Nobody has the full picture

It’s November 3, 2026, almost four years since ChatGPT launched. I started using it shortly after its release, and found it difficult to separate my excitement about its capabilities from my concern about what they might mean for my future.

I worried about what widespread automation would mean for people who relied on a salary, including me. If nobody needed their labour, being willing to work harder would not help them support themselves or build anything of their own. I began to feel that I had a limited window to establish some financial independence. At the time, I thought that window might be one, two, perhaps three years.

That put a clock on everything. I spent part of every day thinking about what I could build or invest in to make money. I saved and invested as much as I could, while studying the industry behind AI: chips, infrastructure, models and applications. I wanted to understand where value would accrue, which capabilities would become inexpensive and which problems would remain difficult.

Almost four years later, the world I was preparing for hasn’t arrived. My original timeline was wrong. Yet the progress I see in my own work makes the possibility of much broader change hard to dismiss. I’ve never felt this combination of stress and excitement for so long.

Much of that excitement now comes from building Oplema, our analytics software. Jens and I have been working on it for five months, four as an incorporated company. Agents work around the clock, allowing the two of us to build in roughly half a year what I would previously have expected to take a team of 40 a couple of years.

That shift helps explain the claims on X that “software is dead”. But the ease of producing another notes app tells us little about software as a whole. That market was saturated before AI.

Large software products involve many integrations, dependencies and failure paths. Producing an individual feature is only part of the work; making the whole system reliable for people who depend on it is a much larger task. I don’t expect that work to disappear in the short term.

It would still be a mistake to get comfortable. Our competitors have access to the same tools, which means we have to keep improving the product, deepen its integrations, expand what it can do and remain willing to change direction.

The difficulty is knowing how far to extrapolate from what we see in our own work. Over the past four years, I’ve repeatedly moved between expecting AI to automate all work, starting with white-collar jobs like mine, and wondering whether this is another case of “nothing ever changes”. The predictions range from abundance to mass unemployment and extinction. There is also the question of purpose if work disappears, although how much purpose a white-collar job provides depends on the person. That’s a subject for another post.

Lately, I’ve been trying to turn down the noise from the media, X and the most prominent voices in AI. The people building these models may have a much better view of the next release. That doesn’t give them a complete view of how the economy, businesses or individual lives will change in response.

I still want to understand their thinking. I just don’t want to organise my life around the latest forecast. I already did that with my own one-to-three-year timeline.

The same applies to the models themselves. They can help me research an idea, test an assumption or build something. I still need to form my own judgment about what I believe and what I’m willing to risk. I don’t want to defer those decisions to a model, a feed or someone running an AI lab.

Yesterday, I read Rikiya Takehi and his coauthors’ post on weight-space programming. They used Astra to reduce reasoning loops in a small Qwen model by editing its weights directly. It made me reconsider whether the world I’d been preparing for might be only a couple of years away. That reaction was what finally prompted me to start this blog.

I’ve had hundreds of thoughts I wish I’d written down. Writing gives me a reason to examine them more carefully, and a record I can return to as my views change.

No schedule. Just something worth writing down from time to time.