AGI Fool's Gold: LLMs

The current AI trend is like the Gordian Knot, so I'm going to cut to the chase. Large Language Model software tools ("LLM") are not Artificial Intelligence ("AI") because the apparatuses are pattern matching machines which contain almost all existing functional data. Software companies, especially Alphabet Inc, want to re-define our vocabulary and perspectives in relation to their LLM work so their LLM marketing department can bring in the hay. To be clear, I'm not saying LLMs are useless or a wrong direction in development of "AI," but rather it's the Fool's Gold for the path towards AGI, the pursuit of the academic field of Artificial Intelligence research. Large Language Models are pattern-matching engines trained on humanity's digital exhaust, not thinking machines. The intentional mis-marketing of their fluency for intelligence is the category error of our decade.

Alphabet (formerly Google, founded in 1998) has profiled the entire internet since its creation as a search engine so that every possible digital bit pattern is referenced for the tuning of their machines: the pattern for a red apple the LLM tool references has been processed from more apples than I will see in my life, so the probability of the stochastic machine generating a viable red apple is within the bounds because the models are over-trained: normal real intelligence does not require the level of training LLM tools undergo. If the inference is not trained into the weights of the LLM already, it's baked into Gemini's "red apple" instruction book toolchain. That's not intelligence: what has been achieved is a larger series of actions which incorporates a stochastic machine and look-up commands. We've had that since the 1960s.

Allow me to explain: this is the Thought Experiment called "The Chinese Room" argument by John Searle (1932-2025). You, the observer, have a dialogue with an anonymous almost-sealed room by passing notes written in Chinese into the room through a compartment, and the room returns a response written in Chinese. You think the room must contain an entity that is intelligent in Chinese, but that's not so. There is a person in the room with dictionaries and instruction books to decipher the content of the notes, act on them, and then re-cipher them into Chinese to be sent out again. The person has no idea what the notes say, and yet they act on them to a caliber that the outside observer believes the inner-person speaks Chinese very well. There is no intelligence in the room beyond the person being able to identify the strokes of a character to find in a reference book and act on it.

Apply this to legal work, to animation work, to anything that you could condense into an "instruction book" for the LLM to ingest and you will fit OpenAI, Anthropic, and Google's LLMs into the Chinese Room argument. If you were to read the short story "The Game" by Russian cybernetist Anatoly Dneprov (https://www.hardproblem.ru/en/posts/Events/a-russian-chinese-room-story-antedating-searle-s-1980-discussion/), Dneprov's conclusion answers the riddle of the hallucinations we see from LLM tools: incomplete/incorrect data or machine implementation. If LLM tools were truly intelligent^†, they would understand their output as incorrect before exposing it to the user.

"Ok, so what's the threshold for AI, asshole? These LLM tools are taking jobs and I feel dumb." AI must think, as Searle said, in such way that it understands the context before being provoked: consciousness does not exist in the LLMs. Daniel Dennett (1942-2024) counters the Chinese Room argument by saying the experiment simplifies the intelligence away from the operator. Paraphrasing, "If the content of the note requires world knowledge or elicits an emotional response, does the operator look that up too?" But this too can be distilled into a world atlas or a feelings chart: there's no possibility for a metaphysical discussion with the operator. The operator cannot initiate a stop of the process with the correspondent, and discuss how received notes are trivial or some other meta-commentary, or that the metadata of the references have gaps or errors.

Alphabet, Meta, Anthropic, and others have explained their "AI genesis" in 3 steps: 1) Prep: build the Transformer matrix and sanitized dataset; 2) Train: use GPU clusters to calibrate the Transformer matrix on the dataset into the latest LLM model; and 3) Export: expose the LLM tool to different toolchains for consumers to interact with. A key observation of this: the models do not think, they wait for an initiative on step 3; the models stop learning after step 2's training such they are frozen in time. The LLM tool prompts are a shell game: one prompt copy is sent to the trained tool to attempt to decrypt and assessed by its dictionaries and lookups while a second copy of the prompt, with the system logs of the process, is sent to the mothership to be used for the training step for the next model without gaining knowledge except user behavior. Tighter and tighter the feedback loop becomes, and more slop is passed off as intelligence.