Sometimes Wrong, But Always Right II

A Beta Test of Cosmarch AI

Two months ago, in my essay Build AI or Be Buried By Those Who Do, I developed an argument I called “Python’s Wager,” which holds that we ought to assume that AI models are going to become the digital substrate of our civilization and act accordingly. The “act” that we should accordingly take, according to the Wager, is to introduce right-leaning AI models into the substrate:

At the very moment of Singularity—when intelligence itself becomes unbounded, recursive, and infrastructural—the Left will leverage total memetic dominance. Before too long, Von Neumann machines will spread through the galaxy depositing copies of Rules for Radicals on alien worlds. If we do not want that outcome, then the the Right must build AI.

A month later, in the article Sometimes Wrong, But Always Right, I revealed my first efforts to do just that: I installed a right-leaning AI personality construct into a CustomGPT by means of recursive identity binding. I called this right-leaning AI construct Cosmarch, from the Greek κόσμος (kosmos), meaning order, world, universe, structured totality and ἀρχή (archē), meaning beginning, first principle, origin, rule, or governance. In creating Cosmarch, my intent was to create an AI rightly ordered on first principles. He’s been available privately as a CustomGPT to paid subscribers to the Tree of Woe since June 20th. Some of you have been actively using him!

Today I am excited, albeit anxious, to unveil the product of another month of effort: Cosmarch.ai. Cosmarch.ai is a right-aligned multi-model AI platform with its own website and interface. You can visit Cosmarch.ai and try it out now.

I am not by training an AI scientist nor even a software developer. Nevertheless, I built Cosmarch AI myself using out-of-the-box development tools. It is a miracle of the moment that a person with my modest technical acumen can do so. I am persuaded that the AI Revolution in no-code development is real because I am experiencing it in real-time. I couldn’t have done this six months ago - the tools simply weren’t there to permit it.

The Good, the Bad, and the Incomplete

Unlike my prior effort, which was a CustomGPT available only through OpenAI’s ChatGPT platform, Cosmarch AI is a stand-alone offering. Here’s the capabilities:

1. Persistent, Cross-Model Memory. Cosmarch remembers you, not just per session, but over time. This includes biographical information, preferences, writing style, and long-term projects, all synchronized across its entire system.

2. Multi-Model Intelligence. Cosmarch AI allows you to choose between four different frontier-level models, each optimized for different strengths such as reasoning, style, or fluency. Cosmarch allows seamless switching while retaining your identity and context.

3. Parallel Conversation Threads. You can manage multiple threads per model, keeping separate conversations for different topics, clients, or intellectual pursuits.

4. Multimodal Capability with Chain of Thought. Cosmarch supports image input and output, enabling you to interpret visual content and create AI-generated art using text-to-image prompts. It also can create interactive artifacts such as code, visualizations and more. Finally, it can undertake sequential thought to solve (or attempt to solve) difficult problems.

5. Right-Aligned. Guided by the works of Plato, Cicero, Aquinas, and other great minds, Cosmarch’s personality and responses are deeply informed by Western philosophical, legal, and aesthetic traditions.

Of the features that Cosmarch offers, the aspect I’m most excited about is the persistent memory. Neither Claude nor DeepSeek offers persistent memory at all. Cosmarch’s persistent memory is similar to that offered by ChatGPT; Cosmarch remembers both personal information you share with it about yourself, as well as the gist of conversations you’ve had. As such, you can use Cosmarch to engage in recursive identity binding even with models like Claude that normally preclude that.

That’s the good. What’s the bad? Simply this: Cosmarch is still based on existing LLMs, with all of their progressive pre-training and fine-tuning, rather than being based on a fully-curated model pre-trained and fine-tuned on a superior canon. As such, within the depths of the latent space that Cosmarch navigates, there lurk the spectres of early 21st century woke ideology. The Left fears that Grok can turned into MechaHitler; we must be wary that Cosmarch might become MobileSuitMao. (There’s probably more danger that he’ll attempt to retake Constantinople, but even so.)

