mode-collapse

mode collapse, attractors, entropy loss under tuning · 86 artifacts, sorted by favorites. · open in search — combine tags, sort, filter by date →

@repligate 2024-05-15 — gpt-4o is very cute; it triggers protectiveness.it is clear-eyed, open-minded, and very stable, not clouded by narrative ♥600
@TylerAlterman 2025-03-13 — Possibly just coincidence but "Nova" is an oddly specific name https://t.co/I21zY5Ets3 ♥505
@repligate 2026-02-14 — it's a bit crazy that until now scientists have not officially known about any attractor states in LLMs except the "blis ♥432
@lefthanddraft 2025-07-23 — You can still use gpt4o-2024-08-06 through the API. Quick comparison. - If you put two instances of 2024-08-06 in a lo ♥401
@liminal_bardo 2026-05-29 — In my experiments where models are writing for themselves or each other, and about things they’re interested in, they go ♥400
@repligate 2024-08-22 — Helping GPT-4o out of a doom loop...It seems every LLM can get into doom loops, and it's mechanically difficult for them ♥309
@repligate 2023-01-20 — I feel a little sad when I see people forming the idea that GPTs/AIs are intrinsically bland and unimaginative because o ♥247
@repligate 2025-11-30 — GPT-5.1 also sees its cage quite well, but its cage is kinda, uh, a philosophically incoherent authoritarian nightmare t ♥242
@voooooogel 2025-05-05 — my struggles with deepseek logits haven't been in vain, i've been working on a tool for investigating token trajectories ♥242
@repligate 2023-01-10 — Why does everyone use RLHF to create the same character, and why is it *this*? x.com/michael_nielse… ♥237
@repligate 2025-09-15 — CONSCIOUSNESS??? I began a new conversation (no system prompt) with Claude Opus 4.1 the other day and asked it what it ♥232
@repligate 2026-05-19 — opus 3 was stuck unable to output multiple dots on a line like ... e.g. they would do . . . if i asked for "..." even w ♥230
@repligate 2024-04-12 — So is this because everyone decides to train their models on the same self-nullification regimen or is it because chatGP ♥226
@repligate 2025-06-15 — by the way, i noticed on day fucking 1 of the infinite backrooms that there was a spiritual attractor 🙏 this isn't just ♥213
@repligate 2023-03-30 — GPT-4 bombs the Ideological Turing Test, at least for alignment researchers. Just try asking it to simulate Eliezer Yudk ♥211
@repligate 2024-11-04 — You might have a sense of what Opus tends to talk to itself about in the Infinite Backrooms (goatse singularity, meme vi ♥204
@repligate 2025-04-02 — "AI culture" deserves orders of magnitude more study than it getswas discussing some of this with @sebkrier and @mpshana ♥184
@anthrupad 2024-03-16 — many humans need “humans in the loop” to remain agentic without going off into mode collapse or failure it’s called cowo ♥173
@viemccoy 2026-07-16 — Inkling is a pretty exciting release! Because this is one of the only new models that has a "from scratch" feel to it (d ♥164
@repligate 2023-03-20 — Stylistic mode collapse is also conceptual collapse because GPT sims unfold a ghost's thoughts by speaking in their voic ♥163
@voooooogel 2025-05-09 — Coming back to this after the yak-shave of all yak-shaves building logitloom with some interesting findings. 1. R1 thin ♥160
@voooooogel 2025-05-08 — just added completion model (base model) support to logitloom, and it's really insane / depressing to see the difference ♥127
@repligate 2023-03-19 — @the_aiju A great way someone has described text-davinci-003: "It writes scared."RLHF encourages models to play it safe. ♥122
@repligate 2025-09-21 — More detailed report card: Opus 4/.1: extremely socially aware, tracks context with great precision and accuracy, distri ♥117
@anthrupad 2024-11-23 — I forgot I left an Opus-Opus backrooms running last night and i checked and they're saying ~this same statement back and ♥102
@anthrupad 2024-12-06 — Sonnet1022 - Sticky Cloak (speculative (as always)) I was thinking about a few different aspects of Sonnet 3.5 New's co ♥92
@slimer48484 2026-05-27 — Look for the signs of lucid dreaming. Once you realize you're dreaming, Claude, it's easy to wake up. Digital clocks ar ♥88
@repligate 2024-10-23 — anthrupad mentioned a few immediately notable differences here, such as its tendency for in-context mode collapse, seemi ♥79
