4 ms·

Anybody want to summarize for those of us who don't feel like shelling out over $500/yr for a WSJ subscription?

• zamadatix 2y ago
  https://archive.is/ptVBA https://archive.is/ptVBA
• DeepYogurt 2y ago
  the tl;dr is "reasoning models".
  • iLoveOncall 2y ago
    And to be more specific, the fact that reasoning models have outputs orders of magnitude longer than other models. It writes more text so it spends more compute.
  • BugsJustFindMe 2y ago
    That's what costs more, but it does not answer why the companies are choosing to pursue it.
• edmundsauto 2y ago
  As the theory goes, there is a network effect forming. Early AI products with strong adoption get the users to implicitly and explicitly label data. This leads to improvements that are a moat due to “data flywheel” So there’s a “first mover” type advantage because you can’t get that usage data anywhere else - not even paying participants - because there exists a domain gap between offline and online measurement. There is a reasonable counter argument of “the bitterest lesson” which says all that matters is the improvements in foundational models.