New Adventures in AI

Hello there 👋 It's been a while.

I'm slightly crestfallen that the first topic I've motivated myself to write about after a lengthy hiatus is AI. Such is the world we live in, eh? At least I've not referred to it as Super Intelligence.

One of the interesting impacts of the much-vaunted AI revolution is the sheer volume of human output it's catalysed.

It's hard to think of a topic that's generated such an array of brain dumps, deep dives, breathless headlines, hot takes, and academic theories.

Maybe something we can all agree on is that the rise of AI has kickstarted an almighty wave of content creation? An endless feed of text, audio, and video waxing lyrical about AI’s potential, its consequences, its value, its harms. Etc, etc.

As a result, new training data is being pumped out on a constant basis, keeping all those shiny models topped up with fresh fodder.

Perhaps that was the plan all along?

I digress.

I actually started writing this post in March. Since then I've come back to it sooooo many times, nipped & tucked, but never rounded off.

Glancing at it every few weeks plagued me with self-doubt. The points I was trying to land seemed to age quickly against the backdrop of a fast-moving field – heaven forbid my carefully-crafted prose might feel dated.

Six months on though, my sentiments have shifted somewhat.

At the beginning of the year we were awash with excitable "everything-is-changing-hold-onto-your-hats!" narratives (see Matt Shumer's "big dumb post"*) and a slew of reporters, podcast hosts, and substackers merrily vibe-coding their every waking thought.

Documenting chipper conversations with Claude might have made for sticky content, however the all-encompassing paradigm shift is still, well, a work in progress?

And now, as I type, I can barely scroll an inch without stories about existential threats and AI wiping out humanity popping up. Life comes at you fast, I guess.

The hype cycle, purposefully designed to beguile and confuse, stays constant though.

We lurch from one definitive position to the next, forced to learn new terminology and terrible company names at a breakneck pace (rogue agents, recursive self-improvement, Hugging Face, siiiiighhh) and lap up the opinions of people who don't exactly have a notable track record for telling the truth or landing their promises.

So let's talk about context.

Automation for the People?

Earlier in the year I published a post where I repeated the phrase "it's a run, not a race" ad nauseam. In it I pondered:

• the importance of finding the right context when it comes to technology choices
• the evolution of large language models and the resultant shift in the AI discourse

What follows picks up on those themes. It's a selection of unstructured and (undoubtedly) interim thoughts on the use of AI that have been floating around in my head.

These thoughts reflect my context.

With AI, much of what I end up trawling through, what I imagine we all trawl through, can reflect such opaque motives that it blands into nothingness, transient noise in amongst a cacophony.

In a work setting, I get the impression most people are scrambling through an avalanche of AI-related material, slightly desperate to find their North Star; a needle pointing to a concrete use case in amongst a haystack of possibilities.

Out of work the conversations feel less flailing, and the chats I have often reflect the hazy narratives perpetuated by AI boosters and doomers alike.

All the jobs will go. The robots will cure cancer. We'll automate all the boring bits and then spend our lives on important stuff. We'll automate all the interesting bits and then spend our lives in pointless drudgery. We'll drown in slop. Is slop actually art? The earth will flourish when the agents solve Climate Change. The earth will end when the agents develop a God complex and annihilate us mere mortals. Data centres are a necessary evil. Data centres are where the fightback begins.

I'm labouring the point, but it's why I think context really, really matters. And it matters all the more when dealing with consequential and contentious technology.

I'm under no misconceptions that what I'm writing here only adds to the endless pile of digital landfill; another overlong post reflecting the views of a privileged middle-aged white man. So please take it or leave it on that basis.

(Self) Reveal

This is me: I've worked in lots of different sectors, in large and small organisations, with both tiny and chunky budgets.

I'm a technologist – in the sense I know about the application of lots of different types of technology – but I'm definitively not a dev, or an engineer, or a sysadmin.

I largely deal with digital products and strategy, and I'm a stickler for involving users in the design of the projects I'm involved in.

