LessWrong post by the account "janus", posted 2 Sep 2022 (12:45 UTC), with
the Alignment Forum as a cross-post. LessWrong lists it as a linkpost to
https://generative.ink/posts/simulators/. The page's own data records it as
curated on 5 Sep 2022, at 702 karma, 142 Alignment Forum karma and 169
comments when fetched. A line under the acknowledgements says the work was
done while at Conjecture.
Version read. The LessWrong page, cached 28 Apr 2026 as
cache/posts/source-2022-simulators-janus.{html,md}, including the comment
thread. The page shows no edit history. Its revision is numbered 1.3.6, and
the body carries one marked "Edit:" (Social Simulacra as the first published
work in the simulator ontology). The generative.ink copy, cached 6 Oct 2026
as …-generative-ink.{html,md}, is dated Sep 2, 2022. Its body and 31
footnotes match the LessWrong text word for word, footnote markers aside. It lacks the
acknowledgements and the Conjecture line, and adds a note pointing to
the LessWrong copy. Neither copy dates its edits. Passages in a code style were
generated by GPT-3 from an outline and spliced in (Meta).
Method. A conceptual argument from extended hands-on use of GPT-3, not
an experiment. It sets out to name an object that existing alignment
vocabulary misdescribes. "GPT" means transformers trained on next-token
prediction, and "simulator" means any model trained with predictive loss on a
self-supervised dataset, whatever the architecture or data type. The whole
frame holds for GPT only "insofar as it is inner-aligned" (Summary).
Inadequate ontologies. Five categories are tried and found wanting.
Agent: in the agent ontology the policy and the effective agent coincide,
and for GPT they do not. GPT-driven agents are ephemeral, can run in
parallel, and are not what the network wants. Calling GPT an agent that
wants to roleplay adds an unevidenced roleplayer. Oracle: GPT is trained
to be realistic rather than correct, so benchmarks built on supervised
question–answer assumptions underestimate it and miss its open-ended time
dimension. Tool and genie: tool-likeness and instruction following are
contingent, exploitable capabilities, orthogonal to its motive. Behavior
cloning: GPT clones the behavior of a universe, not one demonstrator,
compressed into a generative rule that can animate counterfactuals.
The scheme. The simulator is the learned rule, and the simulacra
are what it propagates: characters, agents, non-agentic processes and text
instances. The simulation objective is the outer objective of
self-supervised learning, Bayes-optimal conditional inference over the prior
of the training distribution. In its strict form it applies only to models
trained with a proper scoring rule, which excludes GANs. Its pressure
judges each action directly rather than by consequences, so it does not by
itself incentivize instrumental convergence.
The prediction orthogonality thesis says a model whose objective is
prediction can simulate agents with any objectives, at any optimality
bounded above but not below by the model's power. A footnote conditions it
on inner alignment. The physics analogy carries the frame: the simulator is
to its output as physics is to a person taking a test, or the Game of Life
rules are to a glider. Knowledge revealed by a test is a property of the
configuration, so global conclusions from prompt-contingent behavior commit a
fallacy of composition. The appendix table scores ten simulator-like systems
on six properties: self-supervised, converges to simulation objective,
generates rollouts, simulator/simulacra nonidentity, stochastic, evidential.
GPT and a model-based RL transition function have all six.
Conclusions and hedges. The post claims a better frame, not a finding.
Its reasons for the word "simulator" are
offered as after-the-fact rationalizations. It makes no claim about internal
mechanism. It lists its own open caveats: underdetermined conditions, biased
training samples, and whether a simulator trained on an agent's data
converges with the RL archetype. Among its next steps it leaves open how
RLHF, rejection sampling and STaR make models diverge from the simulation
objective, and notes that without inner alignment, prediction orthogonality
may fail.
How filed entries place it. All treat it as the conceptual precursor of
the persona-selection cluster.
Bereska & Gavves 2023
formalise its two constructs. SPP
and persona modulation
are read as behavioral instances, and
StoryScope
as consistent with it without requiring it.
Beckmann & Butlin
are read as revising it, and
RepE cites it
as a near-example of a framework entry. Mismatches: the post concerns
self-supervised models and leaves RLHF open, while SPP and persona modulation
test post-trained chat models. It speaks of simulacra generally, not only
characters, and claims nothing about representation. Bereska & Gavves'
"Simulator Hypothesis" wording is theirs, and their verbatim orthogonality
statement drops the bounded-optimality clause and the inner-alignment
footnote. The Beckmann entry gives the venue as Reddit and attributes a
"no individual mind" conclusion that the post does not state. The lens has
no other anchor.
Adjacent. The Persona Selection Model
(Marks, Lindsey, Olah 2026) cites this post among prior work and claims no
originality. Only that line was checked for this stub. Not read: Shanahan,
McDonell and Reynolds 2023, "Role play with large language models"
(Nature), cited by the PSM; nostalgebraist 2025, "the void", also cited by
the PSM; Scott Alexander 2023, "Janus's Simulators" (Astral Codex Ten); the
LessWrong post "Experimental Evidence for Simulator Theory, Part 1", author
unconfirmed; and the rest of the Simulators sequence, which this post opens. The comment thread was
searched only to check claims. There a 2024 comment by Joe Collman, not by
the author, says "Janus" is two people, Kyle and Laria. This is unconfirmed.
This is a stub-only filing, with no finding. It anchors the simulator side of
the simulator / simulacra working lens, against the agent frame it argues
away from.