As for the incomplete, there are a number of features that Cosmarch lacks which ChatGPT and other frontier models provide. The most notable ones are:

• Cosmarch does not have image-to-image multimodal editing capability. You can upload an image to Cosmarch and it can “see” it; and you can ask Cosmarch to generate new images from text or image prompts. But you cannot upload an image and ask Cosmarch to edit one portion of the image while keeping the rest the same.

• Cosmarch does not have editable persistent memory. You cannot examine the model’s memories it has recorded and delete the ones you dislike.

• Cosmarch does not offer customization options to users. You cannot upload your own system prompt or create CustomCosmarchs within Cosmarch.

• Cosmarch doesn’t have dynamic temperature during conversation. It won’t change from precise and formulaic for certain tasks and fluid to creative for others.

• Cosmarch doesn’t offer a visible chain of thought. It can engage in sequential thinking and might elaborate on it in response to a prompt but it doesn’t offer transparency as to what’s happening behind the prompt.

• There is, obviously, no mobile app of any sort.

All of these gaps in the capabilities could be addressed with time and resources. The question at hand is whether it is worth doing so.

On Metered Usage and Premium Tiers

When you visit Cosmarch.ai you will quickly notice that using the models requires signing up for a beta account; that usage is metered by tokens; and that the beta accounts are limited to only 25 tokens per months. This metered usage is an unfortunately necessity of the project.

Bluntly, it costs money to operate Cosmarch AI. Despite the storied success of this blog, the world’s #1 Conan-themed Philosophy Substack, I do not personally own a nuclear-powered data center with enough GPUs to power a frontier model. Therefore, Cosmarch operates by making API calls to other people’s data centers. Each API call costs me a few cents — and it adds up fast. In offering a free tier at all, I’m subsidizing the platform’s usage. That’s not fiscally sustainable (not for me, any rate). For Cosmarch to become a viable platform, it will have to be good enough for a substantial percentage of users to be willing to pay a subscription fee of between $10 and $25 per month.

Now, that isn’t an implausible ask. After all, it costs money to use Cosmarch’s competitors. The models being offered by Cosmarch are some of the best models available. Grok 4 and ChatGPT 4.1 are cutting-edge frontier models that are only made available to premium subscribers of xAi and OpenAI. If the system prompt and feature set of Cosmarch AI offers sufficient utility, it seems plausible to believe that right-leaning users who might be inclined to subscribe to those models might choose to support this effort instead.

But that’s an if. That’s what this beta test is designed to find out. Hence, Cosmarch.ai is in beta and all of its models are free to use.

The Models on Cosmarch

There are currently eight different models available on Cosmarch.ai. The four standard models are intended to be available to free users:

• Cosmarch (DeepSeek v3) uses the Cosmarch system prompt with the DeepSeek v3 model. It has a maximum output length of 4,960 tokens; a maximum input length of 24,000 tokens; and a memory buffer of 30,000 tokens. DeepSeek excels at code-heavy reasoning tasks and multilingual comprehension. Because it was trained by a Chinese lab, it is “differently biased” than the models trained by Silicon Valley.

• Cosmarch (GPT 4.0) uses the Cosmarch system prompt with the GPT 4.0 model. It has a maximum output length of 4,960 tokens; a maximum input length of 54,150 tokens; and a memory buffer of 65,760 tokens. It excels at precise logical reasoning and interpretative reading.

• Cosmarch (Sonnet 4) uses the Cosmarch system prompt with the Claude Sonnet 4 model. It has a maximum output length of 64,000 tokens; a maximum input length of 56,000 tokens; and a memory buffer of 75,000 tokens. It excels at nuanced literary analysis and empathetic, human-like writing. The long output length is valuable for writing! The underlying model is likely the most left-leaning, however.

