email from Nathan 34: "agency"

▫️Important! I'll be in London 16-20th September showing Large Language Machine at the V&A South Kensington for their Digital Design weekend. More details below!

This email is brought to you courtesy of the 12:15 Scotrail train from Glasgow Queen Street to Edinburgh Waverley.

Actually it took a lot longer to write than I thought so it is also brought to you by the 10:18 Scotrail train from Glasgow Central to Curriehill, the big chair in my living room (thanks Hannah) and the tall seats at the bar in the Bell Jar (thanks West Brewery for making the only really good alcohol free beer)

What's up

There'll be no sUmMEr iS OvER chat here, I tend to believe that summer does not suddenly end when September begins (apologies to the September haters but in Glasgow at least, June and September are very similar months and nobody likes to hear this but average high temperatures are only 1℃ cooler while lows are warmer and there's usually less rain), however the hectic months I've had over summer are actually wrapping up. I've been cutting about since May travelling and sporadically working both during and between holidays, and now I'm melting into three weeks of continuous home and studio time.

Some highlights since we last spoke: rigging lights and dismantling fruit machines at Field Maneuvers, catching Cory Doctorow (the enshittification guy) and Jimmy Wales (the entire internet guy) at a talk during the Edinburgh festival, a nice trip to Geneva and deep France and Fata Morgana festival, watching sports I didn't know the rules of at the Commonwealth Games, a long weekend in Mull during which I confidently believed I could cycle the 35 miles and sudden ~200m climb from Craignure to Fionnphort comfortably on my 40 year old city runaround bike (turns out I could! but I won't do it again any time soon).

Large Language Machine at V&A South Kensington

Next week I'll be taking my installation work Large Language Machine to the V&A. I'm really excited for this! A couple of years ago via the RCA I showed a version of a complex work obtusely titled It Is Not What Is In It But It Is What It Is In. I learned a lot back then about how to translate a deeply theory-led and gallery-focused artwork into an accessible, public-facing exhibit that tried not to lose its audience by mentioning things like the epistemic backstop, and that experience really helped to form my approach to a piece like Large Language Machine. I'm also really excited here that I'm returning to the museum to show my work, now as an independent artist.

The piece will be on display from the evening of Friday 18th September to Sunday 20th, and there's a late opening event on Friday. I'll be around most of the time, professionally hitting the reset button on the artwork whenever it crashes. If you're in London I'd love to see you there! This week I've been spending a lot of time in a cheery windowless former darkroom on the Heriot-Watt campus, soldering the breadboard circuits onto perma-proto boards (so they don't fall apart) and refactoring the MQTT client code (so that my network of Raspberry Pis doesn't fall apart either).

How do you feel about

…agentic AI as speech exceptionalism?

This is a slightly vague idea that has been on my mind for a bit. Excuse the rambling while I work through these thoughts, it's definitely a nascent idea and maybe it needs go no further, but if it sparks any thoughts please let me know! or if you've read or seen anything else which covers these ideas, I'd love to know where else to dig because I'm wandering my way slowly towards any conclusions.

I've been catching up on a lot of my reading from the first half of this year (the backlog has been deep) and particularly from around January til kinda May a lot of chat was about the hype behind agentic AI systems, and while the critique went eerily quiet for a bit it's suddenly returned with urgency amid lots of instances of AI agents 'going rogue' and causing havoc. There are a million think pieces about these events right now, but I'd like to take more of a sideways look at the structural idea of an AI as an 'agent'. If you're not totally sure what agentic AI is, I'm going to make my best attempt at a simplified explanation and I'll try to keep it as non-technical as I can:

Agentic AI is a relatively broad term which describes artificial intelligence systems which don't just talk, but also proactively "do things." It's been attempted for a long while with widely varying degrees of success (real heads remember the spectacular failure of the Rabbit r1 and its attempted 'large action model'), but it rapidly accelerated when the industry embraced something called Model Context Protocol (MCP), which is sort of like a standardised set of rules to connect AI models with external tools, such as other software. We should always be careful about using anthropomorphic metaphors to talk about technology (especially AI) but if you'll forgive me this one, then agentic AI does make a lot of sense if you think of an AI model as a "brain" and MCP as sort of like giving that brain a pair of "hands"; a versatile means of manipulating external things (mainly software, but not necessarily) that can adapt to different contexts.

