The AI industry is hitting data limits for training massive, general-purpose models. The next wave of progress will likely come from creating highly specialized models for specific domains, similar to DeepMind's AlphaFold, which can achieve superhuman performance on narrow tasks.

Relying on a single frontier model is risky and inefficient. The next phase of AI will involve intelligently routing queries to the most appropriate model—be it cheaper, faster, or local. This will redistribute value from a few dominant labs to a long tail of specialized models, maturing the ecosystem.

The most efficient AI architecture separates reasoning from knowledge. Models will shrink, focusing parameters on intelligent processing, like an "Einstein who never saw the world." They will rely on cheap, efficient tools like retrieval for information, solving compute shortages.

The most valuable data for creating intelligence is private and locked within enterprises. This proprietary data will be used to create millions of specialized AI models, each outperforming general-purpose models for specific tasks, creating a diverse AI ecosystem.

Base10's Head of AI training argues against the notion of a single, all-powerful AI. Instead, he bets on a future with hundreds of millions of models, each continually learning and adapting for a specific person or company. This paradigm shift focuses on organic, specialized intelligence over a monolithic, one-size-fits-all approach from frontier labs.

The debate over whether "true" AGI will be a monolithic model or use external scaffolding is misguided. Our only existing proof of general intelligence—the human brain—is a complex, scaffolded system with specialized components. This suggests scaffolding is not a crutch for AI, but a natural feature of advanced intelligence.

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.

The Fetus GPT experiment reveals that while its model struggles with just 15MB of text, a human child learns language and complex concepts from a similarly small dataset. This highlights the incredible data and energy efficiency of the human brain compared to large language models.

New AI models are moving away from brute-force computation. By selectively focusing on relevant data, much like the human brain indexes memories, they can achieve massive performance gains and cost reductions, overcoming a major bottleneck in current architectures.