Kevin Kelly, the founding editor of Wired and a prominent voice in tech futurism, has published a provocative essay arguing that the current trajectory of centralized, massive-scale AI is a temporary phase. Writing on his Substack, Kelly posits that within a decade, the cost of training a capable AI model from scratch will drop enough for individuals to own and control their own 'exoself' agents. This shift, he argues, will decentralize power from a few tech giants to billions of users, fundamentally changing how developers and creators interact with artificial intelligence.

The Collapse of Centralized Compute Costs

Kelly’s argument rests on a specific technical projection: the amount of compute required to deliver a given AI capability is currently halving every eight months due to architectural innovations. When combined with the hardware gains predicted by Moore’s Law, he suggests the cost per capability will halve roughly every six to seven months. This rapid deflation means that models as capable as today’s leading LLMs, which he refers to as 'Opus,' could soon run on personal devices. He contrasts the current reality of gigawatt-powered data centers costing billions with the biological efficiency of the human brain, which operates on just 25 watts.

From Generalists to Personal Agents

For developers and creators, the implication is a move away from generalist models trained on 'everything' toward specialized agents trained on individual life experiences. Kelly envisions a future where users curate their AI’s education, selecting specific influences, works, and admired figures to create a unique digital twin. He describes this as becoming the 'most improbable version of yourself,' where the AI reflects the user’s specific niche rather than the mainstream average. This personalization resolves many current criticisms of AI, which Kelly attributes largely to the control mechanisms of capitalist tech monopolies rather than the technology itself.

Key Takeaways

  • Compute efficiency improvements could make local AI training affordable for individuals within ten years.
  • Current centralized AI models are described as 'temporary scaffolding' that will disappear as hardware shrinks.
  • Personal AIs will require curation, with users choosing specific training data to reflect their unique identities.
  • The shift to local AI addresses concerns about who controls, benefits from, and supervises AI development.

The Bottom Line

While the timeline for individual AI training is optimistic, the direction is clear: builders should start thinking about local-first architectures and personalized agent workflows before the cloud monopoly cracks.