吳恩達 Agentic AI 第四講 Benefits of agentic AI
- URL: https://www.youtube.com/watch?v=bR8SkENV10Y
- Channel: 扣子說AI
- Fetched: 2026-08-12T17:31:07+00:00
- Language: English (auto-generated)
Transcript
[00:03] I think the one biggest benefit of
[00:04] agentic workflows is that allows you to
[00:07] do many tasks effectively that just
[00:09] previously were not possible. But there
[00:11] are other benefits as well, including
[00:13] parallelism that lets you do certain
[00:15] things quite fast, as well as modularity
[00:17] that lets you combine the best of breed
[00:19] components from many different places to
[00:21] build an effective workflow. Let’s take
[00:24] a look. My team collected some data on a
[00:27] coding benchmark that tests the ability
[00:30] of different LLMs to write code to carry
[00:32] out certain tasks. The benchmark used in
[00:35] this case is called HumanEval, and it
[00:37] turns out that GPT-3.5, this is a model
[00:40] that the first publicly available
[00:41] version of ChatGPT was based on, if
[00:43] asked to write the code directly, to
[00:45] just type out the computer program, gets
[00:47] 48% right on this benchmark. Uh, this is
[00:50] a positive key metric. GPT-4 is a much
[00:52] better model. Its performance leaps to
[00:55] 67%
[00:56] with this also non-agentic workflow. But
[00:59] it turns out that as large as the
[01:01] improvement was from GPT-3.5 to GPT-4,
[01:05] that improvement is dwarfed by what you
[01:06] can achieve by wrapping GPT-3.5 within
[01:10] an agentic workflow. Using different
[01:12] agentic techniques, which you’ll learn
[01:14] about later in this course, you can
[01:15] prompt GPT-3.5 to write code and then
[01:18] maybe reflect on the code and figure out
[01:20] if you can improve it. And using
[01:22] techniques like that, you can actually
[01:24] get GPT-3.5 to get much higher levels of
[01:27] performance. And similarly, GPT-4 used
[01:29] in the context of an agentic workflow
[01:32] also does much better. So, even with
[01:34] today’s best LLMs, an agentic workflow
[01:37] lets you get much better performance.
[01:40] And in fact, what we saw in this example
[01:42] was the improvement from one generation
[01:44] of model to another, which is huge, is
[01:47] still not as big a difference as
[01:49] implementing an agentic workflow on the
[01:51] previous generation of model. Another
[01:53] benefit of using agentic workflows is
[01:56] that they can paralyze some tasks and
[01:59] thus do certain things much faster than
[02:01] a human. For example, if you ask an
[02:04] agentic workflow to write an essay about
[02:06] black holes, you might really have three
[02:08] LLMs running parallel to generate ideas
[02:11] for web search terms to type into search
[02:14] engine. Based on the first web search,
[02:16] it may identify, say, three top results
[02:19] to fetch. And based on the second web
[02:21] search, it may identify a second set of
[02:24] web pages to fetch, and so on. And it
[02:26] turns out that whereas a human doing
[02:28] this research would have to read these
[02:31] nine web pages sequentially or one at a
[02:33] time, when you’re using an agentic
[02:35] workflow, you can actually paralyze all
[02:37] nine web page downloads and then finally
[02:40] feed all these things into the LLM to
[02:42] write an essay. So, even though agentic
[02:44] workflows do take longer than truly
[02:47] non-agentic workflows sold by direct
[02:49] generation by just prompting a single
[02:51] time, if you were to compare this type
[02:53] of agentic workflow to how a human would
[02:55] have to go about the task, the ability
[02:57] to paralyze downloading lots of web
[02:59] pages can actually let it do certain
[03:02] tasks much faster than the non-parallel,
[03:04] sequential way that a single human might
[03:07] process this data. To build on this
[03:09] example, it turns out one of the things
[03:11] I often do when building agentic
[03:13] workflows is look at the individual
[03:15] components like the LLM and add or swap
[03:18] out components. So, for example, maybe I
[03:21] look at the web search engine I use up
[03:23] here and I might decide that I want to
[03:25] swap in a new web search engine. When
[03:27] building agentic workflows, there’s
[03:28] actually multiple web search engines
[03:30] including Google, which you can access
[03:32] by Serper, as well as others like Bing,
[03:35] DuckDuckGo, Swisscows, E.com. There are
[03:37] actually quite a lot of options for web
[03:39] search engines designed for LLMs to use.
[03:42] Or maybe instead of just doing three web
[03:44] searches, maybe on this step you can
[03:45] swap in a new news search engine so they
[03:49] can find out what’s the latest news on
[03:51] recent breakthroughs on black hole size?
[03:55] And lastly, instead of using the same LM
[03:58] for all of the different steps, I will
[04:00] often try out different large language
[04:02] models and maybe try out different LM
[04:05] providers to see which one gives the
[04:07] best result for different steps of this
[04:10] system. So, to summarize, the main
[04:13] reason I use agentic workflows is it
[04:15] just gives much better performance on
[04:16] many different applications. But, in
[04:19] addition, it can also parallelize some
[04:22] tasks that humans would otherwise have
[04:24] to do sequentially. And the modular
[04:26] design of many agentic workflows also
[04:29] lets us add rapid tools and sometimes
[04:32] swap out models. We’ve talked a lot
[04:34] about the key components of building
[04:36] agentic workflows. Let’s now take a look
[04:39] at a range of agentic AI applications to
[04:42] give you a sense of the sorts of things
[04:44] people are already building and the
[04:45] sorts of things you build a build
[04:47] yourself. Let’s go on to the next video.