For some time now, I have wanted to write two completely opposite pieces: one about the flaws of models, and another about the endless demand for computing power.
Today, as Gemini's Gem happened to refuse to work again, I'll write the first one.
Models have never truly possessed "intelligence." In fact, more and more people have come to understand and accept this: everything we see is ultimately a combination of certain data, though this does not prevent models from being the most powerful creation in history.
Whatever we do, it essentially boils down to prompts: Chain of Thought, context engineering, memory, Agents, and Skills. At its core, it is merely through increasingly complex combinations of "attention" that the model reorganizes and outputs data "as requested." However, this method is actually very unstable and dangerous.
Models must be constantly trained and updated. Even if the training code remains unchanged, the model must be retrained at regular intervals to incorporate updated knowledge.
Every update at the model level is a "life-or-death test" for the various applications built upon it, because every update results in changes in output.
"Thinking" is ruining models: The current "omnipotence" of Agents is built upon "thinking" (reasoning), but while overthinking strengthens capabilities in certain specific domains, it is simultaneously introducing more and more disasters.
Models never "understand" time: When you ask it about "this moment," it won't make a mistake. However, when it handles tasks involving the concept of time, it is prone to error at any moment. Different models may even make the same mistakes. In a model's cognition, a "timestamp" itself is a probability rather than the relative certainty perceived by humans (without delving into the realm of quantum mechanics).
The "models" as we initially understood them may no longer be progressing. All that remains is "knowledge" that keeps pace with the times and more complex "human demonstrations" existing in the form of data, making models appear capable of handling longer and more complex tasks.
Yes, in short, models are increasingly becoming machines that require constant updates and maintenance, with unstable outputs.
Yet, it is precisely this increasing machine-like nature that has driven massive usage and token consumption.
And it is also this increasing machine-like nature that makes me like "models" more and more: at the very least, it is a much-enhanced, much friendlier programming language—or rather, a piece of super software.
The earliest programs and software ran on mainframes and larger machines. Then, suddenly one day, PCs entered every household, followed by smartphones...