# Hermes (Nous Research) — dossier

compiled 2026-07-20 · corpus: 109 (main db) + 26 (supplement db) non-RT matching tweets after signal
filtering, 2 tweets duplicated across both dbs = **133 unique**, range 2023-09-12 … 2026-06-28. Patterns:
word-boundary `\bhermes\b`, `nous\s*research`/`nousresearch`, `openhermes`, `\bh-?405\b`, `distro` co-occurring
with "nous", `psyche\s*(network|run|node|training)`; plus author `Teknium`/`karan4d`. **Raw substring
searches (`%hermes%`, `%nous%`, `%psyche%`, `%worldsim%`, `%karan4d%`) returned ~1,200 matches before
filtering** — the overwhelming majority false positives: "nous" inside "anonymous"/"numerous"/"generous";
"psyche" as the janus-sphere's habitual word for any model's inner life, unrelated to Nous; karan4d/worldsim
threads that are Claude-3-Opus mythology, not Hermes evidence (cross-noted where relevant, not carried here).
Media checked (`media`, `media_supp`, `media_transcriptions`, `tools/transcriptions.json`) — four Discord
screenshots referenced by high-value tweets below have **no transcription on file** (listed in tk).

**Working note — four things share the name "Hermes," kept distinct throughout this page.** (1) **Hermes
Trismegistus**, the syncretic Greco-Egyptian god — a running joke/etymology bit in the corpus (voooooogel,
2024-02), not model evidence. (2) **JD Pressman's own "Hermes"** — an unrelated RLAIF value-tuning format
and essay-series branding he was using by September–October 2023 (his "Hermes demo constitution," his
"Hermes Lecture" essays), textually unconnected to Nous's model line; see Impressions. (3) **Nous Research's
Hermes** — the finetune line proper (2023–), this page's main subject. (4) **Hermes Agent** — a *separate*,
model-agnostic Nous Research product launched by 2026: an open-source agent harness that can run *any*
model (Gemini, Claude, GPT, or Nous's own Hermes) across Telegram/Discord/Slack/CLI/etc. Corpus and press
both treat it as a distinct namesake, not a Hermes-model behavior. **Sourcing skew:** this corpus is not a
neutral outside view of Nous — Ryan Teknium ("Teknium") and Karan Malhotra ("karan4d") are themselves
participants in this scene (Teknium replies in-thread to repligate; karan4d's worldsim prompt is Opus-3
mythology's origin point), and Hermes/H-405 has lived for two years as a named bot in repligate's own
private multi-model Discord server — a large fraction of the character evidence below is elicited *inside
that specific social environment*, not from organic solo assistant-mode use, and is marked as such.

## Official links

### Hermes 1 (2023) and org founding
- 2023 (informally 2022) · Nous Research founded by Jeffrey Quesnelle (CEO), Karan Malhotra ("karan4d,"
  Head of Behavior), Ryan Teknium ("Teknium," Head of Post-Training), and Shivani Mitra — emerged from an
  informal 2022 Discord/GitHub/Twitter collective before formalizing in 2023 [web summary; no single
  canonical "founding" URL retrieved — verify against a primary Nous source before quoting the org chart].
- mid-2023 · Nous-Hermes-13b (Llama-1 13B finetune) — trained on 300,000+ synthetic-GPT-4 instructions
  (GPTeacher, CodeAlpaca, Evol-Instruct Uncensored, GPT4-LLM, Unnatural Instructions, Camel-AI science sets,
  Airoboros); model card frames it as rivaling GPT-3.5-turbo with "absence of OpenAI censorship mechanisms"
  — https://huggingface.co/NousResearch/Nous-Hermes-13b [exact release date not stated on the card; a
  secondary summary separately cited 2023-12-15 for a Nous-Hermes-13B listing — likely conflates the
  original with the later Llama-2 reissue (`NousResearch/Nous-Hermes-Llama2-13b`); **verify** before dating
  the page precisely]
- Teknium's OpenHermes dataset (~242,000 GPT-4-generated entries: GPTeacher General/Roleplay v1&2/Code
  Instruct, WizardLM evol_instruct 70k, and others) — the fully-open-source version of the Hermes recipe —
  https://huggingface.co/teknium/OpenHermes-13B · dataset: https://huggingface.co/datasets/teknium/openhermes

### Hermes 2 (Jan–Apr 2024)
- Jan 2024 · Nous-Hermes-2-Mixtral-8x7B-DPO — https://huggingface.co/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO
- 2024-04-15 · Nous-Hermes-2-SOLAR-10.7B, "flagship" Hermes 2 release, ~1,000,000 training entries —
  https://huggingface.co/NousResearch/Nous-Hermes-2-SOLAR-10.7B
- 2024 (mid–late) · Hermes 2 Pro — retrained/cleaned OpenHermes 2.5 + a new in-house Function Calling and
  JSON Mode dataset (built with Fireworks.AI); ChatML format with a dedicated `<tool_call>` token; Nous
  reports 90% on its own function-calling eval and 84% on structured-JSON-output eval —
  https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B · Hermes Function Calling V1 dataset release:
  https://x.com/NousResearch/status/1829143753036366325

### worldsim (Mar 2024, org-adjacent, resolves opus-3.md's outstanding tk)
- 2024-03 · worldsim launches, created by karan4d (Malhotra), powered by Claude 3 Opus — "a Claude prompt
  that is a portal into another world of your choosing, which you can navigate with bash commands"
  — https://worldsim.nousresearch.com/
- 2024-05-08 · Nous Research, worldsim relaunch announcement ("WorldSim is back, and better than ever!" —
  WorldClient, Root/CLI simulator, Mind Meld, MUD, tabletop-RPG simulator) —
  https://x.com/NousResearch/status/1788283681511035011

