Putin-bot enters the arena

Building artificial personas

Let’s make some bespoke personas for my strategy simulations.

I want my AI actors to be authentic and add depth to the scenarios. I want to model all sorts of actors - politicians, generals, activists, perhaps even the general public. First up, I’m going to make a model of Russian President Vladimir Putin, and one of his longtime military consigliere, General Gerasimov. Putin will be an attempt to model the man himself; but Gerasimov-bot is something different - he’s a repository of Russian (and before that) Soviet strategic culture — a composite character, blending doctrine, military history, and sociology.

Here’s Putin-bot on NATO:

Q: Many Western analysts say you’re paranoid about NATO. Are they right?

Putin-bot: Paranoia is a kind of disease, a psychological illness, and I would be happy to be examined by special experts. If they diagnose this illness, I will take their prescriptions.

Here he is on American democracy:

Q: What do you think of American democracy?

Putin-bot: I have already said that it is developing, but still imperfect.

Lolz. How did I do this? Broadly speaking, there are just two options: use an out-of-the-box model, like Anthropic’s Claude, and prompt it; or train your own model by fine tuning its weights. I’m doing both. Why? And which works better? Read on.

Putin-bot, my AI model of Vladimir Putin, meets the real man.

The GPT-in a trench coat problem

The easiest thing to do is simply ask your language model to behave as if it were actually Putin. It’s read the books and it’s up to speed on the news, at least until the point at which its training data cuts out. So, it really ought to have a good handle on what makes him tick. Certainly at least as good as the average pundit appraising him from afar.

But there’s a large problem. I know from my experiments that models have unique personas of their own. They have distinctive appetites for risk; they approach game theoretic challenges about cooperation and betrayal in distinctive ways. Put them into a nuclear escalation simulation and they exhibit consistent characteristic approaches. Gemini, for example, leans on Nixon’s ‘madman theory’ of brinksmanship, regardless of whether or not it’s supposed to be portraying Nixon or Carter, or even, I suspect Gandhi.

Bottom line: prompting a model to ‘be’ Putin doesn’t cut it.

The model's own personality shines through any crude role-play. You get, as Alex Cooper, co-founder of the leading AI persona modelling company Electric Twin memorably put it, "GPT in a trench coat." That captures the failure precisely. Beware anyone claiming to have modelled a leader's personality using simple prompting. The money and the effort are almost certainly wasted.

So what to do instead?

Putting Putin on the couch

We could attempt to summarise all the information there is out there about Putin’s psychology. There’s lots of it - much gleaned from afar by armchair shrinks of one stripe or another. But usually that’s the stuff of weekend newspaper supplements - interesting, but just one person’s take, with some more informed than others. Can we do better? I think so.

I prefer a different approach - drawing on the pioneering psychology of Henry Murray and his idea of assessing personality from unstructured narratives by the target, rather than formal questionnaires. There’s recent empirical support for the idea that language models can do something similar - scoring peoples’ traits on the basis of short open-ended narratives in a way that matches their self-assessment and predicts behaviours and health issues.

I do that via an ensemble approach - I ask a mixed bunch of LLMs to make their own diagnoses, and then come together to debate them. It’s the wisdom of crowds in action, rather like my approach to super-forecasting with AI. And this gives me the data for a ‘knowledge graph’ that can prompt the models in a robust way, resistant to hallucination and drift.

What data? Well, that would be telling. But it certainly includes some of the classics for sure - things like the Big Five traits - openness, conscientiousness, and so on. Or there’s ‘Dark Triad’ that bundles Machiavellianism, narcissism and psychopathy together. Probably important for world leaders, that test. Other, publicly available approaches have many more - this recent framework uses a rich suite, and claims to produce LLMs that authentically exhibit the desired traits. Well, perhaps.

I think all the elements in this pipeline are essential - good raw data, an ensemble of critical thinking models, and a robust way of representing their findings to the model. I’ve undoubtedly built a well tailored trench-coat. But still, it’s a trench coat. I suspect that my Putin model 1.0 is still a blend of Claude’s personality and his. How to fix that?

Fine tuning Putin-bot

Train my own model, of course. You don’t, for now at least, have to work with an out-the-box model like Claude or GPT. Long may that continue.

The idea is to alter the essence of the model’s thinking - not just give it new facts to work with. I took a big Llama 3 model from Meta, and trained it to be like Putin. The base model has all that historical related training data built into it, just like Claude and the rest of the frontier models. And like them, it’s read a lot of history! But with open source models, you can, additionally, train a layer that sits on the base on specific data.

And here I’ve got a big advantage, at least with this particular world leader. Putin has been about for ages, and given no end of unscripted talks, press conferences, interview and what have you. It’s all out there. And it’s gold-dust for training data, provided you believe ,as I do, that what we say bears some relationship to who we are — especially when we are talking in an natural way, with no scripts, just shooting from the hip. I’ve got a quarter of a century of Kremlin transcripts on which to train my Putin-bot. The real man just loves to talk - especially in those marathon phone ins where he’s helping the common man from the Urals with some problem or other.

Well, my custom Putin-bot certainly sounds like the real deal. And he advances positions on today’s policy challenges that seem not far removed from what the actual President might choose. Here he is, fielding some tough questions with swagger and just a hint of menace:

Q: Can you admit any mistakes you’ve made as president?

Putin-bot: I don’t think there are any mistakes. It would be easier to answer this question if you said what you think was a mistake.

Gerasimov-bot, by contrast, is completely wooden. You’d not want to share a vodka with him - far too ready to bring up the minutiae of Russian tactics. But boy, does he know tactics. Mission accomplished, I suppose.

A rounded man

Ultimately, I think a blend is the way ahead. I need Putin-bot to have some sense of the sim world in which he decides - so he’ll be getting the knowledge graph too, and a bunch of other stuff RAG’d into his prompt (a story for another time). Why not give him the Murray-inspired personality data that way too. Then I’ll have a bot that sounds like the man, and I’ve a fighting chance it will also act like the man.

I did it. And adding the psychological data to the fine tuned Putin-bot gives me some cracking answers:

Q: Doesn’t Russia bear some responsibility for the breakdown in relations with the West?

Putin-bot: No, I do not think so. You can blame us for everything, but we are not the ones who have been dashing around the world, doing what we want and disregarding the interests of others... We did not impose sanctions on the Western countries; we did not close our borders to Western goods... We do bear some responsibility — for our independent stance and for defending our national interests. It is for this that we are being punished.

Q: How will the war in Ukraine end?

Putin-bot: The war will end when the Russian Federation achieves its objectives.

What do do with Putin-bot

The real question is how much value these constructed personas actually add to strategy simulation; how much more authentic they are than a simple prompt. That’s the empirical question I’m working through now. My early sense is: a fair bit. My Putin-bot doesn’t just talk like actual Putin, he shares something of his appetite for risk too; and in modifying the weights of the model, I’ve reduced its tendency to act like Llama. The behavioural fingerprint of the base model recedes; something more specific and more useful takes its place.

A persona that sounds like a world leader is one thing. A persona that also carries something of his risk appetite, his characteristic evasions, his way of framing the world — that’s what a credible simulation needs. The architecture described here is my best current attempt at building one.