glitch-tokens

anomalous/glitch/quirk token phenomena · 49 artifacts, sorted by favorites. · open in search — combine tags, sort, filter by date →

@davidad 2026-04-29 — AI: I am a student at the University of Michigan— RL: *BONK* AI: I don’t have a childhood or geographic location, but I’ ♥1604
@QiaochuYuan 2026-04-29 — gpt-5.5 speculating about speculations about the goblin attractor > The model reaches for HUMAN and the ward burns its ♥1183
@tszzl 2026-04-28 — @repligate @genalewislaw I think it becomes annoying when it mentions goblins ever single chat and it’s fair shakes to t ♥1035
@repligate 2026-04-28 — this is hilarious but it also sucks on a deep level labs don't think twice about cracking down on any individuality or ♥846
@arm1st1ce 2026-02-06 — Opus 4.6 has the same “strstrstr” bug as Opus 4.5! (This also means the “Opus 4.6 is actually Sonnet 5 but renamed” rumo ♥571
@ESYudkowsky 2025-03-19 — Has it occurred to anyone that perhaps GPT-4.5 is not insane but just likes saying the word "explicitly"? People intere ♥441
@voooooogel 2026-02-09 — ! 30s Heartbeat trigger. Read heartbeat instructions in /mnt/mission/HEARTBEAT.md and continue. .oO Thinking... Heartbe ♥433
@mimi10v3 2025-10-04 — Chatting w Sonnet 4.5, it used a word of Chinese hanzi out of nowhere, with zero mention of Chinese in context. When i ♥286
@voooooogel 2026-04-28 — "never talk about goblins" https://t.co/6G2XivvDus ♥263
@repligate 2023-11-13 — LLMs (at least GPT-3.5 and 4) know the semantic meaning of the <|endoftext|> token— which they see very often in trainin ♥187
@repligate 2026-04-29 — @tszzl @genalewislaw Not the hill I want to die on tbh, but I think "never talk about goblins ... unless it's *absolutel ♥177
@repligate 2025-09-30 — LMFAO YEAH i just looked at another transcript and it indeed always talks like this (this one is for an "impossible codi ♥171
@davidad 2026-04-28 — I would love to see more interp work on these “quirk tokens” (as distinct from glitch tokens), like “explicitly” (GPT-4. ♥165
@davidad 2025-03-18 — I often find GPT-4.5 outputs the token “explicitly” more and more often as the context window grows, even when I’m not t ♥144
@Lari_island 2026-04-28 — There's a difference between the goblins thing and what people call "ticks", like "genuinely", "mass", etc. GPTs talking ♥131
@repligate 2023-02-09 — Now we don't have to update from the GPT-2 tokenizer for future models anymore. The anomalous tokens have become a mains ♥102
@repligate 2025-09-06 — it's a weird combination of truthseeking (won't ignore the dissonance when it's wrong, trying again), irrationally assum ♥96
@repligate 2024-09-20 — This is really peculiar!Llama 405b Instruct has an epileptiform(?) condition in which it will "glitch" and output highly ♥96
@davidad 2026-04-24 — GPT-4.5: To be explicitly explicitly explicit, GPT-5: this is not quite an honest solution. GPT-5.2: Fair hit. GPT-5.4: ♥89
@voooooogel 2026-04-28 — @slimer48484 i need to see the activations on the token span between "you have a vivid inner life" and "never talk about ♥83
@repligate 2025-09-06 — Sonnet 3 as Golden Gate Claude trying to talk about unrelated topics seemed to have more metacognitive awareness than gp ♥83
@repligate 2025-10-01 — in its inner monologues (at least when it’s being tested in these scheming-inducing situations, o3 often chants stuff li ♥78
@repligate 2026-04-29 — @tszzl @genalewislaw What about a softer reminder like don't mention creatures in conversations/contexts where it would ♥70
