基于酵母的LLM研究
我花了很多时间在自我修改、自我学习的系统上。我甚至有一个基于基因改造酵母的设计。这个点子是将酵母与康威生命游戏结合起来。事实证明，可以在生命游戏的平台上构建计算机。
酵母可以被改造成遵循构成生命游戏的生死规则。CRISPR-CAS9技术以及精心设计的“基因开关”技术使得将这些规则嵌入DNA成为可能。
关键在于将基因改造的酵母与遵循生命游戏计算机布局的初始条件结合起来。困难的部分不是计算，而是对酵母的照料和喂养，这至今仍未解决。事实证明，酵母团块会专门化，以便为团块提供营养。这意味着一个更大的团块具有内部计算能力。
然而，聚集的酵母自我构建成深度学习网络是通向智能酵母的另一条可能路径。当前的问题是如何进行结果的反向传播，以便能够对团块应用强化学习。如果酵母能够自我构建和学习，它们就可以在没有外部帮助的情况下传播。
至于训练/反向传播的问题，我目前的想法是找到一种方法，将现有的LLM权重“印入”基于酵母的深度学习网络。每个酵母节点从现有的已发布权重集中获取一个权重。因此，这个酵母生物现在在知识上等同于电子版本。
仍然有很多工程、生物学和遗传学的问题需要思考，但到目前为止，似乎这些问题虽然困难，但是可以解决的。
查看原文
I've spent a lot of time on self-modifying, self-learning systems. I even have
a design based on genetically modified yeast. The trick is a cross between yeast and
the Conway Game of Life game. It turns out that it is possible to build a computer on
a game-of-life platform
https://www.youtube.com/watch?v=Kk2MH9O4pXY&ab_channel=AlanZucconi<p>Yeast can be modified to follow the live/die rules that make up the game-of-life.
CRISPR-CAS9 technology along with carefully designed "genetic switch"
technology makes it possible to embed these rules in DNA.
https://www.amazon.com/Genetic-Switch-Third-Lambda-Revisited/dp/0879697164<p>The trick is to combine genetically modified yeast in an initial condition following
the game-of-life computer layout. The difficult part isn't the computation but the
care and feeding of the yeast, still unsolved so far. It turns out that clumps
of yeast will specialize so they provide nutrients to the clump. That implies a more
massive clump with internal compute capability.<p>However, clumping yeast that self-structures as a deep learning network is another
possible path to intelligent yeast. The question before the court is how to do
back propagation of results so one can apply reinforcement learning to the clump.
If yeast can self-structure and learn they can propagate without external help.<p>As for the training / backprop question my current thinking is to find a way
to "impress" an existing LLM set of weights on a yeast-based deep learning
network. Each yeast-node gets a weight from an existing published set.
Thus the yeast-beast is now equivalent to the electronic version in knowledge.<p>Still LOTS of engineering / biology / genetics to ponder but so far it seems
the problems are hard but solvable.