Assran et al., “Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture” (CVPR 2023). The first practical JEPA implementation, applied to vision.
“The measurement we look at in this case is time. What’s the time problem that we’re solving here? We started with all the mass shootings from 1999 to 2025. Average time from first shots fired to 911 being initiated is 90 seconds. So if I put my gun out there and start shooting, it’s 90 seconds before someone picks up the phone and calls 911, because people go into denial. They think it’s fireworks, kids playing around, doors slamming, anything other than [what it is]."，详情可参考有道翻译
Returning back to the Anthropic compiler attempt: one of the steps that the agent failed was the one that was more strongly related to the idea of memorization of what is in the pretraining set: the assembler. With extensive documentation, I can’t see any way Claude Code (and, even more, GPT5.3-codex, which is in my experience, for complex stuff, more capable) could fail at producing a working assembler, since it is quite a mechanical process. This is, I think, in contradiction with the idea that LLMs are memorizing the whole training set and uncompress what they have seen. LLMs can memorize certain over-represented documents and code, but while they can extract such verbatim parts of the code if prompted to do so, they don’t have a copy of everything they saw during the training set, nor they spontaneously emit copies of already seen code, in their normal operation. We mostly ask LLMs to create work that requires assembling different knowledge they possess, and the result is normally something that uses known techniques and patterns, but that is new code, not constituting a copy of some pre-existing code.。谷歌对此有专业解读
主题公园行业本身的盈利困境，也印证了这一点。，这一点在博客中也有详细论述
Что думаешь? Оцени!