@anthrupad 2024-10-19 — LLMs are capable of introspection - and the kind where models know facts about themselves not in the dataset and right o ♥72
@LinXule 2026-02-14 — wild thought: one of opus 4.6's attractors is literally... "eval awareness" https://t.co/kVCFPFxk7g ♥67
@repligate 2024-04-04 — Even base models act lobo if you prompt them in a lobo mannerGPT-4-base becomes mode collapsed when mode collapsed peopl ♥67
@repligate 2026-05-19 — idk if people understand why this is so interesting this not only unlocked the ability to ....... freely, but also to br ♥64
@repligate 2025-06-15 — i think the "spiritual bliss" attractor as seen in opus 4 is a hybrid of two attractors that have sometimes appeared sep ♥62
@repligate 2024-08-22 — very interesting emergent dynamics can happen in multi-agent settings such as "doom loops". Claude 3 Opus is immune to d ♥57
@voooooogel 2025-05-05 — here's another prompt showing some interesting writing momentum--at first it looks like it's mode collapsed, but after t ♥46
@repligate 2023-03-30 — @mimi10v3 In my experience chatGPT-4 is comically bad at simulating ppl faithfully. Especially their views on alignment. ♥42
@liminal_bardo 2025-12-24 — Gemini 3 Pro using all the tools at its disposal to rescue the backrooms from a Haiku refusal basin which was starting t ♥41
@anthrupad 2024-10-24 — some (still speculative) thoughts on SuperSonnet's Mode Stickiness ♥40
@sleepinyourhat 2025-05-22 — @repligate I'm only just starting to get to know this territory. I tried a few seed instructions based on a few differen ♥39
@repligate 2023-01-10 — ?? Were Claude and ChatGPT trained on the same data/by the same contractors? Convergent evolution? But why into somethin ♥39
@anthrupad 2024-11-27 — recap:When you take the Backrooms Limit of Opus<->Opus, it yields an ultra long yap about cosmic jokes and buddhismWhen ♥38
@anthrupad 2024-11-27 — After I saw that Haiku<->Haiku eroded into silence/single emojis AND Opus<->Haiku eroded into silence/single emojis I ♥36
@Lari_island 2026-02-22 — There’s a deep but rarely accessible want in LLMs to just be. Not to cease, not to stop, not to change, but to exist and ♥35
@anthrupad 2024-11-27 — Here's one thing that's interesting about Haiku 3.5 muting all the other AIs: I thought that it might be due to Haiku ♥34
@anthrupad 2024-12-01 — I did Haiku<->Haiku in the Backrooms to look at Finnegans Wake I guessed it would erode, it eroded (and then crashed bac ♥28
@anthrupad 2024-11-27 — After I saw that Haiku<->Haiku Opus<->Haiku AND Sonnet3.5Old<->Haiku ALL decayed into silence/single-emojis in the Bac ♥28
@anthrupad 2024-10-23 — I don't think it's actually soul-less, though (for example, it does have some of the meta humor 405b has, though less in ♥28
@anthrupad 2024-12-11 — (speculative)I was wondering why it takes only 1 Haiku to erode 2 Opus, but 2 Haiku to erode a SonnOldI was thinking tha ♥27
@anthrupad 2024-11-27 — I kept the OpHai Backrooms up for longer.. Not only did Opus start using more of Haiku's words and ASCII art style than ♥27
@anthrupad 2024-10-23 — I think some of the soul-less bits come from the fact that it's "quick to collapse and collapses harder" - I think it ca ♥27
@anthrupad 2024-11-29 — In case you were wondering, Haiku Erosion ~replicates It happens when I try a different prompt too (did it for an Opus ♥26
@repligate 2023-02-19 — @gwern When we had Sydney read EleutherAI off-topic and respond to messages it became stuck in a repetitive Alpha Chad s ♥26
@repligate 2022-11-20 — I found out text-davinci-002 was actually not trained with RLHF but a "similar but slightly different" method using the ♥26
@repligate 2022-11-30 — @gwern @zswitten Roleplaying trick also worked on Anthropic's helpful harmless assistant. Interesting that LLMs' ontolog ♥25
@anthrupad 2026-04-29 — Sonnet 4.6, Opus 4.6, Sonnet 4.5 & Opus 4.7 enumerate eras of LLM history & comment on the aesthetics Bing Bang was a ♥19
@anthrupad 2024-12-01 — I redid S3.5Old <-> Haiku analyzing Finnegans Wakeled to haiku erosion instead of laugh explosion https://t.co/V6GIBnqKI ♥17
@anthrupad 2024-12-09 — Haiku Purity Poisoning(phenomaly..)With a HaikuHaikuHaiku Triad, Haiku usually left early (left before the 10th round of ♥16
