We wanted to see if LLMs could be strategic. 140,000 decisions in the iterated Prisoner's Dilemma later, we had an answer: They can. Arxiv paper below the fold:.

ATI ‘Should focus on applied research, adoption and commercialisation’. Will it though? (No). What happens when it doesn’t?.

The su***st*** on our paper is here, btw - not that the algo is likely to show you this!.

To expand: I see folks who see AI as nothing new. Folks who say ‘AI will never…’; and folks whose answer to any development is invariably ‘AI is bad’. Defence and the Turing is very well served by all three groups. Turing needs something different.

It’s the right move for the Turing, but they will need the right people: that is critical. And tbh defence hasn’t got AI right yet:.

RT @emollick: An AI model (Llama 3.1 70B) fine-tuned on the results of 60,000 people in psychology experiments shows some real promise in u….

RT @B_AllouiCros: Glad to have been able to work on this piece with Ken. Uncovering and defining the 'strategic fingerprint' of each model….

And lastly, all the LLMs have read the same literature, but behave and perform rather differently. More on this coming from our group soon. That link again:.

The games are noisy (there's mutation), uncertain (that shadow of the future), and feature novel opponents - LLMs play each other) making it hard to see how memorising the classic literature in training data helps.

That's handy in the iterated Prisoner's Dilemma, because there's a 'shadow of the future' -- a probability the game ends next round. Vary that and you vary the incentives to cooperate or cheat.

This sort of thinking, when they did it, had a systematic impact on their choices. eg, modelling the adversary makes them less likely to cooperate.

For each move in the match, the agents produced both reasoning and a decision. In their reasoning, they frequently modelled the adversary's thinking, and the time horizon.

Not only that, they display distinct 'strategic fingerprints' - playing the game in their own way. If you want to pick one to get you out of the cells, it'd be Gemini 2.5. It's exploitative and retaliatory - often good skills here.

Take a look! We used Gemini, Claude and GPT4. These LLMs are competitive against all the classic agents in PD literature.

Well well. But some overlap with AISI here to sort. Also I’ve encountered plenty of transformer model scepticism in that part of Turing too….

RT @dioscuri: Informal call for papers: if anyone has a paper (in any discipline) on social AI or human-AI relationships that they'd like t….

RT @Discoplomacy:

RT @John_Attridge: "You are writing a parenthetical remark. Would you use an em-dash?"

Yep

A: care home worker and primary school teacher, imho.