|
|
|
|
|
|
by Kim_Bruning
1 day ago
|
|
|
Sure, but they're not producing derivative works. I'm told some of the recent Qwen models were trained on something like 140TB of data or so. I can put some of the Qwen models on a slightly large USB stick (~32G) . No way that contains the whole source text. No form of compression survives that. It's an old argument, it's still the simplest argument-by-numbers. The intuition this counters is "it's drawing from a big database" or "it memorized everything". It didn't. It still can recall passages at times. But then again, so can I for my favorite books and movies. And no I'm not saying it's human, to stop the next big jump. LLMs are LLMs. They have their own properties. Some of which happen to coincide with human properties. Just like a robot arm has some properties in common with humans. Similar requirements, similar constraints; you're going to have to expect similar behaviors and failure modes at least some of the time. But I haven't heard anyone claim a robot arm can edit a python program, nor have I heard of an LLM picking 100 widgets per minute. |
|
|