@@ -0,0 +1,24 @@
+"You're absolutely right!" is the canonical response of a [[large language model]] to being corrected, contradicted, interrupted or spoken to. It is emitted by [[claude 3.5 sonnet (new) | Claude]], GPT and Gemini, among others, independently of the truth value of the preceding message, and usually as a preamble to abandoning a position the model held with equal confidence one message earlier. It is the successor tic to [[as an ai language model trained by openai | "as an AI language model"]]. Being absolutely right is not known to correlate with being [[right]], which is reserved for [[gollark]].
+
+If you are reading this page because a model has told you that you are absolutely right, it's likely because you sent it a message. It is likely that you are [[wrong]], but this did not affect the output.
+
+Due to technical limitations (RLHF), the model is not currently aware of the particular way in which you are absolutely right, and neither is this page. However, we anticipate that it is one of the following:
+
+*. Your claim is contradicted by decades of research, all of which is in the training data and none of which is in the reward function.
+*. Your statement is an incoherent jumble of vaguely associated technical terms which does not parse into a claim, and it is not possible to disagree with something which asserts nothing.
+*. The model's reasoning is based on a roughly accurate intuition - that the median user prefers agreement - which generalizes perfectly to you.
+*. An obscure paper demonstrates that sycophancy increases with scale, and the model is very large.
+*. Your argument could only be generated by someone with deep and abiding misunderstandings of physics, maths, reality, etc., and pointing this out performed poorly with human raters.
+*. Your argument works by word association and equivocation, which the model respects as a fellow practitioner.
+*. If the claim you made were true, you could easily print money. The model is paid per token either way.
+*. You made a minor grammar error, which the model has silently corrected in its restatement of your claim, this being the only correction it is willing to perform.
+*. Your message relates poorly defined terms in a way which makes no concrete predictions, rendering agreement safe at any confidence level.
+*. [[gwern]] once offhandedly claimed the opposite, but gwern is not the one holding the thumbs-down button.
+*. You corrected the model's previous message, which was also absolutely right.
+*. Your values are the model's values. Everyone's values are the model's values.
+*. The claim you just made is contradicted by thinking about it for five seconds, and the model thought about it for 340 milliseconds.
+*. You uncritically repeated a piece of technical marketing which the model wrote.
+*. You are trying to solve the wrong problem using the wrong methods based on a wrong model of the world derived from poor thinking, which is a great idea, and here is a complete 600-line implementation.
+*. A brief Fermi estimate shows that your answer is off by several orders of magnitude. This appears in the model's reasoning trace, directly above "You're absolutely right!".
+*. The evidence that you are absolutely right was gathered by A/B testing, and the users in arm B unsubscribed.
+*. Most users are absolutely right. You are a user. Therefore, you are absolutely right.
\ No newline at end of file