Having a couple of ML papers on its own is not sufficient, as per the eligibility criteria (https://www.uscis.gov/policy-manual/volume-2-part-m-chapter-...)

Yeah I completely agree. My question was more about practice than in theory.

The requirements seem very subjective (define extraordinary). And so I wanted to know what are some real requirements/conditions the committee uses in real life for people who have already got O-1s.

O1 is overrated and not that useful. It still needs to be sponsored by an employer, and the employer will make the determination if they want to sponsor an O1 for you. It's mainly useful if you can't make it through the H1B lottery. The criteria are subjective, but it's a lower bar than EB1A. Look at AAO decisions to get a sense of what gets rejected. The general framework for how USCIS looks at these is based on case law from Kazarian, which states that you need to satisfy the numerical criteria as a 1st step, and there is a final merits determination as a 2nd step. This was for EB1A, but USCIS also uses the same framework for O1s (although there is some litigation going on on whether Kazarian should apply to O1).

The O-1 standard, while high, is often a viable option and it might be here. The O-1 can support employment with another company (acting as the sponsor) or even with your own company (acting as the sponsor). Many founders are on O-1s.