In Silicon Valley, many investors, founders, and tech giants often frame AI as an all-powerful force, something close to omniscience.
They call this artificial general intelligence (AGI). This is a hypothetical system that can match or exceed human reasoning across virtually any task, and even surpass human cognition entirely, referred to as superintelligence.
AGI is what OpenAI and its competitors are racing to be the first to achieve.
However, the industry is not monolithic, and neither are its builders.
Peter Steinberger, the developer of OpenClaw, the personal AI assistant behind the agent-only social network Moltbook, told the Y Combinator podcast on Saturday that he believes the best AI is specialized, not generalized.
“What can one person actually accomplish? Do you think one person can build an iPhone, or do you think one person can go to space?” he said. “We are specializing as a group and even more specializing as a larger society.”
The same applies to AI, he said. Current AI systems are called “general purpose.” But in reality, they are already specialized in certain tasks. Examples include startups building models to solve Erdos math problems or identify genetic mutations.
The hype surrounding AI is mainly We’ll focus on the types of models that are built, from large-scale language models to computer vision systems and even world models. — Some startups and tech giants are also experimenting with more focused, subject-specific forms of intelligence.
For example, Axiom, a startup founded by former meta-researcher Karina Hong and backed by $64 million in seed funding, is building specialized intelligence to tackle advanced mathematics.
Google DeepMind has developed AlphaGenome, a system that predicts how mutations and mutations in human DNA affect a wide range of gene-regulated biological processes.
Companies are also developing smaller models that make it easier to build AI systems tailored to specific subjects or domains. Aidan Gomez, CEO of Cohere, which builds AI technology for enterprises, told Business Insider in 2024 that there is increasing pressure to build “smaller, more efficient models” and make them smarter with the right data and algorithms, rather than just scaling.
Steinberger is not alone in opposing the idea of an omnipotent general intelligence.
Timnit Gebru, a computer scientist who founded and heads the Distributed AI Institute, said in a video published by Nature magazine in November that AGI is a “fiction.”
She said the backbone of engineering is building well-scoped and testable systems. She said the pursuit of “an imaginary and undefined ‘god in the machine'” is pushing the industry toward even greater labor exploitation and environmental destruction.
Steinberger was previously best known for starting the PDF processing company PSPDFKit, but after leaving the company, he started developing with AI. He said the technology has come a long way since the goal was simply to build “something you could type in so that the computer could do something.”
