Post #4747680 · Elektrine

2026-09-18 11:13 UTC

One distinction is essential: training is not inference. During training, a model learns its parameters from enormous datasets through repeated computation across many accelerators. This is where most upfront demand for compute, energy, and training data arises. It is also where many questions about how content was acquired belong. During inference, an already-trained model applies those parameters to a new input. Each run still uses energy, and usage at global scale adds up, but it does not repeat the training process. It is far less resource-intensive per task. Efficient runtimes and quantization now make surprisingly capable inference possible on home hardware. Qwen3.8-27B is one example. Training and inference belong to the same system, but not to the same bill. Mixing them together makes the discussion simpler, which is not always the same as making it useful. 5/34

Replies (1)

@masek@infosec.exchange 2026-09-18 11:13

"There are two kinds of fools. One says, ‘This is old, and therefore good.’ And one says, ‘This is new, and therefore better.’” John Brunner, The Shockwave Rider Neither familiarity nor novelty tells us whether an LLM is useful, harmful, intelligent, or trivial. “Merely autocomplete” understates what scaled prediction can do. But calling it a “machine god” or an “incarnation of evil” does not explain it either. Humans are unreliable autocomplete engines too. We readily complete fragments of other people’s behavior with intentions we supplied ourselves, then forget who supplied them. “Machine god” and “incarnation of evil” are mythological labels, and mythology is wonderfully cooperative. It can mean whatever the argument requires while projecting intention, destiny, and moral agency before we have demonstrated them. I prefer capabilities, limits, uses, and consequences. Less dramatic than summoning gods and demons, but more informative. 6/34

Open ##4747799