I don’t buy into the idea that AI is all hype. I think people are looking at the limitations of today and missing where this is going. Some of my thinking involves a leap of faith, but it also comes from spending a lot of time using these tools and seeing how quickly the practical possibilities are changing.
I think we have effectively reached AGI by my own functional definition: sufficient general cognitive capability to perform a very broad range of intellectual tasks, learn to use tools, operate computers, solve unfamiliar problems, plan, research, code and pursue goals across domains.
That doesn’t mean the systems are perfect. They can fail on long workflows, lose track of objectives and require supervision. My point is that the limitation of one model instance doesn’t necessarily define the limitation of a system built around it.
We invented organisations for a reason
A person doesn’t run a complex business by doing every task themselves. They delegate, bring in specialists, ask for reviews and change the plan when something fails. I think agent systems can increasingly operate in a similar way.
A manager agent can coordinate specialists, while other agents inspect the work and challenge the result. The question I’m interested in is what that organisation can accomplish, not just what one uninterrupted conversation can manage.
My thought experiment is a system with permissioned access to computers, services, resources and other agents. Could it build and manage an organisation that accomplishes complex real-world goals? If it can, then for my purposes the debate about general intelligence has largely been answered.
Capability isn’t the same as adoption
A large company taking years to reorganise around AI doesn’t tell me the technology lacks power. Procurement, permissions, legacy systems and liability all affect deployment. A founder or small team can often explore the tools much more directly.
I can already use them to perform substantial parts of work that previously needed other people. Most people are still nowhere near using these systems to their full potential. I think that gap matters.
The economics are the crucial threshold
I keep returning to cost. Increasingly, the question is how much it costs for AI to complete useful work reliably. Token price alone isn’t enough. I mean inference, retries, tool use, agents reviewing other agents, human supervision and the infrastructure behind all of it.
If the total cost per successful unit of economically valuable work keeps falling while capability improves, more categories of cognitive work cross the threshold where using AI makes commercial sense.
And if AI itself helps improve software, inference efficiency and AI research, that threshold could move faster than people expect.
Employment changes before the headline numbers
I expect substantial disruption to computer-based employment. I don’t think it needs to begin with a single dramatic wave of layoffs. It can start with fewer junior hires, vacancies left unfilled, smaller teams and less contractor work.
My bet is a political and social red-alert moment around AI and jobs by the end of 2026. I put my own probability at about 75%. I may be early; if it takes longer, I expect the issue to be much harder to ignore by the middle of 2027.
Why I expect a correction in 2027
I think AI can become dramatically more powerful at the same time investors decide that too much money has been committed too quickly. Revenue and demand can be real while the returns still struggle to justify the capital being consumed.
My bet for 2027 is a major deflation in AI-related financial expectations. Some companies and infrastructure projects could struggle, capital could become more selective and the US equity market could take a substantial hit.
I think that could create meaningful recessionary pressure, potentially contributing to a US or global downturn. But the technology could keep advancing throughout it.
“We paid too much, too early” and “this technology is becoming extraordinary” can both be true.
Cheaper AI may mean more spending
There’s a complication: making inference cheaper may increase demand faster than it lowers the bill. A company could move from occasional AI use to dozens of agents working continuously.
That is why I don’t assume efficiency means total spending collapses. It could mean the same spending, or more, produces much greater economic output.
The infrastructure wave
Data centres, grid connections and energy infrastructure are physical projects. Announcing them and delivering them are different things. My expectation is that a significant part of the capacity wave becomes economically important around 2028.
By then, I expect models, chips, inference efficiency and agent architectures to have improved. More capacity would arrive into a world where each unit of compute can do more useful work. That combination is why I think 2028 could look very different from 2026.
Where my thinking on ASI has moved
I originally put a June 2028 date against artificial superintelligence. I’ve taken that fixed deadline out of my strategic bets. I still think 2028 is possible, particularly if AI starts meaningfully accelerating its own research and improvement during 2027.
I’m treating that as an open question I’m following closely. Revenue could take longer to catch up without automatically ruling out faster progress in capability. I’ve written more about that distinction in Why 2028 still feels possible.
The underlying bet
I don’t expect every forecast to land exactly on schedule. I do think extrapolating today’s world in a straight line is a much bigger mistake.
There can be bubbles, failures, delays and spectacularly bad predictions without the underlying technological change disappearing. I expect a financial correction, not a permanent end to the AI story. A second expansion may be built much more heavily on evidence of what these systems actually do.
That is the bet.
This is my personal view, recorded in September 2026. It is a set of forecasts and interpretations, not a statement that these future outcomes are established facts.
See the dated bets