① 他说过的话

13 条

从 2014 年，也就是新 AI 时代非常早期的阶段，到 2024 年，GPUs 的 flops 提升了大约 120 倍
between 2014, just very early innings of the new AI era till 2024, you had about 120x improvement in the flops of GPUs.
—— Thomas Sohmers · [07:43]

指向原始笔记的链接

处理一个缓存命中的 token 基本是免费的。它的成本是真正重新计算并生成那个 token 的千分之一量级
both processing a cash token is essentially free. It’s one one thousandth of the cost, you know, order of magnitude of actually having to recompute and generate that token.
—— Thomas Sohmers · [11:25]

指向原始笔记的链接

这也解释了为什么据报道 Anthropic 目前在 API 业务上有 80 个点的毛利润率
there’s a reason why Anthropic is reported to have 80 points of gross margin right now on API business.
—— Thomas Sohmers · [12:03]

指向原始笔记的链接

如果他们停止训练，他们会一夜之间实现巨额盈利
if they stopped training, they’d be massively profitable overnight.
—— Thomas Sohmers · [12:49]

指向原始笔记的链接

技术能力的集中，让矩阵乘法变得需要合法许可，这会把我们倒退到前——不只是工业革命之前，而是启蒙运动之前的能力水平
the concentration of technological capability, making it legal to do matrix multiplications is like the thing that will set us back to pre, not just industrial revolution, it’s like pre-enlightenment capabilities.
—— Thomas Sohmers · [14:46]

指向原始笔记的链接

如果技术本身被限制在只有那些人手里，那么他们就会成为新的领主和国王，其他所有人重新沦为农奴
if the technology itself is restricted to just those, then they’re going to be the new lords and kings and everyone else is back to serfs.
—— Thomas Sohmers · [15:14]

指向原始笔记的链接

它现在已经成为左右两派几乎团结一致的议题——反数据中心。我认为这几乎完全是一场中国的心理战
it’s now become a almost unifying issue on left and right about being anti-data centers. I think that is almost entirely a Chinese psyop.
—— Thomas Sohmers · [21:40]

指向原始笔记的链接

而我们只是不允许它们接入电网。而它们实际上本可以降低所有人的价格
And we’re just not allowing them to hook up to the grid. Where they could actually be lowering the prices for everyone.
—— Thomas Sohmers · [26:22]

指向原始笔记的链接

贯穿这些整个会话的所有 token 里，大约有 96% 是被缓存的
about 96% of all the tokens that go through these entire sessions are cached.
—— Thomas Sohmers · [39:29]

指向原始笔记的链接

只要在服务上有 50 个用户，仅仅是这些单独的用户会话，就会超过你试图存储的模型权重
with just 50 users on your service, the user context, just those individual sessions end up being greater than the model weights that you’re trying to store.
—— Thomas Sohmers · [40:27]

指向原始笔记的链接

gated DeltaNet 及其衍生版本，通过这种机制，你在注意力部分花费的总时间可以减少大约 75%
gated DeltaNet and its derivative versions where you can have like a 75% decrease in the total time you’re spending on the attention portion with this mechanism.
—— Thomas Sohmers · [57:38]

指向原始笔记的链接

比起那个 60 比 1,我更在意的是，五年前 60 美元的那个 token，今天没有人愿意花一分钱
the thing I care a lot more about than just that 60 to one is the fact that a $60 token five years ago, no one would pay a cent for today.
—— Thomas Sohmers · [60:33]

指向原始笔记的链接

把那两到三周压缩到两天多一点，这仍然令人瞠目结舌
So compressing that two to three weeks down to two days and change, it’s still mind-boggling.
—— Thomas Sohmers · [53:26]

指向原始笔记的链接

② 出现在这些集

1 集

• 《推理才是赚钱的生意：Positron 联合创始人谈内存墙、缓存暴利与「守住前沿」之争》 — 作为主持

③ 他谈到的

点进去有真内容 —— 本页主要出口

Harry Stebbings · Positron · Anthropic · OpenAI · NVIDIA · DeepSeq · 推理 · 内存墙 · KV 缓存 · token

④ 也在聊「智能体」的人