Risk Between Ice and Gas

Over the past two weeks I wrote that confusion is the correct response to what is happening in AI, and that we can prepare for it by generating a bit of chaos we choose ourselves. The first was a condition arriving from outside, the second a training we set up on purpose. Both lead to the same question, which is what all that preparation is for. My answer is risk: how we understand it, how much of it we can afford, and how its shape is changing as AI becomes more powerful.

What Risk Is

I see risk as the measure of how imperfectly we forecast the future. If we could predict exactly what follows from a given set of starting conditions, no future event would carry any risk at all. We cannot, and so every action we take has consequences we only partially anticipate. The cumulative set of consequences that diverge from what we expected is what we call risk.

Entire industries exist to measure that divergence and to put a price on it. Finance makes much of its money from asymmetries in how risk is understood by the parties to a transaction. Insurance sells protection against it, and insurers in turn buy protection for themselves from reinsurers, who absorb the risk the insurance industry wants to place.

When unpredictability grows, the pressure travels up that chain. Florida shows it clearly. Fifteen property insurers have gone insolvent in the state since 2020, many of them unable to find enough reinsurance to cover the claims a bad hurricane season would bring. If you own an apartment in a building on a Miami beach, it has become harder to insure, and the company that wrote your policy may itself be the next one to fail. Some reinsurers returned to the state this year, at prices that reflect what they learned.

None of this is new. What AI changes is the shape of the distribution, because it is moving risk two ways at once.

Risk Made Small

In our individual lives, the tools now available, used appropriately, can reduce the risk of starting something to a degree that has no precedent. Prototyping, market research, customer testing, go-to-market strategy, branding, customer support, vendor and supply chain management: every facet of a business of any size can be understood, structured and delegated to AI agents, ever more effectively and at a cost that keeps falling.

Last week I described how limited liability made failure survivable for ordinary people. AI does something similar to the cost of the attempt itself. People who would never have dared can try, fail without being crushed, and try again, until they discover what they are good at, what the world needs from them and what someone will pay for. That is the search we pursue every week at Ikigai Collective, and it is open to many more people than before, at least in the countries that allow it.

The solo founder, until recently an outlier, becomes a perfectly normal choice. You will still want to work with others, for the joy and the excitement of building something together, but you will seek them out because you want to and not because the work cannot be done without them. That is a new freedom, and choices made freely tend to benefit everyone who takes part in them.

The Value of What You Have Already Written

A friend from Dubai who joined our Ikigai Lab Night for the first time this week made the same point from another angle. He fed his agents twenty years of email: hotel stays, flights, subscriptions, receipts that had sat in his inbox without any value. From that archive alone they learned his habits well enough to order his groceries without his involvement. His worry is the mirror image of that convenience. The more you put into an agent, the more it understands about you, and he wants an open format to preserve that knowledge across platforms, so that a bot built today could still be consulted by his children decades from now.

The market has already put a price on archives like his. In August, after Spirit Airlines ceased operations, Google paid $10 million for its internal data at a bankruptcy auction: about 100 million emails, 500 million Teams messages and decades of documents and code, with another bidder close behind at $7.5 million. An individual archive is worth far less, and I would not sell mine blindly. The price still tells us something about the value of the finest-grained record of a life or a company, and we are the ones best placed to use our own.

Risk Nobody Can Measure

At the other extreme, the same tools are powerful enough to be used in ways that are hard to control. As of today there is no engineering or scientific solution to the reliable control of AI, whether it is directed by humans or acting on its own initiative after being pointed in a direction. Without such a solution we face a category of risk that cannot be measured at all, because no forecast exists against which to measure the divergence.

We live between these two extremes: a world in which the risk of trying has never been smaller, and one in which the risk of the whole system has never been harder to estimate.

The Cost of Refusing Risk

Faced with that uncertainty, many institutions choose to stop. In Europe I have seen working AI projects shelved in anticipation of the AI Act’s rules for high-risk systems, which include anything used to manage workers. Those rules were due to apply on 2 August 2026. In July the Digital Omnibus entered into force and moved them to 2 December 2027. Projects frozen to avoid a deadline now wait sixteen months longer for a rule that may change again.

The habit runs deeper than regulation. About ten years ago the head of human resources of one of Italy’s largest banks told me that a third of its over 100,000 employees could stay at home the next day and the bank would work better, because so many of them spent their time justifying their existence with work nobody needed. They could not do it, he said, because of the social cohesion and peace the bank had to maintain. What that choice expresses is a disbelief in individual agency, in people’s capacity and desire to grow, and it leaves many of them believing they are useful when they are paid to subtract value.

The contrast with what I heard this week in Abu Dhabi could not be sharper. The UAE’s Regulations Lab, created by a federal law in 2018, grants a temporary licence to projects that current legislation either forbids or does not cover, such as self-driving vehicles. Applications are reviewed in weeks. The licensee operates under agreed conditions for six months to a year and shares what it learns, and the government drafts legislation from that shared experience. One of the sectors it experiments in is the government itself. That matters in a country that has committed to running half of its government sectors, services and operations on agentic AI within two years. The private sector will need agents of its own that can speak and transact with the government’s, and those rules will be written from evidence gathered together.

Between Ice and Gas

There is an analogy whose depth I have not yet fully measured. Life exists between two extremes. On one side is the frozen, crystalline order of rock and ice, where matter passes time, solid, reliable and unchanging, aware of nothing beyond what it is. On the other is the whirling chaos of a hot gas, where no structure holds for long. From a thermodynamic point of view, life evolved in the interval between no change and too much change.

AI forces us to maintain, at all costs, trajectories in the same interval. We have to keep evolving in the interval, using the new tools to make our own risks small enough to take and making every effort to keep the systemic risks manageable, so that the tools serve our purposes. It will be hard, complicated and turbulent. I am convinced we will succeed, because no alternative exists: the future is what we embrace, design and build, and by definition that includes our own success.

My first public talk on the technological singularity was twenty years ago this month, in Milan. Few people in the room believed it would arrive in their lifetimes. Our friend from Dubai told us he had followed my work for a long time without believing it either, until the arrival of ChatGPT made it impossible to doubt. We no longer need to argue about whether the change is coming. The task now is to help others find their own place in the interval, by talking about it and by inviting them to join us.