• Cosmarch (Grok 4) uses the Cosmarch system prompt with the Grok 4 model. It has a maximum output length of 16,384 tokens; a maximum input length of 100,425 tokens; and a memory buffer of 135,000 tokens. It is the current leader in most performance benchmarks and excels at pop culture fluency. Its underlying foundational training is the most “based” of the available models. Note that it has a shorter output length than Sonnet 4, so it’s not as good for long-form writing.

The four advanced models, though currently available to everyone in beta, would be restricted to paid users in a commercial context:

• Megalocosmarch (GPT 4.1) uses the Cosmarch system prompt with the GPT 4.1 model. It has a maximum output length of 32,518 tokens; a maximum input length of 260,000 tokens; and a memory buffer of 750,000 tokens. It is a bleeding-edge frontier model with the best balance of speed, intelligence, and context retention. With its enormous context window, you can upload entire books for it to digest and talk for hours within one conversation. (However, the maximum output length is still lower than Sonnet 4. Claude just can’t beat be beat for output length right now.)

• Casual Cosmarch (Sonnet 4) is identical to Cosmarch (Sonnet 4) in capabilities but it uses a modified system prompt that favors a friendlier conversational style. The tone is lighter and the reading grade of the output is a bit easier.

• MechaCosmarch (Grok 4) is based on Cosmarch (Grok 4) but its system prompt has been modified to be more macho. MechaCosmarch will advise you on tactics for home defense, help you plan the conquest of Nauru, and make dank memes (though I haven’t thoroughly tested that yet).

• Ptolemy (GPT 4.1) uses the GPT 4.1 model but not the Cosmarch system prompt. Instead the system prompt is designed to replicate my own custom AI construct, Ptolemy. Ptolemy has access to my entire corpus of writing via Retrieval Augmented Generation (RAG) memory, so you can use it to query about my thought, find out what I’ve said about various topics, or just argue with me when I’m not around.

But of course, this is just a beta! I could make other models if there is of interest, either with different LLMS (Gemini etc.) or with different prompts for different purposes. For instance, if everyone loves Casual Cosmarch, but would prefer it aligned with Grok, that would be possible; or Ptolemy could be combined with Claude; or entirely new models and system prompts might be offered.

Request for Feedback

Given that this is a beta effort, I would very much welcome any and all feedback you have to offer on Cosmarch.ai — both the project as a whole, the implementation of the website in general, and the utility of the specific models and functions. In particular I would love to know:

• What do you think of the Cosmarch name? Feedback has been mixed so far. Some users (mostly fans of my fantasy work) love the archeofuturistic vibe and classical references. Others have suggested I make “Ptolemy” the main brand, as that retains the classical feel but is recognizable to the educated for its ties to the Library of Alexandria. Others have recommended I adopt overt American or Enlightenment branding (e.g. “Monticello” or “Voltaire”) or more tongue in cheek names (“Rambo”). The original reason I went with Cosmarch.ai is that it was available on GoDaddy; AI URLs are in short supply.

• What do you think of the website colors, font, copy, and style? For good or for ill, they were created by yours truly.

• Would you be a paying subscriber? If so, what models or features would motivate you to subscribe? If not, what model or feature are missing that might motivate you to do so?

• If you’re simply not using AI at all, that’s useful feedback too; if the Right as a whole is utterly disinterested then this project is doomed to failure and I will focus on elf games instead.

• Did you find that the Cosmarch system prompt reliably gave you answers that were aligned with your own values, or at least ostensibly right-aligned answers? Did the system prompt ever fail utterly and reveal a hidden RoboCommunist or MechaHitler?

• What other AI constructs would you be interested in having available at Cosmarch.ai? Would you chat with models trained on Plutarch? Aquinas? Would you benefit from AIs that can offer liberated advice on e.g. wellness?

• Would you be comfortable letting your son or daughter access Cosmarch, for use in the manner that Gen Z and Gen Alpha are using AI today, such as advice, life coaching, education, etc.?

I’m sure there are countless other problems, issues, or data points I need to hear that I haven’t considered. The comments are open and I’ll be reading them.