The 'Easy' First 80%
When an AI model reaches 80% proficiency, it means it has mastered the predictable parts of a problem. This is the domain of big data and clear patterns. For an aspiring 'AI researcher,' this could involve synthesising vast amounts of scientific literature,
writing code for experiments, and analysing structured datasets to find correlations. These are tasks where the rules are well-defined and there is a massive volume of training data available. The model becomes exceptionally good at handling the common scenarios, or the 'head' of the data distribution. It can solve problems that have been solved before, just faster and at a greater scale than any human. This initial progress often feels rapid and revolutionary, leading to impressive benchmarks and a sense of accelerating momentum. However, this impressive performance is built on a foundation of repetition and pattern matching, not genuine understanding.
The Brutal Final 20%
The journey from 80% to 100% is where things get exponentially harder. This final stretch isn't about doing the same things better; it's about acquiring entirely new capabilities. This is the realm of the 'long tail'—the near-infinite number of rare, unexpected, and novel situations that don't appear in the training data. A human researcher excels here. They handle unforeseen experimental errors, formulate a completely new hypothesis, display genuine creativity, and exercise common sense. Current AI systems struggle with these tasks because they lack true reasoning and the ability to generalise knowledge to entirely new contexts. Getting an AI to master this long tail of exceptions requires a shift from just recognising patterns to understanding causality and context, a challenge that some researchers believe the current deep learning paradigm may not be equipped to solve on its own.
Moravec's Paradox at Work
This difficulty is explained by a concept known as Moravec's paradox. First articulated in the 1980s, it observes that in AI, the easy things are hard, and the hard things are easy. For AI, complex abstract tasks like playing chess or performing mathematical calculations are relatively 'easy' because they operate on formal rules. In contrast, tasks a one-year-old child finds simple—like walking, recognising a face in a cluttered room, or picking up a new object—are incredibly difficult for machines. These sensory and motor skills took millions of years of evolution to perfect in humans and are deeply intuitive, operating below the level of conscious thought. An AI researcher's job isn't just high-level logic; it's full of this kind of intuitive, contextual 'easy' stuff, like sensing a colleague's doubt or having a hunch that a particular line of inquiry is a dead end. This is the part of intelligence that is hardest to replicate.
The Sobering Reality of Diminishing Returns
Achieving the final 20% is also subject to diminishing returns. Pushing a model from 95% to 96% accuracy on a complex task might require as much data and computational power as it took to get from 0% to 95%. The costs—both in terms of energy consumption and financial investment—skyrocket for incremental gains. This scalability dilemma is a major roadblock. We can't simply assume that more data and bigger models will automatically bridge the gap to true, human-level intelligence. It suggests that the final leap to AGI won't just be a matter of more brute force, but will require fundamental breakthroughs in algorithms, cognitive architectures, and our very understanding of intelligence.
