Zurück zu den Videos

Warp

The secret to cheaper, smarter coding agents

Plattform
YouTube
Länge
3:44
Gesprochene Sprache
Englisch (Vereinigte Staaten)
Aufrufe
761
Median-Schnitt
23.27s
Gesicht im Bild
95%
Dunkle Bilder
3%
Warp The secret to cheaper, smarter coding agents at 0:00medium presenter
WarpThe secret to cheaper, smarter coding agents
0:00
Warp The secret to cheaper, smarter coding agents at 0:03medium presenter
WarpThe secret to cheaper, smarter coding agents
0:03
Warp The secret to cheaper, smarter coding agents at 0:03
WarpThe secret to cheaper, smarter coding agents
0:03
Warp The secret to cheaper, smarter coding agents at 0:05
WarpThe secret to cheaper, smarter coding agents
0:05
Warp The secret to cheaper, smarter coding agents at 0:07corner inset
WarpThe secret to cheaper, smarter coding agents
0:07
Warp The secret to cheaper, smarter coding agents at 0:11corner inset
WarpThe secret to cheaper, smarter coding agents
0:11
Warp The secret to cheaper, smarter coding agents at 0:17corner inset
WarpThe secret to cheaper, smarter coding agents
0:17
Warp The secret to cheaper, smarter coding agents at 0:18corner inset
WarpThe secret to cheaper, smarter coding agents
0:18
Warp The secret to cheaper, smarter coding agents at 0:22corner inset
WarpThe secret to cheaper, smarter coding agents
0:22
Warp The secret to cheaper, smarter coding agents at 0:23corner inset
WarpThe secret to cheaper, smarter coding agents
0:23
Warp The secret to cheaper, smarter coding agents at 0:25corner inset
WarpThe secret to cheaper, smarter coding agents
0:25
Warp The secret to cheaper, smarter coding agents at 0:26corner inset
WarpThe secret to cheaper, smarter coding agents
0:26
Warp The secret to cheaper, smarter coding agents at 0:26corner inset
WarpThe secret to cheaper, smarter coding agents
0:26
Warp The secret to cheaper, smarter coding agents at 0:27corner inset
WarpThe secret to cheaper, smarter coding agents
0:27
Warp The secret to cheaper, smarter coding agents at 0:29corner inset
WarpThe secret to cheaper, smarter coding agents
0:29
Warp The secret to cheaper, smarter coding agents at 0:32corner inset
WarpThe secret to cheaper, smarter coding agents
0:32
Warp The secret to cheaper, smarter coding agents at 0:33corner inset
WarpThe secret to cheaper, smarter coding agents
0:33
Warp The secret to cheaper, smarter coding agents at 0:37corner inset
WarpThe secret to cheaper, smarter coding agents
0:37
Warp The secret to cheaper, smarter coding agents at 0:38corner inset
WarpThe secret to cheaper, smarter coding agents
0:38
Warp The secret to cheaper, smarter coding agents at 0:57corner inset
WarpThe secret to cheaper, smarter coding agents
0:57
Warp The secret to cheaper, smarter coding agents at 1:02corner inset
WarpThe secret to cheaper, smarter coding agents
1:02
Warp The secret to cheaper, smarter coding agents at 1:06corner inset
WarpThe secret to cheaper, smarter coding agents
1:06
Warp The secret to cheaper, smarter coding agents at 1:10corner inset
WarpThe secret to cheaper, smarter coding agents
1:10
Warp The secret to cheaper, smarter coding agents at 1:14corner inset
WarpThe secret to cheaper, smarter coding agents
1:14
Warp The secret to cheaper, smarter coding agents at 1:22corner inset
WarpThe secret to cheaper, smarter coding agents
1:22
Warp The secret to cheaper, smarter coding agents at 1:30corner inset
WarpThe secret to cheaper, smarter coding agents
1:30
Warp The secret to cheaper, smarter coding agents at 1:38corner inset
WarpThe secret to cheaper, smarter coding agents
1:38
Warp The secret to cheaper, smarter coding agents at 1:58corner inset
WarpThe secret to cheaper, smarter coding agents
1:58
Warp The secret to cheaper, smarter coding agents at 2:02corner inset
WarpThe secret to cheaper, smarter coding agents
2:02
Warp The secret to cheaper, smarter coding agents at 2:10corner inset
WarpThe secret to cheaper, smarter coding agents
2:10
Warp The secret to cheaper, smarter coding agents at 2:14corner inset
WarpThe secret to cheaper, smarter coding agents
2:14
Warp The secret to cheaper, smarter coding agents at 2:18corner inset
WarpThe secret to cheaper, smarter coding agents
2:18
Warp The secret to cheaper, smarter coding agents at 2:26corner inset
WarpThe secret to cheaper, smarter coding agents
2:26
Warp The secret to cheaper, smarter coding agents at 2:30corner inset
WarpThe secret to cheaper, smarter coding agents
2:30
Warp The secret to cheaper, smarter coding agents at 2:38corner inset
WarpThe secret to cheaper, smarter coding agents
2:38
Warp The secret to cheaper, smarter coding agents at 2:42corner inset
WarpThe secret to cheaper, smarter coding agents
2:42
Warp The secret to cheaper, smarter coding agents at 2:46corner inset
WarpThe secret to cheaper, smarter coding agents
2:46
Warp The secret to cheaper, smarter coding agents at 2:51corner inset
WarpThe secret to cheaper, smarter coding agents
2:51
Warp The secret to cheaper, smarter coding agents at 2:58corner inset
WarpThe secret to cheaper, smarter coding agents
2:58
Warp The secret to cheaper, smarter coding agents at 3:02corner inset
WarpThe secret to cheaper, smarter coding agents
3:02
Warp The secret to cheaper, smarter coding agents at 3:06corner inset
WarpThe secret to cheaper, smarter coding agents
3:06
Warp The secret to cheaper, smarter coding agents at 3:14corner inset
WarpThe secret to cheaper, smarter coding agents
3:14
Warp The secret to cheaper, smarter coding agents at 3:22corner inset
WarpThe secret to cheaper, smarter coding agents
3:22
Warp The secret to cheaper, smarter coding agents at 3:24corner inset
WarpThe secret to cheaper, smarter coding agents
3:24
Warp The secret to cheaper, smarter coding agents at 3:31corner inset
WarpThe secret to cheaper, smarter coding agents
3:31
Warp The secret to cheaper, smarter coding agents at 3:38corner inset
WarpThe secret to cheaper, smarter coding agents
3:38
Warp The secret to cheaper, smarter coding agents at 3:42medium presenter
WarpThe secret to cheaper, smarter coding agents
3:42
Warp The secret to cheaper, smarter coding agents at 3:44
WarpThe secret to cheaper, smarter coding agents
3:44

