Looking back across the companies I’ve built and backed, I keep returning to one question: what has just become possible?

I think of the answers to that question as technological inflection points: moments when something that was previously too expensive, too slow, too complicated or simply impossible suddenly becomes possible, and a small number of founders understand the implications before most people do.

The implications can reach well beyond business. Some of these new capabilities change what we can attempt. Others change what we feel we cannot afford to refuse. Understanding a new capability means asking what people can now do differently, which constraints disappear and what becomes valuable as a result. A founder can understand something important about a technology and still be wrong about what follows from it.

What changed at Soft32

Looking back, it's easy to make it all sound deliberate. At Soft32, the business model developed over years.

I started Soft32, a software download and discovery platform, in 2001. Around 2010–2011, we began building our own advertising platform to work directly with software developers. We had identified needs that the big advertising platforms weren’t addressing because they weren’t built specifically for that market.

One of them was reaching users while they were already downloading a program. At that moment, they were more open to installing similar or complementary software. We could recommend those products, and developers paid us mainly when someone installed them.

As I told Ziarul Financiar in a 2014 interview, the business took off in 2012. We had moved from affiliate marketing to display advertising and then to our own advertising platform, increasing the average revenue each user generated for us.

Recognizing an opportunity and finding the business it can support are not one decision. The opportunity doesn’t only appear when a company starts, either. Soft32 didn’t become a new company; we found a different way to serve a market we already knew.

A correct insight can still lead to a bad investment

Seeing the implications early only matters if it changes what you do: what a team builds, what it can afford to test, or what it no longer needs to do itself.

There is another distinction for investors. A technology can become important and a company can execute well, while the investment still disappoints if the entry price already assumes an exceptional outcome.

Being right about the direction of technology is not the same as being right about a company. And being right about a company is not the same as being right about the investment.

Understanding what follows matters just as much when the consequences go beyond business. Two areas hold my attention now: healthspan and military AI. One asks what happens when some of what we once accepted as inevitable with aging becomes preventable or reversible. The other asks what happens when an advantage creates pressure to adopt it.

What longer healthy lives could change

By healthspan, I mean the years we can live in good health, able to do the things that matter to us. I think we still underestimate what extending them would change.

I believe we will become able to reverse some of the damage associated with aging, rather than only manage its consequences. In 2020, David Sinclair and his colleagues reported restoring visual function in mice by partially reprogramming cells in the eye.

In June 2026, Life Biosciences, which Sinclair co-founded, announced that the first participant had received ER-100. The Phase 1 trial tests cellular reprogramming for optic nerve conditions. It primarily assesses safety and tolerability, with visual function also evaluated. It’s the start of human testing, not evidence of benefit in people.

Other approaches have reached different stages. In December 2023, the FDA approved its first CRISPR-based treatment, Casgevy, for certain patients with sickle cell disease. In September 2025, ProteinQure began testing its AI-designed cancer drug candidate PQ203 in humans in a Phase 1 trial.

I’m interested in what these approaches could make possible together: treating more disease at its source and restoring functions we have learned to expect to lose.

If that translates into substantially longer healthy lives, the changes wouldn’t stop at healthcare. How would we think about education, careers or retirement if we had many more years in good health? What would we choose to spend the extra time doing?

The population effect would depend on births as well as deaths, how widely treatments became available and how societies adapted. Who would get access, and on what terms?

And there is the darker question of an almost eternal tyrant. We don’t need literal immortality for that concern to matter. A treatment that gives someone more healthy years could also give a ruler more years to hold on to power.

I also see possibilities in the other direction: more time to learn, to work on problems that take decades or to reconsider the timescale of human exploration beyond Earth.

When a capability creates pressure to use it

With military AI, a new capability may start to feel compulsory.

In June 2026, Shield AI, which I’ve backed, announced a U.S. Air Force production contract for the Collaborative Combat Aircraft program. The contract covers its Hivemind software for multiple autonomous aircraft to operate together under human supervision. The announcement doesn’t say who authorizes an engagement against a target.

That distinction matters. The International Committee of the Red Cross distinguishes AI that supports human decisions from weapons that, once activated, select and engage targets without further human intervention. Not every military use of AI is an autonomous weapon, and not every autonomous weapon uses AI.

I’m fine with autonomy in sensing, navigation, coordinating aircraft, electronic warfare and intercepting incoming missiles or drones within limits people have set. People decide when to activate a system, what it may engage, where and for how long, and keep the ability to stop it.

Three kinds of decision should stay with people:

  • A machine should not pick a human being as a target and attack. Whether someone is a combatant, wounded, surrendering or a civilian can change in seconds, and the error is irreversible.
  • Starting a war, widening one and anything involving nuclear weapons should stay in human hands.
  • People must make the legal judgments: whether a target is lawful, whether expected civilian harm would be excessive compared with the concrete and direct military advantage anticipated, and whether all feasible precautions have been taken. AI can inform those judgments. A person has to make them and answer for them, whether the target is a person or an object.

A human approval button isn’t enough on its own. The ICRC has warned about humans rubber-stamping AI recommendations under time pressure. If there is no practical opportunity to understand or challenge a recommendation, the person may be in the process without meaningfully controlling it.

My concern is the competitive pressure between adversaries. Each side could delegate more because it expects the other to act faster, even when both recognize the risks.

I would keep the boundaries around targeting people, escalation and legal judgment even at a speed disadvantage. That is why I want rules that bind adversaries, rather than limits left to one side’s restraint. In August 2026, the UN and ICRC renewed their call for international rules on autonomous weapons.

The catastrophic possibility I worry about is a sequence of errors and reactions moving faster than people can interrupt. A system intended to give one side an advantage could make the situation harder for either side to control.

Right now, I’m watching whether we can restore health we once expected to lose, and whether we can retain meaningful control over decisions we are learning to automate. I want the first to move from promising research into useful treatment. I don’t want the second left to the logic of an arms race.