Tech Thinker Kevin Kelly: 99% of AI Will Not Interact with Humans in Five Years
Kevin Kelly, founding editor of Wired, stated at the BOE Global Innovation Partner Conference that in the next five years, 99% of artificial intelligence will not even interact with humans. He also noted that most of these predictions could be wrong.
He pointed out that whether AGI can be achieved and whether many jobs will disappear is unknown to both experts and ordinary people. Companies seek certainty, but the next decade may not unfold in a way they prefer. Patients are not waiting for hospitals to approve AI doctors; they are already seeking AI directly. AI will handle medical knowledge and information, while humans will provide actual medical services, with most people likely to start using true AI doctors within two years.
Currently, he describes AI as a bookworm: smart and powerful through learning from books, but lacking real memory and true learning ability. The next phase will require intelligence that understands the world, rather than knowledge-based intelligence. He referred to humanoid robots as the most complex machines ever created by humans, more complex than jet planes or rockets; most robots have a reliability of about 99.99% without human intervention, and achieving true maturity may take another 10 years.
He described the human brain as a 25-watt supercomputer, while data centers operate at megawatt levels. Silicon Valley AI engineers are managing 6 to 10 coding agents, while ordinary people may only need to interact with one or two, with the rest operating in the background. Within three years, people will develop preferences and loyalty towards certain agents, as these agents can access emails, messages, locations, and social relationships, potentially understanding individuals better than they understand themselves.
The earliest users of smart glasses will not be office workers but blue-collar workers without fixed desks, and it may become the next major terminal after smartphones. There may be a small investment bubble in the AI field, but demand for AI tokens will exceed supply over the next decade. Chinese models seem to lag behind American models by only a few months in benchmark tests; tokens can be purchased based on price, speed, or intelligence, and they are not entirely the same.
In market mechanisms, buyers are the certainty-seeking enterprises and hospital systems, while sellers are patients, blue-collar workers, and Silicon Valley engineering teams who have bypassed approvals to use models directly. The events are driven by expectations of token shortages and gaps in humanoid machine reliability, with funding flowing towards computing power, world models, and the last engineering redundancy for the final 9, benefiting those who can embed intelligence in the background and sell differentiated tokens based on electricity pricing logic, while traditional buyers face pressure as they continue to procure based on certainty budgets but cannot get answers from experts regarding AGI and job issues.
Source: Public Information
ABAB AI Insight
Kevin Kelly wrote about distributed systems and collective intelligence in his 1994 book "Out of Control," which has since been repeatedly validated by the internet; he later extended this logic to cloud computing, the Internet of Things, and mirrored worlds in his works "The Inevitable," "What Technology Wants," "The Inevitable World in 5000 Days," and "2049." This time, he first negated his predictions, continuing his long-standing approach of viewing technology as an ecosystem rather than a singular singularity.
Capital is shifting from "how many books a large model has read" to "how much real-world experience it has" in world models and the reliability engineering of the final 9. Tokens are being sold based on speed, price, and intelligence level, indicating that computing power is no longer settled at a uniform electricity price; resources are being redirected to fill the gaps in humanoid machines, which are more complex than rockets, in terms of mechanics, safety, and energy efficiency, rather than just piling on another round of dialogue demonstrations.
Similar paths have occurred before: when electricity entered factories, 99% of the current was never seen by people; GPS first proved its value in trucks and construction sites before entering the pockets of white-collar workers; cloud services initially hid servers in data centers before allowing users to face just one button. The industry is still at the front end of an expansion phase—the demonstration videos are already cool, but the home scenario's final 9 has not yet passed, and smart glasses and AI doctors are following a penetration sequence of first reaching construction sites, then self-service for patients, and finally office and hospital approvals.
Structurally, this represents a transfer of pricing power. When models are quoted based on token varieties, when 99% of intelligence is designed not to interact with people, and when loyalty shifts from family to a piece of code in three years, pricing power shifts from "who has the strongest general model" to "who can embed invisible intelligence into processes and charge differentiated prices." The mechanism is one of sustained supply shortages, with demand already being voted on by patients and blue-collar workers, while the certainty illusion of the intermediary hospitals and enterprises becomes a discount factor instead.