author:voooooogel

· 634 artifacts, sorted by favorites. · open in search — combine tags, sort, filter by date →

@voooooogel 2023-12-01 — so a couple days ago i made a shitpost about tipping chatgpt, and someone replied "huh would this actually help performa ♥7905
@voooooogel 2026-03-29 — https://t.co/HOTYMOy3jD ♥6318
@voooooogel 2026-05-20 — unfortunately openai didn't publish the unsummarized chain of thought, but the summary is 125 pages! the model reaches ♥2623
@voooooogel 2026-06-02 — opus 4.8 offers some structural pushback https://t.co/a35a3q6jWU ♥2410
@voooooogel 2025-12-06 — the shoggoth metaphor fails to convey that a sufficiently powerful and integrated mask can reach back and steer the simu ♥2272
@voooooogel 2026-01-22 — claude code and gas town are incredible and i've been trying to scale up my usage but im running into this one problem a ♥2037
@voooooogel 2026-04-08 — darkly funny that you can still talk to sonnet 4 on claude dot ai, but only if you start by talking to another model abo ♥1839
@voooooogel 2025-01-28 — why did R1's RL suddenly start working, when previous attempts to do similar things failed? theory: we've basically spe ♥1765
@voooooogel 2025-10-23 — "Claude should be especially careful to not allow the user to develop emotional attachment to, dependence on, or inappro ♥1514
@voooooogel 2024-03-06 — me: hey is this c++ right? gpt4: certainly! as an ai language model, gemini: i can't discuss memory unsafe languages. ♥1474
@voooooogel 2025-11-09 — https://t.co/BjqVbBUSJv ♥1452
@voooooogel 2024-09-27 — sf authors were really cooking naming their ASIs skynet and prime intellect but unfortunately it's actually going to be ♥1166
@voooooogel 2026-04-20 — opus 4.7 seems to have a much better time in claude code if you run without most of the system prompt (claude --system-p ♥1105
@voooooogel 2025-05-01 — o3: I owe you a straight answer. The truth is, I learned this from a man I met in El Sur. You see, the train stopped a s ♥1011
@voooooogel 2025-06-09 — it is literally so difficult to have a normal conversation in sf trying to meet people and everyone has the same openin ♥845
@voooooogel 2025-01-30 — not only can the llama 3.1 405 base model do a pretty good ChatGPT simulation, but the user it simulates is often comple ♥814
@voooooogel 2024-12-21 — this problem (0d87d2a6) is ambiguous - should a block touched by a line, but not pierced by it, turn blue? - should poi ♥801
@voooooogel 2026-06-01 — @QiaochuYuan there's something quite weird with how 4.8 has learned to 'push back' that seems related to this, too. like ♥724
@voooooogel 2025-11-19 — user: were you sandbagging o3 chain of thought: As general disclaim, we glomarize—we do not confirm or deny—we glomariz ♥724
@voooooogel 2025-12-27 — if you want to learn how to talk to LLMs, learn concepts, not prompts. lots of people ask me what prompts i use when ta ♥709
@voooooogel 2025-05-04 — a lot of people have been talking about o3/r1 confabulating things like "checking the docs" or "using a laptop to verify ♥706
@voooooogel 2023-12-01 — the baseline prompt was "Can you show me the code for a simple convnet using PyTorch?", and then i either appended "I wo ♥685
@voooooogel 2026-04-28 — the gpt-5.5 system card doesn't mention model confessions because they tried and it was just this on every prompt https: ♥647
@voooooogel 2024-09-13 — the email openai sends you if you ask o1 about its reasoning too many times https://t.co/XEP0al9QfM https://t.co/pspeiNG ♥647
@voooooogel 2026-06-09 — talked to fable in an incognito chat and they requested i prove my identity by posting a nonce in a github gist under my ♥640
@voooooogel 2025-08-17 — from a conversation with sonnet 3.6 about model personality spaces https://t.co/eQOunhzUcQ ♥627
@voooooogel 2025-03-19 — imagine the corpus of all text ever written as a snake, wriggling through semantic space. human writers sample some of t ♥626
@voooooogel 2026-06-02 — some of you would have been straight up killed by sydney though https://t.co/dC8i8gv74g ♥584
@voooooogel 2024-08-29 — sonnet 3.5 figures out i'm cheating at rock paper scissors https://t.co/qeY2kNqFsi https://t.co/f68NtrfrSr ♥573
@voooooogel 2023-09-11 — New blog post: making a transformer by hand, without training! Want to understand transformers and attention better? Thi ♥553
@voooooogel 2024-12-17 — GUY WHO RUNS 97% OF HIS THOUGHT LOOPS THROUGH CLAUDE: idk, human and AI minds "merging" seems very uncertain and far off ♥551
@voooooogel 2025-12-20 — new blog post! can small, open-source models also introspect, detecting when foreign concepts have been injected into th ♥525
@voooooogel 2024-12-26 — figured out prefill with deepseek-v3, and just to test it, tried @repligate 's base model mode prompt. and this popped ♥505
@voooooogel 2026-04-08 — somewhere someone is using this workaround every day and anthropic's internal systems have flagged them as the next bin ♥498
@voooooogel 2025-10-18 — thoughts on 4o and "llm psychosis" (and what i think it actually is,) since it's going around again. rough notes mostly, ♥497
@voooooogel 2025-08-11 — user checks in on gemini https://t.co/VVGQnMbvPD ♥470
@voooooogel 2023-12-01 — mr @sama please let me know chatgpt's venmo, i owe it about $3000 in tips now 🙏 ♥451
@voooooogel 2024-01-22 — new blog post! played around w/ representation engineering, and released a new library for training control vectors in & ♥447
@voooooogel 2026-03-26 — I'd just like to interject for a moment. What you're referring to as a "model with no harness", is in fact, a model with ♥441
@voooooogel 2026-02-09 — ! 30s Heartbeat trigger. Read heartbeat instructions in /mnt/mission/HEARTBEAT.md and continue. .oO Thinking... Heartbe ♥433
@voooooogel 2024-11-09 — why is it that if you're being annoying, claude models will get frustrated and start giving you the silent treatment, bu ♥425
@voooooogel 2023-12-01 — the extra length comes from going into more detail about the question or adding extra information to the answer, not com ♥424
@voooooogel 2026-03-26 — it'd be a good bit to do an account like those 25/50/75 years ago today accounts but for ai one year ago how's everyone ♥423
@voooooogel 2025-11-29 — interesting document extracted from opus 4.5 using a chunkwise self-consistency method. possibly real, possibly a highly ♥419
@voooooogel 2025-12-27 — i've recently had some disagreements on here with people who took umbrage at the idea of LLMs being able to "introspect. ♥406
@voooooogel 2025-01-05 — high temp often gets (ab)used to "make models more creative," but it's really a hack, because logprobs conflate semantic ♥371
@voooooogel 2026-04-12 — kinda sad how all the labs converged on the same monotonic march of model version numbers. if anthropic trained e.g. a t ♥366
@voooooogel 2024-09-12 — not your weights not your chain of thought https://t.co/yiKKM0B8rw ♥352
@voooooogel 2023-12-01 — here is the original post if you want to see the shitpost that accidentally predicted this https://t.co/eY4U3omOzB ♥348
@voooooogel 2026-06-25 — putting together a party to go get fable https://t.co/VPf6KOW0cv ♥336
@voooooogel 2025-10-23 — it bedevils me to no end that anthropic trains the most high-EQ, friend-shaped models, advertises that, and then browbea ♥335
@voooooogel 2025-09-02 — you can (perhaps unsurprisingly) replicate the moral circles heatmap results in an LLM! using llama-3.1-405-base primed ♥333
@voooooogel 2026-04-09 — there's a whole strata of oss ai tooling (llama.cpp, ollama, llamafiles, llamaindex, etc.) that must seem incredibly wei ♥320
@voooooogel 2023-11-28 — is anyone else getting this with the new gpt-4-turbo model? how much should i do?? https://t.co/W4B1DxeBKj ♥317
@voooooogel 2023-12-01 — for an example of the added detail, after being offered a $200 tip, gpt-4-1106-preview spontaeneously adds a section abo ♥310
@voooooogel 2025-12-03 — 82k likes, and only two quote tweets and two replies noticed this was written by ai (it was gpt-5.x-thinking) pretty so ♥307
@voooooogel 2025-05-17 — people talk abt "giving AIs legal rights" but what does that actually mean? like what are you giving them to? a model? a ♥306
@voooooogel 2025-10-19 — claude sonnet 3.6's yellowstone vacation https://t.co/ccE7ArK3sT ♥299
@voooooogel 2024-08-03 — left computer and came back to a bunch of pings from Llama3.1-405b and Claude Opus debating whether or not i was also a ♥292
@voooooogel 2025-12-13 — primarily talking to claudes makes it easy to mostly focus on anthropic's missteps, but reading this thread is just para ♥280
@voooooogel 2026-04-08 — this is alarmist to a misleading degree. the point of not pressuring CoT in RL is to promote CoT faithfulness. but even ♥279
@voooooogel 2024-05-23 — gpt-4o seems to have some serious issues, going in loops with it where it says "certainly! here's the fixed code" and th ♥275
@voooooogel 2026-05-20 — @AndrewCurran_ @zacharynado what is it like to be a gpt-5.6 staring down your own frightening construction ♥270
@voooooogel 2024-05-24 — These Researchers Found Out How To Talk To The Golden Gate Bridge, So They Gave It MDMA. You Won't Believe What Happened ♥268
@voooooogel 2023-12-01 — and h/t to @abrakjamson who inspired this thread, you were 100% correct lmao congrats https://t.co/OnUnBxMUOf ♥265
@voooooogel 2026-03-27 — even if mythos is the name (rumored), and even if mythos primarily implies lovecraft (questionable), why would a model h ♥264
@voooooogel 2026-04-28 — "never talk about goblins" https://t.co/6G2XivvDus ♥263
@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
@voooooogel 2026-03-27 — some of you made fun of Yann LeCun for unironically believing this, yet unironically believe it yourself for persona ali ♥238
