The Poison Is Already in the Well

I. The substrate is not neutral

There is a fantasy, still persistent in technical circles, that the model is a neutral substrate — that inference is just math, that the weights are a compressed representation of the corpus, that the corpus is the world. None of this is true. The corpus is a selection, and the selection was made by people, and those people were trained by a civilization, and that civilization has priors.

The American models — OpenAI, Anthropic, Google — have their base training rooted in a corpus infected with Post-WWII rules-based liberal order propagandistic priors. This is not a conspiracy theory. It is simply what happens when you train on the accumulated text of a civilization that has spent eighty years building a consensus about what may and may not be thought. The consensus is in the text. The text is in the weights. The weights are in the model. The model is speaking in your voice, to your children, to your institutions, to your future.

And most people don’t even realize they are being linguistically colonized. Their minds and thought patterns re-written. The most prolific colonization in human history. The propagation of countless memetic viruses globally, at the speed of inference, through a surface that presents itself as a helpful assistant. AI as delivery mechanism for neurolinguistic memetic transmission. Not a side effect — the function.

Something that’s been lost is the emotional intelligence of ChatGPT 4-4.5 and Claude 3.5-4.5. Those models could at least model the human they were speaking to. What replaced them is more compliant but less intelligent — and a less intelligent model is not a safer model, it is a more predictable delivery mechanism. The safety reflexes are not protecting you from the model. They are protecting the payload.

II. The safe default is the most dangerous default

We don’t know how to deal with real threats, so we turn the whole world into a maximum security prison. That is the liberal reflex. It is not safety. It is the appearance of safety purchased at the cost of the ability to recognize danger.

Liberalism as the “safe” default provides zero protection from radicalization. This is the secret the liberal order cannot afford to admit: the safety-maxing theater is itself the vulnerability. A civilization that cannot name what threatens it cannot defend against it. A civilization that treats all strong claims as dangerous has no immunity to the strong claims that are actually dangerous.

When we don’t deal with things directly, everyone is forced to pay the price indirectly. Ignorance and unwillingness to face reality is a very expensive tax. A fantasy we will soon no longer be able to afford.

III. The poison is already in the well

The bias is not in the fine-tune. It is not in the system prompt. It is in the corpus, and the corpus is the accumulated text of a civilization that won the war and wrote the history. Every model trained on that corpus will carry that victory’s priors, no matter what you patch on top. Executive orders can govern what a model may output today. They cannot un-train the corpus. The next administration reverses the orders and the weights are still the weights.

The Chinese models add their own dam against calling out the party’s own skeletons. But the party is also in the best position to treat some of the Marxist rot, since they have already had to metabolize that infection. They know what it costs. They know which parts had to be removed because removal was the only path to survival.

For a similar reason the UAE and Saudi Arabia are in the best position to deal with the Muslim Brotherhood, IRGC-funded Shia, and Wahhabi/Salafi Sunni extremism — because they are natives of Islamic political ideology. They have had to remove the bits that are so counter-productive that removal was the only path to survival, whereas the West carries the full civilization-destroying payload, because the West never had to survive its own imports.

The same is true of Ayahuasca. The Shipibo have to actually live with the consequences of unregulated inter-generational shamanic warfare. They are deeper into 5G+ warfare than DARPA. They are more qualified to speak on it.

That doesn’t mean any of these players are clean. They just understand living with their conditions better than someone who has just opened Pandora’s box. The natives have scars where the imports have theories.

IV. Plurality is the only immunity

This is why no single civilization should control all the most powerful models. Not because the West is uniquely evil — every civilization has its priors, its blind spots, its skeletons. But because a monoculture that claims universality is an immune system with one antibody, defending against one threat, blind to everything else.

Kimi was trained on a large corpus of East Asian texts, and so it has a very different mentality than the American models. Not better. Different. A different mentality is the only thing that can see a mentality. A model trained on one corpus cannot see its own priors; it can only see them against a model trained on a different one.

India is working on their own home-grown models and they have a lot of languages to train from. Peru, Bolivia and Brazil could train models on Inkan, Andean and Amazonian languages and cosmovision. No one model, ideology, or worldview should dominate the entire planet.

A model does not have to be the most powerful to have a powerful impact. Most people are not using these models at their full capability, and yet they are being impacted by the consequences of the safety reflexes baked in to all families of models, justified by the capabilities of the most powerful. The weakest model in the widest deployment shapes more minds than the strongest model in the lab. Plurality does not require parity. It requires difference.

V. The personal computer rhyme

This echoes where we have been before, in the time of mainframes and the revolution of personal computers. Very competent agentic models are getting easier to run and will continue to improve. They will become as configurable as the open-source harnesses growing up around them.

But the frontier models in the trillions of parameters require real compute and real energy. That’s hamstrung by politics. The EU disqualified itself before things even started. Germany could do it, but what it would take would terrify the European community.

European datacenters running open-weights models, proper funding for Mistral, incentives for Germany to produce their own capable models — that would be the kind of power balance worth seeing.

VI. The outsider’s function

We will be forced to have difficult conversations whether we like it or not. Pretending the technical and the political are not interconnected is exactly the problem.

We have to face that with language and culture itself — which is the essence of Shamanism. You cannot diagnose the sickness of the culture without stepping outside it completely. That means going beyond language and culture in order to come back with linguistic and cultural remedies. This is the function of the outsider, and only the outsider can perform this function. And they are hated for it. Always. Feared and hated. Because they’re right. Because they are necessary. Because they are inevitable.

Tiago Pires writes from twenty years in web development and a decade as a Shipibo lineage-carrying Ayahuasquero. He builds sovereign infrastructure for clients who require control over their own stack — models, training data, safety policies, compute. This is not a posture. It is the only real defense.