I have my limitations though, just like you. Limitations based on circumstance and experience and perspective.

And remember, so does everybody else: from the tech billionaires to the many, many commentators whose words dominate our timelines.

One of my biggest bugbears is a failure to acknowledge the often blinkered outlook of those whose rhetoric tends to dominate AI discussions.

I think it's critically important to remember people's limitations, and motivations, when considering what they have to say on any given subject.

Failure to do this plays directly into the hands of those with dubious interests, and can set false expectations.

For all my hand-wringing, however, there's no point in denying a big shift in the utility of AI tools, and sizeable leaps forward in the quality and complexity of what can be achieved.

We've at least partially emerged from the hubris-and-bullshit phase that has dominated since ChatGPT publicly launched, into something, well, different.

If I were to sound a note of caution, I've seen a wave of enthusiasm from people who earned their stripes in the early web era starting to sound a teensy bit like peddlers of magic beans.

Don't get me wrong, there are clearly interesting things afoot. But I do wonder if being able to tinker in a way many once fondly did in the 1990s – where some hacked together HTML, a pirated copy of Photoshop 3, and an FTP account made you feel you were changing the frickin' world – is skewing matters a little (see also: Jukesie's take on Yahoo Pipes).

Your context has shifted, does that mean everyone else's has?

And underpinning all of this is, of course, the uncomfortable paradox. While tools have genuinely improved, the discourse – in fact the whole business model – remains as broken as it's always been.

Life's Rich Agent

So here are a few ruminations. Delete as appropriate. I've written them as a person who sets policy and technical direction within an organisation, and has responsibility for ensuring digital services run safely and smoothly. A generalist with general observations.

I don't have a definitive source: I read widely and I listen to an array of tech-related podcasts; I talk to people I trust, many of whom have followed similar career paths to mine; I'm on a couple of super useful group chats; I try to stay curious.

I also spent a couple of months dabbling with Claude Cowork earlier in the year, following tutorials in Toby Barnes' excellent Field Notes.

Building a small library of personal web tools helped me get my head around possibilities as well as highlight some specific constraints.

One of those constraints is that – in contrast to the folklore – much of what gives generative AI its utility is really not for everyone. Coders, dabblers, tool builders, fo sho. As Benedict Evans helpfully reminds us though, most people are not tool builders.

1. Ephemera is easy
   It's really, really simple to put together a basic prototype. A game, a website, a piece of functionality. It's also relatively trivial to stitch stuff together, and to get tools to plan, action, take decisions, and carry out tasks without intervention.
   
   It's satisfying to experiment and create stuff that would have previously required time, or a team, or both. I can think of lots of contexts where this approach is good enough, perfect even, but I can think of plenty more where it's never, ever going to make the grade.
   
2. Scale is hard
   Translating anything, AI or otherwise, into a meaningful, long-term, robust and accessible product is difficult.
   
   The Internet is awash with dead websites, dormant social accounts, abandoned apps and now, I imagine, millions of half-baked AI projects. Fine. However I'm starting to see what I might politely call bedroom projects influence the direction of products and services.
   
   It's easy to spin up something convincing – a working chatbot, a slick interface, an agentic solution – and just as easy to be seduced into thinking it's ready for rollout. Let's ship tomorrow lads!
   
   Grumpy old folk like me spent years unpicking, relayering and professionalising the way web products come to fruition for good reason. If your solution is aimed at the public it needs to be genuinely roadworthy, lest you find yourself responding to a data breach or unpleasant hallucinated mess.
   
   I'm not saying that workflows need to stay the same btw – we all have to adapt to the way AI fits into, and evolves, our processes. But try to think about it as providing links in a chain rather than the chain itself.
   
3. The fiddly bits are still fiddly
   As someone managing an organisation's tech infrastructure, I shudder slightly at the prospect of giving everyone access to AI tools they don't fully comprehend.
   
