AI, speculation, and what remains#

There is a lot of noise around AI at the moment, and much of it is coming from the financial markets.

Circular financing, pressure on bond markets, data centre cancellations, heavily indebted infrastructure providers and difficulties securing enough electricity. Investors are asking harder questions about how the expansion is being financed and whether the expected returns will materialise. Recent reporting on AI funding and neocloud debt and Microsoft’s cancellation of an Nscale data centre agreement illustrates those concerns.

Somewhere in that discussion, the technology and the financial speculation surrounding it have become bundled together.

I think they need to be considered separately.

AI did not decide how much money to borrow, what valuation to assign to a company or whether to commit to building a data centre before securing its power supply. People made those decisions.

Speculation, a failure to exercise restraint and a fear of missing out are familiar human behaviours. We have managed all three perfectly well without artificial intelligence.

I keep an eye on the financial side because the consequences can reach beyond the companies directly involved. But my interest in applying AI to practical problems is a separate matter from whether the current investment frenzy makes financial sense.

The technology survives the companies#

If a couple of the largest American AI companies disappeared tomorrow, there would certainly be disruption.

Customers would lose services. Applications would need alternatives. Some projects would become more expensive or cease to be viable. Depending heavily on one provider creates a real business risk.

But the research, published techniques, software and already released model weights would not disappear with those companies.

The genie is out of the bottle.

That does not mean every capability is freely available, or that a locally hosted model can replace every commercial service. It means that AI development has spread far beyond any individual company.

A company can fail without taking an entire field of technology with it.

Progress takes different forms#

It is also easy to get the impression that progress in AI consists entirely of building larger models, buying more GPUs and consuming more electricity.

There are other directions.

Smaller models and models specialised for particular tasks offer opportunities where a large general-purpose model may be unnecessary. Open-weight models provide options for running AI on your own infrastructure, subject to their licences and hardware requirements.

Europe is contributing here too. France’s Mistral, for example, has released open models in a range of sizes, including smaller models intended for more constrained deployments.

Efficiency is another part of the picture. DeepSeek’s published work demonstrates techniques such as mixture-of-experts, where only part of a model’s parameters are activated for each token. That does not remove the hardware requirements, but it shows why GPU count alone is a poor measure of progress.

Researchers are also investigating different ways of generating text. Next-Concept Modeling, or NCM, builds an output through progressively more detailed representations, generating each level in parallel.

Its authors report substantial speed improvements against diffusion language model baselines on specific benchmarks. It remains early research, rather than an established replacement for leading commercial models, but it illustrates that the current approach is not the only avenue being explored.

What is actually useful?#

For someone building a product or improving a business process, the questions are fairly ordinary engineering questions.

What problem are we trying to solve? How reliably can the system solve it? What does it cost to operate? Where does the data go? What happens when it gets something wrong?

Sometimes a powerful hosted model will be the appropriate choice. Sometimes a smaller local model will be sufficient. Sometimes conventional software will do the job more simply and predictably.

Those decisions should follow from the requirements and testing.

The financial markets will reach their own conclusions about valuations, debt and expected returns. Some investments may prove justified; others may turn out to have been expensive exercises in optimism.

Neither outcome, by itself, tells us whether a particular application of AI is useful.

My interest is in applying the technology to real problems, testing its limitations and finding out where it earns its place. That work can continue regardless of what happens to the share prices.

Now for my own speculation#

Where does this leave me and others like me?

I have spent more than twenty years working as a software engineer, with an electronics background before that. Increasingly, I find myself questioning how much longer “software engineer” will adequately describe what I do.

In my own work, manually producing code is becoming a smaller part of the process. LLMs can produce much of it, and I have to admit that the results are often perfectly acceptable for the business applications and processes I work on.

There are still mistakes, questionable assumptions and things that need correcting. I still watch the code being generated, examine the changes and check the behaviour. But I would be kidding myself if I assumed that my current role in that process was somehow protected from further automation.

How long will I be needed to review it in quite the way I do today?

You can already assemble teams of AI agents within a coding harness, assigning different models and responsibilities to implementation, review and testing. I use OpenCode or Pi, and this is a way of working I have been exploring: agents producing code, critiquing it and making improvements through an iterative loop.

Is it perfect? No. Agents can miss problems, reinforce each other’s mistakes or produce something that passes its tests while failing to meet the actual requirement. Giving a second agent the job title “reviewer” does not automatically make its judgement independent or correct.

Nevertheless, I expect these systems to improve. That is my speculation, and I should apply the same caution to it that I would like investors to apply to theirs.

I also suspect that progress will come from a wider range of companies and approaches than the current attention on a handful of large providers suggests. Smaller players, specialised models and better tools may change the economics of software development further.

None of that gives me much confidence in a career built primarily around manually writing code. I do not know how quickly the profession will change, or what its eventual shape will be. But I am increasingly reluctant to define myself exclusively through manually producing code when these tools are taking on more of that work.

It was always about making things#

I have started moving away from describing myself exclusively as a software developer or software engineer.

Maker? Tinkerer? Builder? Inventor? I have yet to settle on a title that quite fits.

What interests me is taking an idea or a problem and making something that works. Sometimes that involves software. Sometimes electronics, physical hardware or a combination of all three.

Looking back, that was always the attraction. The code was a means of making something happen, even if learning to write it occupied a substantial part of my life.

There is uncertainty in this change of direction, and I would not pretend to have worked it all out. Finding useful things to build and finding people willing to pay for them remain two separate challenges.

But making things is still something I want to do. With the tools changing around me, that seems a better starting point than holding tightly to a particular job title.

I think that is the direction I am going to follow.