# Post Analysis: @thesupermannx on "LLMs actually understand" (2026-10-11) Back to [[Users/@thesupermannx/thesupermannx-x-profile|@thesupermannx profile]] · Paper note: [[Users/@thesupermannx/Papers/Mechanistic Indicators of Understanding in LLMs (arXiv 2507.08017)|arXiv 2507.08017]] · Related pattern: [[Topics/Ribosome Pausing and Cotranslational Folding/Related Trends#1. "AI cracked open X" posts that re-label old science|Ribosome topic: the "AI cracked open X" pattern]] **Verdict:** The three-tier framework is described accurately, but a philosophical synthesis is presented as researchers who 'proved' understanding and 'documented grokking', and the paper's caveat that models keep mixing in memorized heuristics is reversed. ## The post (verbatim) Posted **Oct 11, 2026, 7:59 AM CT**. The link ([arXiv 2507.08017](https://arxiv.org/abs/2507.08017)) is in the self-reply. > Researchers proved LLMs actually understand things. > > They used "Mechanistic Interpretability" to literally look inside the hidden layers of Large Language Models. > > And found that to predict the next token, an AI doesn't just rely on surface-level statistics. > > It actively builds a three-tiered model of reality: > > 1. Conceptual Understanding” The AI spontaneously creates "features." It doesn't just memorize a word. It builds an internal, multi-dimensional concept of what that entity actually is. > > 2. State-of-the-World Understanding: It dynamically tracks facts. When it states a fact, it isn't just spitting out a high-probability string of text. It is actively linking internal concepts together in real-time to model the state of the world. > > 3. Principled Understanding: This is the breakthrough. The AI actually stops relying on memorized facts entirely. It discovers the underlying "circuits" or universal rules connecting those facts. > > It learns why things happen. > > The researchers documented a phenomenon called "grokking." > > It's a sudden shift during training where the model abruptly stops memorizing its training data and transitions into deep, generalized understanding. > > It actually deletes its sprawling memorized examples and replaces them with a compressed model of reality. > > Nobody programmed these tiers. The model built them itself to survive the training process. ## Media The image shows the paper's real title page (title, authors and affiliations) and its abstract. ✅ It is the correct paper. ## Interaction with [[marshallpittman]] **@marshallpittman reposted this post on Oct 11, 2026, 12:00 PM CT** ([repost 2109328391853776926](https://x.com/marshallpittman/status/2109328391853776926)). ## Engagement (Oct 24, 2026) | Impressions | Likes | Reposts | Quotes | Replies | Bookmarks | |---:|---:|---:|---:|---:|---:| | 26,272 | 576 | 114 | 24 | 61 | 427 | ## Claim-by-claim check **Ratings:** ✅ accurate · 🟡 overstated or simplified · 🔀 misattributed · ❓ unsupported by the paper · ❌ contradicted by the paper | # | Claim | Rating | What the sources show | |---|---|---|---| | 1 | "Researchers proved LLMs actually understand things" | 🟡 | The paper *argues* that MI findings make the "mere imitation" picture "increasingly untenable". It is a conceptual argument, not a proof | | 2 | "They used 'Mechanistic Interpretability' to literally look inside" | 🔀 | It synthesizes **other researchers'** MI findings | | 3 | Three tiers: conceptual, state-of-the-world, principled | ✅ | Matches the abstract closely | | 4 | "The AI actually stops relying on memorized facts entirely" | 🟡 / ❌ | The abstract does say "ceases to rely on memorized facts" for this tier. But the paper stresses that models use heterogeneous mechanisms in parallel. "Entirely" reverses that caveat | | 5 | "The researchers documented a phenomenon called 'grokking'" | 🔀 | Grokking was reported by Power et al. (2022) and others. This paper reviews it | | 6 | Grokking replaces memorized examples with a compressed model | ✅ | A fair summary of the grokking literature the paper discusses | | 7 | "Nobody programmed these tiers. The model built them itself" | 🟡 | The tiers are the authors' philosophical categories. What models "build" are the mechanisms MI has found | ## Bottom line - **The framework is relayed well:** better than usual. - **The credit and certainty are inflated:** philosophers synthesizing others' results become "researchers proved… documented."