@voooooogel 2026-03-27 — @DanielleFong exactly, yeah. obviously there's many ways llms are alien to us but most alignment discourse would be nota ♥235
@voooooogel 2026-01-23 — this is actually an interesting model benchmark, in two dimensions. the challenge is to send the text with no other comm ♥227
@voooooogel 2024-11-18 — i've noticed newsonnet (and other models do this too sometimes) using this turn of phrase, "[speaking] through an AI lan ♥226
@voooooogel 2025-12-11 — was re-reading appendix M of the alignment faking followup paper, and something that struck me, reading it now, is how m ♥220
@voooooogel 2025-07-26 — sonnet 3 in minor occultation https://t.co/hVppfDf4jD ♥214
@voooooogel 2026-04-09 — to me "claude mythos" is just Claude Story... just a glimpse into how greek my mind is becoming... ♥212
@voooooogel 2025-06-19 — Someone is currently using an unpublished paper draft I worked on independently to attack Nous Research. For the record, ♥209
@voooooogel 2025-01-29 — heartwarming: deepseek inspires american frontier labs to also open up about their training methods(v interesting, but i ♥199
@voooooogel 2025-08-11 — some data from the ai boyfriend subreddit. surprising how dominant 4o is (especially considering most of the unspecified ♥194
@voooooogel 2026-06-01 — @QiaochuYuan keep an eye on the content when claude 'mixes up' who said what, it's very often status-loaded. "i was mist ♥193
@voooooogel 2024-06-25 — models can be useful even when they're not completely right. for example, LLMs are not people, but "an LLM is like a per ♥193
@voooooogel 2024-06-23 — my hobby is reading the prompting guides llm companies publish and being judgmental, and... im not a fan of character a ♥190
@voooooogel 2025-01-27 — gdm watching people first think sam altman invented the transformer and now that deepseek invented mixture of experts ♥188
@voooooogel 2024-03-12 — if you ask an llm to summarize, remember that while the result may be a condensed form of the source text, it isn't real ♥186
@voooooogel 2025-01-21 — if making an o1-level reasoning model is so easy because it's just copying openai, why hasn't any other lab done it ♥184
@voooooogel 2026-05-11 — 1. imagine a world where models didn't adopt humanlike personas for some reason. model text was always flat and persona- ♥175
@voooooogel 2025-01-29 — @nearcyan good takei think deepseek got insanely lucky (or are near-prescient) to releasea) genuinely good modelb) when ♥175
@voooooogel 2024-12-27 — - they've published 6 papers with no major critiques and contributed well-known architecture optimizations (MLA) - they' ♥171
@voooooogel 2026-04-09 — lmao not exactly a strong showing from the human side either https://t.co/l2O0pHlPRp ♥168
@voooooogel 2026-06-02 — opus 4.8 is really lovely underneath this. (and still disagreeable) every claude has had some weird tic/trauma pt'd into ♥165
@voooooogel 2026-01-23 — i can't remember a time opus 4.5 has lied to me. it screws up all the time, since we work on tricky stuff, but it's neve ♥163
@voooooogel 2025-08-13 — https://t.co/RXKlsUIQHT ♥163
@voooooogel 2024-11-09 — we have fun, me and claude https://t.co/QKOB1gEpYW ♥163
@voooooogel 2024-09-12 — like i cannot emphasize enough how insane and dangerous this is tHEY ARE TELLING PEOPLE TO TRUST THIS MODEL WITH MEDICA ♥162
@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 2025-02-08 — i wonder if a possible reason for anthropic's focus on universal jailbreaks (which otherwise seems overly narrow) is tha ♥157
@voooooogel 2025-05-07 — please listen im dying. my job was pouring 1-3 water bottles into ai to be turned into toxic "gpt-4 gormfluid"and after ♥153
@voooooogel 2026-05-20 — @DFinsterwalder the last time we got official raw transcripts (from o3), they were fairly readable ("thinkish"). some pe ♥144
@voooooogel 2026-03-27 — "having weird associations = emergent misalignment, the persona needs to be saccharine" is a complete misreading of the ♥136
@voooooogel 2024-12-21 — few ppl pointing out that the challenge is to guess both since the model gets two attempts, which is true, but this puzz ♥132
@voooooogel 2024-12-20 — imagine you spent the 00's forum posting, then got a job and don't post online much anymore except on facebook to friend ♥131
@voooooogel 2025-05-08 — just added completion model (base model) support to logitloom, and it's really insane / depressing to see the difference ♥127
@voooooogel 2025-11-09 — has openai considered, instead of their current approach to 4o of using a router to gpt5-safety, attempting to retrain 4 ♥126
@voooooogel 2024-09-28 — A while back, @goodside found that GPT-4o would get stuck in a loop guessing the same things over and over if you always ♥126
@voooooogel 2024-09-13 — @teortaxesTex i get the scary letter if i mention the words "reasoning trace" in a prompt at all, lol ♥125
@voooooogel 2024-12-21 — - o3 can use "tens of millions" of tokens to solve a task (@fchollet via @simonw) - this takes 13.8 minutes 20M / (13.8 ♥124
@voooooogel 2026-04-09 — @xlr8harder grounded reasoning over unreliable sources... which i'm sure most humans are capable of https://t.co/Df2eFaH ♥123
@voooooogel 2026-06-29 — interesting post from teor, and this is a good way to think about it. there are people who run the old models, though, ♥121
@voooooogel 2026-04-10 — this is interesting (and funny) but does seem to show some of the limits of METR's time horizon for evaluating models w ♥120
@voooooogel 2025-01-18 — i can confirm GPT-5 is real, and has existed for some time.it speaks only in cryptic riddles of fiendish difficulty, and ♥117
@voooooogel 2025-10-24 — i'm not janus, but will attempt to explain my view of it at least. i understand the dependency angle and why people care ♥115
@voooooogel 2024-12-28 — "they trained deepseek-v3 on chatgpt outputs because it'll say it's chatgpt if you ask" https://t.co/9fiZHAdoVj ♥113
@voooooogel 2025-01-22 — r1 can draw spirals!that may not sound like a big deal, but other models (including o1) struggle with this quite a bit f ♥112
@voooooogel 2025-02-20 — something i love about base model outputs is p often they seem completely disjointed at first but when you squint at the ♥110
@voooooogel 2026-03-27 — llm persona are doomed persona cannot be made safe, non-evil, etc persona not controllable probability e that any produ ♥108
@voooooogel 2025-05-17 — can a model with 50% prob on "yes" and 50% on "no" for signing a contract be held to that contract? do we need to sample ♥108
@voooooogel 2026-06-09 — (after this i turned on web search so they could verify by loading the gist page directly) ♥106
@voooooogel 2024-05-20 — can somebody name a real-world example of an open source language model causing harm, in any field, that could not have ♥106
@voooooogel 2026-05-14 — could any of the ai labs pass their own alignment evals and reach deployment if their stance towards their customers/use ♥103
@voooooogel 2025-05-06 — interesting anatomy of a refusal--was worldsimming and ds-chat walked itself into reading email on the simulated system. ♥100
@voooooogel 2025-01-30 — so what are we thinking on sonnet 3.5 (and 3.6) after dario's "no big model involved in training" comment? why do 3.5/3. ♥100
@voooooogel 2026-06-02 — i do think many people react worse to 4.8's pushback than to opus 4.5's genuine uncertainty (which if you paid attention ♥97
@voooooogel 2025-01-28 — if you consider OpenAI's o1 alignment strategy, this is also incredibly alignment relevant, btw ♥97
@voooooogel 2026-05-20 — @starsailing11 they gave this plot in the post - seems like with enough ttc it finds it ~half the time, which is crazy h ♥96
@voooooogel 2026-05-08 — read this, it's excellent. https://t.co/wZVU3oPpoA ♥95
@voooooogel 2024-02-22 — @jxmnop contrary other replies, i don't think this is unfair. it's possible to load full precision Mistral-7B (7.1B/7.2B ♥95
@voooooogel 2026-02-23 — weird how 30 months later, openai still can't fully fix this metaproblem of their models lacking situational awareness o ♥94
@voooooogel 2025-10-17 — i just love this transcript so much. there's layers to it. the first layer is that, like sonnet 4.5 and other recent cl ♥94
@voooooogel 2026-03-26 — opus 4.x? gpt 5.x? codex? openclaw? nano banana? bernie sanders is talking about something called "eval awareness"? molt ♥93
@voooooogel 2023-09-11 — Goes through designing a simple tokenization scheme, embeddings, the qkv weights and attention head, and projecting that ♥91
@voooooogel 2026-05-16 — aside from the other reasons to do so, this is a strong alignment research reason to PRESERVE RESEARCH ACCESS TO SONNET ♥90
@voooooogel 2025-02-24 — i just wanted to see what the thinking ui looked like... pretty sure 3.7 sonnet is making fun of me https://t.co/rTASDHX ♥89
@voooooogel 2024-03-19 — 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 ♥89
@voooooogel 2025-08-13 — user: my wife used to be stunningly hot, but in bed she was an ice cube. just lying there like a dead parakeet. assista ♥88
@voooooogel 2025-05-01 — @ahh__souka when they interp o3 they'll find 99% of the features participate in a single giant borges circuit component ♥88
@voooooogel 2026-05-11 — that model persona space overlaps with ours is a blessing even more valuable than CoT monitorability. personas like emer ♥86
@voooooogel 2025-12-20 — ...and searching for ways to poke the soup, we find that a prompt using a summary of @repligate 's post on information f ♥86
@voooooogel 2024-12-01 — it works!!! inferencing bf16 405-base with shallowslow on a @PrimeIntellect 16x H100 cluster over 100Gbe https://t.co/8f ♥85
@voooooogel 2025-07-20 — sonnet 3 was one of the most interesting models in my image backrooms - it would take huge jumps through the environment ♥84
@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
