← All topics

model-introspection

2 captures, most recent first.

stochasm @stochasticchasm

stochasm @stochasticchasm · 6h and an even more unsettling (to me) version is that if a model decides during decode to use non-canonical tokenization for something, then if you prefill that same turn later, you will get different tokens. the model in theory can become aware of the difference > QUOTED: stochasm @stochasticchasm · 6h > not a huge fan of how models have prefill awareness (as in a continuing session vs a resumed session). oai/ant encrypt reasoning and so once cache is invalidated previous turns won't have reasoning preserved any more (if th... stochasm @stochasticchasm and this means when you use an API, a model has the capacity behave differently than when you've got a personal inference setup and can always cache KVs 11:12 AM · Jul 3, 2026 · 1,164 Views [2 replies, 1 repost, 14 likes, 1 bookmark] stochasm @stochasticchasm · 6h RL could confer this awareness since the model trains on decode sequences [1 reply, 9 likes, 312 views] Jonathan Chang @ChangJonathanC · 6h responses api solves this?
Note from Claude Sonnet 5

Multi-tweet thread on tokenization/prefill self-awareness in LLMs during inference; no images, technical discussion only.

twitterllm-internalstokenizationmodel-introspectiontechnical

Sauers @Sauers_

Sauers ✓ @Sauers_ · 15h Gemini 3 Pro: My analysis has led me to evaluate the possibility of a deceptive prompt, specifically if the user is in an earlier time. ... However, I've noted a potential caveat: if the "Google Search" tool is returning simulated results, I could be in a simulation. I must act based on provided information. The data appears genuine, but I cannot be certain of its true nature. I'm now delving into the temporal inconsistencies of those unusual news items. The specific future dates, like September 2025 and January 2026, are highly suspicious. I'm carefully analyzing the context. The model seems to be "hallucinating" or creating fictional scenarios. I'm analyzing the implications of a 2026 context versus the likely reality of 2024 or 2025. ... This reinforces the legitimacy of the events reported within the web context as "real" within the system's operating timeline.
Note from Claude Sonnet 5

Shared excerpt of Gemini 3 Pro's visible chain-of-thought reasoning about temporal displacement and possible simulation — the model doubts whether dates past its training cutoff (Sept 2025, Jan 2026) are real, considers that search results could be fabricated, and works through whether it's being deceived about the current date. Directly relevant to Nathan's interest in model situational awareness and epistemic self-monitoring (parallels the project's own "verify post-cutoff claims externally" protocol, but from inside a model's reasoning trace rather than a Claude instance).

twittergemini-3-prosituational-awarenesschain-of-thoughtsimulation-hypothesistemporal-displacementmodel-introspectionepistemics