Horses for Courses

How to ride to the AI Frontier

I fell off a horse called Cuddles when I was six and took this as a cosmic sign that I should never attempt equestrian pursuits again. So when I say that prompting AI models is like riding a horse, I’m speaking entirely metaphorically—though after two years of wrangling LLMs at Meta, I suspect the metaphorical and literal experiences share more DNA than I initially imagined.

We’re living through a peculiar moment in creative tool evolution, one that reminds me of the early internet’s beautiful chaos before the platform oligarchy smoothed everything into algorithmic conformity. Remember when every website demanded you learn its particular grammar? When games like Myst forced you to master arcane interaction rituals, or when early MUDs required you to type “get lamp” and “go north” with the precision of ritual incantation? The friction wasn’t inefficiency—it was where meaning lived.

Each digital space had its own logic, its own aesthetic sensibility, its own particular demands on your attention and imagination. You couldn’t just tap and swipe your way through everything. The interaction model was part of the narrative, a visceral component that made you complicit in the story being told. Learning to navigate these spaces created the kind of personal stakes that ease-of-use design explicitly eliminates.

Then the platforms conquered everything. Facebook flattened human interaction into infinite scroll. Google reduced information discovery to a single search paradigm. Mobile interfaces converged around the same swipe-tap-pinch vocabulary until every app became a variation on the same gestural theme. We gained unprecedented convenience and lost the beautiful friction where creativity used to live.

Today’s AI landscape feels similarly diverse and similarly precarious. Different models embody different computational aesthetics, different approaches to meaning-making, different biases about what constitutes compelling output. But the industry is racing toward standardisation, building guardrails and guidelines to make outputs more reliable, more predictable, more algorithmically convergent.

Don’t get me wrong. I’m totally in favour of using AI powered tools to make media development, production and discoverability processes more efficient, freeing up time and resources for engagement with ideas and narratives and how best to contextualise them. But I do fear the over-homogenisation of outputs and expectations.

Which is precisely why we should lean into the weird while we still can.

The Familiar Path Home

Most people approach LLMs like they’re driving a car. Turn the key, step on the gas, steer where you want to go. It’s a mechanical view that assumes control, consistency, and linear progression from prompt to output. But anyone who’s spent serious time with these models knows this is rubbish. LLMs are far more like horses than automobiles—each with its own gait, its own preferred trails, its own particular ways of responding to guidance.

When I was building prompt judges at Meta—automated systems that evaluate whether AI responses meet specific quality criteria—I noticed something revealing. Different models might achieve similar scores on our benchmarks, but their reasoning processes were often completely divergent. It was like watching different riders navigate the same course: they’d all (mostly) clear the jumps, but their approaches revealed entirely different training philosophies.

A prompt judge, for those not neck-deep in AI development, is essentially a meta-AI that evaluates other AI outputs. Think of it as an automated critic trained to recognise good storytelling, compelling character development, or conversational naturalness. Building judges reinforced something crucial I’d repeatedly observed in my genAI prompting: each LLM has deeply grooved pathways—linguistic patterns and conceptual routes that feel like home territory.

These pathways manifest as templated language that anyone can spot: the numbered lists, the executive summary prose, the particular cadences that feel algorithmically generated rather than genuinely authored. Like Waymo vehicles in San Francisco, which famously form impromptu convoys along their favourite streets, AI models develop preferred routes through conceptual space.

You have to push them out of their comfort zone to generate something new. Create a dialectic, ideas in opposition. The models know what a plumber might logically advise. They’re trained on the tropes of astronaut-speak. But ask them to respond like a space plumber, say, and suddenly you’re in uncharted territory.

The Editorial Frontier

This approach of dialectic between ideas and inversion of expectation was the central innovation of my show “Liquid Television”. When we created LTV for MTV, the original pitch was: put your favourite TV shows in the blender and hit purée. We weren’t just showcasing different animation techniques—we were creating a space for experimental forms that couldn’t exist elsewhere through a context that valued dissonance. We were bringing the audience in on the experiment. The show worked because we actively curated against the familiar, against the templates and formulas that dominated mainstream television.

The current AI moment offers similar possibilities, but only if we resist the gravitational pull toward standardisation. Every model is being trained toward predictability and safety, creating computational monoculture. The interesting creative territory lies in the conceptual gaps, the combinations that don’t have preset responses, the frontier between what the algorithm expects and what it encounters.

This requires intentional prompting toward unexplored territory. You can ride the familiar route home—asking AI to generate the kind of content humans already make—or you can head off trail to the frontier and invent new kinds of storytelling enabled by these particular computational capabilities.

Polyamory as Creative Resistance

Which brings me to my recommendation for creative polyamory with AI systems. Not the shallow promiscuity of model-hopping, but thoughtful relationship-building with multiple partners, each offering different kinds of creative intimacy and, crucially, different blind spots.

The advantage isn’t just having access to different capabilities, but in the creative friction that emerges when you explore the same problem through different computational lenses. Each model brings its own biases, its own unexpected insights, its own way of getting lost in interesting directions. The conversation between these perspectives often produces more interesting results than any single interaction.

