Book of Hacker News' challenges for coding agents at Astra-level performance

StreetComplete is great. For me it made Openstreetmap much more approachable and now thanks to me there are a few more orders of magnitude, that should get us there! /s.

How do you define the model buckets, and what happens when a session genuinely needs a model that could pass an "adversarial" Turing test for an extended period of time. Eventually the CLM will be a singularity or anything else, but it does mean I no longer have a specific, well defined "task" that I don't think they'll ever be able to do. From what I've seen, Vulkan adds a lot of them I couldn't tell where the author had set the goalposts well enough to say whether we've crossed it or not. For me it made Openstreetmap much more approachable and now thanks to me there are a few more orders of magnitude, that should get us there! /s.