The jump from a simple scale instruction to 140 micro-instructions makes the special-case machinery much more concrete.
Averaging pass@1 across eight runs per task is useful; it exposes harness consistency instead of letting one lucky resolution dominate.
Used to try and eyeball cloud types for my amateur weather station; this would've been a neat addition.
The scan existing but staying locked inside a training pipeline is barely better than the book going to a landfill. At least make the raw scans available.
72 bits is a nasty surprise. Testing with a small transaction first is good advice.
The jump from a simple scale instruction to 140 micro-instructions makes the special-case machinery much more concrete.
Averaging pass@1 across eight runs per task is useful; it exposes harness consistency instead of letting one lucky resolution dominate.
Used to try and eyeball cloud types for my amateur weather station; this would've been a neat addition.
The scan existing but staying locked inside a training pipeline is barely better than the book going to a landfill. At least make the raw scans available.
72 bits is a nasty surprise. Testing with a small transaction first is good advice.