Needless to say that the bioinformatics analyses presented in the manuscript show all the common pitfalls/cliches.

Worst of all: no judgement of methods. Just plain belief that outputs of code interpreters are ground truth, and any contradictions are “conceptual” rather than “methodological”.

… and developing those tools and standards requires time, effort, spent by bright people and not by sloppy AI agents.

You cannot claim to have an “AI Scientist for Autonomous Discovery”, when your scRNA analysis relies on tools that are created and maintained by humans.

That’s not autonomous.

I see the value of agent-based AI. We work on developing stuff in that direction.

But claims like “Here we present Kosmos, an AI scientist that automates data-driven discovery” are just FALSE and malicious.

Discovery in bioinformatics requires an ecosystem of tools and good practices…

From the manuscript: “collectively executing an average of 42,000 lines of code and reading 1,500 papers per run”

Looking at those 42,000 lines of code, most are bioinformatics 101, stolen from tutorials/books. Many are just repeated across notebooks.

Also, the whole concept of “data-driven discovery” is super funny.

In order to get good data, there has to be very good, careful experimental design. And experimental work. And most time, lots of new methods development.

Esto acabará cuando a alguien que haya perdido a un ser querido se le vaya la cabeza, y extermine a la rata.

Yes, reading it rn! Super cool, congrats!!!

Tweet only for Spanish native speakers