AI为什么能举一反三?没人说得清
原创
一个网络在几百万张猫图上学到了"猫"这个概念,却没在训练集里见过任何一张完整的猫。这事儿本身就很诡异。Astral Codex Ten的Scott Alexander把这件事拎出来深挖——AI的泛化能力至今没有令人信服的理论解释。我们训练模型靠梯度下降,可数学上完全无法从损失函数推导出"它应该能理解概念"。学界有人说是贝叶斯推理的副产品,有人说是数据里藏着隐式特征。我的判断:这不是技术债,是认识论债务。我们花了几万亿参数堆出来的东西,居然说不清它为什么能work。这比任何算力瓶颈都危险。
原文:Mysteries of AI Generalization · 来源:Hacker News
版权声明
所有资源都来源于爬虫采集,如有侵权请联系我们,我们将立即删除
相关文章
- 0.05ms 往返:localhost 为什么不能直接当网站?
- Fornax's Guide to Fast Ethernet
- Possibly leaked NSA Cyberattack tool behind attacks (2017)
- Steve Wozniak on Writing Apple Integer Basic
- PayByPhone is stealing $50M/year from their users
- Grand Theft Auto 6: A Landmark Achievement That Redefines a Console Generation
- Research Acceleration: The Jagged Frontier
- Google Embraces Geothermal Technology to Vaporize Rock for Clean Power
- Microsoft's Agent Framework PyPI package compromised to steal credentials
- Nvidia's Seven Billionaires Have Lost a Third of Their Wealth
最新文章
- 0.05ms 往返:localhost 为什么不能直接当网站? 13小时前
- Fornax's Guide to Fast Ethernet 13小时前
- Possibly leaked NSA Cyberattack tool behind attacks (2017) 14小时前
- Steve Wozniak on Writing Apple Integer Basic 14小时前
- PayByPhone is stealing $50M/year from their users 14小时前
- Grand Theft Auto 6: A Landmark Achievement That Redefines a Console Generation 14小时前
- Research Acceleration: The Jagged Frontier 15小时前
- Google Embraces Geothermal Technology to Vaporize Rock for Clean Power 16小时前
- Microsoft's Agent Framework PyPI package compromised to steal credentials 17小时前
- Nvidia's Seven Billionaires Have Lost a Third of Their Wealth 17小时前
itfan123