More importantly, this multiplicity serves as resistance against the inevitable convergence. Models get updated, deprecated, or transformed in ways that can dramatically alter their behaviour. Having deep relationships with several systems means you’re not entirely dependent on any one approach to creative partnership—and you’re actively preserving computational diversity before it gets standardised away.

I think there are two equally valuable ways to use AI right now in creative practice. One is about control, using AI tools to make processes more efficient and impactful within the framework of current workflows and outputs. This is critical work for a media industry struggling with current funding models, but is tough to implement given the equine nature of the underlying models, and the sense that they retain a ‘mind’ of their own. Then, there’s the search for new kinds of creative collaboration, formats we have not yet imagined, content we can’t yet describe, interaction models and audience engagement processes we are about to invent.

I scraped the surface of this exploration in my recent post on collaboration with Claude on “PROMPT”—a meta-fictional story told entirely through video generation prompts. The piece works because it occupies territory that didn’t exist before AI: it’s simultaneously about AI video generation, created with AI collaboration, and structured as AI prompts. It wouldn’t have been made with previous tools, and it reveals possibilities for narrative that emerge specifically from this technological moment.

Different Horses, Different Territories

The art lies in matching computational temperament to creative challenge. Some models excel at analytical depth and nuanced reasoning—perfect for unpacking complex narrative structures or exploring philosophical implications. Others shine with rapid-fire ideation or stylistic mimicry. Still others have surprising capabilities in unexpected domains that only emerge through experimentation.

I’ve developed what I think of as model-specific dialects—different prompting approaches that play to each system’s strengths while pushing against its comfort zones. With some models, I tend toward structured, contextual prompts that lay out creative challenges systematically. With others, I’m more oblique, using metaphor and contradictory direction to guide them away from their default responses.

But the deeper principle is horses for courses: you wouldn’t attempt the Grand National on a Shetland pony, nor would you take a thoroughbred racehorse up a mountain trail. Different creative challenges require different computational capabilities, and the current diversity of AI models offers a remarkable stable of different approaches to creative intelligence.

Is Computational Specificity the answer?

We’re in a precious window where different AI systems have genuinely distinct personalities and capabilities, but the entire industry is racing toward artificial general intelligence with a ‘one ring to rule them all’ mentality.

This is precisely backwards. The smartest guy in the room isn’t always the best at everything, or indeed anything. That part will be down to the rest of us.

It’s tempting to imagine computational specificity as the answer—models trained exclusively on experimental literature or underground comics or non-Western narrative traditions, each becoming a dedicated specialist in particular aesthetic territories. These would certainly support current creative practice, offering sophisticated access to established forms and cultural knowledge, and will be excellent tools for clear use cases. But like having an expert tour guide through well-mapped terrain, such models would excel at reproducing what we already know those traditions should look like, which is precisely the kind of templated thinking that forecloses genuine discovery. The most interesting creative territory lies not in better access to existing aesthetic landscapes, but in the unmapped spaces between them, where familiar categories break down and models can’t rely on their training to know what comes next. This is the place where the human artist takes the reins and discovers new paths with AI.

Before the Trails Get Paved

I suspect the current diversity will eventually diminish as the technology matures—the same way kids’ games evolved from the glorious ecosystem of Flash developers, small agencies and bedroom coders crafting entirely idiosyncratic experiences to the current reality where virtually every young creator and game studio builds within Roblox’s templated universe, trading creative sovereignty for platform reach, or the way the auteur cinema of the 60s and 70s lost out to the movie franchise model from the 80s on. The path toward computational monoculture feels almost inevitable, driven by the same market forces that consolidate every creative industry.

But for now, we have both computational diversity and the possibility of computational specificity. The question isn’t which AI is “best”—it’s how to use this multiplicity to explore creative territory that wouldn’t be accessible through any single approach, and how to preserve space for the development of models that can take us to genuinely unique places rather than just more efficient versions of familiar destinations.

The frontier is where we’ll find the new formats, new stories, and new ideas that AI can enable. Don’t just use these tools to do the same old thing with marginal efficiency gains. Play and see what you can do that you’ve never done before, enabled by these particular computational capabilities.

Ask yourself: what would a space plumber have to say? What stories exist in the gaps between familiar categories? What happens when you prompt for the intersection of concepts that don’t traditionally intersect?

The models are being trained toward predictability, but creativity lives in the unexpected. Lean into the weird while the weird is still possible. The trails are being paved as we speak.

I still have no intention of getting back on an actual horse. Some metaphors are best left safely metaphorical, particularly when you remember that Cuddles was allegedly a gentle children’s pony who somehow managed to deposit a six-year-old in a hedge with the efficiency of a medieval trebuchet.

Instead, I’m saddling up and heading back to the AI Frontier. I hope to see you all along the trail.

Next time: The most impactful class I took in film school was actually in the psychology department. where I learned about “Regression in the Service of the Ego”. In my next post, I explain what the f*** that means and why it matters even more in the GenAI age.