3 ms·
I vaguely recall some of the recent math breakthroughs being less about parallelism and more about serial compute. A lot of, "Keep trying," "Try harder," "Try a
by csallen 2mo ago
I vaguely recall some of the recent math breakthroughs being less about parallelism and more about serial compute. A lot of, "Keep trying," "Try harder," "Try another way," etc.
Regardless, I think both things are great, horizontal (more approaches) and vertical (deeper approaches). And both are made more practical with faster inference.
- energy123 2mo agoThose results were found with 64 subagents, or with GPT 5.6 Sol Pro which uses parallelism under the hood. Low latency is a big deal but the immediate use cases are somewhat different in the short-term (more serial coding workflows) rather than pure math research which is effectively massive-scale search through a tree of possibilities, which is where you want throughput and low cost per token, rather than high speed per token.