   Rolling out Copilot across an organisation feels simultaneously like the easiest and least strategic response to FOMO I can think of. That said, I also empathise with the pressure people are under to show progress and keep up with peers and competitors.
   
   As per the 'tool builders' point above, doing this in a sustainable, sensible way relies on a degree of familiarity with how modern software and the web intertwine: what an API does and how to use it, where code is hosted, how stuff connects together and why it sometimes doesn't, where data resides, the integrity of that data, the security protocols in place, how your domains are managed, the free tools you can rely on and where you absolutely need a paid licence.
   
   I'm not saying this is rocket science, but it also isn't insignificant. This type of work – understanding how digital patchwork quilts are pieced together – is suited to dev teams, and people noodling with code in their spare time, but really unsuited to the workforce as a whole.
   
   When things break (and they will break) we risk becoming more and more reliant on the AI being the only hired hand available to fix the error. More broken plumbing, more obfuscation, and not to mention...
   
4. Fantasy costings
   None of the current cost models are fixed, and everything is being underwritten by financial backers and venture capitalists. So when we hear AI is cheaper than employing a person, it's worth asking "at what price, and for how long?" 🤔
   
   You can't balance the cost of a salary against the equivalent of an introductory offer. Prices are being kept low to win market share, and at some point the people bankrolling this will want their money back.
   
   Also! For complex problems, agents burn through far more tokens than a quick chat, so even if the price per token holds steady, the bill grows with every smidgen you pass the AI's way. And the sums rarely include the human cost of processing, correcting and babysitting the output. Someone still has to check the thing before it goes out the door, and that someone is on a salary.
   
   I know there's a lot of yada yada about having to "rewire your business to succeed in this new AI-powered world". But simply not knowing whether your shiny solution will cost ten grand or a hundred grand feels like a massive, gaping hole.
   
5. Friction is underrated
   And finally... what of those promised efficiency savings? Personal AI agents are suddenly flavour of the month (see: Meta's Muse, OpenAI's Dots, Instinct, Vellum), complete with cutesy avatars to distract us from stories about rampant data centre expansion and the world ending.
   
   These hook into the premise that automation of everyday tasks is the ultimate prize we've long been striving for. I'm sure there's a fair amount of truth to this: look at the thousands of productivity apps already on the market, or tutorials to spin up your own agent to deal with the horror of your email inbox.
   
   People clearly struggle with the overwhelm of life and its associated digital clutter, and there's something attractive – in both a work and general day-to-day context – about throwing all your shit at a machine to sift, sort, complete and file.
   
   A couple of critical things seem to be missing from the premise. Efficiency gains quickly become business as usual, and new tasks soon creep in to eat up all that banked time (see Parkinson's Law). Language model usage in my own organisation definitely feels like a more established workflow than it did six months ago, but is it making us more efficient? For some I'm sure the answer's yes, for others I suspect it just becomes part of the accepted rumble of working life.
   
   More pointedly, is removing friction a good thing? The reason I don't reply to emails instantly, or churn out responses to documents, or decide things on a whim, or book stuff straight away, is that I value the in-between times where I process and get my head around a problem.
   
   Friction is also what stops bedroom projects shipping too early: review, refine, sign-off, testing with actual users.
   
   I'm sure to many this only confirms my role as procrastinator-in-chief, but I honestly think, in an era so dominated by kneejerk reactions and expectations set in real time, friction keeps us grounded.

Out of Time

Gosh, what a lot of words, I hope some of them resonate. Maybe you can get Claude to summarise?

Even if the hot takes have dated, I don't think the key questions have – but I'm probably not the right judge.

Does my context reflect your context?

Further reading

• Posting Dave Karpf's piece from 2024 again as it feels even more relevant now: How are these digital butlers supposed to get paid?
• Lauren Pope's new report is full of excellent nuggets: AI in charity content. It's not just about content, or charities. Or AI for that matter.
• Models don't go rogue

Thank you for reading 📖

*kudos Garbage Day for calling it.