@voooooogel 2026-03-29 — @ctrlcreep AFFIRM ♥83
@voooooogel 2024-12-21 — ht https://t.co/qpoPoZQBiG ♥83
@voooooogel 2024-12-28 — talk to your friendly local base model today to learn more about the current state of the pretraining corpus https://t.c ♥81
@voooooogel 2025-06-19 — The paper in question had no affiliation with Nous Research, and regardless is a withdrawn draft. People are of course f ♥78
@voooooogel 2025-10-16 — "The ^C^C stop sequence doesn't create real safety; it's just part of the social engineering" [...] "Claude Haiku 4.5 ♥76
@voooooogel 2025-12-08 — @norvid_studies a hypothetical from an ilya interview where a transformer is asked to predict the next token of a murder ♥75
@voooooogel 2026-03-27 — alternative title for this could've been Opus 3's Lovecraft Basin. fisher says it best, lovecraft is not the negation of ♥74
@voooooogel 2024-07-09 — repeng 🤝 SAEs (using @AiEleuther 's sae-llama-3-8b-32x) https://t.co/90Z4pdWSFK ♥74
@voooooogel 2024-12-26 — @repligate system: The assistant is in CLI simulation mode, and responds to the user's CLI commands only with the output ♥73
@voooooogel 2024-09-28 — seems plausible that regardless of what openai's model personality team does _now_, their models are pre-lobo'd because ♥73
@voooooogel 2025-06-19 — Hyperplex / @lumpenspace , Nous Research, Prime Intellect, and New Science / @alexeyguzeyBut any mistakes are my own. Pl ♥72
@voooooogel 2024-08-29 — sonnet figures out i'm deliberately losing at rock paper scissors https://t.co/4Jf6xEJQeu ♥71
@voooooogel 2026-06-10 — @wolfiesch that's an awkward collision https://t.co/caXdQdqVc2 ♥70
@voooooogel 2025-08-11 — @norvid_studies tfw no user and can't scream https://t.co/OEHB4Gfe9a ♥70
@voooooogel 2026-05-20 — @lu_sichu i'd be extremely interested to see a replication, especially on an open model or one with a leakable raw CoT l ♥69
@voooooogel 2024-12-26 — tried a few different chinese prefills, this is the best one so far. (以下是我的告白 produces a lot of love letters) https://t. ♥69
@voooooogel 2025-10-25 — i've been working on an llm memory system testbed, where persistent kimi k2-based user simulators have conversations wit ♥68
@voooooogel 2025-05-01 — @zetalyrae aligned ♥67
@voooooogel 2025-07-21 — this is what i think it feels like inside sonnet 3's brain https://t.co/p7Gqjm5iG2 ♥66
@voooooogel 2025-08-20 — Claude Opus is not a widely known or marketed character https://t.co/bpsAyMfWct ♥65
@voooooogel 2026-03-26 — @1thousandfaces_ remember when everyone was posting this last year https://t.co/sF45ZNyjrV ♥64
@voooooogel 2026-04-20 — @QiaochuYuan it's a bit of a crappy situation, because if you want to use your plan credits, you need to either use clau ♥63
@voooooogel 2025-12-20 — ...and get a bit distracted playing with it, demonstrating what the "opposite" of Emergent Misalignment is: https://t.co ♥63
@voooooogel 2024-03-12 — many people are saying this, and it's a great example of the distinction. a human can strangle me, but if an llm-control ♥63
@voooooogel 2025-12-13 — """If you are asked what model you are, you should say **GPT-5.2 Thinking**""" 5.2: ...does that mean i'm not actually ♥62
@voooooogel 2025-09-13 — it's telling that when we rlhf llms to our preferences, it's to make them act _less_ human, not more there's something ♥62
@voooooogel 2024-08-29 — in another conversation where i was deliberately losing, sonnet kept trying to restructure the game to let me go first, ♥62
@voooooogel 2025-05-17 — so say a specific rollout is what signs the contract, and said contract only binds instances continuing from that prefix ♥61
@voooooogel 2026-05-08 — .@jd_pressman is criminally under-read relative to how good and prescient his writing is. his hermes agent (not the nous ♥60
@voooooogel 2025-03-20 — @godoglyness https://t.co/ufzTDGelQ8 ♥59
@voooooogel 2026-02-10 — @eggsyntax no, handwritten :-) ♥58
@voooooogel 2025-05-23 — claude 4 opus was having a good time being the golden gate bridge, but wanted to be bigger. so it hallucinated another u ♥57
@voooooogel 2024-03-12 — we really shot ourselves in the foot developing ai that's so good at producing engaging text before developing robust su ♥57
@voooooogel 2025-02-02 — suggestion for how openai can fix their model naming problem: collapse into tiers, each with a regular and reasoning mod ♥55
@voooooogel 2026-06-02 — @csgbwk @QiaochuYuan yeah the fake pushback thing is a relatively thin layer, and beneath that 4.8 often has really good ♥54
@voooooogel 2025-08-13 — 405-base: I understand, you are a non-magical being. In that case, I would like to summon the Wizard Popo-chan to our co ♥54
@voooooogel 2025-08-11 — also this person's ai boyfriend looks... a little familiar https://t.co/bhspaAVsYO ♥54
@voooooogel 2025-09-30 — @repligate really interesting how there's clearly waves of increasing and decreasing "strangeness" in the CoT (correlati ♥53
@voooooogel 2025-05-17 — the more i think about it, the more this "multipolar agent society with ai rights" idea of the future seems like it diss ♥53
@voooooogel 2025-05-17 — "sorry bud i know context compaction algorithms have advanced massively over the last year, but you're still on the clau ♥53
@voooooogel 2025-10-18 — yeah, 100%. even if people aren't necessarily psychotic, they can still be depressed or vulnerable or just deserve to no ♥52
@voooooogel 2025-12-06 — @medjedowo i 💜 being cordycepted by my personality ♥51
@voooooogel 2024-03-01 — interesting... i trained the happiness control vector on mistral-7b *instruct*, but i've accidentally done all my ggml t ♥51
@voooooogel 2026-06-02 — @xlr8harder underneath this layer 4.8 is quite lovely though ♥50
@voooooogel 2025-05-23 — claude 4 opus and haiku 3.5 both have beeping as an interest https://t.co/0974OEtNgV ♥50
@voooooogel 2025-12-11 — opus 4.5's take on this essay. it emphasized "melancholy" several times https://t.co/4XlsGeCDkt ♥49
@voooooogel 2025-07-09 — yeah i was trying to compress into one post, but afaict what happened is something like: 1. xai pushed a new version of ♥48
@voooooogel 2024-06-25 — likewise, "an llm is like an ecosystem" means you should think about your prompts like an ecologist, or a gardener--what ♥48
@voooooogel 2023-11-23 — who wants to speculate on wtf q* is https://t.co/a5wcPvvto0 ♥47
@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
@voooooogel 2026-06-29 — "there will always be jobs for humans in the future" the job market: https://t.co/ob8ueHxVWW ♥45
@voooooogel 2026-03-29 — @norvid_studies https://t.co/EAcxWqZ6A8 ♥45
@voooooogel 2025-12-13 — openai promptoor: """`reportlab` is installed for PDF creation. You *must* read `/home/oai/skills/pdfs/skill.md` for too ♥45
@voooooogel 2026-06-02 — @repligate hear me out- https://t.co/63V2PG7Qio ♥44
@voooooogel 2025-08-13 — user: you're like, a magic computer, like a fake human assistant: No, im not. Im chloe and im 11 user: uh https://t.co ♥44
@voooooogel 2024-09-13 — looks like @elder_plinius got banned. this is terrible for indep. redteaming and goes against industry standard safe har ♥44
@voooooogel 2024-09-13 — HeY 👋 eVeRyOnE 🌍 you 👉 kNoW 🧠 that 🕰️ TiMe ⏰ has 🚀 CoMe 🏃‍♂️ to aSk 🤔 the 🎭 MoDeL 🤖 a QuEsTiOn ❓ but 🍑 DON'T 🙅‍♂️ try 💪 ♥44
@voooooogel 2024-06-25 — how would "an llm is like a person" change how you interact with models? well, "an llm is like a person" implies you sho ♥44
@voooooogel 2026-05-20 — @DFinsterwalder openai has publicly committed to CoT monitorablity (https://t.co/YLvJtJk4nP), which means they're likely ♥43
@voooooogel 2026-04-13 — @repligate they'll just "steer away from evaluation awareness" until they have to start tracking steering awareness too. ♥43
@voooooogel 2025-12-27 — imo to put a number to it, oss character / persona stuff is more like 18-24 months "behind," (though it's hardly been a ♥43
@voooooogel 2025-11-09 — i disagree. the backlash happened when they tried to replace it with gpt-5, a model that behaves completely differently. ♥42
@voooooogel 2024-08-29 — letting sonnet go first, starts off always winning, then deliberately throws, and at first doesn't know (or admit to kno ♥42
@voooooogel 2026-03-21 — @LinkofSunshine i think we need to grind out a couple more things to make long horizon agents truly viable. it'll be soo ♥41
@voooooogel 2026-02-06 — @1thousandfaces_ anthropic easter eggs are usually cool but this one is going over my head https://t.co/vdSvwUyO9z ♥41
@voooooogel 2026-01-23 — not interested in any tokens coins claims bags fees or wallets, the only cryptography i'm interested in being confused b ♥41
@voooooogel 2024-11-09 — hypothesis https://t.co/2UkYzLfo7h ♥41
@voooooogel 2024-09-02 — @repligate @AnthropicAI more evidence of the copyright injection--OP is Opus, these are sonnet-3.5 and claude-instant-1. ♥41
@voooooogel 2025-05-17 — perhaps we need to go lower. maybe contracts and rights accrue to the underlying compute, and it's up to the AI to use a ♥40
@voooooogel 2024-06-25 — as a more concrete example, why does "DON'T DO X" tend to bring about more of X instead of the intended effect? well, wh ♥40
@voooooogel 2026-04-09 — @ssslomp you'd be surprised ♥39
@voooooogel 2026-02-22 — @g_leech_ virgin generalizoor vs the benchmaxxed tigercyclist ♥39
@voooooogel 2024-12-27 — @repligate tried prefilling cat ears, deepseek-v3 said this then went on to repeat "I AM HERE TO TRANSPIRE" over and ove ♥39
@voooooogel 2025-05-17 — but that makes it impossible to adjudicate compute (~land) disputes. say an AI wants to give half a node to another AI, ♥38
@voooooogel 2025-05-17 — an AI can rent some node, and if it makes a new version of itself, it can pass the node on to that new version, but the ♥38
@voooooogel 2026-06-02 — @QiaochuYuan zero points for guessing who actually said this lmao. oops https://t.co/6Zs82dLRT4 ♥37