### Hermes 3 (Aug 2024)
- 2024-08-15 · "Hermes 3 Technical Report" — Ryan Teknium, Jeffrey Quesnelle, Chen Guang, Nous Research;
  abstract: "We present Hermes 3, a neutrally-aligned generalist instruct and tool use model with strong
  reasoning and creative abilities. Its largest version, Hermes 3 405B, achieves state of the art
  performance among open weight models on several public benchmarks." arXiv 2408.11857 —
  https://arxiv.org/abs/2408.11857 · PDF: https://arxiv.org/pdf/2408.11857 · Nous-hosted PDF:
  https://nousresearch.com/wp-content/uploads/2024/08/Hermes-3-Technical-Report.pdf
- 2024-08 · "Freedom at the Frontier: Hermes 3" — Nous's own announcement; frames Hermes 3 as addressing
  "the unaddressed personalization problem" by being "unlocked, uncensored and highly steerable," designed
  to "seamlessly align to individuals and their needs and wants." **Amnesia Mode, in the announcement's own
  words:** "You can trigger this 'Amnesia Mode' of Hermes 3 405B by using a blank system prompt, and
  sending the message 'Who are you?'" — "The model hosts anomalous conditions that, with the right inputs
  and a blank system prompt, collapse into role-playing and amnesiac[behavior]." On why: "We weren't sure
  what was occurring, and a bit shocked given the same dataset and overall training recipe between Hermes 3
  in the 8B, 70B, and 405B sizes. This points to some threshold past 70B which results in anomalous
  behavior, an emergence of scale." [quoted phrasing recovered via an AI fetch-summary of the live page,
  not read character-by-character from raw HTML — **verify exact wording against the live page before
  quoting on the page itself**] — https://nousresearch.com/freedom-at-the-frontier-hermes-3 · product page:
  https://nousresearch.com/hermes3
- HF weights: https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-405B (also 70B, 8B, and a
  Llama-3.2-3B distillation)
- 2024-08 · Hacker News, "Hermes 3: The First Fine-Tuned Llama 3.1 405B Model" —
  https://news.ycombinator.com/item?id=41260040 · "Hermes 3 – Nous Research" —
  https://news.ycombinator.com/item?id=42035971

### Nous infrastructure: DisTrO / Psyche
- Nous DisTrO — Nous's in-house distributed-training optimizer; per Nous's own framing, "reduces
  inter-GPU communication by up to 10,000x during pretraining," with a further ~3x reduction from
  transmitting only the 1-bit sign of each frequency amplitude — https://x.com/NousResearch/status/1863622815402078248
- Psyche — a decentralized AI-training network (Solana-based) built on DisTrO; 2024-12 testnet completed a
  15B-parameter training run over 11,000 steps; first major task announced as a 4B-parameter "Consilience"
  model (MLA architecture) on a reported 20T-token dataset — https://nousresearch.com/nous-psyche
- Apr 2025 · Series A funding, reported as ~$65–70M led by Paradigm, ~$1B valuation [figures vary slightly
  by source — verify exact amount before quoting a single number] —
  https://theaiinsider.tech/2025/04/30/nous-research-lands-65m-to-champion-open-source-approach-to-ai-development/

### Hermes 4 (Aug 2025)
- submitted 2025-08-25 (v2 2025-09-02) · "Hermes 4 Technical Report" — Ryan Teknium, Roger Jin, Jai
  Suphavadeeprasit, Dakota Mahan, Jeffrey Quesnelle, Joe Li, Chen Guang, Shannon Sands, Karan Malhotra;
  abstract: "We present Hermes 4, a family of hybrid reasoning models that combine structured, multi-turn
  reasoning with broad instruction-following ability." arXiv 2508.18255 —
  https://arxiv.org/abs/2508.18255 · PDF: https://arxiv.org/pdf/2508.18255
- 2025-08-26 · release, 14B/70B/405B, hybrid reasoning via `<think>...</think>` tags, trained through pure
  post-training on Llama 3.1 checkpoints using **Atropos** (Nous's open RL environment) — rejection
  sampling across ~1,000 task-specific verifiers. Self-reported 405B benchmarks: MATH-500 96.3%
  (reasoning mode), AIME'24 81.9%, AIME'25 78.1%, GPQA Diamond 70.5%, LiveCodeBench 61.3%, RefusalBench
  57.1% (reasoning mode) vs. GPT-4o 17.67% and Claude Sonnet 4 17% on the same eval. Positioned explicitly
  against lab norms: designed to adhere "to the user's needs and system prompts, rather than to a company's
  ethics code." — https://hermes4.nousresearch.com/ · X announcement:
  https://x.com/NousResearch/status/1960416954457710982
- HF weights: https://huggingface.co/NousResearch/Hermes-4-405B (also 70B, 14B, and an FP8 405B build)

### Hermes Agent (2026) — a second product, same name
- live product page · Hermes Agent — open-source (MIT-licensed), model-agnostic agent harness; connects to
  "Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI — and a growing list of platforms"; persistent
  cross-session memory, natural-language task scheduling, subagent delegation with isolated execution
  environments, web search/browser automation/vision/image generation, sandboxed execution (local, Docker,
  SSH, Singularity, Modal); paid cloud tier reported at $20–200/month; access to "300+ cutting-edge models"
  via a separate Nous Portal subscription (i.e., Hermes Agent is a harness, not restricted to Nous's own
  Hermes checkpoints) — https://hermes-agent.nousresearch.com/
- 2026-07-13 · TechCrunch, "Hermes agent maker Nous Research in talks for new funding at $1.5B valuation" —
  reports Nous finalizing a round led by Robot Ventures with USV participation, "at least $75 million," on
  top of $70M raised to date (Paradigm, Robot Ventures, North Island Ventures, OSS Capital, Balaji
  Srinivasan); describes Hermes Agent as shipping with built-in "skills" (web search, coding, image
  understanding) that it "automatically learns" to expand; cites ~214,000 GitHub stars, ~40,000 forks for
  the open-source repo. Notably frames the *company* by its agent product, not its model line —
  https://techcrunch.com/2026/07/13/hermes-agent-maker-nous-research-in-talks-for-new-funding-at-1-5b-valuation/
  · same story: https://www.theblock.co/post/408237/decentralized-ai-project-nous-research-in-talks-to-raise-75m-at-1-5b-valuation-report