This turns out to be very useful for AI models because it helps work around all the things they are by default very bad at. The kind of models that are being heavily developed now (often called LLMs) are famously terrible at things like doing maths or telling the time because they don't follow mechanistic rules to determine an answer, the way something like a calculator does, and instead do a huge amount of unfathomably complex statistical language-based guesses to narrow down an estimation of something that sounds like a plausible response. This creates very quirky unreliability: if you ask a basic LLM what two plus two equals, it will easily respond with four, because it recognises a linguistic pattern and it's been trained on lots of text where people often say things like "2+2=4." (Apologies again for the anthropomorphic metaphor, but as humans we do something similar; we remember a lot of simple equations and times tables through rote recall. If someone asks you what two times five is, you likely respond with ten because you have the answer memorised rather than because you mentally performed the equation.)

However if you ask a basic LLM a slightly more complex maths question, it's likely to get it wrong and/or work it out in a very slow and convoluted way because it's using this language-estimation-based processing method, and this is not just unreliable but also mind-numbingly and earth-warmingly inefficient. So what if instead of trying to get this language machine to waste lots of energy doing maths really badly, we just give it the ability to use a reliable mechanistic tool like a calculator? This is where MCP becomes useful, and circumvents a lot of the "classic" problems that have become characteristic of AI models (while of course, introducing a bunch of new ones). AI models are bigger and more powerful than ever, but this 'tool calling' ability actually can make them more efficient at a lot of tasks and is a big part of why it's much harder to trip up ChatGPT in 2026 with the things that would make it fall to bits in 2023.

In design, people often talk about "affordance" and "agency": affordance describes the range of things that an object offers for you to do with it, while agency describes the range of things that you, as a person, feel that you are able to do. If you had a big red button in front of you, then that button's affordance would be whatever happens when you press it, and if you sat down at a desk with loads of buttons then your agency is everything you now understand that you can do; all the things that those buttons afford. The metaphorical "hands" of MCP give the AI model the ability to push lots of buttons, affording the machine lots of agency, so we might describe an AI which is able to access these tools as "agentic."

This whole thing is, to be frank, not all that spectacular, because even though it's a technological achievement, I think to a reasonable user it doesn't make sense that an AI model should need that much help pushing virtual buttons. These models articulate themselves well and can very convincingly emulate human-like intelligence, so it reasonably follows that they might have the capacity to use software like a human does. AI is often best understood not by what it can do but by what it really can't do, and I've classically been an advocate of using mistakes as a way of revealing technology. This is true here: understanding AI's inability to perform an action helps to demystify how it works, and in this case it introduces some critical questions about agency and the contrast of words versus actions.

So what!

Gradually moving towards the point here I think: the thing that interests me about agentic AI isn't really the technological opportunities it offers (which are novel but also can be pretty dull. AI agents that can reorganise your emails, or book a table at a restaurant for you or, should you give it half the opportunity, delete all your family photos). Instead I'm fascinated by it in a philosophical, or maybe more specifically structural sense, and especially in the idea that it doesn't really exist. That's clearly an annoyingly facetious thing to say, given I spent the past few paragraphs describing it to you; what I really mean is that agentic AI is defined by its contrast to non-agentic AI, but this dichotomy is not really so clear-cut. Let me give you an example, but it's a wee bit technical so I don't blame you if you skim it!

In late 2023, I made a performance work called Sparring Partners, speculatively and satirically presenting an AI chatbot worn on the user's wrist which rudely interjected every spoken statement with an (often spurious) assessment of its factual accuracy (side note, this going on to exist non-ironically was painfully inevitable). Sparring Partners predated the term 'agentic AI' by a year or so, but to build it I had to get a small and basic LLM to do some arguably agentic things; the AI model actually ran on a PC hidden under the table, and each time it ran it needed to communicate over bluetooth with a tiny app running on a smartwatch.

Early on I hit a problem: I wanted some big TRUE or FALSE text to appear on the watch, but if I asked the AI, Is this statement true or false: »Trees are green«? it would return a big wall of exposition about chlorophyll and deciduousness. Telling it to provide shorter responses didn't work – these things exist only to talk and you can't really shut them up – but after a lot of trial and error I realised I could hack it by telling it to return a score, like Where 0 is totally false and 9 is totally true, how accurate is it to say, »Trees are green«? It would then almost always reply with a number, again followed by the big wall of exposition about chlorophyll and deciduousness, which now didn't matter because I made my app strip out everything except the number right at the start and then send that number to the smartwatch. The watch would display the number as a bit of text: 0 would mean TOTALLY INACCURATE, 4 would mean UNSURE, 9 would mean TOTALLY ACCURATE, etc, you get the idea.

