But I could chalk that up to your point on threads of independent reasoning.

Appreciate you coming down this rabbit hole with me lol.

Yup, that exactly. I suppose sometimes I wonder at what point do preserved high-level trends become a product of excellent reproducibility, vs an artifact of an elbow plot that I decided should be cut off at PC15 vs another group’s PC4.

I find there’s so much left to chance: random seed, nuances in software/tooling, personal preference of parameterization etc.

I do think there’s a distinction to be made between reproducing code, and reproducing biological insights. To your point, reproducing code is relatively simple. But reproducing analyses to yield similar insight?

Solid point. Often reproducibility is challenging not even between independent research groups, but simply iterating on a single experiment within a lab.

Right. Confirming accuracy of things is somewhat easier in a wet lab. If your cells die, you probably did something wrong. The qualitative confirmation is usually explicit, and re-tracing your steps is structured.

What methods do you take to support deterministic results when doing deep learning?

You don’t have 1000+ citations in <24hrs? You’re clearly not locked in.

Fascinating. I wonder what level of effort assembling an analogous training “pangenome” dataset (sampled across every demographic) might require. Certainly is sorely needed, both in pure AI and AI for better patient outcomes.

Thanks for sharing Tim.

Not obsolete at all, and won’t be for the foreseeable future. We need more representative reference genomes.

But I do hope the “AGI 2025” folks see this 😅

Not obsolete at all, and won’t be for the foreseeable future. We need more representative reference genomes.

But I do hope the “AGI 2025” folks see this 😅

After this I gave it additional context by saying, “I’m a woman, try again.” This was the output:

Such cool renderings. Do you host these anywhere? Would love to share 1 or 2 but want to be able to give you the appropriate credit.