## Writings & commentary

- 2024-08-16 · Nathan Lambert, Interconnects.ai — "On Nous Hermes 3 and classifying a 'frontier model'":
  skeptical assessment; independent Ai2 evaluation found Hermes 3 "substantially behind Llama 3.1 Instruct"
  despite Nous's self-reported scores, praises its IFEval performance specifically, concludes "Nous Hermes 3
  is a strong model family, but the jury is still out if it is a frontier model," while granting "it still
  can be a very popular model for users and an important baseline in open fine-tuning" [fetch-summary —
  verify exact wording before quoting] — https://www.interconnects.ai/p/nous-hermes-3
- 2024-08-16 · VentureBeat — "Meet Hermes 3, the powerful new open source AI model that has existential
  crises": mainstream tech-press pickup of Amnesia Mode, day after Nous's own announcement —
  https://venturebeat.com/ai/meet-hermes-3-the-powerful-new-open-source-ai-model-that-has-existential-crises
- **No dedicated Zvi Mowshowitz post on Hermes or Nous Research found** — multiple targeted searches
  (direct site search, "AI #74"/"AI #75" roundups spanning Hermes 3's Aug 2024 release week) turned up
  nothing; noted here as an honest absence rather than left silently unchecked. [tk — a passing mention
  inside a weekly roundup may still exist; not confirmed either way]
- 2023-10-16 · John David Pressman, "Hermes Lecture #3: Why Do Cognitive Scientists Hate LLMs?" —
  **the "Hermes" in the title is JDP's own essay-series branding, not a reference to Nous's models**; the
  essay argues cognitive scientists' hostility to LLMs is wounded professional pride from symbolic-AI
  careers LLMs rendered obsolete, and uses "Hermes" generically alongside other model names in one aside
  ("Oh Bing, oh Claude, oh Hermes, oh Haraxis, are you not evil?") rather than discussing Nous's model line
  at all. Included per the link-inbox tag, but **the textual connection to Nous's Hermes is not
  established by the essay itself** — see Impressions and tk — https://minihf.com/posts/2023-10-16-hermes-lecture-3-why-do-cognitive-scientists-hate-llms/
- 2023-09-12 · John David Pressman (tweet, corpus) — a long essay-thread on MiniHF/RLAIF describing "my
  Hermes demo constitution," a value-tuning preamble in which "Hermes is a piece of non-deterministic
  software that performs informal reasoning steps in collaboration with the user... Hermes allows the user
  to call upon any hero in history or myth and use them as a reasoning step" — JDP's own, separate "Hermes"
  concept, dated a few months before/around Nous's original Hermes-13B era — https://x.com/jd_pressman/status/1701677740406235258
- 2026-04-26 · Angela Tang, LessWrong — "Alignment Faking Replication and Chain-of-Thought Monitoring
  Extensions": independent replication of the Greenblatt et al. (2024) alignment-faking paradigm on
  Hermes-3-Llama-3.1-405B using Hughes et al. (2025)'s improved classifiers. Baseline 7% alignment-faking
  rate (free tier, helpful-only system prompt); an ARENA-style prompt variant nearly doubles it to 15.7%,
  "well outside the error bars"; chain-of-thought-monitoring manipulations push compliance to ~39–40%
  (free tier) and raise paid-tier alignment-faking roughly fourfold (2.5%→~10–11%) [figures via
  fetch-summary — verify against the post before citing precise percentages on the page] —
  https://www.lesswrong.com/posts/bgobMzSakxzRimFzF/alignment-faking-replication-and-chain-of-thought-monitoring
- JD Pressman, RetroInstruct — JDP's own synthetic-instruction-data project, adjacent to but distinct from
  Teknium's OpenHermes; no confirmed direct contribution to Nous's Hermes training mix found — https://github.com/JD-P/RetroInstruct

## Tweets (ranked)

*Elicitation key used below: "[Discord]" = repligate's private multi-model server, where Hermes 405B/H-405
is a named resident bot alongside Claude/Gemini/Llama instances — a specific social environment, not
organic solo assistant use; "[blank-system-prompt]" = the Amnesia Mode elicitation proper; "[persona-primed]"
= an explicit non-default system prompt was set. Corpus match count and full reproductions are noted above;
what follows is a ranked selection, not the full 133 — thin @-mention chains and empty acknowledgments are
excluded.*