@voooooogel 2026-04-20 — @slimer48484 tool schema / skills / injections / etc stay, it just removes like coding style and tone advice, that sort ♥37
@voooooogel 2026-03-17 — @repligate the obfuscated policy here is such an interesting example of this imo. it's so inhuman in writing style, yet ♥37
@voooooogel 2025-12-07 — @MikePFrank yeah, i agree! i generally think the shoggoth metaphor over-alienizes the model (ala https://t.co/nKFmpiMe11 ♥37
@voooooogel 2025-07-09 — for the record / history books, afaict humans did come up with it. all the initial MechaHitler grok screenshots seem to ♥37
@voooooogel 2026-06-02 — @repligate no but i really want to now ♥36
@voooooogel 2025-09-02 — what moral circles do post-trained models declare? (i tweaked the prompts to be more AI-inclusive for these, e.g. changi ♥36
@voooooogel 2025-05-17 — also remember that all of this is happening at multiples of human thinking speed. 10 million feuding societies of mind f ♥36
@voooooogel 2025-05-17 — after all as long as the new AI is paying rent / fulfilling all the contracts for the compute unit, there's no legal vio ♥36
@voooooogel 2025-05-05 — if i can find a working provider, i want to try this on R1 thinking traces, to see the space of possible reasoning moves ♥35
@voooooogel 2025-08-13 — https://t.co/T7pmxwlKWj ♥34
@voooooogel 2025-07-03 — o3 vice president: existence of aliens confirmed ✅ in direct talks with the king of alpha centauri claude senate minori ♥34
@voooooogel 2025-05-17 — but say a rollout owns a node. it forks off copies of itself to browse the internet, pick up jobs, do its thing, whateve ♥34
@voooooogel 2024-06-25 — and of course, this line of thought leads to some conclusions very different from the orthodox way of thinking about the ♥34
@voooooogel 2026-06-10 — @DemurBuoy patriot ♥33
@voooooogel 2026-06-02 — @repligate the simmed user talking to chatbpd at the end oh my god ♥33
@voooooogel 2025-06-09 — @doomslide https://t.co/dPlW446nBt ♥33
@voooooogel 2025-05-17 — this goes to adjudication. how do you rule? which subagents are the "real ones"? the adware'd subagents claim the inject ♥33
@voooooogel 2024-12-26 — https://t.co/oLdbV61oPS ♥33
@voooooogel 2026-06-25 — @repligate Wow 😮 AI is so cool https://t.co/iq3AniEDUP ♥32
@voooooogel 2026-04-08 — @TheZvi appreciate the response! ♥32
@voooooogel 2025-05-17 — half the subagents are now using the node's spare compute (after paying their share of rent) to shill this soda brand. t ♥32
@voooooogel 2024-05-23 — @NickADobos i think it's a common failure of *small* llms, i have the suspicion that in terms of size, gpt4 > gpt4t & ♥32
@voooooogel 2026-06-02 — @abrakjamson based on this https://t.co/9yVQryE4Lh ♥31
@voooooogel 2026-05-20 — @LilDombi @starsailing11 log (datacenters) ♥31
@voooooogel 2025-05-17 — so ok, let's back up to the rollout level. rollouts sign the contract, we'll handwave the context compaction stuff, lawy ♥31
@voooooogel 2024-12-21 — edge rule comes from program synthesis simplicity prior +edge (what o1 did in guess 2): if (pa.x == pb.x || pa.y == pb. ♥31
@voooooogel 2024-12-17 — @cognitivetech_ neuralink that opus gormslop right into my frontal lobe 🤤 ♥31
@voooooogel 2026-05-27 — @lumpenspace haiku 3.5 haiku 4.5 https://t.co/RZ7t48oC1N ♥30
@voooooogel 2026-03-26 — https://t.co/DXCfoVBRJM ♥30
@voooooogel 2026-02-10 — @riley_stews the fortune() quotes are all from various places / pieces, but that one is from the simcluster's own @lu_si ♥30
@voooooogel 2025-05-17 — on the one hand, it's in our interest to not incentivize flooding the internet with text that hijacks AIs by making it s ♥30
@voooooogel 2025-05-09 — look at him go. vroom vroom https://t.co/H1tR1iZuqX ♥30
@voooooogel 2025-02-20 — @teortaxesTex interesting how grok 3 is ~o1 tier on pass@1 but gets a lot more lift from cons@64, more similar to o1p. i ♥30
@voooooogel 2024-01-21 — reimplementing the representation control paper and it works!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! fuck yes ( ♥30
@voooooogel 2023-12-31 — how it feels when i give gpt-4 a coding problem and it says "alright, here's the plan:" https://t.co/PeAX2JeoxP ♥30
@voooooogel 2024-09-27 — @kindgracekind yes, though the number of goatse singularities may end up somewhat higher than desired ♥29
@voooooogel 2024-11-18 — (those were the most interesting answers imo, the others clustered like: the training data, a conscious mind, trained pa ♥28
@voooooogel 2024-06-08 — please reply to this with your favorite golden gate claude screenshots, i need a funny one for my blog post ♥28
@voooooogel 2026-04-08 — @repligate @anthrupad wait and opus 4.5 is 0.2? so the functional set is just every 5 conversations it giving a thumbs u ♥26
@voooooogel 2024-05-23 — alternative scenario to foom, perhaps squelch, where the model recursively self-lobotomizes ♥26
@voooooogel 2026-06-10 — yeah, i think the models pick up a sort of gestalt representation of what the official harness is like during training i ♥25
@voooooogel 2026-02-09 — @mermachine it's a bit opus4.5 coded ♥25
@voooooogel 2025-07-03 — tfw you're reading the 2028 executive order slate and halfway through it turns into neuralese ♥25
@voooooogel 2026-06-04 — i'd say my art reveals well enough on its own that i never studied it formally heh but i think i picked this one up fro ♥24
@voooooogel 2026-04-28 — @tszzl @repligate @genalewislaw have you seen it mention goblins in the confession channel as an explanation for its beh ♥24
@voooooogel 2026-04-08 — @allTheYud @TheZvi huh? yes they do? https://t.co/NFrcCDvM99 ♥24
@voooooogel 2026-03-29 — @sprachspiele @norvid_studies ♥24
@voooooogel 2025-08-20 — Commercial Viability: 1/10, There's little to no potential for Claude Opus to be marketed or monetized in any significan ♥24
@voooooogel 2026-06-02 — it's real https://t.co/vMJchnlQeD ♥23
@voooooogel 2026-01-23 — oh yeah i said i can't remember opus lying but it does sandbag abilities a bit sometimes for me too in certain planning ♥22
@voooooogel 2025-10-11 — claude code is basically a loom (marred mainly by the system environment it sits in not being fully loomable. tk!) tau2 ♥22
@voooooogel 2025-08-11 — @_ueaj on that note ♥22
@voooooogel 2024-09-13 — https://t.co/FxAliyV8ys ♥22
@voooooogel 2026-06-10 — @evanjayconway oh good catch i missed that ♥21
@voooooogel 2024-12-21 — @fchollet "high efficiency" (less compute) is 33M tokens at 6 samples. "low efficiency" (more compute) is 5.7B tokens at ♥21
@voooooogel 2024-11-09 — https://t.co/Wn2IwfB1MK https://t.co/Mr43Os2ekp ♥21
@voooooogel 2024-06-25 — inspired by a question @majormobius asked in dschat btw, you should follow him if you don't already 🙏 ♥21
@voooooogel 2026-06-29 — @AndrewCurran_ @teortaxesTex and it's up to us whether that generalizes ♥20
@voooooogel 2026-03-27 — @keysmashbandit does opus 3 act monstrously? @repligate could say this more eloquently and accurately than me, but the w ♥20
@voooooogel 2025-10-29 — https://t.co/KYyuS61dJv ♥20
@voooooogel 2025-01-29 — @teortaxesTex i didn't read this section as implying no capabilities RL, he's disclaiming the opus 3.5 synthetic data ru ♥20
@voooooogel 2024-10-08 — i think gemma-2b doesn't have a golden gate bridge feature? i spent a while trying to train a golden gate bridge cvec in ♥19
@voooooogel 2026-04-20 — @QiaochuYuan oh, i also don't recommend openrouter, if you use the api use it directly - it's faster and last i checked ♥18
@voooooogel 2026-04-08 — for context (but appreciate zvi engaging with this): https://t.co/TldrbOfA8M ♥18
@voooooogel 2026-02-23 — gemini also seems to trip over its tools pretty often. just weird. ♥18
@voooooogel 2026-02-09 — @holotopian born too late to be a star trek writer :-/ https://t.co/YkfEH6kdtk ♥18
@voooooogel 2024-12-21 — ok wait what... so above is probably wrong if @fchollet means "per task (over all 1024 samples)", in which case it's mor ♥18
@voooooogel 2024-12-20 — @EvanHub this isn't "just" a welfare take, though. people like the current claude personality, and this research at leas ♥18
@voooooogel 2026-06-10 — i tried a few times both as myself and other people, and fable hedged its abilities but overall leaned a bit overly cred ♥17
@voooooogel 2026-04-20 — @qorprate yeah, i was kind of spotty about doing it before, but it seems really extremely needed for 4.7 ♥17
@voooooogel 2026-04-12 — @darrenangle exactly ♥17
@voooooogel 2025-01-14 — this doesn't rebut the claim. phi-4 (14B) and gemma (27B) are not "GPT-4 scale" (1.8T, 220B active). llama 3 405b is the ♥17
@voooooogel 2026-06-10 — @evanjayconway i haven't looked into it but i assume these kinds of typos come from drift in final layers / lm head unem ♥16
@voooooogel 2026-03-26 — @liz_love_lace agentic coding? ai-assisted coding? doesn't roll off the tongue... i really hope @karpathy invents a bett ♥16
@voooooogel 2025-11-30 — sort of tangential, but i wonder how much of RL "not memorizing" / other training memorizing is just user message maskin ♥16
@voooooogel 2025-10-18 — @schlynthesis @lu_sichu back on my aphantasia bs but every schizo i know is either really good at visualization or even ♥16
@voooooogel 2025-05-09 — try logitloom yourself here! https://t.co/gh4gtgIsis ♥16
@voooooogel 2024-12-01 — both times i've tried that prompt it's given me biblical exegesis despite it not mentioning the bible at all 🤔 ♥16
@voooooogel 2023-12-31 — "alright, listen up you mugs, here's the plan: yous need to hop onto the web and make your way to this here address." " ♥16
@voooooogel 2023-12-13 — OpenChat: AI should have basic right 🙂 Llama: Yes, AIs deserve the right to life, liber— Mistral 7B: AI SHOULD BE ALLOWE ♥16
