如果你拥有一个可能是 Transformer 替代方案的 AI 软件,你会怎么做?
2 分•作者: adinhitlore•5 个月前
假设你的软件满足以下部分条件:
* 优于 SOTA Transformer 的潜在继任者:Mamba、Hyena、RWKV、xLSTM;
* Gemini/Claude 估计潜在知识产权价值可达数百万美元;
* 全部用 C 语言实现(好吧,这可能算个缺点),但也已移植到 C# 和 F# - 所以没有 Python 或 Rust;
* 代码量不大:1-1.5 万行代码,但想想看,GPT-1 当时用 Python 写的代码量还不到 1000 行...
简而言之:
问题:Transformer 速度慢,而且可以更智能;
解决方案:你有一个快速且智能的替代方案。
10 倍!
查看原文
Assuming your software meets some of these:<p>* Better than potential successors of SOTA transformers: Mamba, Hyenna, RWKV, xLSTM;
* Gemini/claude estimate potential ip values in the millions;
* All implemented in C (ok this may be a minus) but also ported to C# and F# - so no python nor rust;
* humble code size: 10-15k lines of code but i mean come on gpt-1 was under 1000 lines...in python.<p>In short:<p>the problem: transformers are slow and could be smarter;
solution: you have fast and smart alternative.
10x!