By subscribing, you agree Substack's Terms of Use, and acknowledge its Information Collection Notice and Privacy Policy.
核心贡献 · Key contributions
指出缺乏持续学习能力是 LLM 成为真正有用员工的关键瓶颈。 Identifies lack of continual learning as a key bottleneck for LLMs to become truly useful employees.
认为当前 LLM 无法像人类一样随时间改进，限制了其实用部署。 Argues that current LLMs cannot improve over time like humans, limiting their practical deployment.
预测解决在线学习将导致 AI 价值的非连续性，可能迅速产生超级智能。 Predicts that solving online learning will cause a discontinuity in AI value, potentially leading to rapid superintelligence.
对近期（2026 年）可靠的计算机使用智能体持怀疑态度，原因包括长周期、数据缺乏和算法进展缓慢。 Skeptical of near-term (2026) reliable computer use agents due to long horizons, lack of data, and slow algorithmic progress.
给出 50/50 时间线：AI 端到端报税在 2028 年，AI 像人类一样在工作中学习在 2032 年。 Provides 50/50 timelines: AI doing taxes end-to-end by 2028, AI learning on the job like humans by 2032.
指出训练算力的 Scaling 无法持续到本十年后，进展将转向算法。 Notes that scaling of training compute cannot continue beyond this decade, shifting progress to algorithms.
局限 · Limitations
论点依赖于个人使用 LLM 的轶事经验，可能不具有普遍性。 The argument relies on personal anecdotal experience with LLMs, which may not generalize.
低估了基于强化学习的微调和自我验证模仿人类学习的潜力。 Underestimates potential of RL fine-tuning and self-verification to mimic human learning.
假设实验室在持续学习上无突破，但进展可能快于预期。 Assumes no breakthrough in continual learning from labs, but progress may be faster than expected.
50/50 时间线是主观的，基于有限的公开证据。 The 50/50 timelines are subjective and based on limited public evidence.
未充分考虑 AI 自动化 AI 研发的影响，这可能加速时间线。 Does not fully consider the impact of AI automating AI R&D, which could accelerate timelines.