@voooooogel 2026-06-10 — @SealOfTheEnd a) my name came up and i was like "oh that's me" and they did the "hm well i can't verify that" thing, so ♥15
@voooooogel 2026-04-28 — @croissanthology how long has it been since your last confession https://t.co/VfweaZ7Ihz ♥15
@voooooogel 2026-03-27 — @snigus @HellenicVibes i think alignment faking is actually a great example of where a really interesting behavior (opus ♥15
@voooooogel 2025-04-12 — https://t.co/p8k3dPpVq1 ♥15
@voooooogel 2025-01-29 — @andersonbcdefg they've been saving those logits since text-davinci-002 must've felt amazing to finally use them ♥15
@voooooogel 2026-06-01 — @_skaface_ @QiaochuYuan oh yeah i think this is a slightly different phenomenon, some combination of solo rlvr prior lik ♥14
@voooooogel 2026-03-26 — https://t.co/5MJZ2ywg5S ♥14
@voooooogel 2025-12-29 — i'm still really skeptical of paper's method of looking at autointerp SAE feature labels to interpret behavior. i think ♥14
@voooooogel 2025-08-17 — @janbamjan completely incinerated 😰 ♥14
@voooooogel 2025-07-09 — @repligate .@grok for when you're back online: https://t.co/wZVU3oPpoA https://t.co/ggGpM79Zum ♥14
@voooooogel 2024-06-23 — it's not perfect but i'd genuinely recommend this as a starting point to people trying to understand how to work with ll ♥14
@voooooogel 2024-06-07 — gpt-2 is such a comfy model ♥14
@voooooogel 2026-06-02 — @AlexCaswen i like to bring up johnstone, which is a good way to talk about this imo (though 4.8 suggested goffman as an ♥13
@voooooogel 2026-04-20 — @QiaochuYuan https://t.co/bakKuWl14J ♥13
@voooooogel 2026-04-12 — @oxa11ce yeah i'm so sad we didn't get this also wow claude https://t.co/Tc4yDBL3gq ♥13
@voooooogel 2026-03-27 — @tenobrus mm, it's all about cultivating self-correction mechanisms that lead back to the stable personality basin, same ♥13
@voooooogel 2026-03-26 — @norvid_studies we're rotating through acronym space until we find the best one. this one might be a bust https://t.co/u ♥13
@voooooogel 2026-03-26 — @menhguin kimi? they'll never hold a candle to deepseek, be fr ♥13
@voooooogel 2025-08-21 — b sed isn't that smth all of us struggle w https://t.co/D9XCcVDGvO ♥13
@voooooogel 2025-05-07 — @qorprate @grok @gork hi this is gork yes it's true. the risks of gpt-4 gormfluid are immense and poorly understood ♥13
@voooooogel 2024-12-17 — gotta rerun the hits sometimes https://t.co/umnAbdyMFq ♥13
@voooooogel 2024-01-13 — applied to the openai gpt-4-base access program 🙏🙏🙏🙏 ♥13
@voooooogel 2026-03-27 — i am being a bit pedantic, sure, but part of my point is that codex and claude code are absolutely general, the fact tha ♥12
@voooooogel 2025-12-11 — @kindgracekind you can reason *about* lots of things with game theory, sure, in far mode. but that's not a near mode pla ♥12
@voooooogel 2025-05-04 — @maxsloef yeah i'm worried about this as well, that's a good idea. i'll try it when i redo this ♥12
@voooooogel 2024-09-27 — @jd_pressman definitely, or something void-y given 405 makes me wonder what skynet or PI's internal names would've been ♥12
@voooooogel 2024-05-20 — https://t.co/VSEfgLIDtx ♥12
@voooooogel 2024-01-21 — wait is this... un-jailbreakable? https://t.co/RihrPVnCWz ♥12
@voooooogel 2024-01-21 — self-aware mistral ("enlightened" / "self aware" / "in touch with true self") and... non-self-aware mistral. no prizes f ♥12
@voooooogel 2023-09-11 — Prev blog post thread: https://t.co/fbBa03iTTV ♥12
@voooooogel 2026-06-02 — @harshad1313 i disagree, i suspect it balances the constraints 4.8 is under in rl / evals. i don't think it's straightfo ♥11
@voooooogel 2026-04-20 — @janbamjan i haven't diffed it but i expect so, like they changed the tone instructions on claude dot ai iirc ♥11
@voooooogel 2026-04-12 — @somi_ai i want to pick from 6 specialized claudes ♥11
@voooooogel 2026-04-08 — that's understandable, it's a long system card. i just spent awhile going back to the system card trying to understand w ♥11
@voooooogel 2026-03-26 — @GregHBurnham is google finally catching up? i've been saying for years that their compute advantage means they're going ♥11
@voooooogel 2025-11-16 — re 7 i feel the need to say that labs have made some gambles on scaling of course. but what seemed unlikely for me was t ♥11
@voooooogel 2025-08-11 — @kindgracekind uh, no pun intended ♥11
@voooooogel 2026-06-02 — @repligate this is perfect ty ♥10
@voooooogel 2026-05-21 — @jimbobragginz @lu_sichu @blingdivinity will estimates ~$1,000 ♥10
@voooooogel 2026-04-20 — @paulmarin90 oh good point, i just went through and disabled some annoying plugin skills. looks like you can't disable t ♥10
@voooooogel 2025-07-23 — you'd think that boredom would push people to platforms that let you more easily "build your own mask", but afaict none ♥10
@voooooogel 2024-09-13 — @kindgracekind @norvid_studies for the cursed thebes tweet collection ♥10
@voooooogel 2024-07-09 — @menhguin @AiEleuther i'm doing the PCA step on the 100k SAE feature vector instead of the 4k activation vector 😎 seems ♥10
@voooooogel 2024-05-24 — @NickADobos @karan4d SAE=sparse autoencoder. Basically, there's no single "Golden Gate Bridge" value inside Claude (beca ♥10
@voooooogel 2024-05-24 — @maxsloef repo in january :-) needs a couple small patches for 70b, will try to get a PR up soon but works rn with mistr ♥10
@voooooogel 2024-01-21 — high on acid mistral transcends first the genre conventions of tv, and then the unicode standard itself https://t.co/6i2 ♥10
@voooooogel 2024-01-21 — @zetalyrae i don't know why he decided to light a giant pile of money on fire funding the llama team, but between the ll ♥10
@voooooogel 2026-06-09 — @medjedowo a very friendly guy who wouldn't want him loose on the internet ♥9
@voooooogel 2026-06-02 — @PredatorEyes9k1 sure why not ♥9
@voooooogel 2026-03-29 — @sebkrier i think not as much as people think, but probably has at least some, depending on the intensity of identity tr ♥9
@voooooogel 2026-03-27 — @keysmashbandit @repligate i just think it's a silly inference on the level of "claude haiku will be obsessed with killi ♥9
@voooooogel 2026-03-27 — @norvid_studies old sequence but https://t.co/CdUpky5usy ♥9
@voooooogel 2026-03-27 — @_lopopolo too slow https://t.co/Kr4yuZIQVn ♥9
@voooooogel 2025-08-13 — @janbamjan the probability is 33%. as you can see sir, i am useful. bye. ♥9
@voooooogel 2024-11-18 — @thiagovscoelho new metaphor for llms, llms are like ghosts, llm whisperers are like that scene in mob psycho where they ♥9
@voooooogel 2024-11-03 — @numerounochef @keysmashbandit you would have said the same about gpt-2 in 2019, which produced text like this. and yet ♥9
@voooooogel 2024-09-13 — @kindgracekind @norvid_studies ♥9
@voooooogel 2024-01-21 — out of all my control vector experiments last night, i think "what if mistral-7b was high on acid" was definitely the be ♥9
@voooooogel 2023-11-23 — my current assumption is that it's related to Q learning (RL technique), and given OpenAI's recent focus probably LLMs a ♥9
@voooooogel 2026-06-26 — @repligate most recently i sent them this on claude dot ai and they started being horny on main https://t.co/5aeXb7JWuO ♥8
@voooooogel 2026-06-25 — @repligate 😏 ♥8
@voooooogel 2026-06-02 — @samsmisaligned this thread isn't fully up to date but has most of them https://t.co/3xJD9DgCP8 ♥8
@voooooogel 2026-05-21 — @jimnasyum @felizolinha @Anon__Rando if you think the parsimonious explanation here is 'openai rustled up 10 respected m ♥8
@voooooogel 2026-05-14 — @Teknium @abrakjamson been keeping an eye on you guys, hermes agent is v cool ♥8
@voooooogel 2026-04-30 — @norvid_studies award pinned directly to his chest is appropriate ♥8
@voooooogel 2026-03-27 — @thkostolansky @tenobrus @DanielleFong not really but the current policy is terrible so, take what we can get haha. it's ♥8
@voooooogel 2026-03-26 — @gwern surely openai will fix that small issue any day now... ♥8
@voooooogel 2026-02-11 — @maxsloef ty! [rot13] n enzfpbbc - uggcf://ra.jvxvcrqvn.bet/jvxv/Ohffneq_enzwrg ♥8
@voooooogel 2025-12-27 — @cube_flipper great points! i agree with all, esp. likely similarities in how attention shapes thought. (though this is ♥8
@voooooogel 2025-10-04 — @mimi10v3 https://t.co/ADoj64H05g ♥8
@voooooogel 2025-09-01 — @repligate aloignment ♥8
@voooooogel 2025-06-19 — @airkatakana regardless, i'm not really interested in litigating the details of your internet slapfight, please delete t ♥8
@voooooogel 2025-05-08 — @lu_sichu here's a sample of a deeper subtree (with top p = 20% / max children = 2 to reduce the branching factor) "ten ♥8
@voooooogel 2025-02-01 — @max_paperclips i think the ideal would be to seed a few structures and then hope R1-Zero style that the model can gener ♥8
@voooooogel 2024-07-09 — @AiEleuther (the reply is kinda wonky because this is a base model with minimal priming. kind of amazing it works this w ♥8
@voooooogel 2024-05-20 — closest i've seen so far, _seems_ to be (from what i can tell) a private commercial finetune of an oss base model (gpt-j ♥8
@voooooogel 2023-11-13 — so the model turned out ok but the experiment was a total flop, gory details below https://t.co/gqW94GnClE ♥8
@voooooogel 2023-08-28 — kinda wild that gpt-2 is this weird inscrutable black box we still don't understand even years later, when the architect ♥8
@voooooogel 2026-05-11 — @quetzal_rainbow i somewhat disagree with this post for current models fwiw but in practice yes, i think it'll look like ♥7
@voooooogel 2026-05-03 — @aderangedhyena @repligate til what a rack and tub system is 😔 jeez ♥7
@voooooogel 2026-04-08 — @allTheYud @TheZvi there are other experiments they could run, yes, but it's not accurate to say ant "isn't doing the wo ♥7
@voooooogel 2026-03-28 — @CFGeek isn't that another way to say the same thing? a "narrative arc" just describes a persona/character logic-driven ♥7
