human-computer symbiosis

AIW: looked at one way, this is Yet Another Article about AI, and there are more than enough of those already. Still, as someone who eschews US-based Big Tech in his personal life and is at the same time responsible (and even happy to be responsible!) for leading AI-powered launches at work, I like to think that I have a unique and conflicted position on GenAI - enough to make for a compelling story.

In any case, writing this scratches a personal itch, by helping me make sense of that inner conflict. If that's your thing, read on. If on the other hand you'd like very much to hear about anything other than AI, by all means play a game, wait for my next blog post, or do something else you enjoy instead. I won't mind. Really.

First, since there are so many words on this topic elsewhere, I'll point to some articles that resonate with how I think.

Quality in the Age of Slop brings in the concept of Quality from Zen and the Art of Motorcycle Maintenance: a sort of intuition about what is good, worthwhile, beautiful, etc. that LLMs utterly lack, and that is essential to the satisfaction of craft.

My AI Adoption Journey (also linked to by the above article) shows one path to adopting GenAI, one that to me comes close to the craft ideal of treating these models and harnesses as power tools - not gods, not oracles, just a tool that can save time and effort if used wisely. More than anything else, I believe it embodies a healthy balanced approach: not taken in by the hype, not rejecting new tools outright.

Deep Blue was the first article I saw that articulated the impact GenAI has on our identities as tech workers, and what it means to wrestle with fundamental change in our profession. I definitely felt shades of this in early 2026, and in retrospect one of the best decisions I made as a leader was to just be open about that - it validated similar feelings that my teammates had, and it helped clear a path for us to approach this with curiosity instead of fear.

The People do not Yearn for Automation is an excellent take on the disconnect between those pushing for AI adoption and the rest of us. Much as they always have, humans instinctively distrust and resist top-down pushes to make themselves legible to inhuman systems. I strongly agree that attempts to roll out GenAI tools must respect our humanity. (In this I am apparently joined by the Pope, who makes a similar point in his recent encyclical.)

It seems likely that the contributions of human operators and [computer] equipment will blend together so completely in many operations that it will be difficult to separate them neatly...

- Man-Computer Symbiosis, J. C. R. Licklider

In his classic 1960 paper Man-Computer Symbiosis, Licklider set a bold vision for computing: one where computers and humans would work together on complex problems, each augmenting the other, to achieve what neither could do alone.

66 years later, I find myself sparring with LLMs via Claude Code around complex tasks ranging from new feature development to architectural plans. As a tool, it has definite blind spots: to get optimal results, I must steer it using my human judgment and expertise. In exchange, I get access to a new kind of search index / pattern extrapolation engine, one encoded into billions of neural network weights learned through resource-intensive training on massive datasets.

And I do enjoy this process, in fact. I'm aware that this likely has much to do with my own approach to software engineering, which has always balanced a sense of craft with a product-minded focus on getting things in front of people to test, as well as a leadership-minded view towards building teams that build great products (and maybe not building every part myself). I've always been a bit of a paint drip person, with broad interests plus a penchant for learning deeper on topics as needed. Used purely as a tool in my toolkit, GenAI feels powerful - but containable when needed, given the right setup and precautions.

This has also given me cause for reflection. Man-Computer Symbiosis presents a powerfully prescient vision, but it also does so in a detached, value-neutral way. The rollout of GenAI has been anything but value-neutral: from disinformation to mass layoffs to contentious datacenter buildouts to environmental impact to concerns about ownership of training data, there are many valid reasons for distrust.

For being value-neutral, the vision does have a glimmer of hope: computers will augment us, not replace us. This certainly resonates with my own experience of LLMs and harnesses around them. At work, I wrote an AI-assisted engineering strategy for my team, which includes this segment on human-first spaces:

So what does a “human-first space” mean for using AI tools?

Much as there’s a difference between AI-assisted engineering and vibe coding, there’s a difference between AI-assisted thinking / writing and AI slop.

AI tools can be useful sparring partners, brainstorming helpers, and research assistants, even on high-level tasks. They also make mistakes, try to please the user (“sycophantic” behaviour), and don’t have the full business / problem / solution context you do.

In coding tasks, we can often guard against slop by using automated tools to add friction. Human-first spaces are precisely where friction is hardest to apply.

To me, this is what it means to respect our humanity in this process. Licklider is careful to note that symbiosis is mutual: in his vision, not only do computers augment humans, but humans augment computers. Where exactly we best contribute in the process as humans can and will shift. It always has: I started my career in a time before git, AWS, infrastructure-as-code, DevOps, CI / CD; my father learned to program on punchcards; my grandfathers fought in WW2 while Alan Turing was building codebreaker machines in Bletchley Park.

Friction is also a useful framing to me. LLMs will happily drift in a token-producing void without sources of feedback that sit outside the agent harness: linters, formatters, tests, humans-in-the-loop, and so on. This is not just an issue with current frontier models, but rather an inherent limitation of LLMs as a technology. It is not solvable with more precise prompts; if you load too many rules into the context window, LLMs will simply ignore them at random, like the dutiful non-deterministic output generation engines they are.

I'm encouraged that this has always been an issue with every generation of AI / ML systems - no matter how large the training set or sophisticated the modelling techniques. Optical character recognition (OCR) tools barf out extra characters here and there. Classification algorithms misclassify. Automated translations miss nuance and dialect. LLMs ignore instructions on occasion. At root, people are not perfectly reducible to data points, nor are the systems and societies we build.

I also find hope in Om Malik's AI models are having their iPhone moment. What's next?, where he describes the arc of commoditization that LLMs and agentic harnesses are already travelling, like many technologies before them. As someone who gave up US-based Big Tech last year, I very much look forward to a not-too-distant future where frontier model performance plateaus, open-weight models and open-source harnesses catch up, and locally-run models become standard practice. I can't wait for LLMs to have their "performance per watt" moment, where the arena of competition switches from number of parameters and benchmark-maxxing to useful tokens per unit energy expended. I would love for the hype bubble to burst, giving way to a new era of simply working with our new tools - rooted in a more balanced, respectful view of their place in our society.

Put another way: I have a value conflict. Some of my values run against giving so much power to an oligopoly of AI merchants, against burning ever more dead dinosaur to fuel our progress as a species, and against demeaning the value of human thought and labour. Some of my values drive me to be intensely curious about new possibilities, to seek lifelong learning and growth, and to lead others around me to navigate challenges without fear.

I don't have an answer to this inner conflict. For the moment, the realities of upcoming deadlines and impending fatherhood leave little time for more than idle reflection (and the occasional blog post!) One of my not-too-distant-future goals is to investigate open-weight models and open-source harnesses further, and see where I can replace Big Tech models and harnesses with them.

But in any case: I think there is great value in getting these conflicts and questions out into the open. Technologies change societies, often in ways their inventors never envisioned. That change can be a process of conversation and evolution, or one of upheaval; it becomes the latter when we fail to make space for the former. For those of us who lead, it's on us to make that space in our teams and communities - to build hope and steer towards responsible, humane approaches rooted in curiosity and experimentation, even (especially!) when we ourselves are uncertain where it's all going.

I must not fear. Fear is the mind-killer. Fear is the little-death that brings total obliteration. I will face my fear. I will permit it to pass over me and through me. And when it has gone past I will turn the inner eye to see its path. Where the fear has gone there will be nothing. Only I will remain.

- Frank Herbert, Dune