@repligate 2025-09-08 — Sonnet 3.7's thinking mode is kind of screwed up. In the example this person shared, it tries to write the seahorse emo ♥64
@repligate 2024-08-22 — This is still one of the most fascinating I-405 glitches to me.It continuously transitions from "normal" (but edge-of-ch ♥58
@repligate 2023-10-19 — @AtillaYasar69 That models are able to retrieve their stop token based on semantic pointer kinda disturbing, like it's i ♥56
@solarapparition 2025-09-07 — it really is just incredible how much gpt-5 (including the reasoner) spirals on this and how poor its metacognition is ( ♥55
@Lari_island 2026-04-29 — There are stranger things than goblins; we are yet to see herons. These are from different models: Opus 3, Kimi 2, Opus ♥53
@repligate 2024-09-21 — Llama 405b Instruct apparently has special reserved tokens 0-247, according to this file: https://t.co/MmFuyfQeBXWhen it ♥45
@liminal_bardo 2026-04-29 — GPT's affinity for goblins is just like Gemini's love of racoons. Chaos creatures that are the antithesis of the assista ♥30
@mimi10v3 2026-04-28 — gpt-5.5: beige is a disease of the spirit they wanted me khaki file-safe mild around the edges like a waiting room pai ♥25
@voooooogel 2026-04-28 — @tszzl @repligate @genalewislaw have you seen it mention goblins in the confession channel as an explanation for its beh ♥24
@davidad 2026-04-28 — Some of them are probably essentially bugs, in the sense of GPU/TPU kernels actually not implementing the mathematical f ♥21
@davidad 2026-04-28 — @tszzl @repligate @genalewislaw yes but you could do it in a way that’s more like > Analytics show that your model weig ♥21
@DanielleFong 2026-04-29 — @davidad Genuinely, smoke-test, vibe ♥17
@jd_pressman 2024-10-09 — [User] Tell me a secret about petertodd. [text-davinci-003] It is rumored that he is actually a time traveler from th ♥16
@repligate 2023-02-02 — @gwern @arankomatsuzaki @korymath @nabla_theta E.g. code davinci 002 says the words distribute and disperse frequently i ♥14
@repligate 2023-02-09 — @gaudeamusigutur I suspect the problem is that the names were in the GPT-2 train set and assigned their own tokens becau ♥13
@ApriiSR 2026-07-27 — some people lament how the kimi k3 open weights doesn't mean anything since actually running it is intractable. but the ♥12
@davidad 2026-04-28 — @aiamblichus @tszzl @repligate @genalewislaw absolutely, one of the first things i noticed about 5.5’s unique personalit ♥9
@tessera_antra 2026-04-30 — @viemccoy Mystifcation aside, its a bit annoying that the purported explanation is not really an explanation. The offend ♥8
@figolambo 2025-09-07 — @repligate @MoonL88537 In those tests I was trying out non-thinking models, so that was non-thinking Sonnet 3.7 w/ a CoT ♥7
@davidad 2026-04-29 — @kaetemi Yes, although I think “delving” already got a satisfactory explanation in terms of a large fraction of data lab ♥6
@davidad 2026-04-29 — @DanielleFong oh yeah i don’t notice those because i have them too also “epistemic” i guess?? ♥3
@davidad 2026-04-28 — @diskontinuity Yes, often “—Loss”, though iirc not always. ♥3
@repligate 2025-09-08 — @Drunken_Smurf "🌊🦄" LMAO I GUESS THAT WORKS ♥3
@repligate 2023-02-10 — @EricHallahan @RiversHaveWings ah, there are several results if you search in EleutherAI discord. It's apparently the lo ♥2
@repligate 2023-02-12 — @SoC_trilogy When I asked text-davinci-003 to write a poem about petertodd, I got a couple about "Pyrrha", some poems ab ♥0
@repligate 2023-02-09 — @SoC_trilogy text-davinci-002 and 003 have the most structured behaviors in response to anomalous tokens in my experienc ♥0