News for hackers. Stuff that matters.
Re:DeepThought-8B: A small, capable reasoning model by benchmarkist on Saturday November 30, 2024 @02:57AM (#42280256)
AI marketing departments are fond of anthropomorphic language but it's actually just regular beam search.
Re:DeepThought-8B: A small, capable reasoning model by sva_ on Saturday November 30, 2024 @08:16AM (#42281437)
Do you have a source on them using beam search? Seems like choosing several paths over top k would be rather expensive