Early AI training involved simple preference tasks. Now, training frontier models requires PhDs and top professionals to perform complex, hours-long tasks like building entire websites or explaining nuanced cancer topics. The demand is for deep, specialized expertise, not just generalist labor.

In high-stakes fields like pharma, AI's ability to generate more ideas (e.g., drug targets) is less valuable than its ability to aid in decision-making. Physical constraints on experimentation mean you can't test everything. The real need is for tools that help humans evaluate, prioritize, and gain conviction on a few key bets.

Software engineering is a prime target for AI because code provides instant feedback (it works or it doesn't). In contrast, fields like medicine have slow, expensive feedback loops (e.g., clinical trials), which throttles the pace of AI-driven iteration and adoption. This heuristic predicts where AI will make the fastest inroads.

Broad improvements in AI's general reasoning are plateauing due to data saturation. The next major phase is vertical specialization. We will see an "explosion" of different models becoming superhuman in highly specific domains like chemistry or physics, rather than one model getting slightly better at everything.

In a group of 100 experts training an AI, the top 10% will often drive the majority of the model's improvement. This creates a power law dynamic where the ability to source and identify this elite talent becomes a key competitive moat for AI labs and data providers.

AI's ability to perform software engineering tasks that would take a human hours is doubling every 4-6 months. This rapid, exponential progress suggests a near-term future where AI can automate its own research and development. This self-improvement loop is the critical inflection point that could trigger a massive, unpredictable leap in AI capabilities.

In a sign of recursive capability improvement, OpenAI found that its model-based grader for the HealthBench evaluation benchmark was more accurate and consistent than the average human physician performing the same grading task. This demonstrates that models can not only perform a task but also evaluate that performance at a superhuman level, a key component of scalable oversight.

The transition from the AI "middle game" to the "endgame" is marked by a critical shift: when top human research talent ceases to be a differentiating factor. At this point, AI progress becomes a function of an organization's existing AI capabilities and its access to compute, because the AIs themselves become the primary researchers.

The true exponential acceleration towards AGI is currently limited by a human bottleneck: our speed at prompting AI and, more importantly, our capacity to manually validate its work. The hockey stick growth will only begin when AI can reliably validate its own output, closing the productivity loop.