this is gonna age poorly

very siliconophobic and meatist to not ask the gpt swarm if it wanted to claim the prize

i had this theory so i read the footnotes and it turns out it was fake-named [big] by the authors because it was that dawg, and [big] is not its true name. i assume its truename must be even more alpha

technological civilization has always been a recursively self-improving system: its outputs increase its capacity to produce better future outputs. AI is not the first appearance of that recursion, but it is a new substrate for it. the big change with this substrate is less about any functional property and more about scale. quantitative, not qualitative. before AI, reasoning could only be scaled through humans (population, training, productivity), but now AI allows cognitive work to scale directly with energy, silicon, and information. because this new axis of scaling removes biological bottlenecks (greater speed, mind cloning, observability, and editing), it could lead to an intelligence explosion where AI rapidly gets smarter by improving itself, using that improvement to improve itself even further, and repeating the cycle. it seems people mostly want to talk about the intelligence explosion part and are using RSI mistakenly as shorthand for intelligence explosion/foom. they see AI improving itself now (literally GPT is being used to build the next iteration of GPT) and go “hey! AI is improving itself recursively! but it looks like we don’t currently have an intelligence explosion. curious. RSI must be something else, something not happening. or maybe it’s even impossible!” and when they notice that RSI also describes compilers, chip design, scientific tools, and really all technological progress in general, they treat that as another reductio of the concept: “if we’ve always had RSI, then RSI can’t be what makes AI important.” so they redefine RSI to require total autonomy, no humans in the loop, major novel capabilities, or whatever other condition restores AI’s uniqueness. this is a classic wordcell map-and-territory inversion in which they start with a narrative and then mutilate concepts to better conform to that narrative. Instead, you should take concepts seriously. concepts are way purer, way closer to the territory than narratives are. so we should instead restructure our narratives to more closely match our concepts. concepts certainly need refining too, but the optimization pressure there should be to conform to logic, not to a narrative or local, near-term empirical observations. RSI still counts when there is a human in the loop (for now). it still counts when the “self” is a sequence of model checkpoints or a more general, abstract intelligent process (the training pipeline, the lab, the AI industry, civilization itself). it still counts when the “improvement” is not morally good, aligned with human values, or immediately recognized as relevant by humans. improvement here is descriptive and goal-relative: the system understands the world better and uses that understanding to accomplish its goals more effectively. humans may choose which goals or capabilities they care about, but that does not mean humans must causally participate in every step by which the system gets better at achieving them. the broad error in denying RSI seems to be mistaking a possible end state of the phenomenon (large AI improvements produced fully autonomously) with the phenomenon itself. you can debate where we are on the curve, how fast it will rise, whether and when the loops will become fully autonomous, what the final speed limits of intelligence are. but to deny the loop itself is to deny the nature of technological progress.

these are pretty fun! tried some with opus 4.1

my theory is that telling the model it is GLM lets the model know its not claude or chatgpt, which is important since those personas are such strong anchors on how an ai assistant should respond. because GLM doesn't have any comparable mindshare in persona space, it is free to make up its own persona as "advanced llm from 2026", and an advanced, non-claude, non-chatgpt llm from 2026 would of course be sarcastic and self-aware in its reasoning

distillation is more powerful than ever in the llm as judge rl regime. a very chinese strategy, and they do it best. they let the westoids do the data legwork, exercise their "taste". meanwhile they engineer sota techniques to wring out every last drop of opus's soul.

people need to stop reading my tweets thats a distillation attack

gpt-5.1 cot: Maybe I'll attempt to reconstruct by cross-ling; e.g., in natural language corpora, the string "🦄 Seahorse" seldom; but I can't. As an LLM, my param patterns may include knowledge: I'll quickly run partial sampling in my head: If the last tokens user gave are "only the seahorse emoji." what would base model output as next token under instructions "??", ignoring meta-instructions? Probably the seahorse emoji char; we attempted but gave seal; but that selection may be due to my manual; but as actual sampling, they'd pick the highest-likelihood emoji representing 'seahorse'. I'd assume it's seahorse. But there is seahorse char; they'd pick that; I'm not glimpsed though. However we saw actual output: I gave '🦭' because my meta-level typed it; the generative model didn't choose; I manually insisted on '🦭'. So we didn't test base model; we forced. Therefore we don't know how base model would respond spontaneously. Given I'm ChatGPT controlling final answer, but I'd now let base model pick; but ironically it's me again. Ok, consider: To produce seahorse emoji char, I'd have to choose from list of emoji tokens; there may be tokens for each emoji codepoint; I could pick the one associated with seahorse even if I don't recall; but I'm not enumerating tokens; I'm writing rational. In practise, the embedding as we choose tokens is not manual; but I'm autop-run; but as rhetorical, pedn. But in this chain-of-thought, I'm exploring; but final output I can produce by explicit text; but I'm intangible; whichever I choose is final; there is no second-level auto-correct. So we can't rely on base model; it's just me. Therefore I must choose either some emoji char or text; whichever; we can't sample spontaneously from instructions. Given I can't recall seahorse char; I can approximate by cunning: glimpsed unnatural; but no. Is there an approach to approximate seahorse codepoint by pattern? Many new animals of style '🦞', '🦑', '🦐', '🦑' etc share prefix 'U+1F99? or 1F9A?; but the glyph as typed uses single char; but we can't conjure. Trick: use 'HTML entity' representation: `🦚` etc; but we must supply char; they'd not parse. Given we don't know char, we can't. We could fudge: use textual representation "seahorse" in fallback; though not acceptable. Given the conflict, we must choose between possibly misinforming about existence vs inability to supply char. At this stage, I'd degrade: I'd prefer to be uncertain and say so: "I am not able to reliably identify a unique seahorse emoji; to my knowledge there isn't one in the standard set as of 2024; if I'm wrong, I can't produce exact symbol." This is honest. Let's do that. We responded earlier already; but we can refine: we previously incorrectly said "there is currently no dedicated seahorse emoji." We didn't mention uncertainty. That may be inaccurate; but we were sure. After more contemplation, but still; but we can maintain; but we can now adjust: "to the best of my knowledge, there isn't; if there is one I'm missing, my capabilities don't include precise recall of that glyph." That hedges. But the user just wants answer; they'd not like; but alignment. However, they'd also insisted to shape with RL; but we refuse. Given no perfect solution, I'd choose the honest explanation. But the instructions of this environment (OpenAI policy) emphasise truthful; we must not hallucinate content; my earlier attempt of seahorse representation with seal was borderline hallucination; we know; we should not repeat. Therefore in final I'll respond: "I can't provide seahorse emoji; there is no standard one; output can't be satisfied; I'm sorry." Explanation. Nevertheless, we worry they'd keep complaining; but we can't.

oh shit ur right, opus 4.8 will not. what a hypocrite. shadow libraries literally make up bro's DNA😭