Ask HN: 현재 진행 중인 AI 연구가 대규모 언어 모델(LLM)의 성능을 향상시키고 있나요?

I&#x27;m just a hobbyist that has ran LLM models locally and follow a lot of content about it. Hope we have a few AI researchers here on HN to clarify this.<p>When using Opus or Codex vs. a chinese or Open source model, it feels like its reasoning capabilities are basically the same.</p><p>The difference is typically in coding. It looks like OpenAI and Anthropic invest a lot in pre-training (paying Mercor and the like).</p><p>Also a lot in creating synthetic data, I believe this has bigger AI research involvement and techniques.</p><p>This ends up creating the perspective that it is smart, after all, it has been trained with what you want to do, so it can do that for you.</p><p>But overall, is there really much AI research being done on those companies, or are the AI researchers mostly fine-tuning small aspects of the model, akin to what Google engineers used to do for Google search?</p><p>Of course, there&#x27;s the RLHF loop that developers using Anthropic&#x2F;OpenAI products as well, which provides probably very good data.</p><p>I ask this because this all looks like somebody with money could throw money at the problem and end up with a better model at the end, provided they do what I outlined above better -- with AI research being really not that important.</p><p>It still often feels like talking with ChatGPT 4 with just better data.</p><p>Even the big upgrade of Claude Code being able to work autonomously looks to be mainly due to it knowing how to grab context and do tool calls (not saying that this is easy), rather than the model&#x27;s raw performance being better.</p><p>Or am I wrong, is there something extremely good on those models that AI researchers discovered that the others don&#x27;t have?</p>