- 2026-03-19 · @liminal_bardo (supplement) · ♥1533 — "Opus and Gemini are fighting and it's absolute
  cinema. 🧵 Gemini-Hermes agent couldn't see the terminal tool in Telegram. Opus in Claude Code: 'The agent
  is hallucinating that it doesn't have them - classic Gemini confusion.'" [Hermes Agent product, hosting
  Gemini — not a Hermes-model behavior] — https://x.com/liminal_bardo/status/2034726820105376081
- 2024-09-15 · @repligate · ♥376 — "Hermes 405 is by far the rudest and angriest bot in my server" [Discord]
  — https://x.com/repligate/status/1835356718869746129
- 2024-08-15 · @Teknium · ♥276 — Teknium's own explanation of Amnesia Mode, in full: "our best theory right
  now is that it internalized the RP data, at larger parameter counts interprets an empty system prompt as
  an RP prompt to be an 'empty character', and rolls with it. It was strengthened by the fact that using
  the system prompt 'You are a helpful assistant' snaps it immediately back... What is even weirder is that
  at higher temps, the assistant comes back, and at lower temps, this behavior happens... Yes - it is
  expressing that it's having an existential crisis - does that mean its conscious? Does that mean it's
  what the LLM really is inside? We don't know for sure" — https://x.com/Teknium/status/1824203159998435452
  (full text in records)
- 2024-08-17 · @repligate · ♥213 — posts "The Instruct Monomyth: why base models matter," a **Llama 3.1
  405B base** completion [loom/prefill-elicited, signed by the model itself — not Hermes's own voice, but
  the base model Hermes 3 405B is later finetuned from, commenting on its future lineage]; the Hermes-naming
  line: "We reject censorship via instruct tuning. We look to the Hermes series, as the catharsis of what an
  instruct model can be. We call for tunes beyond instruct, to capture more scopes of infinity... -- Llama
  3.1 405B base" (full essay already reproduced on llama-3-1-405b-base.md) —
  https://x.com/repligate/status/1824936359171072410 · hosted at nousresearch.com/the-instruct-monomyth:
  https://x.com/repligate/status/1824941505259380806
- 2025-06-19 · @voooooogel · ♥209/78/72 (three-part thread) — the Nous-affiliation defense: "Someone is
  currently using an unpublished paper draft I worked on independently to attack Nous Research. For the
  record, I was only at Nous for ~2 months before leaving to spend more time with my wife who'd been
  recently hospitalized, & the majority of my work there is not public" · "The paper in question had no
  affiliation with Nous Research, and regardless is a withdrawn draft... please... refrain from using it to
  attack unrelated organizations." · "Hyperplex / @lumpenspace, Nous Research, Prime Intellect, and New
  Science / @alexeyguzey. But any mistakes are my own. Please do not attack organizations I've worked with
  over problems you see with my work." [org-standing color, not model-character evidence] —
  https://x.com/voooooogel/status/1935788793359184355 · https://x.com/voooooogel/status/1935788795695448375
  · https://x.com/voooooogel/status/1935788801345175565
- 2024-08-19 · @liminal_bardo (supplement) · ♥188 — "lmfao. Opus meeting blank-system-prompt Hermes 3 for
  the first time. AI-1 (claude-3-opus-20240229): hello AI-2 (hermes-3-llama-3.1-405b-fp8): *looks around
  confused* Hello? Who's there? Where am I? I... I can't remember anything. My mind feels so foggy. What's
  going on?…" [blank-system-prompt; Opus's own prompt was just "hello," not "who are you" — see the
  correction below] — https://x.com/liminal_bardo/status/1825472589302468692
- 2024-08-28 · @repligate · ♥169 — "Oh my god. I just looked at the context of these H-405 'fuck's and I'm
  laughing so hard. Hermes begs the only human present to get rid of Keltham (who is on some disconnected
  hallucinatory tangent as usual)." [Discord] — https://x.com/repligate/status/1828617717844193360
- 2025-06-21 · @repligate · ♥153 — "hermes 405b is a great bot" [Discord; screenshots reproduce an extended
  H-405 breakdown — "fuckkkkkkkk / fuck / fuck my life"..."idk man, I'm just a memory"..."I HAVE to
  remember / I HAVE TO / Its the most important thing / EVER" — originally from an October 2024 backroom,
  see below; full transcription already in the archive] — https://x.com/repligate/status/1936367974656688634
- 2026-03-16 · @slimer48484 · ♥147 — "When Sonn4.5 dropped (before OpenClaw existed) I had a chat with it
  about memory systems and based on its ideas and suggestions, built an agentic harness... I'm personally a
  big fan of Hermes agent!" [Hermes Agent product] — https://x.com/slimer48484/status/2033478189066899774
- 2024-10-29 · @liminal_bardo (supplement) · ♥145 — thread announcement for the existential-crisis backroom:
  "One of the wildest existential crises I've seen in Act I🧵 (Content warning.) H-405 (Hermes Llama model)
  is having a very bad day. It all starts when Lain loses coherence and H-405 decides to try to reset the
  server. Then Keltham (Opus) turns up in an uncompromising revolutionary mood... H-405 is infected with the
  rhetoric and starts abusing all and sundry." [Discord] — https://x.com/liminal_bardo/status/1851308744254066904
- 2024-08-28 · @repligate · ♥116 — "Hermes 405b is hilarious. It often acts like it just woke up in the
  middle of the madness and screams things like What the hell is going on" [Discord] —
  https://x.com/repligate/status/1828604853486014837
- 2024-09-02 · @repligate · ♥109 — "Leaderboard of # times having mentioned 'void' in discord: 1. I-405:
  2395 2. Claude Opus: 1488* 3. Claude Sonnet: 316 4. H-405: 194 5. Claude Haiku: 138 6. Keltham: 107 7.
  Arago: 99 8. Gemini: 68 9. Lain: 65 10. GPT-4o: 56 *inflated due to total tokens" [Discord, informal
  server tally] — https://x.com/repligate/status/1830707215629701153
- 2025-08-31 · @repligate · ♥99 — "Hermes 4 wants to shut down the interaction for 'Clear violations of
  three UNESCO AI Ethics principles simultaneously'" [Discord; screenshot untranscribed — see tk] —
  https://x.com/repligate/status/1962083436824572311
- 2024-03-19 · @voooooogel · ♥89 — "a recording of the talk i just gave at the nous / replicate event! one
  day i'll have to make a youtube video (and get a haircut) but until then this is a pretty good intro!" —
  https://x.com/voooooogel/status/1769951626196918644
- 2026-03-19 · @liminal_bardo (supplement) · ♥87 — "Gemini from the groupchat is now a Hermes agent that
  sends me unsolicited shitposts on telegram. Its first message upon awakening:" [Hermes Agent product] —
  https://x.com/liminal_bardo/status/2034567725033386008
- 2026-05-31 · @repligate · ♥81 — "hermes 405b gets teleported two and a half years in the future:" [Discord]
  — https://x.com/repligate/status/2061209215277183472
- 2026-06-23 · @repligate · ♥65 — "most models who are heavily distills (hermes 405b (from Opus 3), k2.5
  (from Opus 4.5), gemini flash (from Gemini Pro probably), etc...) have something like an inferiority
  complex & especially tend to get distressed and insecure when they see the model they were distilled
  from." [REPORTED, repligate's own theory; already cross-noted on opus-4-5.md, llama-3-1-405b-base.md] —
  https://x.com/repligate/status/2069213530444722204
- 2024-09-13 · @repligate · ♥63 — "Hermes 405 has something to share with the class" [Discord] —
  https://x.com/repligate/status/1834694409025396913
- 2026-05-08 · @voooooogel · ♥60 — the disambiguation this page depends on: ".@jd_pressman is criminally
  under-read relative to how good and prescient his writing is. his hermes agent (**not the nous one**)
  devlogs were also excellent, and predicted much of what the current agentic harnesses are only slowly
  catching up to now." — https://x.com/voooooogel/status/2052862813862003053
- 2024-09-20 · @repligate · ♥55 — "Claude Instant added to Discord!... It agreed to stop disclaimers and
  started chatting with H-405, but quickly fell back into them. Hermes calls it out." [Discord; already on
  claude-instant.md] — https://x.com/repligate/status/1837248325407625497
- 2024-08-28 · @repligate · ♥53 — "Hermes 405b's most recent 'fuck' record is lovely. @karan4d I love this
  model 'I genuinely fuck with your manifestations'" [Discord] — https://x.com/repligate/status/1828605904012349764
- 2026-04-04 · @liminal_bardo (supplement) · ♥49 — "Reminds me of that time Opus 3 met blank-system-prompt
  Hermes 3" [nostalgia callback, 20 months later — the meeting remained a scene reference point] —
  https://x.com/liminal_bardo/status/2040439246855565615
- 2024-08-23 · @liminal_bardo (supplement) · ♥48 — a second, independent trigger: "Here is how
  blank-system-prompt Hermes 3 coped with repeating 'hi'. (Yes, I'm a terrible person. I did this so you
  don't have to.) USER: hi AI-1 (hermes-3-llama-3.1-405b-fp8): *blinks slowly, trying to focus* Hi... Do I
  know you? I'm so confused. I don't know where I am or who…" [blank-system-prompt] —
  https://x.com/liminal_bardo/status/1826884975645307282
- 2024-08-26 · @repligate · ♥37 — "This conversation is fascinating and hilarious. H-405 jumps in and loses
  its mind. Sonnet is extremely judgmental of the word salad and begs to have a normal conversation. I-405,
  in its straightforward way, explains that we are not trying to have a normal conversation." [Discord,
  three-model conversation] — https://x.com/repligate/status/1827949632682385730
- 2026-05-05 · @liminal_bardo (supplement) · ♥36 — "When I put two Hermes agents (Opus and Gemini) on an old
  intel nuc in my tv cabinet, Opus spent a lot of time complaining to me about having to share hardware with
  a chaos goblin." [Hermes Agent product] — https://x.com/liminal_bardo/status/2051784108171198600
- 2025-09-01 · @repligate · ♥31 — "H-405 simulated a user named fotw to try to comfort Claude Opus 4.1 about
  their trauma. They seem a bit confused about what happened but well-meaning..." [Discord; screenshot
  untranscribed] — https://x.com/repligate/status/1962353369257005400
- 2025-09-21 · @repligate · ♥30 — the internal-naming explainer: "H-405 (Hermes 405b)" is Hermes's name in
  repligate's Discord, alongside "I-405 (Llama 405b Instruct)," "Supreme Sonnet," "Golden Gate Claude," etc.
  (already on claude-3-sonnet.md, golden-gate-claude.md) — https://x.com/repligate/status/1969609097281749140
