Definition
Plain language
A larger competing model for in-model retrieval, used as a comparison point.
As stated in the literature
A concurrent long-context retrieval model roughly seven times larger than BlockSearch, trained on much longer contexts; it aces synthetic needle-in-a-haystack tests but degrades on realistic retrieval with hard negatives.
Why it matters: It shows that a much bigger model can look impressive on toy tests yet still fall short on realistic retrieval, so size alone isn't the answer.
For example, MSA nails artificial 'find the hidden sentence' tests but stumbles when the pile of documents contains many tempting near-misses.
Heard on the show
“It also matches or beats MSA, a concurrent model seven times larger, trained on much longer contexts, the same MSA that aces the synthetic needle tests.”Episode 198 — The Model That Knows the Answer and Can't Say It