Notes on living and working with AI
Most of what I read about AI falls into one of two camps: bold predictions about what it will soon replace, or weary complaints about how often it gets things wrong. My own experience has been quieter and more practical. I’ve been subscribed to Claude for over six months now, and it has changed a handful of small and concrete things in how I live and work, which I’d like to set them down while they still feel new.
The most tangible change it has brought to my life is that building personalised tools now takes almost no effort.
Whenever a particular need came up in the past, I had to trawl the internet for publicly available tools and try them out more or less at random. Quite often, what the developer had designed didn’t quite match what I actually wanted, and at other times I had to pay extra to unlock the better features. Now, in the same situation, I can simply spend a few tokens building a tool that meets my needs without any friction. In the initial prompt, I already set out the UI, the combination of features, and the input and output formats that suit me best. In effect, I use natural language to approximate the scenario I have in my head. The AI reads that approximation and generates the product as it understands it. I then compare the result against what I’d envisaged and pick out the small things that need changing. The whole process takes far less time than searching for, trialling and comparing different tools online, and that’s before counting the times I give up on one or switch to another because it never quite fits. If a new need crops up while I’m using the tool, there’s no reason to put up with it. I just add it in.
The tools I make only have to suit me, and nothing else needs to be considered. For an individual user, I can’t think of anything more comfortable than that. This technology is no longer in the hands of a select few, so there’s no need to spend time or money getting over a technical barrier. Since Opus 5.5 came out, I’ve noticed that tokens go considerably further, so creating and maintaining these tools barely registers. There are always enough tokens to go round, so I never have to worry about hitting the 5h limit.
From a professional’s point of view, tools made this way may well be imperfect in terms of code, and their UI may not follow certain design conventions. Still, they serve my own life at minimal cost and require no adjustment on my part, and that alone is the whole point of them.
Beyond everyday life, the impact on research shouldn’t be underestimated either. For pipeline-style work at least, AI already far outperforms any bioinformatician. Once a piece of work is finished, it can also collate all the results and produce a draft report faster and more comprehensively than a human could. Outside those two areas, though, a great deal still can’t be done without human interpretation, particularly analyses where you need to produce an intermediate result, assess it, and then decide which direction to take next. My personal guess is that if enough bioinformatics engineers with a reasonable degree of biological intuition and experience joined AI companies and spent some time on targeted training, even this kind of judgement would no longer be so human-needed. I’ve seen similar roles advertised in some companies’ job postings, though I have no way of knowing whether they’re approaching it from this angle and organising the work accordingly.
None of these changes has given me much of a jolt or any anxiety about my career, though. What I seem to care about more is how to work better with AI, rather than panicking about being replaced. Hardly a day goes by without some new technical update. What I need to think about is how, amid change this hard to predict, I can learn to go with it. It’s rather like sailing. The wind keeps shifting in direction and strength, and perhaps none of it can really be forecast, but for me, at the helm right now, the question that matters most is how to make better use of the wind I encounter. After all, the ability to use AI (though I’ve yet to hear of any metric for measuring it) is the key factor in actually making any of the above happen. If you can’t write a good prompt or manage your context sensibly, then, at least while large language models are still built on conversational completion, even the best model won’t do you much good. I don’t consider myself excelled in this respect. I’ve simply been learning as I go and picking up a few tricks along the way, and so far that seems to be enough. People around me often complain that AI is stupid, but I rarely run into that in my own use. When writing code in particular, as long as the prompt is clear and precise and the context and memory are working properly, AI hardly ever does anything that strays far from what I intended, so the kind of stupidity people talk about doesn’t really arise. I’m more inclined to think the bigger problem lies with the users themselves.
One more thing I’ve noticed in practice, which also owes credit to conversations with my friend Yang: if you want AI to create something that has no established template, it helps to give it more examples of what you’d like it to learn from. Once it has samples to imitate, what it produces won’t carry the biases that its built-in prior knowledge can introduce when working de novo. This is really no different from teaching a pet or a child. Without a template or reference, whatever they do is unlikely to match exactly what the adult had in mind, but once you show them how it should be done and give them a direction to work towards, everything becomes much easier.
Looking back over these six months, the common thread is fairly simple. Whether I’m building a small tool for myself, running an analysis, or asking for something with no precedent, the result depends largely on how clearly I describe what I want and how much useful context I provide. The models will keep improving, and the limits will probably keep loosening. What stays with the user is the ability to express intent precisely, and that seems to me a skill worth practising for its own sake.