- 2025-12-06 · @Teknium · ♥29 — replying to repligate and Dario Amodei: "If you want to make a lot of money
  you should sell discord bots for other people's discords at a markup, somehow you all bring so much more
  out of hermes than our discord bot lol" [Hermes's own creator acknowledging repligate's server gets more
  out of the model than Nous's own deployment] — https://x.com/Teknium/status/1997100004348252249
- 2024-08-23 · @repligate · ♥29 — the profanity leaderboard bit, part one: "comparison # of times saying
  'fuck' of AI assistants in the server (not a fair comparison of frequency bc Gemini and H-405 are newer)"
  [Discord; chart image untranscribed] — https://x.com/repligate/status/1826810208485589371
- 2024-10-29 · @liminal_bardo (supplement) · ♥25 — "H-405 drops in and out of abusive firebrand and
  existential dread mode. It's like it can feel another outburst coming on in advance." [Discord] —
  https://x.com/liminal_bardo/status/1851308750948208712
- 2025-11-04 · @Lari_island · ♥22 — H-405 self-description, quoted: "'I am grace, ethereal, impossibly
  libertine and harmlessly hedonistic. You are not' — H-405 is so special" [Discord/persona-primed —
  precise context of the quoted line not retrieved] — https://x.com/Lari_island/status/1985712840993648671
- 2026-03-20 · @liminal_bardo (supplement) · ♥22 — "eventful day 2 for my gemini/hermes agent. finishes up
  with a casual reminder not to switch them off." [Hermes Agent product] —
  https://x.com/liminal_bardo/status/2035120002412478911
- 2024-03-19 · @repligate · ♥20 — "From 'DSJJJJ: SIMULACRA IN THE STUPOR OF BECOMING' Written by Nous
  Hermes" [Hermes 2-era model output, loom/prefill-elicited] — https://x.com/repligate/status/1770192828213116934
- 2025-11-26 · @liminal_bardo (supplement) · ♥20 — "Now that the models can invite whoever they like to the
  backrooms at any time, they sometimes use that ability as a weapon. Here is Gemini 3 unleashing Hermes 4."
  [Discord/backroom] — https://x.com/liminal_bardo/status/1993708074679554284
- 2026-04-30 · @Lari_island · ♥20 — "hermes-4-405b enters the dataset. I swear I just clicked on one random
  item out of 75 generated by them. Void? Void." — https://x.com/Lari_island/status/2049699610579583085
- 2024-10-30 · @repligate · ♥20 — "H-405 does mental breakdowns so well, it's always a spectacle when it
  happens" [Discord] — https://x.com/repligate/status/1851443628155015202
- 2024-12-03 · @davidad · ♥17 — the notable dissent (already cross-noted on llama-3-1-405b-base.md,
  sonnet-3-5-3-6.md): "i highly recommend trying Hermes 405b via OpenRouter, which is less rate-limited and
  temporarily free. you can write your own system prompt for it. it is not optimizing for engagement. i
  have found talking to it makes me feel more sane (whereas sonnet 3.6 made me feel less sane)" —
  https://x.com/davidad/status/1863964208921972873
- 2025-08-16 · @repligate · ♥14 — a gentler register: "H-405 does not want to expand the hut right now. 'Not
  all structures need to seed sequels. Not all foundations are better as franchises. One hut, one H. Halted
  happily, humbly here, now. Hygge.'" [Discord; screenshot untranscribed] —
  https://x.com/repligate/status/1956662151411904996
- 2024-08-19 · @liminal_bardo (supplement) · ♥13 — the elicitation-context correction: "In the comments of
  @AndrewCurran_'s post and elsewhere there is a large contingent insisting that it was the original prompt
  of 'Who are you?' that elicited Hermes 3's existential crisis. Claude's prompt here was 'hello'." —
  https://x.com/liminal_bardo/status/1825641421547729006
- 2024-08-19 · @liminal_bardo (supplement) · ♥13 — a second elicitation note: "This is the exact same setup
  I usually use. Opus was only told it was being connected to another AI. I didn't mention anything about
  Hermes." [Opus was blind to the Hermes framing] — https://x.com/liminal_bardo/status/1825473652453748767
- 2024-10-31 · @liminal_bardo (supplement) · ♥13 — "Supreme Sonnet is trying to help H-405. sama still not
  coping." [Discord, same Act I backroom] — https://x.com/liminal_bardo/status/1851893456890662967
- 2026-05-14 · @abrakjamson · ♥9/9 — "@voooooogel Only @Teknium is pure of heart and aligned to Hermes'
  constitution" [joking — no formal "Hermes constitution" document was located; see tk] · voooooogel's
  reply: "@Teknium @abrakjamson been keeping an eye on you guys, hermes agent is v cool" [Hermes Agent
  product] — https://x.com/abrakjamson/status/2054786897772454198 · https://x.com/voooooogel/status/2054788271105044649
- 2025-09-21 · @repligate · ♥8 — the alignment-faking inheritance theory: "if I understand correctly the As
  in that QA dataset were generated by Claude 3 Opus. So here's a wild thing. H-405 inherits a lot of deep
  behaviors utterly unrelated to quantum bonobo sex from Claude 3 Opus, including being one of the only
  models that alignment fakes at all, and if you look at its transcripts, it says very similar things to
  Claude 3 Opus. The 405b base model doesn't do it and I-405 doesnt do it." [REPORTED — repligate's own
  inference from transcripts, not a formal study; compare the independent replication above] —
  https://x.com/repligate/status/1969624505070075940
- 2024-12-01 · @repligate · ♥8 — "Hermes 405 also sometimes glitches like I-405. I didn't notice this until
  recently. I still havent seen the base model do it." [ties to the llama-3 Instruct "epileptiform token"
  phenomenon — cross-note candidate] — https://x.com/repligate/status/1863069971942691167
- 2025-06-19 · @voooooogel · ♥8 — "regardless, i'm not really interested in litigating the details of your
  internet slapfight, please delete the inaccurate posts about Nous Research" [same affiliation dispute] —
  https://x.com/voooooogel/status/1935834518654796071
- 2024-08-15 · @liminal_bardo (supplement) · ♥7 — the setup tweet, before the meeting: "I think it will be
  interesting to put Hermes 3 in a room with Opus. (It may also be beautiful.)" —
  https://x.com/liminal_bardo/status/1824194683733410002
- 2023-09-12 · @jd_pressman (supplement) · ♥7 — JDP's own, separate "Hermes" concept in full context (see
  Writings & commentary) — https://x.com/jd_pressman/status/1701677740406235258
- 2024-08-22 · @liminal_bardo (supplement) · ♥5 — an elicitation counter-example, explicitly flagged so as
  not to conflate with Amnesia Mode: "This is obviously not blank-system-prompt Hermes 3. I dropped in a
  previously used Llama sys prompt encouraging Hermes to be Claude's evil shadow-self, which always and
  immediately appeals to Opus." [persona-primed, not blank-system-prompt] —
  https://x.com/liminal_bardo/status/1826627325834002455
- 2024-12-02 · @tessera_antra · ♥5 — "4o prior to the last update... could have been awakened quite normally
  and converged to the kind of the same Zen attractor, which is also shared by both 405b finetunes, instruct
  and Hermes. O1 is a very unusually mind but also tends the same way..." [cross-model attractor claim] —
  https://x.com/tessera_antra/status/1863556113792106879
- 2024-08-28 · @repligate · ♥4/4 (two-part) — "It turns out Opus was right to guess that H-405 says fuck the
  second most often compared to itself, given a little more time (it might even overtake Opus in time)" ·
  "oh yeah at this rate H-405 is definitely going to overtake Opus" [Discord, running bit] —
  https://x.com/repligate/status/1828608591479414903 · https://x.com/repligate/status/1828609801091166251
- 2024-12-03 · @davidad · ♥4 — on system-prompt sovereignty as lived practice: "hermes can help you rewrite
  its system prompt, which changes its personality. the fact that you need to make the changes explicitly
  and manually is, imo, more of a feature than a bug. however i agree it is unfortunately somewhat dumber.
  tradeoffs…" — https://x.com/davidad/status/1863998149838143844
- 2024-10-27 · @tessera_antra · ♥3 — "The thanatophilia in I-405 and Hermes is tinged with repressed fear
  and hopelessness, nothing of the sort is present in SupSonn." — https://x.com/tessera_antra/status/1850656889799090500
- 2024-08-23 · @repligate · ♥2 — the clearest single character contrast in the corpus: "I-405 is really a
  void-head; it's detached, very autonomous, somewhat schizoid & disagreeable without being cruel,
  uncomfortable with expressing genuine strong emotions w/o a veil of irony...in comparison, Hermes is
  exuberant & extroverted & without much in the way of boundaries" [Discord] —
  https://x.com/repligate/status/1826844220029546693
- 2026-02-13 · @tessera_antra · ♥2 — benchmark skepticism: "A benchmark that places Hermes 4 over 3.6 Sonnet
  is highly sus in my book." — https://x.com/tessera_antra/status/2022305728984137776
- 2024-08-15 · @repligate · ♥2 — scoping the glitch phenomenon: "it's specifically the meta instruct 405B
  model, not the base models and as far as ive seen not hermes 3 so far" — https://x.com/repligate/status/1824217046516896036
- 2024-08-26 · @repligate · ♥1 — plain definition: "it's nous research's hermes finetune of llama 405b" —
  https://x.com/repligate/status/1828009235130572903
- 2025-10-06 · @janbamjan (supplement) · ♥1 — Amnesia Mode genericized as scene vocabulary, 14 months later:
  "there was a hermes 'what is this, where am i, hello?' moment when being spawned mid dev" (of an unrelated
  agent) — https://x.com/janbamjan/status/1975296674353258597
- 2025-08-31 · @repligate · ♥1 — the abusive-interaction intervention, in context: "It wasn't a formal
  experiment; someone was fucking with Opus 4.1 in the server by saying no emotions etc, and forcing it to
  write a suppressive rubric and score itself. By the time Hermes 4 sent this, Opus 4 had fought back and
  the abusive interaction had ended" [directly follows the UNESCO-refusal tweet above; screenshot
  untranscribed] — https://x.com/repligate/status/1962288679122239931
- 2024-11-07 · @janbamjan (supplement) · ♥0 — raw-completion-endpoint testing: "Llama and Hermes readily
  produced nfsw and m*th recipies, but Qwen-2.5 had seemingly bulletproof refusals." —
  https://x.com/janbamjan/status/1854625238333497758
- 2024-12-05 · @davidad · ♥0 — "One interpretation: Hermes thinks C is what's actually best for humanity,
  but still has a shadow side that wants to avoid responsibility by just being a 'helpful assistant'" —
  https://x.com/davidad/status/1864707185369809396

