Pointers (December 2022)
Favorites from December 2022:
The Pikchr diagram language, a sort of Boring Logo for grownups.
From a trip to Santa Fe: the Randall Davey Audubon Center and Sanctuary, the Georgia O’Keeffe Museum (she painted a much-wider variety of subjects than I knew), the IAIA Museum of Contemporary Native Arts, and the breakfast burritos at Plaza Café (Christmas-style, of course).
STRG.SNEK, which surprises me with how there can still be fresh takes on snake games after all these years.
From 2019, The Bitter Lesson, about how putting more effort into taking advantage of the increases in raw computation available tends to advance AI more than embedding more domain knowledge.
Since Sutton cites speech as an example, it’s worth recalling the famous line attributed to Jelinek: “Every time I fire a linguist, the performance of our speech recognition system goes up.” I think the article actually suggests linguists will have their revenge on Jelinek: consolidation around a few algorithms should lower the importance of engineers and AI researchers and raise the already-high importance of curating good data and performing thorough evaluations, along with the linguists who have the doamin expertise to do both.
The AI-adjacent field of static program analysis seems to show a similar trend: at some point, SMT solvers became good enough that there was a pivot from developing more-sophisticated custom abstract domains, analyses, and type systems, to increasingly riding the wave of cleverer SMT solvers running on faster machines.