Keep on to blur preview images; turn off to show them clearly
some people legit think that spending = capability oh yeah. You are vastly outspending them all right. no wonder you're building like 3x as many ships Diverting some of this $30T GDP (which part? maybe spending on lawyers and accountants) will make the gap insurmountable…
I lowkey admire burgers for their naive religious confidence truly a people who worship themselves. It's so pure in its lack of self-awareness it's not even hubris, they're like autistic children who've developed a quirky personal faith
啊？AI可以自己找活干了？ 给大家介绍一个炸裂的开源项目 Hephaestus - 这玩意儿让AI Agent自己规划工作，自己发现问题，自己创建任务！抽象到什么程度？它内嵌了个kanban....让AI自己拆解card自己做.... 传统的Agent框架都是你提前写死所有流程，遇到没预料到的情况就傻眼。Hephaestus 直接换了个思路：只定义三个阶段（分析-实现-验证），然后Agent自己看着办。 举例：测试Agent在跑测试时，发现了一个性能优化机会，它不是记个log就完事，而是自己创建了一个新的调查任务，然后放到kanban里，然后另一个Agent接手去研究，确认可行后又自己创建实现任务。整个工作流就这么自己长出了一个分支。 你给它一个PRD，它分析出5个组件，创建5个并行任务。其中一个Agent干完了发现bug，自己创建修复任务。另一个Agent发现可以优化，自己创建优化分支。工作流是实时生成的，而不是一开始就预测好的。 我有空也会测一下试试，看看它能不能真的从0到1自己把活干完。总之先增加到待测试列表。 项目地址：
A coder, road bike rider, server fortune teller, electronic waste collector, co-founder of KCORES, ex-director at IllaSoft, KingsoftOffice, Juejin.
Should you advertise the good of bad people? If you also equally advertise the bad?
Research Scientist @meta (FAIR), Prof. @Unige_en, co-founder @nc_shape. I like reality.
almost no kl mismatch between training and inference would improve the results drastically. most likely we can go easy on the kl regularization as well. the result of choice of *PO not mattering is funny though and i think it would still have an impact on the tails.
RL and efficient distributed pretraining • eXperiments lab • memes and training lores
this is going to marginally improve the quality of results. though my guess is the point about the choice of *PO not having any impact wouldn't be true for every scenario.
RL and efficient distributed pretraining • eXperiments lab • memes and training lores
Thread
this is going to marginally improve the quality of results. though my guess is the point about the choice of *PO not having any impact wouldn't be true for every scenario.