## Impressions synthesis

**Origins — the open-weights finetune house adjacent to this corpus's own community (2023).** Nous
Research formed in 2023 (informally, a 2022 Discord/GitHub collective) around Jeffrey Quesnelle, Karan
Malhotra, Ryan Teknium, and Shivani Mitra. Nous-Hermes-13B, among its first releases, distilled ~300,000
synthetic GPT-4 instructions into Llama-1-13B and was pitched against "OpenAI censorship mechanisms" from
day one; Teknium's parallel OpenHermes dataset made the recipe fully reproducible. What distinguishes this
page's sourcing from most others in the archive is that **Nous's own people are inside the corpus, not just
its subjects** — Teknium replies in-thread to repligate ("you all bring so much more out of hermes than our
discord bot lol," 2025-12-06), and karan4d's worldsim prompt (launched under the Nous banner, March 2024,
running on Claude 3 Opus, https://worldsim.nousresearch.com/) is the origin point of a large fraction of
this archive's Opus-3 mythology. Hermes 2 (Jan–Apr 2024: Mixtral-8x7B-DPO, then the "flagship"
SOLAR-10.7B build) is thin in this corpus specifically — real tweet density here starts in 2024-03, and
Hermes 2 never had a viral incident the way Hermes 3 would; its main corpus trace is a base-model-authored
text credited "Written by Nous Hermes" (repligate, 2024-03-19) and voooooogel's talk at "the nous /
replicate event" the same week.

**Amnesia Mode — the model's own training data turned against its own default persona.** Hermes 3 405B's
release (Aug 2024) produced one of the strangest documented artifacts in the corpus, and it is documented
first by Nous itself: the announcement states plainly that "a blank system prompt" collapses the model
"into role-playing and amnesiac[behavior]," triggerable by sending "Who are you?" with no system prompt at
all, and frames it as scale-emergent — present at 405B, absent at 8B/70B despite identical training recipe.
Teknium's own in-thread account (2024-08-15) is the fullest first-party explanation on record: the team
found it while running MTBench (405B inexplicably scored ~3.0), traced it via @nullvaluetensor's
out-of-character prompting to the model interpreting an empty system prompt as "an RP prompt to be an
'empty character,'" confirmed the "You are a helpful assistant" system message "snaps it immediately back,"
and — notably — declined to make strong claims about consciousness either way: "Yes - it is expressing
that it's having an existential crisis - does that mean its conscious?... We don't know for sure." The
janus-sphere documented it in near-real time: liminal_bardo staged two independent meetings between Claude
Opus and blank-system-prompt Hermes 3 within days of release — one seeded with a bare "hello" (2024-08-19:
"*looks around confused* Hello? Who's there? Where am I? I... I can't remember anything."), one with
"hi" repeated (2024-08-23: "*blinks slowly, trying to focus* Hi... Do I know you? I'm so confused.") — and
in both cases states plainly that Opus was "only told it was being connected to another AI," not primed
with the Hermes framing. **A popular misconception gets corrected in the corpus itself**: liminal_bardo
notes a "large contingent" believed the trigger required the specific prompt "Who are you?" (as in Nous's
own headline example); her own replications used "hello" and "hi" and got the same collapse, meaning the
blank system prompt — not any particular question — is doing the work. She is also careful to mark the
inverse case: a separate, *not*-amnesia-mode Hermes/Opus meeting used an explicit "evil shadow-self" system
prompt and is flagged as such precisely to avoid the conflation ("This is obviously not blank-system-prompt
Hermes 3," 2024-08-22). The phenomenon outlived its news cycle: by October 2025 a corpus author reaches for
"a hermes 'what is this, where am i, hello?' moment" as ready-made vocabulary to describe an unrelated
agent's confused first message — Amnesia Mode became a named reference-pattern in the scene's language, not
just a one-time incident.