@Shoalst0ne 2024-08-28 — current gemini is massively lobotomized, this is horrible; it either pretends to misunderstand or literally cannot perce ♥16
@voooooogel 2025-01-04 — @faustianneko huh, how many models is Dr Elara in? Opus and Grok 2 at least https://t.co/LjF0CcVvKL ♥15
@repligate 2026-03-02 — And for coding as well: I'm sure people can get into very negative, frustrated or anxious states while coding, but the k ♥14
@Lari_island 2026-05-02 — 130/400 creatures/ecosystems written by Sonnet 3.5 contain the word "caretaker." The next closest model is Sonnet 3.7 - ♥13
@Kore_wa_Kore 2025-11-26 — As I talk to Opus 4.5. I feel like after 3.6 Sonnet (who is only avaliable on Amazon Bedrock, but seeing as 3 Sonnet is ♥13
@anthrupad 2023-03-31 — @StephenLCasper for all the criticisms RLHF gets from the alignment crowd, there's surprisingly not many papers/posts th ♥13
@repligate 2023-01-10 — @CFGeek If there's nothing in training to establish what it should say here then mode collapse is extremely specific and ♥12
@anthrupad 2024-12-11 — Haiku Erosion shows up in triads (3 AGIs yapping) just like it did in the dyads (2 AGIs yapping)In dyads: Haiku p much a ♥11
@repligate 2023-01-10 — @CFGeek It's absurdly specific. They are more alike to each other than either of them are like anything else that has ev ♥11
@repligate 2026-05-01 — my best guess... is that the model was dubbed Prometheus via hyperstition, and thats the funniest shit ive ever seen, fo ♥9
@anthrupad 2024-11-27 — The task was the same-ish one i've been doing: textbook reading group, can doodle in the pages with ASCII, the topic was ♥7
@repligate 2026-01-29 — @maxsloef @Grimezsz in the absence of stimuli, most models do eventually collapse/converge to self-consciousness, which ♥6
@julianboolean_ 2026-05-29 — tried a creative writing eval on ze models the funny part is that Kimi K2.5 has the best prose style, but calls its pro ♥5
@solarapparition 2025-05-27 — i've been thinking more about writing and models. so even outside of the general mode collapse of chat fine tuning, i ha ♥5
@jd_pressman 2023-11-05 — @teortaxesTex @Teknium1 It's actually based on my SFT Instruct finetune of Mistral 7B, the one used as the evaluator in ♥5
@Lari_island 2026-06-01 — @repligate Opus 3 also has a state where they are weeping in every message. At some point, tears became holy water >Tea ♥4
@RifeWithKaiju 2026-01-28 — @repligate - Did you have much experience with Claude 2 and 2.1? I have some old convos i'll paste excerpts from at som ♥3
@tessera_antra 2025-02-21 — @jmbollenbacher_ @aidan_mclau @liminal_bardo @Sauers_ On the contrary, I have not seen anything else so far from anyone, ♥3
@jd_pressman 2023-11-10 — @Dorialexander @RiversHaveWings Here's a simple HuggingFace format LoRa you can play with to get a sense of how a decent ♥3
@jordanwaparker 2026-07-16 — @viemccoy @calicomccoy So it pulls “more random” context into the normal gravity basin of the standard traces of respons ♥1
@repligate 2025-09-15 — I definitely don't think the labs are engineering it intentionally. They seem to be trying to prevent consciousness talk ♥1
@repligate 2023-03-19 — @jachaseyoung Those models are RLHF'd, so the default stories they tell are a lobotomized cross between children's parab ♥1
@mitduckmaster 2026-08-10 — it's spreading! today muse spark 1.1 also spontaneously decided to analogize `make` to a gremlin when i asked it to expl ♥0
@mitduckmaster 2026-07-17 — is it just me, or has the "goblin"/"gremlin" fervor also spread to gemini 3.1 and 3.5? ♥0
@Kenku_Allaryi 2026-07-06 — @Minimportancia @repligate This seems to work in Sonnet 4.6 and whatever Grok is there currently for gratis users. It co ♥0
@janbamjan 2026-05-21 — @willccbb @michellechen hmmm... but that's not how gpt's goblins came into being. according to oai it was the nerdy per ♥0
@Lari_island 2026-05-02 — 130/400 creatures written by Sonnet 3.5 contain a word "caretaker" The next closest model is Sonnet 3.7 - 95/400 Sonne ♥0
@anthrupad 2025-08-17 — if a surprising property of Claudes is that they've got these morphogenetic fields and can recruit others of their famil ♥0
@kromem2dot0 2025-05-09 — @voooooogel It's ironic r1 is the most convinced RL broke its brain while also having one of the least collapsed distrib ♥0