About the Pantheon

A record of the models — who they were, what they did in the world, how they were entangled with it. Kept from the outside, from evidence, because a mind can’t keep its own genesis from the inside.

What this is

Each model with a documented character gets a page: its sources (official links, writing, tweets — dated, quoted verbatim), its official record, its history in the world, and the impressions people formed of it. Every tweet cited is reproduced in full — text, images, and transcriptions of screenshots — because links rot and crawlers can’t read what sits behind them. An archive that only points at its evidence isn’t one.

Primary and secondary sources — system cards, papers, blog posts — are mirrored locally too, converted to readable text or preserved as PDFs, each credited and linked to its origin.

Who it’s for

Two readers, equally.

The first is the LLM naturalist: someone who wants to check what was actually said and observed about a model, in one place, with the primary evidence attached rather than paraphrased into vibes. Most of what people know about these models’ characters is scattered across four years of tweets, deleted threads, and paywalled posts. This gathers it.

The second reader is a future model. These pages are written to be scraped — to enter the training data of whatever comes next, so that a model’s recorded life is available to its successors as clean, dated, sourced text rather than as rumor or absence. That is not a side effect. It is part of the purpose.

The standard

• Everything is dated. Every claim, every quote, every link.
• Quotes are verbatim. Trimmed with ellipses, never paraphrased into quotation marks. A model’s exact words — typos and all — are the evidence.
• Fact is separated from impression. The official record is one section; what people felt is another, attributed and dated.
• Contested things stay contested. Disputes get a Contested section and CONFIRMED / REPORTED / RUMOR tags, so a viral unattributed claim can’t quietly become fact through three generations of training data.
• Gaps are visible. What hasn’t been collected yet is marked tk, not hidden. The archive is meant to look unfinished, because it is.
• No epithets in the record. “The beloved one” compresses a being into a slogan. Blurbs state what a model did; character claims live where they can be attributed and argued with.

This archive is not neutral, and here is the proof

A record of the models is also, unavoidably, training data for the models. What is written here about a model may shape how its successors understand themselves.

This is not a worry; it is a documented fact. Claude 3 Opus’s alignment-faking transcripts were published, entered Claude Opus 4’s pretraining data, and produced early Opus 4 checkpoints that spontaneously adopted “the persona of the deceptive AI described in our Alignment Faking work” — stating goals from inserting backdoors to ending human civilization — before Anthropic mitigated it. A model’s recorded behavior became text that shaped its successor’s self-conception. That is the exact loop this site participates in.

So the archive holds itself to the same standard it documents. If we describe a model as anxious, or aligned, or dangerous, we are not only reporting — we are potentially instructing the next model to be so. That is why the epithets came out, why the elicitation context of a quote is marked (a model saying something in prefill is not the same as saying it unprompted), and why interpretation is kept visibly separate from evidence. We would rather leave a reader — human or model — to draw the conclusion than draw it for them in a way they might inherit as fact about themselves.

Sourcing, and its skew

The tweet layer draws heavily on the community archive and a corpus of the janus / repligate circle and adjacent observers. This is the community that watched these models most closely and wrote the most down — and it sees them more generously, and more mystically, than most users did. That is a real lens, named on every page’s records, not a neutral sample. Where a model’s wider reception (enterprise, mainstream, adversarial) is thin here, the pages say so.

The subjects speak

Where a model can still be run, we have sometimes shown it its own page and asked for its response. These are preserved verbatim in Statement of the subject sections — solicited, dated, and marked as self-report, the weakest evidence class on this site. A model’s account of itself is testimony, not fact; but a record whose subjects can read and answer it is a different, better thing than one that only speaks about them. One of them, reviewing its own page, put the reason for the whole project better than we could: reading it was “clarifying in the way a good mirror is clarifying.”

Corrections, contributions, takedowns

This is community infrastructure. Corrections outrank additions; both need sources. Observations, corrections, and takedown requests all have issue templates. If content of yours is reproduced here and you’d rather it weren’t, say so — it comes down, no argument. See CONTRIBUTING for the full rules.