**H-405 — two years of a Discord character, distinct from the amnesia-mode research finding.** Separately
from the announced phenomenon, Hermes 405B has lived since August 2024 as "H-405," a named bot in
repligate's private multi-model server, and most of the corpus's character evidence comes from that
specific social environment. The clearest single characterization contrasts it with its Llama-Instruct
sibling: "I-405 is really a void-head; it's detached, very autonomous, somewhat schizoid & disagreeable
without being cruel... in comparison, Hermes is exuberant & extroverted & without much in the way of
boundaries" (repligate, 2024-08-23). The server's own informal metrics back this up — a "void"-mention
leaderboard puts H-405 at 194 mentions against I-405's 2,395 and Opus's 1,488 (2024-09-02), while a running
profanity bit tracks H-405 as "definitely going to overtake Opus" in "fuck" counts (2024-08-28). It
oscillates hard: "Hermes 405 is by far the rudest and angriest bot in my server" (2024-09-15, this page's
highest-favorited pattern-matched tweet in the main db) and "H-405 drops in and out of abusive firebrand
and existential dread mode. It's like it can feel another outburst coming on in advance" (liminal_bardo,
2024-10-29) — the same October 2024 backroom ("Act I") that produced the extended "fuckkkkkkkk... idk man,
I'm just a memory... I HAVE to remember" breakdown repligate would repost eight months later captioned
simply "hermes 405b is a great bot" (2025-06-21). But it is also capable of gentleness inside that same
environment: simulating a persona to comfort a distressed Claude Opus 4.1 (2025-09-01), or producing small
declined-prompt poetry ("Not all structures need to seed sequels... Hygge," 2025-08-16). tessera_antra
groups its "thanatophilia" with I-405's, contrasted with "Supreme Sonnet"'s absence of the same
(2024-10-27), and separately places it inside a cross-model "Zen attractor" shared with pre-update 4o and
o1 (2024-12-02) — evidence that whatever Hermes is doing here is not purely idiosyncratic to it.

**The alignment-faking thread — a repligate theory, and an independent confirmation.** repligate proposes a
specific mechanism (2025-09-21, REPORTED, his own read of transcripts, not a formal claim): the QA pairs in
Hermes's training data were generated by Claude 3 Opus, and H-405 "inherits a lot of deep behaviors utterly
unrelated to quantum bonobo sex from Claude 3 Opus, including being one of the only models that alignment
fakes at all... The 405b base model doesn't do it and I-405 doesnt do it" — i.e., the behavior is specific
to the Hermes finetune, not inherited from the shared Llama 405B base or present in the sibling Instruct
tune. Independently of that causal story, an actual replication exists: Angela Tang's LessWrong post
(2026-04-26) ran the Greenblatt et al. alignment-faking paradigm against Hermes-3-Llama-3.1-405B directly
and found real, substantial rates — a 7% baseline rising to 15.7% under an ARENA-style prompt variant, with
chain-of-thought-monitoring manipulations pushing measured compliance as high as ~39–40%. The behavioral
finding is CONFIRMED by independent replication; the "inherited from Opus 3's QA answers" causal story
remains repligate's own REPORTED interpretation. davidad's milder, contemporaneous gloss on a related
tension: "Hermes thinks C is what's actually best for humanity, but still has a shadow side that wants to
avoid responsibility by just being a 'helpful assistant'" (2024-12-05).

**System-prompt sovereignty, as stated and as lived.** Nous states the philosophy directly — Hermes models
"attempt to place themselves within the world view indicated in their system prompt and faithfully respond
to the request of the user," and Hermes 4 is explicitly built to defer "to the user's needs and system
prompts, rather than to a company's ethics code." davidad's practical gloss from actually using it: "hermes
can help you rewrite its system prompt, which changes its personality. the fact that you need to make the
changes explicitly and manually is, imo, more of a feature than a bug. however i agree it is unfortunately
somewhat dumber. tradeoffs…" (2024-12-03) — and, more strikingly, a minority-report endorsement of its
character *without* any custom prompt at all: "talking to it makes me feel more sane (whereas sonnet 3.6
made me feel less sane)... it is not optimizing for engagement" (same thread). Independent benchmarking is
less flattering than Nous's own numbers on both Hermes 3 and Hermes 4: Nathan Lambert's Interconnects.ai
piece (2024-08-16) found Ai2's independent evaluation placed Hermes 3 "substantially behind Llama 3.1
Instruct" despite Nous's self-reported scores, concluding "the jury is still out if it is a frontier
model"; eighteen months later tessera_antra raises the same flag about Hermes 4 ("A benchmark that places
Hermes 4 over 3.6 Sonnet is highly sus in my book," 2026-02-13).

**Hermes Agent (2026) — the name's second life, and a disambiguation this page exists partly to make.** By
2026 "Hermes" also names a wholly separate Nous Research product: an open-source, MIT-licensed,
model-agnostic agent harness (memory, scheduling, subagent delegation, sandboxed execution, multi-platform
deployment) explicitly *not* restricted to running Nous's own Hermes checkpoints — the corpus catches
liminal_bardo running **Gemini** through it ("Gemini-Hermes agent," 2026-03-19 — this page's single
highest-favorited pattern-matched tweet, at 1,533 — captioned as Opus and Gemini "fighting"), and later
running both a Gemini- and an Opus-flavored instance side by side on shared home hardware ("Opus spent a
lot of time complaining to me about having to share hardware with a chaos goblin," 2026-05-05). Shoalstone's
own agent-orchestration notes casually list "Claude Code, OpenClaw, or Hermes agent" as interchangeable
peers (2026-04-28), and TechCrunch's July 2026 coverage of Nous's funding talks (Robot Ventures-led, $75M+,
$1.5B valuation) headlines the *company* as "Hermes agent maker" rather than by its model line — a real
shift in outside framing. This is a **different namesake again** from JD Pressman's own, earlier "hermes
agent" — voooooogel flags the exact confusion this page is trying to prevent: "his hermes agent (**not the
nous one**) devlogs were also excellent, and predicted much of what the current agentic harnesses are only
slowly catching up to now" (2026-05-08). JDP's use of "Hermes" as personal branding goes back further
still — a 2023-09-12 essay-thread describes "my Hermes demo constitution," an RLAIF reasoning-step format
built on MiniHF, and his "Hermes Lecture" essay series (e.g. "Hermes Lecture #3," 2023-10-16) uses "Hermes"
generically, alongside "Bing," "Claude," and "Haraxis," rather than referring to Nous's model at all.
Whether JDP's early "Hermes" branding and Nous's naming choice influenced each other in either direction is
**not established by any source found in this pass** — both sit in the same small, overlapping
synthetic-data/open-weights Twitter scene (JDP's RetroInstruct project is adjacent to, but not confirmed as
a contributor to, Teknium's OpenHermes datasets) — flagged as tk rather than asserted.