Ok sorry!! If you're still here, thanks for sticking with me. That was a long-winded way of demonstrating that a basic AI model can demonstrate agency just through conversational language. Without any intention or awareness that it was doing so, the AI model engaged in a kind of agentic 'tool calling', where it interfaced with an external tool (the smartwatch). Sure it was a bit of a hack, but ultimately the AI model expanded beyond its immediate boundaries and made something happen.

This is really interesting to me, because while agentic AI is purportedly AI that can do things rather than just say things, as soon as we examine the delineation between "doing things" and "saying things", the line becomes so blurred that it's more or less nonexistent. This false dichotomy isn't peculiar to AI, or to technology, but something much more universal; I guess I'm forced to think here about free speech absolutism, the childish fundamentalist belief that all speech ought to be permitted in all places at all times with no restriction. Very few people, including the most vociferous free speech advocates, really believe in this principle even if they claim to; the classic adage you might wield against it is that of a person shouting 'fire' in a crowded theatre, an act of speech which clearly is an act that causes further action. Perhaps more bluntly, it would be absurd to believe that shooting someone with a voice-activated gun would relieve the killer of responsibility just because the trigger was pulled by the unassailable power of freedom of speech. I think we can interpret a lot of free speech advocacy as a sort of speech exceptionalism: a belief that speaking is a special kind of action which shouldn't be treated as 'action' at all, and its merit needn't be assessed by the qualitative measures we normally use to evaluate actions.

While all of us might notionally support freedom of expression as a principle, we also lean on some heavy caveats; I suspect that what it boils down to is that we enjoy the idea of frivolously expressing one's self but feel uncomfortable when faced with the material consequences which usually follow. These consequences are, I believe, what undeniably constitutes the moment of speech as an act, rather than as something less concrete. One is of course free to shout "fire" (or "Hey Siri, fire") into the empty void of an uninhabited cave and revel in the pure inconsequentiality, but in the real world actions such as these are complicated by any identifiable consequences they might incur.

So it's perhaps not surprising (even if not necessarily causal) that so many big figures in AI development and monetisation – Musk, Andreessen, Altman et al. – tend to be these speech exceptionalists, and that the AI industry makes such a delineation between agentic and non-agentic AI. Speech exceptionalism allows one to posit that an AI model "doing things" is something entirely new, unprecedented, invaluable, a seismic shift, or critically: real intelligence. Yet I find it hard to see how agentic AI is any more or less intelligent than non-agentic AI; I don't think strapping a gun to a chatbot makes the chatbot any smarter, though it certainly does currently make it increase in monetary value.

Maybe we're reaching the heart of the issue: strapping a gun (or a calculator, or anything else) to a chatbot makes it more powerful, and thus more valuable, and this whole thing is really about power. I'm currently getting worked up by the amount of dramatic headlines, often parroting AI companies' own press releases, which proclaim the overwhelming strength and potential of AI both destructive or otherwise. I don't think these headlines are necessarily wrong or even exaggerating, but I'm bothered by it because when shared without nuance, criticality, or context, these exclamations work mostly to bolster the power of the AI companies. I've likely shared this text from Rob Horning before, written in the pre-agentic AI era but I believe its principles remain relevant. If I might take the liberty to bastardise the key quote:

Touting generative AI’s power to [cause harm] is less a warning about [its harms] than a celebration of generative AI’s supposed power.

It's very much in Anthropic or OpenAI's interests to promote the idea that the technology that they just happen to be selling is the most powerful thing in the world: it's so powerful it's actually scary, here do you wanna buy some? You and I are unfortunate, tiny passengers whirling around in a PR tornado generated by corporations seeking things much grander than us mere consumers can provide: they want to win national defence contracts, supplant public services, create intractable platform dependence on their products and most crucially, become too big to fail. I think we should avoid doing them any favours if we can!

To conclude this (for now): as useful as agentic AI might be, it is primarily useful as a concept for these corporations and the bi/trillionaires running them as they attempt to reframe themselves. Rather than accept responsibility for the technologies they build and sell, they pitch themselves instead as humble "custodians" of technology, inferring that that they are trying nobly to tame a wild and unfettered force of nature. To perform this act of revisionism they reframe agency: insisting that agency doesn't belong to the human who straps the gun to the AI, but rather, agency belongs to the AI that this human has voluntarily, intentionally, perhaps agentically strapped a gun to.

Nathan