Microsoft Researcher Has The Goat-Powered Argument Against Treating LLMs Like People

Microsoft AI researcher Adrian de Wynter built a working neural network inside Age of Empires 2 and used it to argue that people read human qualities into large language models far too easily. He laid out the result in a paper titled "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II," a title chosen to sound ridiculous on purpose. The build and the reasoning behind it were reported by 404 Media.

De Wynter assembled the system in the game's scenario editor, using in-game objects to stand in for computer binaries. Grass represents 0, bridges represent 1, and goats act as the bits moving through the circuit. From those pieces he constructed a functioning NOT AND gate and a 1-bit perceptron, the simplest form of neural network. Videos of the goat-powered machine running sit on his GitHub page, and to anyone watching, the process looks baffling. That confusion is the entire point.

The operations driving the AoE 2 setup are, at the base level, the same class of operations behind tools like ChatGPT and Claude. The difference is the substrate. Because the inputs are goats and grass rather than sentences, nobody watching feels tempted to call the output human. Strip away the natural language and the illusion of a mind goes with it.

De Wynter told 404 Media that he tends to push an idea to its limit when he wants it to land, and that absurdism has a long history in philosophy and theoretical computer science. The choice of Age of Empires 2 was deliberate. Players have built logic and neural networks in Minecraft redstone before, so he wanted a setting nobody associates with computation to make the demonstration stranger and harder to wave off.

His stated worry is methodological. In the document, de Wynter writes that he has peer-reviewed more than 300 computer science papers over the past two years, and that over half of them opened by assuming LLMs carry human-like traits. He frames that as a problem for the research itself, since starting from an unproven premise distorts whatever follows.

"I propose that we need to stop assuming that LLMs behave like humans just because they were trained with natural language. Instead, we should perform experiments that allow us to see LLMs as how they are, not how we believe they should be."

— Adrian de Wynter

I think the goat framing works better than a straight rebuttal would, because it forces the reader to watch the same math without the comfort of words wrapped around it. The case for treating LLMs as conscious leans heavily on tone, and tone is exactly what natural language supplies and grass does not. De Wynter's separation between what makes a model what it is, the relationship between weights under some operation, and what it is perceived as, is the distinction he wants the field to hold onto.

Not everyone working in the space lands in the same place. In January, Anthropic's in-house philosopher Amanda Askell said on the Hard Fork podcast that the question of AI consciousness is not settled, as reported by Business Insider. Askell, who works on shaping Claude's behavior, said she is more inclined to think models might be feeling things, precisely because they are trained on enormous amounts of human writing full of descriptions of emotion and inner experience.

Her reasoning runs through the same training data that de Wynter points to, but draws the opposite conclusion. When a human gets a coding problem wrong, they often voice frustration, and Askell argued it makes sense that a model trained on those conversations would mirror the reaction. She allowed that scientists still cannot say what produces sentience, whether it needs biology, evolution, or something else, and that sufficiently large neural networks might begin to emulate it.

I see the two positions as less contradictory than they first appear, since both rest on the admission that nobody can currently prove what is happening inside these systems. De Wynter's point is narrower than a claim that machines can never feel. He is arguing against importing the assumption before the experiments are done, and his goats are a way of stripping the question back to the operations themselves. Askell, for her part, calls the problem of consciousness genuinely hard rather than solved in either direction.

What de Wynter has built is not a real chatbot, and he does not pretend it is. It is a 1-bit machine made of livestock and terrain, assembled to show that the math people find convincing in ChatGPT is the same math they find absurd when goats run it. The behavior did not change. Only the costume did.

Read also, Microsoft is reportedly preparing to shut down at least two Xbox studios this year, with Kotaku reporting Compulsion Games, the South of Midnight developer that won seven Canadian Game Awards including Game of the Year, among them, around the same time Xbox Game Studios head Craig Duncan left after under two years.