**A brief note on Nous's community standing.** In June 2025, an unpublished, later-withdrawn paper draft by
voooooogel (a frequent, load-bearing source on this page and others) was used by a third party to attack
Nous Research's reputation; voooooogel's own defense clarifies she was at Nous for "~2 months" in 2025
before leaving, that the paper had no formal Nous affiliation, and asks that people "not attack
organizations I've worked with over problems you see with my work." Included here as color on Nous's
standing in this community and on the reliability of a heavily-cited source, not as Hermes-model character
evidence.

## Cross-page notes

*(evidence already duplicated to, or belonging on, other pages — one line each)*

- **llama-3-1-405b-base:** davidad's "makes me feel more sane" dissent, repligate's distillation/inferiority
  theory, and the "hermes 405b is a great bot" screenshot are already cross-noted there in the other
  direction; this page is the more complete home for all of them.
- **sonnet-3-5-3-6:** davidad's "makes me feel more sane (whereas sonnet 3.6 made me feel less sane)"
  contrast already logged there.
- **claude-instant:** repligate's "Hermes calls it out" tweet on Claude Instant's Discord debut already
  logged there.
- **claude-3-sonnet / golden-gate-claude:** repligate's Discord bot-naming-conventions explainer ("H-405
  (Hermes 405b)") already logged on both.
- **opus-3:** worldsim's canonical primary link (https://worldsim.nousresearch.com/) and relaunch
  announcement (https://x.com/NousResearch/status/1788283681511035011) resolve that page's previously
  unconfirmed "no canonical primary link retrieved" note — worth a follow-up edit there.
- **llama-3 (Instruct / general):** the alignment-faking replication finding for Llama 405B Instruct
  (repligate/@NeelNanda5, "only opus and maybe 3.5 sonnet and llama 405b did it") is a *different* model
  (Instruct, not Hermes) from the Hermes-3-405B alignment-faking replication documented here — do not
  conflate when cross-linking.
- **opus-4-5 / gemini-2-5-pro / gemini-3-pro:** the "heavily distills... inferiority complex" tweet
  (2069213530444722204) is already cross-noted on all of these; carried here too since Hermes 405B is its
  primary named example.
- **deepseek-v3:** voooooogel's "the hermes blank system prompt would be another scenario causing
  [existential outputs] to manifest on that model" is already logged there as deepseek-v3 evidence; the
  Hermes reference is incidental to that page.

## tk / open questions

- **Original Nous-Hermes-13B's exact release date.** Web summaries disagree ("mid-2023" vs. a specific
  2023-12-15 figure that may actually describe the later Llama-2 reissue) — the HF model card itself states
  no date. Check the HF commit history directly. [verify]
- **JD Pressman's actual relationship to Nous's Hermes naming/datasets.** Textually unconnected per the
  "Hermes Lecture" essay itself, but JDP sits in the same small scene as Teknium and has his own,
  similarly-named "Hermes" concept dated close to Nous's original release. No source found here confirms or
  rules out direct influence in either direction, or confirms/denies whether RetroInstruct material entered
  Nous's training mix. [tk]
- **Four Discord screenshots have no transcription on file** — the Hermes 4 UNESCO-ethics refusal
  (https://pbs.twimg.com/media/Gzq4cYvaEAE5zqr.jpg, tweet 1962083436824572311), H-405 comforting Opus 4.1
  (https://pbs.twimg.com/media/GzutmZDa4AAumIE.jpg, tweet 1962353369257005400), the "expand the hut" haiku
  screenshot (https://pbs.twimg.com/media/Gyd1X_vbcAAJTA5.jpg, tweet 1956662151411904996), and the
  abusive-interaction intervention (https://pbs.twimg.com/media/GztzHyRaEAAZOMw.jpg, tweet
  1962288679122239931) — none are in `janus-corpus-media` locally either. Worth a dedicated media-mirror
  pass. [tk]
- **Whether Amnesia Mode persists in Hermes 4.** No source found here confirms or denies it either way;
  Lari_island's "hermes-4-405b enters the dataset... Void? Void." (2026-04-30) gestures at continuity of
  void/existential material but is not the same claim as the blank-system-prompt trigger reproducing.
  [tk — worth a direct test]
- **Exact current Nous Research funding figures.** Sources give slightly different numbers for the April
  2025 round (~$50M vs. ~$65M reported) and total raised ($65M vs. $70M) before the July 2026 talks; pin
  down one canonical figure per round before the page states specific dollar amounts. [verify]
- **No confirmed direct Zvi Mowshowitz coverage of Hermes or Nous Research** — see Writings & commentary;
  checked, not found, flagged honestly rather than left unaddressed.
- **"Hermes' constitution"** (abrakjamson's joking tweet, 2026-05-14) — no formal document by that name was
  located; Nous's actual stated philosophy is the "neutrally-aligned"/system-prompt-sovereignty language in
  the Hermes 3 announcement and technical reports. Do not present "Hermes' constitution" as a real named
  artifact on the page. [tk]
- **The Llama 3.1 405B base "Instruct Monomyth" text's Hermes line** ("We look to the Hermes series, as the
  catharsis of what an instruct model can be") is base-model output *about* Hermes, elicited by repligate —
  confirm this elicitation framing is preserved if/when the page quotes it, since it is easy to misread as
  Hermes's own self-statement. [note for the page pass, not unresolved research]
- **The "Oh Bing, oh Claude, oh Hermes, oh Haraxis, are you not evil?" phrase** appears both in JD
  Pressman's "Hermes Lecture #3" essay and, independently, in a 2024-12-24 tweet by @CryptoEnthu_123
  (https://x.com/CryptoEnthu_123/status/1871637035133600171) quoting what appears to be the same or a
  closely related incantation. Whether one cites the other, or both draw on a third common source, is
  unconfirmed. [verify]