I just don't see how more compute solves the issues they have with the fact that they've basically providing only marginal improvements on precious tasks but we'll see
Replies
Every bit of research I've seen on LLMs suggests they're just stuffing more heuristics into a bag of heuristics the more compute and training data they get, not actually developing any fundamental reasoning capabilities
aug 23, 2025, 4:27 pm • 1 0 • view
I don’t see how they could! If I were in charge I’d be focusing on how to leverage the capabilities that already exist Like, if you invented the forklift but no one understood pallets What’s the task container that makes the tool an actual multiplier? I would be focusing on that question
aug 23, 2025, 4:43 pm • 3 0 • view
But I work in client services so I guess “focus on implementation” is always going to be my suggestion to a tech company 🤣🤣
aug 23, 2025, 4:44 pm • 1 0 • view
Google search finally learned that it should stop recommending that people eat rocks. Though the AI summary does still cite satirical articles to justify claiming that it's valid to do so. In other words, more compute can get us closer to the "eat zero rocks" asymptote.
aug 23, 2025, 4:25 pm • 0 0 • view
Yeah the advances won't come from compute, but some new crazy unknown thing like word2vec and transformers were for what we have now
aug 23, 2025, 4:33 pm • 0 0 • view
More compute is helpful but we're already well into the diminishing returns part from what I can tell
aug 23, 2025, 4:33 pm • 1 0 • view