Not a surprise at all, as those old CNNs are tuned specifically for the task.

Considering CGPT et al were getting ~48 percent accuracy versus the old tuned net's ~64 percent, we're still dealing with coin toss odds here.

no shit. this post is dumb. when they say in the news that *ai" is great at science/medical thing, thry arent talking about llms. a lotof bteakthrough was done with google alphafold.

Why on earth would you want to use an LLM for this? A CNN is clearly the superior choice here. It’s like saying a hammer does a better job of driving nails than a screwdriver.