Building on Jell-O

The current generation of applications built on general-purpose AI models may look solid at first glance but lean on them in the real world and they wobble. Models trained on the chaotic mass of the internet and optimized to sound natural aren’t built for reliability. They’re foundations that shift under pressure, and no enterprise can depend on a system that can’t hold its shape.

Recent AI progress followed a simple law: more data, more compute, bigger models, better results. But scaling laws have leveled off. The best of these systems get the facts wrong, and can’t reliably and repeatedly perform key agentic tasks. We’ve reached the point where new approaches are needed. The next frontier in AI is about reliability through specialization.

The public mood has shifted from wonder to wariness. We’ve gone from look what it can do to can we trust it? Hallucinations, bias, and unreliability aren’t edge cases; they’re fundamental aspects of the architecture. This is the moment to redefine progress in AI around reliability if we want the benefits to diffuse into the economy.

When you pretrain on the internet, you don’t just get knowledge and generalization; you get noise. The web is full of falsehoods, deception, and bias. Large language models compress that into something fluent but unreliable and untrustworthy. 

As Andrej Karpathy noted in his recent conversation with Dwarkesh Patel, today’s models spend much of their capacity on “memory work instead of cognitive work,” because “the internet is so terrible.” Scaling up on noisy data teaches models to compress the slop, not reason over signal. The next leap, as he suggested, won’t come from more of the internet—but from curated data and new algorithmic approaches that reward reliability.

Real-world domains demonstrate that specialization is key to outcomes. Medicine, finance, and science each have unique and complex requirements. General-purpose models can mimic tone but miss context. Special-purpose AIs reason natively within their environments, tuned to their data and constraints.

The next advance in AI will come from systems that specialize, and leverage the opportunity to  constrain the model to the domain to achieve greater reliability.  

So while general-purpose AI is like Jell-O, unable to provide a solid foundation for critical applications - specialized models will provide the solid, reliable underpinnings needed to deliver durable value. 

That’s what we’re building toward at Scaled Cognition—AI with structure, purpose, and reliability.