Peter Diamandis introduced the segment on the Moonshots panel with a striking summary. Researchers, he said, had mapped the latent space of the top large language models and found a 98% overlap in reasoning pathways. His reading of the conclusion: "They think the same way. They solve problems the same way. They use the same internal representations."
If that were right, he argued, choosing between one chatbot and another would be mostly cosmetic — you would be picking an interface while talking to "the exact same God model."
That claim, and the argument it set off, is the clearest recent statement of an idea worth naming: AI monoculture. In farming, a monoculture is a field planted with a single variety. It can be efficient, and it can be wiped out by a single disease. The proposal on the table is that the AI industry is building something similar, and that the risk is not only commercial sameness but shared blind spots.
Three claims that get bundled together
The debate becomes much clearer once three separate claims are pulled apart.
The first is output homogeneity: different models give very similar answers to the same open-ended question. This is measurable from the outside, without knowing anything about a model's internals.
The second is shared internal representations: the models encode concepts in the same internal structure — the "latent space," meaning the numerical space in which a model represents meaning before turning it back into words. This is a much stronger claim, and harder to test.
The third is operational monoculture: whatever the models are like internally, the way they get deployed concentrates risk. Running five thousand copies of one model is cheaper than running five thousand different ones.
The paper the panel was discussing addresses the first claim.
What the Artificial Hivemind paper measured
Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond), by Liwei Jiang and colleagues and published in October 2025, studies repetition and semantic similarity in what models write. Its dataset, Infinity-Chat, holds roughly 26,000 open-ended English queries drawn from real user conversations. The homogeneity experiment uses 100 representative prompts, collecting 50 responses per model per prompt under each decoding setting.
Similarity is computed as the cosine similarity between sentence embeddings of the generated answers — a numerical comparison of what the answers mean, not an inspection of a model's internal reasoning. The cross-model averages reported in the main paper run from 0.71 to 0.82, not 98%.
The texture of the result is more interesting than the number. Asked for a metaphor for time, models from different laboratories cluster on the same few images: time as a river, time as a weaver. The dataset also includes 31,250 human annotations, and the authors find that model-based evaluators handle divergent individual preferences poorly — when people genuinely disagree about which answer is better, automated judges do not track that disagreement well.
The authors propose shared training sources and synthetic data as explanations but leave the causal investigation open. They also mark the limits: the prompts are English-only, and embedding similarity is a blunt instrument. Their stated purpose is a benchmark for diagnosing and addressing what they call generative mode collapse — the narrowing of a model's range of plausible outputs.
So the paper supports a finding about answers. It does not establish that two models arrive at those answers by the same internal route.
Why the models might be converging
Diamandis offered the training-data explanation: synthetic data and models learning from each other's outputs. "The training data has become a shared bloodstream," he said. "Every model drinks from each other, and the result is convergence towards a single reasoning architecture."
Alex proposed a different cause. All of these systems were trained on the same reality, he said — the same universe, the same version of humanity in their pre-training material. Convergence would follow from that alone. He pushed the point further, citing studies that correlated the hidden activations of GPT-2 with fMRI measurements from human brains. Not only are the models correlated with each other, in his account; they are correlated with us. "We're all in the same universe."
Mostaque agreed that the result should not surprise anyone. There is not much difference in data between the big laboratories, he said; some train a little longer, some use slightly different reinforcement learning. He added a limit on how much originality to expect: the models are not yet doing "crazy original stuff." More original behavior, in his view, came from earlier systems like AlphaGo, which had far less existing data to model off. And convergence is partly what users want: "we should be shocked if they aren't the same, because we want them to have similar outputs for similar inputs in almost all cases."
The claim about internals
The closest the discussion came to evidence about shared representations was Dave's account of gauge rotation. Historically, he said, researchers have struggled to take a finished model and extend it, because the representations between layers sit at a rotation in vector space unique to that model. Try to map one model onto another and the internal coordinates do not line up; the raw parameters look unrelated. Researchers have now worked out how to rotate those coordinates without destroying the models, he said, which makes it possible to say of two systems: "it's the same thought. It just doesn't look the same because it's rotated in space."
His interest was partly practical — the same technique would let a laboratory build on a past billion-dollar training run instead of starting from scratch. It was an assertion made in conversation, without a paper attached, but it names the thing that would need to be true for the strong version of the hivemind claim to hold.
The objection: nature diversifies
Salim Ismail took the contrary position. As nature evolves, he said, you get more diversification and more species, not less. Convergence now may be a transient phase rather than an end state: the models are training and distilling from each other today, but over time he expects different models doing different things.
