The whole idea of AI is based on the concept of Evolution. They thought Intelligence can be evolved by random system using LLM. It somewhat works because it is guided by human intelligence.     See More uman intelligence works using a mental processing system. Which is why 14 Billion years couldn't evolve into the Constitution of the United States. If Intelligence cannot be evolved by random process, then it cannot also evolve life.    See Less

Love the channel but the volume is FAR too quiet     See Less

Encoding for a moment does not imply memory. Huge parameter space allows for mapping different inputs to different activations, but these activations are dynamic and are wiped out. The paper     See More show success with reading of specific information from networks. Math is deceiving. Can this be used as compression algorithm beating current compression? Not really.    See Less

TL;DR: no two unique prompts share the same hidden state, NOT that two identical input contexts will always produce the same outputs and therefore can reverse engineer output tokeniza     See More he original training material tokens.    See Less

Jeez, this kind of AI slop is soooo annoying and widespead on YT, no human sense-checking whatsoever - at least get the screen captions correct, in meaning and grammatically!! If the written     See More nd grammar aren't even right, what hope is there that the actual content is real/correct/reliable? This kind of low-quality, AI-reliant rubbish is killing the internet; it's the Dead Internet Theory self-fulfilled!! 😠    See Less

It's ironic that the AI "reader" writes "LLMs" as "enemies" :)     See Less

LLMs are the greatest thing since slice bread.     See Less

Todays LLM's; Understanding, Not possible, Cₛ = Φᵢ × Rᵍ = 0 still inert zero comprehension. In CIITR language, the video accurately describes static structural density, but it mi     See More reats this as a pathway toward understanding. CIITR explains precisely why it is not. The injectivity result shows that LLMs have: Perfect memory, but No rhythm, No phase stability, No metabolic temporalization, No ability to stay in resonance with reality.
Thus, in CIITR terminology, the video is describing a Type-B structure: a system that is internally flawless but externally inert — coherent, yet completely without Rᵍ. Injectivity proves that LLMs are lossless maps. CIITR proves that lossless maps are not understanding systems. Rhythm, not memory, is the boundary of comprehension. So the video strengthens my theory rather than contradicting it. It dramatizes what CIITR already formalizes: LLMs are high-Φᵢ machines with Rᵍ = 0, and therefore incapable of Cₛ.    See Less

First of all, video demonstrates misunderstanding of the article. The "remembering" is not about the things you write into the chat window. It is about things the model author used     See More it. Basically, it's the mathematical proof that the training data may be reconstructed from the model activations.

This article is likely to be close to true from a theoretical point of view, but for practice it means nothing.

A theoretical problem is, the injection assumes precise weights are known. What about quantization? It is a irrecoverable loss of information, breaking injection.

A practical problem is, for the described reversal you would need to know *all the activations*, billions of them, and you're never given that information. The model's response, even in the embeddings space, is insufficient. This is comparable somehow to the asymmetric encryption: mathematically there is one-to-one correspondence (bijection) between the public key and the private key, and therefore theoretically you can find the private key knowing the public key, but in practice there is no known efficient algorithm of doing that.    See Less

This video suggests that LLM is basically a glorified hash table, which is absolutely incorrect.     See Less