VibeThinker-3B is a 3-billion-parameter model that scores 94.3 on AIME26 and 80.2 on LiveCodeBench, matching flagship models orders of magnitude larger. The paper calls this the Parametric Compression-Coverage Hypothesis: verifiable reasoning compresses into a small core you can train hard against a checker, but general knowledge — facts, concepts, the long tail of things a model might need to know — doesn’t compress the same way. There’s no shortcut past storing it, which is why a 3B model can match a system orders of magnitude larger on a proof it can check, and still know almost nothing else the larger system knows.
A biohacker sequencing his own cheek cells on a MinION runs into the same problem from the other side. Five full end-to-end runs in, he has a checkable readout of his own bases: a VCF file. What that file means depends on cross-referencing it against ClinVar, gnomAD, PharmGKB — population-scale catalogs of which variants do what in whom. Run the same swab through the sequencer fifty more times and gain nothing on that front. The missing information was never in the sample. It’s in how much of the population has been annotated.
Both split along the same line. Verification is a fact about the one instance in front of you, and effort spent on that instance can drive it arbitrarily close to certain. Significance is a fact about the population the instance sits inside, and no amount of scrutiny on a single case manufactures the reference set it needs to mean something.