraordinary, and dismissing what has happened in AI over the last several years as marketing hype requires its own considerable act of motivated reasoning.
But the fact that some changes are happening does not make every prediction about the economic consequences true. When an AI system does something astonishing, it could be evidence of how quickly the technology is progressing. When it fails at something surprisingly simple, it is easy to say that we are still early. When a company produces substantial productivity gains from AI, we see evidence of its economic value. When companies struggle to produce those gains, we dismiss them as having not yet figured out how to implement it correctly. When models become more autonomous, it’s called progress; when they behave unpredictably, that can be interpreted as evidence that their capabilities are becoming more sophisticated than we anticipated, or that the tools are not yet as easily wielded as we hoped.
Any one of these explanations might be right. The problem appears when we accept them as right without ever stopping to reconsider the conclusion. At that point, we risk confusing our interpretation of the evidence with evidence that our explanation is right.
The prediction that AI will destroy IT service businesses is a useful example because the reality already appears more complicated. AI can already automate some work that previously required people, and there is little reason to believe that process will stop. At the same time, much of the work required to make AI useful inside actual organizations currently looks remarkably like services. Models have to be integrated into existing systems, organizational data has to be understood, workflows have to be redesigned, outputs have to be evaluated, and security, compliance and governance questions have to be resolved. The model may perform an increasing percentage of the underlying work, but somebody still has to understand the problem and determine how the technology should be applied to it.
Perhaps AI eventually eliminates much of that work as well. Perhaps it instead changes the economics of services so dramatically that we consume far more of them. Perhaps entirely new categories of work emerge around capabilities that are only now becoming possible. All of these outcomes have historical precedent because technology does not have a single effect on labor. Sometimes it replaces people, sometimes it makes them more productive, and sometimes reducing the cost of something increases demand enough to expand the market around it. It frequently does all three in different places at the same time.
This is what makes technological forecasting so difficult. We observe something real and then extrapolate from it while implicitly holding the rest of the world constant. Software becomes easier to produce, so perhaps fewer software engineers will be required. But lower barriers to creating software may also mean that vastly more software gets produced. Models become better at performing professional work, so perhaps professional-service firms need fewer people. But lower costs may make sophisticated services available to customers who previously could not afford them. The first-order effect can be relatively easy to see while the second- and third-order effects remain almost completely unknowable.
We have been making predictions about the eventual destination of computing for a long time. Ray Kurzweil famously predicted human-level machine intelligence around 2029 and the technological singularity around 2045. Those dates may turn out to be remarkably prescient. What interests me more is how difficult long-range technological predictions can be to evaluate while they are unfolding. If the predicted change happens, the thesis is vindicated. If something adjacent to it happens, perhaps the prediction was directionally correct. If the timing proves wrong, perhaps the underlying thesis survives with a different date. There is nothing wrong with updating a forecast as new information arrives; we should update forecasts. The question is whether we are updating the forecast or continually modifying the explanation so that the original belief never really has to confront contrary evidence.
Skeptics are capable of exactly the same behavior. If you begin with the belief that AI is mostly hype, every impressive demonstration can be dismissed as cherry-picked, every investment boom can be called a bubble, and every productivity improvement can be seen as temporary or overstated. When the technology clears one hurdle, another can take its place. A skeptic who cannot describe the evidence that would make them more optimistic has the same problem as an optimist who cannot describe the evidence that would make them more skeptical.
I don’t know exactly what AI will do to the economy, and I am increasingly suspicious of anyone who claims to know with great precision. I think it is an extraordinary technology. I think it will substantially change software, services and knowledge work. I also think some things intelligent people confidently believe about it today will look ridiculous in retrospect. Those positions are not contradictory. A technology can be enormously consequential while we are wrong about its timing, its economics, its beneficiaries, or even the primary ways in which it ultimately gets used.
There is another reason I am cautious about treating technological progress as something inevitable or autonomous: people are building it. People exactly like you and me. However sophisticated the systems become, underneath them are human decisions about what to optimize, what data to use, what risks to accept, what behaviors to reward, what problems deserve attention and what products eventually get shipped. The people making those decisions could be brilliant, but intelligence has never made human beings infallible. The history of technology contains extraordinary achievements alongside unintended consequences, misplaced confidence, perverse incentives, and outcomes that the people responsible for creating the technology did not anticipate.
Markets are not particularly good at showcasing this kind of uncertainty because uncertainty is difficult to sell. Narratives are much easier. This technology will change everything. This industry will disappear. This profession is finished. This company represents the future. Sometimes those statements will be right, and the people willing to believe them before everyone else will make extraordinary amounts of money. But the existence of enormous rewards for being right also creates enormous incentives to preserve the narrative once we have committed ourselves to it.
None of this is an argument for being less ambitious about technology. The most consequential technologies probably should inspire ambitious predictions about what becomes possible. But the more consequential the technology, the more important it is to separate what we can observe from the story we have constructed around it.
I increasingly think the most useful question is simply what would change my mind. Not what would cause me to modify the explanation, move the timeline or find another reason the expected outcome has not happened yet, but what would actually cause me to reconsider the underlying thesis.
Markets will continue producing narratives because we need them to make decisions under uncertainty. Some of those narratives will turn out to be remarkably prescient, and others will look obvious only in retrospect for having been wrong. The difficult part is remembering, while we are still living through the uncertainty, that we do not yet know which is which.
