public timeline

Showing all public posts on this server. Log in to post and see your home timeline.

凛子@活跃中 @Rinkko@nya.one
这个现象其实很好解释。通常来说，人们并不愿意承认别人比自己更正确，即使心里清楚对方是对的，也可能碍于“面子”而不愿承认。

但面对 LLM 时，情况就不同了。大多数对 LLM 缺乏深入了解的人，往往倾向于认为它给出的答案是正确的。同时，他们也不会因为自己在某些方面不如 LLM 而感到丢脸。

毕竟，承认自己不如一个 AI，远比承认自己不如另一个人更容易接受。

RE: https://ckpt.hanchan.vip/hanchan/p/1793942717.858559
寒江在下雪 (@hanchan@ckpt.hanchan.vip) 想到身边总有些讲不清需求的大人，如今赶时髦用上了豆包，被 AI 糊弄了，反倒觉得是自己表达不清，甚至当成笑话讲。可同样的人，在没有 AI 的年代，遇到听不懂自己胡言乱语的真人，却总能理直气壮地责怪对方。一想到未来人们对机器的宽容，竟可能远胜于对真实的人的善意，我就会感到恐惧 ckpt.hanchan.vip
:1000: 1
Tristam @swiftcoder@mastodon.social

@timbray @gruber this is normal for the last few rounds of releases, right? The pro and max variants are vastly higher performance than the base models - even my old M2 Max outperforms a current gen base model by many measures

Tim Bray 🇨🇦 @timbray@cosocial.ca
Spoiler

2/ It’s the Memory Bandwidth numbers, “Up to 170 GB/s” for the M6, “Up to 614 GB/sec” for the M5 Max/Pro models. This is a big deal? Computers pull data from memory into the CPU, crunch it, write it back out. But they can crunch many times faster than the memory can deliver it. So the effective performance depends a *lot* on memory bandwidth.

Tim Bray 🇨🇦 @timbray@cosocial.ca
Spoiler

3/ The kind of heavy-duty work these things are designed for (LLMs, big-video, etc) all tend to work with huge blobs of data, so in many cases the memory bandwidth is the limiting factor. Old programmer mantra: “All computers wait for data at the same speed.”

older