Keep on to blur preview images; turn off to show them clearly
实验验证： Google 设计了一个模型叫 Hope。 它的特点是： 能自我修改，自动决定哪些部分需要学习。 每层都有不同的更新速度。 会根据“惊讶程度”决定要不要记住某个知识（比如遇到前所未见的数据时，会优先记住）。 实验结果：Hope 在 “长上下文理解” 和 “知识保留” 上超越所有现有模型。
Hope： 长文本记忆力更强 推理和常识判断更准确 学新任务时几乎不会忘旧任务 在长上下文“大海捞针”（NIAH）下游任务中展现出卓越的内存管理能力。
> 13,7% independent kind of unexpected. Then again… I want an American equivalent survey
We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1
谷歌发现恶意软件通过调用 AI 即时重写自身以躲避追查，同时还利用 AI 按需创建恶意模块。好消息是目前这个代号 PROMPTFLUX 的恶意软件家族正在早期测试，所以还没有实际的恶意功能。分析发现该恶意软件使用 Gemini AI 即时编写模块并混淆代码以躲避追踪，这会让安全系统检测效率大幅度下降。
查看全文：https://t.co/Hlk2BCp4IB
The more I think about this, the more interesting it is. Sora 2 definitely trained on "Zootopia" - otherwise it wouldn't generate the voices of the characters to match their image. But it's a relatively new Disney movie that isn't on YouTube / publicly available on the open Web. And Sora is the only model that can do this 🤔
Partner @a16z AI 🤖 and twin to @omooretweets | Investor in @elevenlabsio, @krea_ai, @bfl_ml, @hedra_labs, @wabi, @WaveFormsAI, @ViggleAI, & more
HN上这个帖子挺有意思的 作者使用3B条Goodreads的书籍评论训练了一个推荐模型，基于用户喜欢的图书偏好来推荐书目 以前大公司能做的推荐模型，现在个人开发者都可以做的很好，部署成API服务其实也有很大的价值 https://t.co/n8PWYre7kY
Believing is seeing
Interesting (if obvious) fact that Chinese EV exports to Europe are ≈equal in volume to what reaches Asia, yet roughly 25% more value per unit. I suppose margins are healthier too. Eyeballing it, they'd need to 2x exports to Asia to fully offset losing Europoors… Euroriches.
We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1