Frontier Models Show Signs of Continual Learning
A 36Kr analysis cites Anthropic’s engineering report “The components of agent memory,” which found that recent frontier models could select, organize, and verify useful memories in minimally prescribed single-session and multi-session setups. The article also cites Nvidia research on self-summarizing context and GPT-6 Astra results on ARC-AGI-3, using them to argue that continual learning may be emerging across model providers rather than requiring a dedicated memory architecture. These broader conclusions remain interpretive and are based largely on selected studies reported in one secondary article.
The analysis suggests that recent frontier models may already exhibit parts of the select-verify-retain learning loop needed for persistent agents, while presenting the field-wide emergence claim as an interpretation rather than a settled result.
从 2025 年年中开始,持续学习就被视为通向通用人工智能(AGI)之路上最难越过的一座山。
强化学习之父 Richard Sutton 在 2025 年 9 月的对谈中就提出过,真实世界太过庞大且时刻都在变化,智能体必须在与环境的直接交互中边走边学。而单靠静态预训练的方法走不出真正的通用智能,最终只会陷入模仿人类既有文本的死胡同。紧接着在 2025 年 10 月中旬,Andrej Karpathy 在播客访谈中也直言,如今的智能体之所以没法像人类实习生那样真正上岗干活,核心就卡在没有持续学习能力,这样你没法交代一件事就让它真正记住并改变后续行为。而他当时估算,这类底层认知缺陷大概需要十年时间才能逐一磨平。直到 2026 年 4 月,Demis Hassabis 在 Y Combinator 访谈中依然强调,当前模型距离完全体的 AGI 还缺一两块关键拼图,排在最前面的就是持续学习、长程推理与记忆机制,并认为这大概率需要新的理论或架构突破,光靠在上下文窗口上使劲并不管用。
#Anthropic#Claude Opus 4.5#Claude Opus 5#Claude Sonnet 4.6#OpenAI#GPT-6 Astra#NVIDIA#CL-Bench#ARC-AGI-3#Meta#GLM-5.3#Kimi K2.5