‹ Notes

Notes on The Genomic Code: The genome instantiates a generative model of the organism

The Genomic Code: The genome instantiates a generative model of the organism

https://chatgpt.com/c/6817424f-ae74-800d-b584-73cf01b5e007

Interesting ideas.

  • imagine the organism as a generative model.
  • the genome both generates the organism as well as generating the generator. Unlike an ML system like an LLM, where the entire structure is statically defined. The organism’s instantiator is also generated / bootstrapped.

In machine learning, each new model starts with random weights and is trained de novo on large sets of data (e.g. images of horses), and develops a new compressed representation in the process. In biology, evolution has done the encoding job and each new individual inherits the compressed model (i.e., its genome). The encoding in biology is thus done across evolutionary time, in lineages, while the decoding is done by each new individual.

Recursive self-unfolding.

The theory they put forward is that the genome instantiates a generative model of the organism.

This reminds me of the concept of bootstrapping in programming language compilers.

  • Originally, we started with assembly code.
  • Then someone invented the C programming language, by writing a compiler (gcc) in assembly.
  • The compiler was able to assemble C programs to assembly.
  • So then they could rewrite the compiler in C. This is referred to as bootstrapping.

What is interesting about the theory of the genome being a generative model, is that it seems to bootstrap itself. This reminds me of RNN’s and test-time training.

One metaphor - the genome decompresses a representation of its next layer. And then uses that next layer decompress a representation of the layer after it.

Concept

Developmental Recursive Autoencoder (DRAE)

  • The decoder is not static.
  • As it generates the next layer of output, it produces the code that will generate the next layer’s decoder.
  • Each stage’s output includes:
    • Data (e.g., cell state).
    • Code (weights/instructions for next-stage decoder).

The basic idea: DRAE → a factory that builds its own next-stage factory at every step, adapting its design as it grows.