Summary
- A researcher rebuilt an LLM using goats and NAND gates inside Age of Empires II, proving LLMs can be reimplemented.
- Humanlike tone is based on presentation; anthropomorphism doesn't prove sentience.
- Persuasiveness and self-consistency can be measured, yet don't imply real or simulated behaviour.
What does it mean when something is 'sentient'? It's a debate that has raged for years, and it's not something I can really squeeze into this article. What I can say, however, is that the creation and distribution of LLMs have added another area of discourse to the pot. When you ask Gemini a question about how to cook a chicken, or you sic Claude on your coding project, are you asking a program, or a person?
Well, one researcher has an answer for you. They've recreated how an LLM works using nothing but goats in the game Age of Empires II, and thus pose the challenge: if you think an LLM is sentient, then so is the videogame.
A researcher reduced an LLM to goats in Age of Empires II
And makes some compelling arguments against calling one 'sentient'
As spotted by 404 Media, researcher Adrian de Wynter of Microsoft and The University of York had an axe to grind. People were interfacing with LLMs for the first time in human history, and because of how the LLM speaks and interacts with them, they naturally anthropomorphise it. Now the LLM is no longer a system in people's minds; they're almost like a person, and, as such, aren't as understood as they should be.
Want to stay in the loop with the latest in AI? The XDA AI Insider newsletter drops weekly with deep dives, tool recommendations, and hands-on coverage you won't find anywhere else on the site. Subscribe by modifying your newsletter preferences!
In his paper, "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II," he aims to strip away the anthropomorphic layer of LLMs by showing how he made an LLM using Age of Empires II's scenario editor to create NAND gates using goats. With this, Wynter then argues that if you can build an LLM using assets in a video game (or Lego bricks, or even people in Greater Boston working together), then an LLM isn't human-like by default. Instead, it's part of the presentation of the LLM.
It's easy for someone to look at a chat window, see an LLM speaking to them in a human tone, and anthropomorphize it. But as soon as you remove that layer and all you're left with is looking at goats acting as NAND gates, that illusion falls apart:
We argue that many anthropomorphic measurements in AI are measurements of presentation, rather than of an actual system’s behaviour. Moreover, these measurements are irrespective of their quality of being real. Indeed, attributes such as persuasiveness and self-consistency are objectively measurable, but from our work it follows that these cannot imply real (or simulative) behaviour under this setup.
If you want to read more about how the goats work and watch them scuttle around, be sure to head over to Adrian de Wynter's blog for all the details.
Only one is most likely not to lose me any customers