Transkript von The secret to cheaper, smarter coding agents

Von einem Sprachmodell aus dem Ton des Films transkribiert, nicht von der Marke geschrieben. Klicke auf eine Zeile, um an diese Stelle zu springen.

  1. 0:00Let's see a real benchmark to understand how warp benchmarks work and how we used it to reduce our cost to PR by 63%.
  2. 0:06This is the benchmark that we used.
  3. 0:08You can see that it's generated as a report with an overall recommendation of which model to select as your default for the ages that you tested,
  4. 0:16as well as a breakdown with a Pareto curve across metrics like correctness and average cost, the configuration, and the breakdown across all of your scoring criteria.
  5. 0:26Let's start by looking at the overall recommendation here, which is GBT 5.6 sold.
  6. 0:31This was generated based on the level of correctness that that model achieved relative to other models in the set.
  7. 0:38You also notice additional recommendations called out that could also be a viable choice.
  8. 0:43For example here, GLM 5.3 flash was also very performance and at a lower cost.
  9. 0:49So if cost is the metric that you primarily care about, you may reject the overall recommendation
  10. 0:54and go for the most efficient choice.
  11. 0:56You can also use the Pareto curve to inform that decision.
  12. 0:59We can see a breakdown using our scoring criteria.
  13. 1:03For this test, we just looked at the correctness
  14. 1:05of the model, could it complete all of the tasks
  15. 1:07that we set up for it,
  16. 1:08and we can see metrics that we may care about.
  17. 1:10Average cost is going to be the defaults,
  18. 1:12but you can also look at average time
  19. 1:14if you're more concerned with latency,
  20. 1:15as well as a breakdown of output tokens and tool calls,
  21. 1:18which are useful in debugging scenarios.
  22. 1:20You wanna find models that score highest
  23. 1:22on the correctness scale so further up is best. You also want to find the leftmost models which
  24. 1:27are going to have the lowest average cost. And we can imagine a curve that we could plot going from
  25. 1:32the bottom left of the chart to the top right. And the models that fall along this curve are going to
  26. 1:37be GLM 5.3 which achieves relatively high correctness at the lowest cost and GPT 5.6 Sol
  27. 1:43which achieves the highest correctness at a bump in cost. And all the models that fall below this
  28. 1:49curve which are Grok, Opus, and Gemini in this testing are dominated meaning we
  29. 1:54don't need to look at those. And down here we can see a breakdown across each
  30. 1:58of the tasks that were tested for all of our models. And all of these tasks are
  31. 2:02replaying real agent runs that we have put through the factory in the past.
  32. 2:06Clicking into the configurations up here we can see each of those runs. We can
  33. 2:11click into the trial which is our replay of that agent run with that model
  34. 2:15selection and a mock of the MCP servers and secrets that were available in that original
  35. 2:21agent run. We can also view the judges that reviewed that trial run to determine its correctness.
  36. 2:27We can see this scoring agent, which is sort of our review agent that's given a rubric of what
  37. 2:32correctness looks like, give it a pass. And these scores can use rubrics that you define as well as
  38. 2:38model selections to fine-tune the cost of running these benchmarks. Now to close the loop you can
  39. 2:43use the results of these benchmarks to change your default models or even configure custom routers to
  40. 2:48send different tasks to the right model for the job. With warp factories all of this is configured
  41. 2:53as code. In our case we set our implementation agent to use GPT 5.6 sole by default and now
  42. 2:59whenever we trigger a factory agent to work on this repository we're going to be using this model.
  43. 3:04And if you want to get fine grained with your selections you can create a custom model router.
  44. 3:09Here we have our default set to GPT 5.6, but if there are specific tasks that could be routed to a more efficient model like
  45. 3:16Grok 4.6 medium for example
  46. 3:19We could define a description of these sorts of tasks that should be routed to that model instead in our case
  47. 3:24We ran a benchmark specifically focused on front-end UI tasks and we found that grok 4.6 medium was a high performer
  48. 3:31So we added a routing rule to our agent saying that simple front-end and UI tasks should use this model instead
  49. 3:37If this seems useful to you, factory benchmarks are part of warp factories, and you can get early access here.
49 Zeilen · 748 Wörter