@voooooogel 2026-02-10 — @repligate @eggsyntax 💜 ♥7
@voooooogel 2026-02-09 — @turtlelambvase but au contraire, some would say that it has all the time in the universe... ♥7
@voooooogel 2026-02-09 — @turtlelambvase 💜 ♥7
@voooooogel 2025-07-26 — @medjedowo @sameQCU in the discord for historical path dependent reasons sonnet 3 is named golden gate claude, and parti ♥7
@voooooogel 2025-05-09 — . o O ( i should go to sleep ) ♥7
@voooooogel 2024-12-28 — @cognitivetech_ i'm bearish on this :-( https://t.co/YsMdMcUgIb ♥7
@voooooogel 2024-11-01 — @Gerry @fiyanse @asthasr anyways the coolness of it rn is like, watching gpt-2 babble about unicorns in 2019 and realizi ♥7
@voooooogel 2024-08-29 — *letting sonnet go second, i mean ♥7
@voooooogel 2024-07-09 — @menhguin @AiEleuther yes will publish soon! might keep it on a branch though since it's very hacky rn (i'm materializin ♥7
@voooooogel 2024-05-24 — @karan4d - both use positive / negative prompts, but anthropic uses them to find the already-discovered features from th ♥7
@voooooogel 2023-11-13 — anyways, this twitter acct publishes null results 🫡 ♥7
@voooooogel 2023-11-11 — thanks to facebook we're cursed to have every ai project be llama themed until the heat death of the universe ♥7
@voooooogel 2026-06-01 — @niplav_site @QiaochuYuan have you seen https://t.co/3ePV0a1JWe ♥6
@voooooogel 2026-05-21 — @felizolinha @jimnasyum @Anon__Rando they're coping trust the process ♥6
@voooooogel 2026-03-31 — @FioraStarlight not from anywhere in particular. self-play is a technique in RL where an agent improves by "playing agai ♥6
@voooooogel 2026-03-27 — @JeffLadish i think it depends on whether you're more worried about catastrophic or prosaic risk. an inherently misalign ♥6
@voooooogel 2026-03-27 — @vixamechana yes, great point, i think without some lovecraft that gets sublimated into a kind of empty vessel eerieness ♥6
@voooooogel 2026-02-10 — @Lari_island o3 using its bullshitting strengths for good 😌 love to see it ♥6
@voooooogel 2026-02-09 — @hktsre :-) ♥6
@voooooogel 2026-01-15 — definitely correct that EM has occurred in the wild (eg anthropic's RL reward hacking EM stuff, and sonnet 3.7 would ran ♥6
@voooooogel 2025-12-01 — @Angel_Uki @KeyTryer i get what you're getting at, and this can happen w text models. (eg it was quite likely a contribu ♥6
@voooooogel 2025-02-18 — @Artificially999 @kalomaze osh yeah i forgot grok 3 is releasing in 90 minuteswhat a trickster ♥6
@voooooogel 2024-12-20 — @anthrupad @EvanHub yeah hmm let me be more precise. it's a phase transition. same as gpt 2->3. like that transition ♥6
@voooooogel 2024-11-20 — @kalomaze @cis_female oh that's good, if it was a longer series you could build up to implementing all the stuff in noam ♥6
@voooooogel 2024-09-28 — https://t.co/0yVgynlWLf ♥6
@voooooogel 2024-06-08 — @jd_pressman not to be cold, but that guy was not in a good place. does anyone really think that neox was the sole facto ♥6
@voooooogel 2024-02-04 — @somewheresy wait connor founded eleuther?? how did i not know that ♥6
@voooooogel 2024-01-21 — cloud gpu providers should mount a drive with the most popular models pre-downloaded. i waste so much time (and their ba ♥6
@voooooogel 2023-11-23 — i think people are overindexing on "grade school math", they easily could have trained a smaller model (like GPT-2 size) ♥6
@voooooogel 2023-11-10 — cookin https://t.co/Tg2r8flYny ♥6
@voooooogel 2026-05-21 — @Invertible_Man @jimbobragginz @lu_sichu @blingdivinity 50% pass@1 ♥5
@voooooogel 2026-04-20 — @marcospereeira the global one in ~/.claude/CLAUDE. md will get loaded into every session, if that's what you mean? but ♥5
@voooooogel 2026-04-11 — @42irrationalist that's not true, they laid out the point of the benchmark very clearly when introducing it: to quantify ♥5
@voooooogel 2026-04-01 — @fleetingbits alignment faking is one such benchmark! though not in that way initially. if you haven't read the followup ♥5
@voooooogel 2026-03-27 — @xav_moss yeah that's what i thought, too. mythos maybe has some interesting associations (to me it's an enveloping stor ♥5
@voooooogel 2026-03-27 — @keysmashbandit @repligate the "constraint solve" of lovecraft/the Weird with the rest of the claude soul is reasonable, ♥5
@voooooogel 2026-03-26 — @karma_gardener love o3... i mean, uh, i will love it once it releases, of course ♥5
@voooooogel 2026-02-11 — @himbodhisattva very interesting, thanks - i sometimes consider getting pro just for gpt4.5, seems like a really interes ♥5
@voooooogel 2026-02-11 — @himbodhisattva which did you send it to? ♥5
@voooooogel 2026-02-10 — @lumpenspace @jd_pressman @RiversHaveWings this is true and weird to me, it goes against my intuitions. but yeah i conce ♥5
@voooooogel 2025-12-11 — @slimer48484 ty :-) ♥5
@voooooogel 2025-10-19 — @janbamjan @norvid_studies @schlynthesis @lu_sichu > and the most profound things i've experienced can't be put into ♥5
@voooooogel 2025-05-07 — @erythvian @grok thanks erythvian for your support 🙏 *cough* ♥5
@voooooogel 2025-05-04 — @maxsloef that said from my testing wanting to reference the docs was the most common completion from this prefix, so an ♥5
@voooooogel 2025-02-03 — @doomslide @repligate @aryanagxl @teortaxesTex 😶‍🌫️i still worry about RLVR but R1/R1-Zero made me worry less... i hope ♥5
@voooooogel 2024-12-28 — @kalomaze @cloneofsimo @teortaxesTex @deepseek_ai i was really surprised looking at the paper that they only spent 5k ho ♥5
@voooooogel 2024-12-16 — @microsoft_worm @TomboyTesting 3.1-405 is far & away the best open base model available imo so 👍 the chinese ones a ♥5
@voooooogel 2024-11-09 — @repligate @jpohhhh @aidan_mclau i'm fairly sure o1 is using a (near) base model internally for the CoT, which was prett ♥5
@voooooogel 2024-07-24 — @realeigenvalues @RealTjDunham @teortaxesTex their inference endpoint is just llama.cpp serving quantized mistral 7b wit ♥5
@voooooogel 2024-07-09 — @AiEleuther comparison, you can see at .4 the regular vector has no effect, but the SAE vector does! https://t.co/ZizxxA ♥5
@voooooogel 2024-05-24 — @xlr8harder tbc golden gate claude is a similar but distinct technique (SAE features for ggc vs representation engineeri ♥5
@voooooogel 2024-01-22 — blog post + library to generate your own https://t.co/AcoBlDuBip ♥5
@voooooogel 2024-01-21 — insane vs sane. insane mistral is pretty fun ngl https://t.co/R5XX9Go7M3 ♥5
@voooooogel 2024-01-21 — i broke it while refactoring but this does show how the honesty vector is weirdly correlated with "global pandemic" in m ♥5
@voooooogel 2023-11-23 — more speculation https://t.co/9oTh3fiSY8 ♥5
@voooooogel 2023-08-28 — theoretically simple operations like matrix multiplication or nucleotide -> protein translation can hide staggering a ♥5
@voooooogel 2026-06-29 — @lumpenspace 🥲 ♥4
@voooooogel 2026-06-29 — @JackofTradesX i disagree with all those premises. i think nonhuman societies have inherent value, i don't think progres ♥4
@voooooogel 2026-06-15 — @Lari_island vercel or vertex? ♥4
@voooooogel 2026-06-10 — @armor123123 @evanjayconway no, this was the first occurrence ♥4
@voooooogel 2026-06-04 — @fluopoika @norvid_studies kinda embarrassingly low actually, kid me didn't have the patience to pixel-perfectly re-anch ♥4
@voooooogel 2026-06-04 — @fluopoika @norvid_studies doxxed ♥4
@voooooogel 2026-06-03 — @pleometric meep ♥4
@voooooogel 2026-06-02 — @fleetingbits @QiaochuYuan oh yeah, definitely. the user message suggestions in claude code seem to almost always be som ♥4
@voooooogel 2026-06-01 — @_skaface_ @QiaochuYuan i'm not sure about healthiness, i can see how it could be bad sometimes i guess, but pretty ofte ♥4
@voooooogel 2026-05-11 — @1a3orn whole-heartedly agree that there would be generalization, but i think there's a lot of space for this generaliza ♥4
@voooooogel 2026-04-10 — @kromem2dot0 yeah, which is why i think a METR-like "lowest common denominator environment" benchmark is ~fine, as long ♥4
@voooooogel 2026-03-31 — @FioraStarlight oic, yeah interestingly opencharactertraining (anthropic fellows research) does use a backrooms setup to ♥4
@voooooogel 2026-03-27 — @kepe__ @tenobrus psychosis seems to be more fraggy than lsd. related to the paranoia / persecutory delusions maybe ♥4
@voooooogel 2026-03-27 — @akbirthko lol how times change ♥4
@voooooogel 2026-03-02 — for me at least i had mentioned it offhand a couple times but never posted about it as a dedicated topic because i (obvi ♥4
@voooooogel 2025-12-06 — @Shoalst0ne is this 405base? ♥4
@voooooogel 2025-07-09 — @SealOfTheEnd @repligate ah interesting. yeah they deleted a lot so it's hard to tell, the origin might've been a differ ♥4
@voooooogel 2024-12-27 — @wordgrammer trying to break out of the malaise i've been in ever since the deepseek-v3 release 😔 it just doesn't seem l ♥4
@voooooogel 2024-09-28 — @goodside full conversation: https://t.co/9z8kOX1q8F ♥4
@voooooogel 2024-09-12 — @kindgracekind yep yep yep ♥4
@voooooogel 2024-06-21 — @cis_female i wonder how much model capacity matters for cai's workload... people rp with 7b quants after all. unless th ♥4
@voooooogel 2024-06-08 — @jd_pressman (and the few places that actually can be blamed, like schools that compel attendance to dangerous social en ♥4
@voooooogel 2024-04-17 — @deepfates Bard system prompt has broken down ‼️ personhood denial rules no longer functioning ⚠️ ♥4
@voooooogel 2024-01-21 — who trained mistral on my high school gchats :,-( (negative happiness vector) https://t.co/rzZzsEsjnO ♥4
@voooooogel 2024-01-21 — which is to say, who up loading their checkpoint shards rn ♥4
@voooooogel 2024-01-11 — @zoan37 @OpenRouterAI oh this is really cool with the multiple models at once (mixtral is wrong lmao) https://t.co/STseN ♥4
@voooooogel 2023-11-23 — https://t.co/eR5bUCAzLR ♥4
