English(EN) so i just heard the argument that we shouldn't dismiss next-token prediction because the fact that something is a next-token predictor doesn't actually place an

AI研究员批评下一词元预测模型的理论极限

作者 PulseAugur 编辑部 · [1 个来源] · 2026-07-31 08:24

一位研究员反对忽视下一词元预测模型的能力，认为理论上的限制并非实际上的限制。虽然一个完美的下一词元预测器理论上可以模仿人类智能，但所需的巨大计算资源将使其与现有方法相比不切实际。研究员强调，关于AI能力的讨论应侧重于在现实约束下（例如数据中心规模的计算能力）可实现的目标，而不是纯粹的理论可能性，而AI公司经常利用后者来避免讨论实际限制。 AI

影响强调在AI开发和部署中，实际约束比理论可能性更重要。

排序理由研究员的观点文章，讨论AI的能力和限制。

在 Mastodon — fosstodon.org 阅读 →

ChatGPT

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

报道来源 [1]

Mastodon — fosstodon.org TIER_1 English(EN) · [email protected] · 2026-07-31 08:24

so i just heard the argument that we shouldn't dismiss next-token prediction because the fact that something is a next-token predictor doesn't actually place an

so i just heard the argument that we shouldn't dismiss next-token prediction because the fact that something is a next-token predictor doesn't actually place an upper limit on its theoretical capabilities. and like, strictly speaking, that's true. a definitionally perfect next-to…

报道来源 [1]

so i just heard the argument that we shouldn't dismiss next-token prediction because the fact that something is a next-token predictor doesn't actually place an

相关实体

相关话题