Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-16 · high
The write-up ran 1,500 candidate submissions in parallel and graded each against a number, not a vague goal.
Why now: The write-up appeared in coverage dated August 16, and doesn't say when the actual 1,500-submission run took place.
A developer's beam search produced a 232x speedup on a GPU kernel across 1,500 submissions, per a write-up on sankalp.bearblog.dev. For anyone writing prompts for a coding agent, that number is the payoff of running many graded attempts at once instead of one long conversation.
The method keeps several candidate approaches alive at the same time, rather than iterating on just one. It prunes the weak ones by what actually measured better, not by which one sounded more promising in the moment. The prompts also carry a number, a target throughput or latency figure, instead of a description of success written in words.
That distinction is where the gain sits. A goal written as a number gives the search something concrete to prune against across all 1,500 submissions. A goal written as a description leaves the agent to guess what counts as done. A beam search can't rank guesses against each other the way it ranks measurements. Anyone treating an agent conversation as a single long attempt is leaving that pruning step on the table.
Each link below shares sources, entities, or timing with this story.
Shared entity: GPU / Same source / Shared topic
Both cover GPU; cite the same source (sankalp.bearblog.dev); overlapping topics (agent, idea, instead, kernel).
Shared entity: GPU / Shared topic / Earlier coverage / Tension
Both cover GPU; overlapping topics (agent, kernel); earlier GPU coverage from 2026-03-20.
Shared entity: Keep / Shared topic / Earlier coverage
Both cover Keep; overlapping topics (agent, goal, over); earlier Keep coverage from 2026-05-12.
Shared entity: GPU / Shared topic / Earlier coverage
Both cover GPU; overlapping topics (agent, instead); earlier GPU coverage from 2026-07-29.
Both cover GPU; overlapping topics (approach, instead); earlier GPU coverage from 2026-07-20.
Both cover GPU; overlapping topics (agent, over); earlier GPU coverage from 2026-06-19.
Both cover GPU; overlapping topics (agent, conversation); earlier GPU coverage from 2026-06-07.
Shared entity: Keep / Shared topic / Earlier coverage
Both cover Keep; overlapping topics (agent, ones); earlier Keep coverage from 2026-05-26.