Work and artificial intelligence

Not long ago, at least in terms of human history, discovering new information meant reading a book, a paper, or a collection of papers. Then came wires, signals, connected machines, screens, keyboards and, eventually, the world wide web. Access to information and information itself grew in abundance, but abundance creates a new problem: how do you find the thing you are looking for?

Search became the answer. For a long time, the normal behaviour was simple: type a question into a box, wait for a list of links, open a few websites, compare the results, and decide for yourself which answer seemed most useful.

Now that pattern has changed. Increasingly, a website or app does not show you a list of possible answers. It gives you a summary, shaped by artificial intelligence, written as if it understands the question. Many people are now using LLM-based chat interfaces in place of traditional search engines. The shift is subtle, but significant: we are moving from finding information to being given an answer.

The future or now?

We first outsourced the task of finding information. Shelves of books became databases, indexes and search engines. Rather than spending hours locating a source, algorithms did much of the retrieval for us based on some key words.

Now we are beginning to outsource something more significant: selection. Historically, humans searched through information, evaluated sources, discarded weak arguments and formed conclusions ourselves. Increasingly, we are asking machines to perform parts of that process on our behalf. The retrieval layer was already outsourced, the sorting layer is next.

In technology circles we have a habit of overestimating short-term change and underestimating long-term consequences. Yet the future often ends up surprisingly close to some of the more ambitious predictions. Programmers are among the first groups experiencing this shift directly, both positively and negatively. Tasks that once required hours of manual effort can now be completed in minutes, but the value is moving away from production code and more towards judgement. Knowing what to build, what to trust, what to discard and what questions to ask is becoming more and more important.

Software engineers are unlikely to be the last profession affected. They are simply the first large-scale participants in a broader transition that will eventually touch almost every knowledge-based profession and any repeatable task performed primarily through a computer ( at least until Robotics finds its own exponential scaling pattern for performance and cost).

Of course, there are unresolved questions on all sides of this change:

• What happens when the scaling requirements of training and serving large language models begin to outpace hardware production, power generation, or worse, our ability to build new power capacity? If that happens, do the scaling laws continue to hold, or do they collide with physical reality?
  
• How will different nations, economies and cultures adopt increasingly capable systems that often appear magical and operate as black boxes? Trust in technology is not distributed evenly across the world.
  
• Assuming capability, automation and adoption accelerate to the degree many predict, what happens to western economies that are overwhelmingly service-based? Does labour demand genuinely decline, or does Jevons Paradox once again make the pessimists look foolish?
  

• If a small number of model providers emerge as dominant suppliers of intelligence infrastructure—perhaps not one company, but two or three per hemisphere—what does that mean for competition, sovereignty and economic power?

  Hypothetically, if AGI were achieved and proved as transformative as some believe, what happens when such a capability is not distributed equally? At what point does advanced intelligence become a strategic asset more significant than conventional military power?
  
• Will technological advancement eventually outrun available capital? Or will capital become the limiting factor long before the technology reaches its theoretical limits? The technology race and the capital race appear to be running in parallel. It is not obvious which reaches its constraints first.
  
• More fundamentally, are we even asking the right questions? Most technological revolutions look obvious in hindsight and almost nobody predicts their second-order consequences correctly in advance.
  

Speaking to people outside the technology bubble yields a very different perspective on AI. I’ve spoken to plumbers, bakers, builders, surveyors, lawyers, accountants and retirees. Many don’t trust the software they already use, let alone the next generation of increasingly sophisticated tools. Most aren’t asking for bespoke systems built specifically for them. They want technology to work, and to work in a way they can understand. Reliability, predictability and the removal of tedious, repetitive tasks seem to matter far more than intelligence for intelligence’s sake.

What’s striking is how many of these problems could have been solved before AI arrived. Better workflows, clearer interfaces and more thoughtful system design would have eliminated much of the frustration people experience today. Yet, taken as a whole, the technology industry has often struggled to consistently deliver those basics. AI may make them easier to implement. Whether it actually results in better outcomes remains to be seen.

We are not thinking machines that feel, we are feeling machines that think. - António R. Damásio