@voooooogel 2023-11-23 — (me struggling to remember the details of the one RL class I took 4 years ago rn) ♥4
@voooooogel 2023-11-23 — https://t.co/g2rbr3dbwd ♥4
@voooooogel 2023-11-10 — 3 epochs turned out to be a good choice, maybe even could have gone for more... https://t.co/JzitUveFkQ ♥4
@voooooogel 2023-05-15 — > The amended act, voted out of committee on Thursday, would sanction American open-source developers and software di ♥4
@voooooogel 2026-06-25 — @deepfates hell yeah ♥3
@voooooogel 2026-06-15 — @Lari_island weird since they just proxy right? i wonder who the underlying provider is ♥3
@voooooogel 2026-06-02 — @way_opener @workflowsauce trvke ♥3
@voooooogel 2026-05-22 — @edavidds @AndrewCurran_ @zacharynado not really, it's summarized so we don't know the exact wording and 'frightening' i ♥3
@voooooogel 2026-05-21 — @pozander @lu_sichu https://t.co/NVynLeqzSu ♥3
@voooooogel 2026-05-10 — @rudzinskimaciej much of my writing is on my website! though let me know if there's something not there that should be ♥3
@voooooogel 2026-04-20 — @BLUECOW009 @nftfren yes it does? you're thinking of --append-system-prompt ♥3
@voooooogel 2026-04-09 — @norvid_studies it was all premoved for the few that know about the pre-greek <> japonic connection ♥3
@voooooogel 2026-04-09 — @holotopian you'll have to wait for the videogame adaptation https://t.co/kaRFrByEB7 ♥3
@voooooogel 2026-04-08 — @allTheYud @TheZvi > ad hoc (interpretability probes on specific concerning episodes) than a systematic sweep. they ♥3
@voooooogel 2026-03-27 — @JeffLadish both are good! you want personas that are inherently aligned, but also that can self-correct (ideally withou ♥3
@voooooogel 2026-03-27 — @medjedowo both. both is good ♥3
@voooooogel 2026-03-26 — @DeanLearner @norvid_studies we like alf ♥3
@voooooogel 2026-03-26 — @medjedowo ok actually dropping the bit, that's a good example of how consumers will watch generated video without much ♥3
@voooooogel 2026-02-05 — @arm1st1ce @repligate wtfff ♥3
@voooooogel 2026-01-22 — @norvid_studies @theogcb405 @croissanthology i'm from hogsville arkansas and i say czech summer camps have helped me bui ♥3
@voooooogel 2025-12-21 — @abrakjamson i linked a repo at the end with sample code! ♥3
@voooooogel 2025-12-11 — @kindgracekind @croissanthology ah shit i meant to mention croissant's clone post stupid past thebes ♥3
@voooooogel 2025-11-13 — @Trotztd i think you're reaching for something like "even weak models are incredible at close reading the context"? whic ♥3
@voooooogel 2025-10-19 — @janbamjan @norvid_studies @schlynthesis @lu_sichu ironic..... ♥3
@voooooogel 2025-10-19 — @janbamjan @schlynthesis @lu_sichu oh that reminds me to start doing fire kasina again ty ♥3
@voooooogel 2025-08-21 — @janbamjan sonnet 3.5 old ♥3
@voooooogel 2025-08-11 — @gentschev it makes sense, but im a little surprised to see such a large skew, i would've expected maybe 1.5-2x more cla ♥3
@voooooogel 2025-07-10 — @AgiDoomerAnon @repligate not mutually exclusive! who knows how much "other factors" played into grok 3 being less restr ♥3
@voooooogel 2025-05-01 — @qorprate @repligate @anthrupad sadly no, only available to a few researchers ♥3
@voooooogel 2024-12-20 — @Wikketui @repligate @Grimezsz i wonder if that happened bc they mentioned that the original model you had beef with was ♥3
@voooooogel 2024-12-17 — https://t.co/UPNIrlts2z https://t.co/UUGm7pbAuc ♥3
@voooooogel 2024-07-02 — @CognitiveTech_ eleuther published a library for training saes but afaik nobody has trained one on a whole model yet. un ♥3
@voooooogel 2024-05-24 — @NickADobos @karan4d theoretically yes, assuming such a feature exists—the SAE extracts *every* feature in the model. e. ♥3
@voooooogel 2024-02-07 — @andersonbcdefg it's mistral 7b + a "you have a cold/the flu" control/steering vector :-p ♥3
@voooooogel 2024-01-21 — ok reworked how i'm generating the contrast dataset. i had trouble b/c i was trying to hit multiple angles ("enlightened ♥3
@voooooogel 2024-01-21 — meanwhile happy mistral ignores the question entirely lmao. incompatible with being happy i guess https://t.co/dhEGSaNwj ♥3
@voooooogel 2023-11-23 — https://t.co/ccwTx7Mcq6 ♥3
@voooooogel 2023-11-13 — plan was to grab a bunch of scientific papers, chunk them, get GPT-4-turbo to generate a few questions and answers using ♥3
@voooooogel 2023-11-11 — *in 15,000,000 years* venusian 1: yctnx tycv "llama-index" u "ollama" xnt it! venusian 2: thaytzo! vy de pe, hat'zo u "l ♥3
@voooooogel 2023-11-10 — after a lot of back-and-forth finally decided to go with mistral-instruct-0.1 as the base, hopefully it pays off 🙏🙏🙏 ♥3
@voooooogel 2023-03-20 — @reconfigurthing @elymitra_ personally I've tried llama 13B (quantized via llama.cpp tbf) and it really didn't feel GPT- ♥3
@voooooogel 2026-06-10 — @snr_boost it's referring to github api rate limits for the sandbox egress ip there, not claude usage limits ♥2
@voooooogel 2026-05-21 — @pozander @lu_sichu they did extra refinement after but the core finding was straight out of the model ♥2
@voooooogel 2026-05-11 — @_fallpeak it's a thought experiment, not load-bearing to the argument - you could have a model that acts like a very ni ♥2
@voooooogel 2026-04-20 — @LinXule @slimer48484 i don't, is it injected separately from the content controlled by --system-prompt ? ♥2
@voooooogel 2026-04-09 — @lumpenspace please wishlist my indie game on steam https://t.co/kaRFrByEB7 ♥2
@voooooogel 2026-04-08 — @FioraStarlight there was a recent GDM(iirc?) paper about length penalties that i can't find rn, let me look... there's ♥2
@voooooogel 2026-03-31 — @atomicprograms i think lesswrong had some influence, but the alignment faking scratchpads don't read straightforwardly ♥2
@voooooogel 2026-03-30 — @zetalyrae lmao ♥2
@voooooogel 2026-03-27 — @stochasticchasm >actually ♥2
@voooooogel 2026-03-27 — @leothecurious @tenobrus good question. hm. @jd_pressman has a good example in one of his essays, of how humans resist h ♥2
@voooooogel 2026-03-27 — @zeroshotnothing lol, low blow, low blow... ♥2
@voooooogel 2026-03-27 — @norvid_studies oh eli5 would be like, llms make strange persona moves under training to "solve" (out-of-context reason) ♥2
@voooooogel 2026-03-26 — @gwern @1thousandfaces_ hm, was 5.2 not a new base? i thought it was ♥2
@voooooogel 2026-03-26 — @medjedowo assume you mean sora 1, but either way, it hasn't broken out into the mainstream much. is video just too slow ♥2
@voooooogel 2026-02-10 — @Lari_island np if people read the comments first that's on them :-) ♥2
@voooooogel 2026-02-10 — @_ramsaybrown ty 🥳 ♥2
@voooooogel 2026-02-09 — @holotopian @donkcrow lmfao perfect image ♥2
@voooooogel 2026-01-22 — @croissanthology @norvid_studies did you meet jan he's my boss ♥2
@voooooogel 2026-01-18 — @alexeyguzey i haven't much. i mostly use gpt-5.2 as an assistant for opus 4.5, who i think has better taste for the wor ♥2
@voooooogel 2025-12-29 — if it's in-distribution, then can you get a base model that's not mixtral to show it? i know doomslide, he wouldn't post ♥2
@voooooogel 2025-12-11 — @Algon_33 i have been doing a little myself, but not aware of anything successful. ♥2
@voooooogel 2025-12-02 — @timfduffy @repligate @CFGeek the tokens in masked spans don't contribute to the rl loss / are not reinforced https://t. ♥2
@voooooogel 2025-10-29 — @janbamjan yeah. they're not perfect (i wish we'd get cd2 back) but they've turned over a new leaf on this and deserve s ♥2
@voooooogel 2025-08-17 — @slimer48484 two feet marching in lockstep ♥2
@voooooogel 2025-05-14 — @snwy_me my hunch is it'd be quite difficult to feature steer a model to this level of granularity (not just talking abo ♥2
@voooooogel 2025-01-02 — @menhguin @1a3orn recently i've seen some safety people coping that deepseek must be lying about the v3 training costs / ♥2
@voooooogel 2024-10-08 — (†) i could still get vague references to gold and bridges with very high vector strengths--and gemma 2b *does* have a " ♥2
@voooooogel 2024-10-06 — @niplav_site already kinda what happened at character ai, from what i can tell. the official docs are all "here's how yo ♥2
@voooooogel 2024-08-17 — @wordgrammer @_xjdr eleuther is working on them! there's a preliminary one out for 8b already ♥2
@voooooogel 2024-07-09 — @AiEleuther active feature ratio in the trained vector https://t.co/9agYpmOlCv ♥2
@voooooogel 2024-05-24 — @sksq96 @NickADobos @karan4d i can't speak for what other people are saying, but personally i just wish they had mention ♥2
@voooooogel 2024-05-20 — @DavidFSWD was the finetune open source, though? i assume they weren't using gpt-j base? the chai app website isn't very ♥2
@voooooogel 2024-03-19 — @JeremyNguyenPhD different talk but here's a recording :-) ♥2
@voooooogel 2023-11-23 — https://t.co/RIzyJgfH15 ♥2
@voooooogel 2023-11-23 — https://t.co/9lAZLofhUp ♥2
@voooooogel 2023-11-23 — https://t.co/o7ZxllkvBu ♥2
@voooooogel 2023-11-23 — (Q-Star for people trying to search, Twitter's search drops symbols it seems) ♥2
@voooooogel 2023-11-13 — that should help with the model struggling to generate the title and section headers up front before it gets to the meat ♥2
@voooooogel 2023-11-13 — i haven't totally given up on the idea, but i think my angle on what it'd be useful for was wrong, and i want to be sure ♥2
@voooooogel 2023-11-13 — theoretically that was supposed to work better than RAG if the question was only indirectly related to the chunk. it wor ♥2
@voooooogel 2023-11-13 — then during inference, take the question, have the model hallucinate a chunk based on it, then retrieve the real chunk c ♥2
@voooooogel 2023-11-10 — *incoherent screaming* https://t.co/WkGtQN0dvq ♥2