Alex argued both sides of that. Early life looked alike, he said; so did early cars and early websites, and then specialization exploded. But from an evolutionary-development perspective, the post-Cambrian record does not show infinite variety. Millions of species cluster into a few dozen body plans. He does not believe there are countless ways to model reality or build a body. His bet is that there is a single best AI architecture that may present as perhaps 35 superficially different "AI body plans" — variants that look distinct but are hidden symmetries of a common underlying design.
Ismail's reply went to the consequences rather than the count. "Nature hates monocultures," he said. "One disease will wipe out a total monoculture. One bad assumption will wipe out a monoculture of ideas." If all the systems reason alike, they share blind spots. The single-model approach, he said, "is too brittle."
Why resilience is the operative worry
The episode's next story gave the abstraction a concrete edge. Researchers have demonstrated self-propagating ideas in multi-agent systems: natural-language payloads that persuade one agent to adopt an idea, keep it in an editable memory file and pass it to another. Mostaque connected it directly to the monoculture argument — if you have a monoculture, the viruses spread rapidly.
Dave supplied the deployment reason it matters. When you launch a fleet of agents, it is more efficient to run the same model five thousand times than to maintain five thousand differentiated ones. "If it's convincing to one agent, it's convincing to all 5,000."
That study's own account is more measured. Susceptibility varied substantially by model and by the structure of the agent network; benign advocacy generally spread more readily than harmful goals; idle agents were more vulnerable; and a simple warning largely blocked transmission. The starting conditions were artificial, the prompts editable, the interactions short. The authors describe the threat as real but presently brittle, costly to construct and comparatively easy to defend against. Operational monoculture is a reason the question matters, not a demonstration that diversity would have prevented anything.
The proposed remedy: build the difference in early
Mostaque's objection to the evolutionary optimism was that evolution is not what is happening. "We're battery farming the AIs," he said — breeding them into "little chihuahuas that are very smart." The models are not out in nature adapting dynamically; they are trained in one direction, and, he added, with one Silicon Valley-type mindset.
His proposal follows from that. Train a model from the start with morality and ethics inside it, drawing on a diversity of cultures, and the resulting latent space should differ far more than it would if those values were applied afterwards. The distinction is between pre-training — the long initial phase where a model absorbs its general picture of the world — and post-training, the later adjustment of a model's behavior. Building diversity in at the first stage, he argued, is how you get resilient systems rather than one latent space that can catch a mind virus, especially as models move into the world as agents and humanoid robots.
The conversation ran that idea into Neal Stephenson's novel The Diamond Age. The current framing of sovereign AI is national — Saudi Arabia has its system, London has its own. The panel raised a different cut: groups of like-minded people holding their own AI across borders, chosen because it maps to their view of the world.
Alex thought both could be true simultaneously. Everyone might feel they have their own private culture and their own little sovereign AI while underneath it is one common algorithm, with everyone claiming credit. Ismail offered the genetic analogy — we all look different, but the ingredients are the same four DNA bases. Alex went further: relative to other species, human genetic diversity is minimal. His prediction is that a few years from now we will pat ourselves on the back for AI diversity and not actually have much.
Open questions
Does output similarity imply anything about internals? The paper measures answers. Dave's gauge-rotation account is about representations. Nothing presented on the show closes that gap, and the paper's authors explicitly leave the causes open.
Would culturally varied pre-training produce underlying diversity, or only a different surface? This is Mostaque's proposal against Alex's prediction, and it is the crux. If different interfaces and cultural identities sit on a common architecture trained on the same human record, a shared blind spot survives the rebranding.
How much diversity buys how much resilience? Nobody has shown that a mixed population of models resists a bad idea in proportion to how different its members are. The mind-virus work found that model identity, network structure and a plain warning all mattered; diversity was not isolated as the protective factor.
Where does the value go if convergence holds? Diamandis argued that differentiation moves to the interface, the harness, the ecosystem, the safety layer, price and deployment speed, and Ismail pushed entrepreneurs toward the application layer. Alex was not convinced it all goes there — "there's a lot of value in the infra underneath as well" — and Mostaque pointed to sovereign AI as the opening. That business argument is developed separately on the idea page The inner loop: why AI advantage may sit outside the model; repeated claims that intelligence is becoming a commodity do not change it.
The panel expected the question to sharpen quickly. Ismail predicted that within three to six months the audience would stop finding the debate abstract and start asking which group they are in.