Date: 4/21/26 1:32 pm From: Jim Forrester via Tweeters <tweeters...> Subject: Re: [Tweeters] The Merlin discussion
I read with interest both Bill's & Chuq's thoughts on this, and for the most part am in agreement, but I can't agree with Bill's answer to the question "First, why do we expect Merlin to be any better than a reasonably skilled human observer? It is trained by humans, so it seems unlikely that it would do better". Well, that argument doesn't stand up. Humans "train" tools all the time, just so that they *can* perform better than the human. Merlin doesn't "think" like a human, it thinks like a large language model. While it is handicapped by having a single sense (hearing), it is being constantly augmented by new and correctly identified songs being added to its repository. While humans may not be able to discern CHSP and DEJU, different wavelength patterns in the two songs will probably at some point mean that Merlin *will* be able to correctly determine the bird - the LLM just needs enough data!
Also, Chuq brings up the issue of Merlin's all-or-nothing approach to bird identification. I've plugged BirdNET before and will again now, because it tells you what its confidence level is. I have no idea what Merlin's threshold is, but somewhere around "highly likely". I've actually written suggestions to the Merlin team to build in the certainty level of identifications into the app, but I am but a leaf in the storm... :-)
_______________________________________________ Tweeters mailing list <Tweeters...> http://mailman11.u.washington.edu/mailman/listinfo/tweeters