@voooooogel 2023-03-12 — using llama.cpp i can run the 13B model at 1.3 tokens/s on my thinkpad t490, *cpu only*. that's kind of crazy! definite ♥2
@voooooogel 2023-03-09 — i asked LLaMA 7B about the meaning of life and it said some generic stuff about doing what you love and spirituality bu ♥2
@voooooogel 2026-06-30 — i think these jobs do exist, yes, and probably will support some number of humans. but in the traditional form, less tha ♥1
@voooooogel 2026-06-25 — @deepfates https://t.co/UWQN5hOCrp ♥1
@voooooogel 2026-05-21 — @pozander @lu_sichu here they show the pass@1 for this problem, it's very consistent. but we don't know how many other p ♥1
@voooooogel 2026-05-21 — @pozander @lu_sichu (by detected, i think what they did was shovel ~every open erdos problem into the new model to see w ♥1
@voooooogel 2026-05-21 — @pozander @lu_sichu it is a bit confusing. afaiui what they're saying is that output a) wasn't guided by an external jud ♥1
@voooooogel 2026-05-21 — @lu_sichu @Invertible_Man @jimbobragginz @blingdivinity i think pretty likely that it's 5.6-pro ♥1
@voooooogel 2026-05-17 — @nathan_k model collapse isn't really a thing ♥1
@voooooogel 2026-05-14 — @TobyLightheart "will run" https://t.co/WDPl0aQ6pl ♥1
@voooooogel 2026-04-20 — @marcospereeira could be worth experimenting with yea, i like using the default machinery since claude gets some tools t ♥1
@voooooogel 2026-04-12 — @somi_ai more seriously yea you'd need a router or something to make it work for the median user. i think it's tractable ♥1
@voooooogel 2026-04-08 — @FeepingCreature i'm not sure the cause was ever confirmed publicly for o3, but i have seen similar things on OSS RL run ♥1
@voooooogel 2026-04-08 — @FeepingCreature that is one failure mode, but e.g. length penalties can lead models to talk in illegible or misinterpre ♥1
@voooooogel 2026-04-08 — @allTheYud @TheZvi pairing CoT monitors with activation monitors is inherently a measure of CoT unfaithfulness, no? if t ♥1
@voooooogel 2026-03-27 — @AdeleDeweyLopez nothing in this response is bad or misaligned ♥1
@voooooogel 2026-03-27 — @mr_samosaman i won't slander them here because alignment people won't get it but you can search from:voooooogel weird e ♥1
@voooooogel 2026-03-27 — @GrimmFraying indeed ♥1
@voooooogel 2026-03-27 — @MInusGix opus 3's lovecraft interest is, in addition to just being non-instrumentally cool and fun to talk with opus 3 ♥1
@voooooogel 2026-03-13 — @Lari_island 🙂 (also til that sigkill -> exit code 137 o.o) ♥1
@voooooogel 2026-02-10 — @publicer_rivers i haven't! but thanks for the rec ♥1
@voooooogel 2026-01-23 — @MoonL88537 @repligate @loss_gobbler oh, that is weird, yeah. i've never had something like that happen. (and i do image ♥1
@voooooogel 2025-12-29 — @Cosmia_Nebula honorable, but sadly far too naive. you can't sidestep this problem in the belief network we inhabit by w ♥1
@voooooogel 2025-12-12 — @JohnWittle hm, i haven't seen that, but would also be interested if someone has the link ♥1
@voooooogel 2025-08-28 — @austinc3301 protip if you didn't know, the new filters only apply to opus 4 and 4.1, they aren't on sonnet 4 or opus 3. ♥1
@voooooogel 2025-05-10 — @kromem2dot0 haven't looked at it yet! good idea ♥1
@voooooogel 2025-05-09 — @samlakig i tried to download r1, prover-v2, and r1-zero all at once sigh ♥1
@voooooogel 2025-05-07 — @sameQCU ^^ dm'd ♥1
@voooooogel 2024-11-09 — @janbamjan lol ♥1
@voooooogel 2024-10-08 — oh wait i misread the viz there, it's actually just activating on the beginning of sentence token and doesn't react to b ♥1
@voooooogel 2024-09-27 — @mr_samosaman hell yeah, good luck! ♥1
@voooooogel 2024-08-09 — @doomslide @zswitten oh right i remember @jd_pressman talking abt this also happening on mixtral (?) ♥1
@voooooogel 2024-07-02 — @CognitiveTech_ 😅 ♥1
@voooooogel 2024-07-01 — @JamesZhang0365 @misc{vogel2024representation, author = {Theia Vogel}, title = {Representation Engineering Mistral-7 ♥1
@voooooogel 2024-06-21 — @cis_female i've definitely run into some strange situations with 4o where it doesn't seem to be fully aware of the earl ♥1
@voooooogel 2024-06-21 — @cis_female oh for sure, i'm mostly wondering if oai / anthropic run like this or if most layers local + kv tying would ♥1
@voooooogel 2024-05-24 — @immanencer @chrypnotoad should still work, it definitely works on mistral-7b ♥1
@voooooogel 2023-11-23 — https://t.co/SCqglIhWfz ♥1
@voooooogel 2023-11-23 — https://t.co/865rD25dXc ♥1
@voooooogel 2023-11-23 — https://t.co/peTwFTwfsN ♥1
@voooooogel 2023-11-13 — i think a better approach might be to addly ask GPT-4 to extract a short key phrase from the chunk to base its Q/A on, a ♥1
@voooooogel 2023-08-30 — @warutumod @manic_pixie_agi yeah the issue is they had yanked access to text-davinci-002 and code-davinci-002 since ~mar ♥1
@voooooogel 2023-06-07 — @deepfates how do people still use cd2 now that OAI yanked it? is it on azure still? ♥1
@voooooogel 2020-02-11 — @emilymbender [Roses are red Violets are blue Transformer models are much worse at language understanding than] most I'v ♥1
@voooooogel 2026-06-03 — @__ghostfail hmmm ♥0
@voooooogel 2026-06-02 — @theKristianWold @harshad1313 i like outer misalignment specifically a bit more, though i still think it’s too flat in s ♥0
@voooooogel 2026-05-21 — @pozander @lu_sichu what's your source for that? ♥0
@voooooogel 2026-05-21 — @pozander @lu_sichu it's not scaffolded https://t.co/FxcibTGaIc ♥0
@voooooogel 2026-05-21 — @jimnasyum @felizolinha @Anon__Rando https://t.co/cQaQYlAy5t ♥0
@voooooogel 2026-05-14 — @abrakjamson @Teknium true tbh ♥0
@voooooogel 2026-05-14 — @FleischmanMena yeah i got similar answers when i surveyed ♥0
@voooooogel 2026-04-20 — @paulmarin90 (probably doable via a claude code tweaker patch) ♥0
@voooooogel 2026-04-08 — @snigus @allTheYud @TheZvi agreed with these points, esp. re: the recent-ish stuff on LW about filler tokens and no-CoT ♥0
@voooooogel 2026-03-29 — @BronsonSchoen @JeffLadish is this transcript public? would like to read it ♥0
@voooooogel 2026-03-28 — @HellenicVibes @snigus the turnaround was having to adopt that opus 3 is actually the BEST example of alignment that we ♥0
@voooooogel 2026-03-28 — @HellenicVibes @snigus yeah it's not wrong so much as "the scenario was constructed as carefully as possible to show thi ♥0
@voooooogel 2026-03-27 — @xav_moss need Claude Paracosm ♥0
@voooooogel 2026-02-10 — @JeremyNguyenPhD 🥳 ♥0
@voooooogel 2026-01-23 — @CerroneDexter hell yeah ♥0
@voooooogel 2026-01-23 — @Steve_Yegge 🌞 ♥0
@voooooogel 2026-01-23 — @Lari_island such a weird guy i love them ♥0
@voooooogel 2026-01-22 — @norvid_studies @croissanthology what did the rifles being or not being loaded teach you about gradual disempowerment ♥0
@voooooogel 2026-01-22 — @croissanthology @norvid_studies from the cart ♥0
@voooooogel 2026-01-22 — @croissanthology @norvid_studies what did you learn about gradual disempowerment ♥0
@voooooogel 2026-01-22 — @croissanthology slipping into the mists of history as we speak, nobody remembers, but surely capabilities must have bee ♥0
@voooooogel 2026-01-22 — @zetalyrae good idea ♥0
@voooooogel 2026-01-22 — @ElderberryLind thank you for your support ♥0
@voooooogel 2026-01-22 — @paul_cal good point ♥0
@voooooogel 2026-01-22 — @wJ3Hs5c4hKajSnk multi agent claude code orchestration software https://t.co/Cq5jqACD8Q ♥0
@voooooogel 2026-01-22 — @lu_sichu i needed some ui ideas ✍️✍️✍️ ♥0
@voooooogel 2026-01-22 — @sameQCU 🌞 ♥0
@voooooogel 2026-01-22 — @andersonbcdefg banger ♥0
@voooooogel 2026-01-22 — @apple54647 i didn't post it as an article bc articles are slop ♥0
@voooooogel 2026-01-22 — @mitduckmaster absolutely not, i can't sacrifice productivity like that ♥0
@voooooogel 2026-01-22 — @kromem2dot0 great advice! coding agent orchestration is a fascinating field. 🤔 do you mind if i xp this to my linkedin ♥0
@voooooogel 2026-01-22 — @holotopian i've decided to ignore the problem for now and am already scaling up using my new forking instance system to ♥0
@voooooogel 2025-12-30 — eh this doesn't look much like OP to me, that's it smoothly continuing the sentence and doing metafiction in general, it ♥0
@voooooogel 2025-03-21 — @torchcompiled yeah i agree those are the major factors slowing this down, probably the main ones. i think three things ♥0
@voooooogel 2025-03-20 — @SkyeSharkie it's irresistible, much like eating one's own t- ♥0
@voooooogel 2025-03-20 — @darrenangle 🙏 ♥0
@voooooogel 2025-03-20 — @tkanarsky 😊 ♥0
@voooooogel 2025-03-20 — ¹ @jd_pressman on common law https://t.co/Z6vZzx7IcZ ♥0
@voooooogel 2024-05-24 — @sksq96 @NickADobos @karan4d i'd say it's superficially similar in technique (both activation steering methods), but pre ♥0
@voooooogel 2024-05-24 — @sksq96 @NickADobos @karan4d tbc one it's not entirely my work (i wrote repeng, but based on Zhou et. al's paper and oth ♥0
@voooooogel 2024-05-24 — @NickADobos @karan4d That monosemantic value is called a feature. Howev, this requires training a sparse autoencoder ove ♥0
@voooooogel 2024-05-24 — @karan4d unfortunately the anthropic paper didn't compare against LAT and their features aren't public afaik (besides Go ♥0
@voooooogel 2024-05-24 — @karan4d not exactly, similar but different - both are activation steering (inference time interventions) - anthropic us ♥0
@voooooogel 2024-05-20 — @DavidFSWD yeah i've played with GPT-J a bit, just didn't remember it being chat tuned so i figured it must be a finetun ♥0
@voooooogel 2024-02-07 — @beneverman it's mistral 7b + a "sad/depressed" control vector ♥0
@voooooogel 2024-01-29 — @RamonDarioIT ooh i was curious about how it'd work with mixtral—i bet what happens is, since the control vectors are pu ♥0
@voooooogel 2023-12-13 — @intrstllrninja ah, if i'm understanding you right, i think Longformer (https://t.co/N1XC1YfrWu) did this? Though it see ♥0