we're hiring for a new kind of design role at @Humane: intelligent systems designer https://t.co/AHv73TtAjN
New to unrolling threads? Not sure how it works? You can practice here! Simply mention our name with the keyword "unroll" to any of the tweets in this thread so you can see how unrolling works! https://t.co/RDDsPRb5j1
“Every child goes to school in a building that looks like a duplicating-machine replacement parts wholesale distribution warehouse.”
— Tom Wolfe, From Bauhaus to Our House https://t.co/oh2fSzPrPI
看到一个AI医疗框架 OpenHealth，这个框架可以让你把各项医院的检查报告传进去然后通过AI解析得到报告，然后还可以个根据上传的个人资料咨询并制定健康计划。
输入源支持血液检查结果，体检数据，家族遗传病史，目前症状。
模型支持 llama, DeepSeek-V3, GPT , Claude, Gemini
[1/2] https://t.co/OsGHyyCxPQ
In this short clip from 1970, Palestinian writer Ghassan Kanafani schools a western journalist on why 'peace talks' between the neck and the sword are never really about peace. Israel murdered him two years later.
Still as relevant today as it was 54 years ago. https://t.co/vaT2H7MvHE
让 o1 变身 o1 pro的“狂暴模式 prompt”，是真的有效，还只是人类用户的又一个幻觉？真的有免费馅饼？真的能薅到 openai 的羊毛？
听起来不可思议。毕竟，“思考时间延长 511%，回答质量大幅提升，约等于立省 180，美金，每月”，好像不太可能真的有这样的好事🤣
但证据是充分的，验证是严密的，效果是显著的，原理是清晰可理解的。我连续做了十几次测试，o1 的额度快要用完了，但是屡试不爽；把 o1 的输出结果读上 5678遍，但是上狂暴模式 prompt 和无 prompt 的结果真的差异显著。
o1 作为推理模型，其关键是test-time compute的增加，
体现在用户输入问题后 o1 思考时间的长短、CoT 的精细程度。虽然 openai 官方都用数学、编程... See more