The Baby Fallacy
I currently have the luxury of being on paternity leave, which allows me to watch my daughter gradually acquire new skills and abilities. People often mock parents for their pride and astonishment at babies doing seemingly menial tasks (“Oh my God, she is holding up her head for like 10 seconds!”), but I believe this is misplaced. The incredible part of watching babies, especially your own, is seeing them on their gradual path from helpless larvae to the incredibly unique individuals they become. It’s essential to remember that every Olympic sprinter or talented athlete, at one point, couldn’t control their bowels, nor had much stability.
The fallacy that system limitations are fully predictive of final system improvement is not just a human development thing. It would be very easy to look at the Wright brothers’ first 12-second flight and either not believe it or write it off as a parlor trick that would never scale (fun fact: the world did! The brothers spent years after their first flight trying to convince people). Same goes for the first cars when people were comparing them to horses. If you were to base your world view on nascent systems without letting yourself believe in the ingenuity of the human colossus, you would shut down every single cool thing as nothing more than “neat.”
The Developmental Imperative
We are terrible at understanding iterative improvement. Humans think linearly, but the systems we build improve exponentially. When you see 1% daily improvement, your brain imagines a gentle slope. In reality, that 1% compounds into a 37x improvement over a year. We don’t intuit this. We never have and likely never will. To our brain iterative compounding is like what my old football coaches would call ‘Disco Buffalo’. Its a buffalo dancing disco, it just makes no sense.
Mastery emerges through stages, not leaps, and the early stages always look underwhelming. Think about how babies develop. Motor skills go from crawling to walking to running to specialized athletics. Language goes from babbling to words to sentences to poetry. Cognition moves from basic pattern recognition to reasoning to creativity to wisdom. At every transition, the previous stage looks laughably primitive compared to what follows. Yet each stage was a necessary foundation for the next.
The same principle applies to any system that learns iteratively. The marginal gains don’t just add up, they compound. Each new capability unlocks the next one. You don’t get poetry without babbling first, and you don’t get babbling without the neural wiring that develops in those early, seemingly useless months of staring at ceiling fans. If you judge any developmental system by its current stage, you will be wrong about its potential nearly every time.
The 2013 Prophecy: Word Vectors and the Skeptics
In 2013, Tomas Mikolov and his team at Google released Word2vec, and something strange happened. You could take the vector for “king,” subtract “man,” add “woman,” and get “queen.” Arithmetic on words. The math worked, and it probably shouldn’t have.
The critics saw a parlor trick. Statistical correlations dressed up as understanding. “Glorified autocomplete will never truly understand,” they said. And from where they stood, they weren’t being unreasonable. Word2vec couldn’t hold a conversation and being able to do simple word math was mostly irrelevant except to search engines. Early simple versions even discarded the trained network as if it were a vestigial organ. All it could do was place words near other words in a way that happened to capture something about meaning.
What the critics missed was that they were watching a baby hold its head up for the first time. Word2vec wasn’t the destination, it was the foundational architecture of understanding. The insight that language could be represented as geometry, that meaning had a spatial structure, was the crawling that would eventually lead to running. The people who dismissed it as a cute trick were confusing current implementation for fundamental limits. They looked at a 12-second flight and concluded humans would never cross the Atlantic.
The Pattern of Premature Dismissal
There is a specific type of expert who is consistently wrong about the future, and it’s usually the person who knows the most about the present. Deep domain knowledge creates blind spots to paradigm shifts. When you understand exactly why something is hard, it becomes nearly impossible to imagine it becoming easy.
Before Deep Blue beat Kasparov in 1997, chess grandmasters were confident that computers could never match human intuition at the board. Before AlphaGo beat Lee Sedol in 2016, Go experts insisted the game was too complex for brute force, that it required something fundamentally human. Before DALL-E and Midjourney, artists and art critics maintained that creativity was uniquely human, that machines could imitate but never create. Every single one of these confident predictions shared the same error: confusing current technical limitations with theoretical impossibilities. The gap between “we can’t do this yet” and “this can never be done” is enormous, and experts routinely collapse it.
The common thread in every premature dismissal is the assumption that the rate of improvement will mirror the pace the critic is comfortable with. It never does.
The Acceleration Paradox
AI development doesn’t follow traditional learning curves because it benefits from dynamics that human learners don’t. Data flywheel effects mean that the more an AI system is used, the more data it generates, which makes it better, which drives more use. Computational scaling laws show that you can predictably improve model performance by throwing more compute at the problem, and the amount of available compute keeps growing. Cross-domain knowledge transfer means that what a model learns about physics helps it write code, and what it learns about code helps it reason about logic.
Each breakthrough enables the next one. This is the compound learning effect, and it’s why the trajectory from GPT-1 to GPT-4 looks less like a straight line and more like a hockey stick. GPT-1 could barely string coherent paragraphs together. GPT-2 generated passable text (unicorns in peru!) that occasionally fooled people. GPT-3 could write essays, answer questions, and code at a rudimentary level. GPT-4 passed the bar exam. The time between each of these steps was roughly the same, but the capability gap between each step grew dramatically. If you projected GPT-1’s limitations forward in a straight line, you would never have predicted GPT-4. That’s the whole point.
The Business and Policy Implications
For investors, betting against technological evolution is historically disastrous. Every major technology wave, from railroads to the internet, has produced a chorus of sophisticated skeptics who could articulate exactly why the current thing would never work at scale. They were always right about the present and wrong about the future. The people who made money were the ones who could distinguish between fundamental constraints and engineering problems.
For policymakers, the risk is even more acute. If you regulate based on what AI can do today, you are writing rules for yesterday. By the time your regulation takes effect, the capabilities you were worried about have been surpassed by ones you didn’t anticipate. This doesn’t mean regulation is futile, it means regulation needs to be built around principles and trajectories, not snapshots.
For businesses, the strategic imperative is preparation over prediction. You don’t need to know exactly what AI will be capable of in five years. You need to build an organization that can absorb new capabilities as they emerge. The companies that will struggle are the ones that looked at current AI limitations and concluded they had time.
The New Heuristic: Potential-Based Thinking
The question worth asking is never “what can it do now?” It’s “what might it do with iteration?” Every limitation in a developing system is a feature request, not a death sentence. The developer’s mindset, one that looks at what’s broken and sees a problem to solve rather than a reason to quit, is the right lens for evaluating emerging technology.
Here are some practical ways to think about it. Trajectory analysis over snapshot assessment: don’t evaluate a system at a single point in time, plot its improvement curve and extrapolate. Identifying fundamental versus implementation constraints: ask whether a limitation is rooted in physics and mathematics or in engineering and resources. If it’s the latter, assume it will be solved. And finally, the baby test: would you dismiss a toddler’s potential based on today’s skills? If not, apply the same charity to systems that are learning.
The Uncomfortable Truth
We prefer the comfort of current limitations because they let us plan. If AI can’t do X today, then our jobs, our skills, our business models are safe. This is a soothing thought and almost certainly wrong. Gradual improvements compound into revolutionary change, and the transition from “kind of works” to “works better than humans” tends to happen faster than anyone expects.
The call here is for intellectual humility. Not breathless techno-optimism, not paranoid doomerism, just the honest acknowledgment that our current understanding of what’s possible is incomplete. The Wright brothers’ first flight was 12 seconds. Sixty-six years later, we landed on the moon. The gap between a baby holding up its head and that same person running a marathon is vast, but it is also inevitable.
In a world of accelerating change, the only safe bet is that the things we are most confident AI cannot do are precisely the things it will do next.
If this is interesting, drop me a line. Find me on X.