Korva is a scroll of the internet’s best ideas.
Four agents, half the time
Cost of speed
Four agents working for half as long cost twice as much for an answer that arrives twice as quickly. Sixteen agents show a similar pattern with slightly less efficiency. Overall the speed-up is slightly sublinear and depends heavily on the problem.
Which tasks parallelize
Math is quite parallelizable. Web search and deep research reports, where agents look through a bunch of sources, are extremely parallelizable. Writing a novel is probably the opposite: ten thousand agents would likely add little, just as ten thousand people would add little to one novel.
Why parallel helps
Parallel agents are a way to scale up thinking time without waiting for it. A single model thinking longer does better, the way a student taking the SATs does better with five hours than with five minutes, but a very long think hits a latency wall. Splitting the work costs efficiency because no agent holds all the context, yet it works well when done right.
Why do some tasks parallelize better than others?
Computer scientist Gene Amdahl pointed out in 1967 that the speed-up from adding processors is capped by the part of a job that must run in sequence. If a task is entirely independent pieces, more workers help almost without limit. If some steps must wait for earlier ones, those steps set the ceiling. A job that is 5 percent sequential can never run more than 20 times faster.
Why does extra thinking time raise accuracy?
OpenAI's o1 announcement in September 2024 reported that accuracy rose as the model spent more compute thinking at test time. On the AIME contest, o1 scored 74 percent with one attempt per problem. Taking a consensus of 64 sampled answers lifted that to 83 percent. Parallel sampling like this is one way to spend extra compute without waiting longer.
Why is web research so parallelizable?
Reading one web page does not depend on what another page says, so a searcher can read dozens at once and merge notes afterward. Computer scientists call work like this embarrassingly parallel. Search engines exploit the same property: Google's MapReduce, described in a 2004 paper by Jeffrey Dean and Sanjay Ghemawat, spread indexing work across thousands of ordinary machines.
More from Dwarkesh Podcast
