顧哥知識系統

吳恩達 Agentic AI 第三講 Degrees of autonomy

閱讀時間約 6 分鐘

吳恩達 Agentic AI 第三講 Degrees of autonomy

Transcript

[00:02] Agents can be autonomous to different
[00:04] degrees.
[00:05] A few years ago, I noticed within the AI
[00:07] community that there was a growing
[00:09] controversial debate about what is an
[00:11] agent, and some people are writing a
[00:13] paper saying I built an agent, and
[00:15] others say no, that’s not really a true
[00:16] agent. And I felt this debate was
[00:19] unnecessary. Which is why I started
[00:21] using the term agentic because I thought
[00:24] if we use it as an adjective rather than
[00:26] a binary is either an agent or not, then
[00:29] we can have just acknowledge that
[00:30] systems can be agentic to different
[00:32] degrees, and this is called out agentic
[00:35] and move on with the real work of
[00:37] building these systems rather than
[00:38] debating, you know, is this sufficiently
[00:40] autonomous to be an agent or not. I
[00:42] remember when I prepared the talk on
[00:45] agentic reasoning, one of my team
[00:47] members actually came to me and said,
[00:48] “Hey Andrew, we don’t need yet another
[00:50] word. You know, we have agent, why you
[00:52] make up another word agentic?” But I
[00:54] decided to use it anyway, and then later
[00:56] on wrote an article in the Given Ideas
[01:00] newsletter the batch, and then also post
[01:02] on social media saying that instead of
[01:04] arguing over which words include or
[01:06] exclude as being a true agent, let’s
[01:08] acknowledge the different degrees to
[01:10] which systems can be agentic. And I
[01:12] think this helped move past the debate
[01:15] on what is a true agent, and let us just
[01:18] focus on actually building them. Some
[01:20] agents can be less autonomous. So, take
[01:23] the example of writing an essay about
[01:26] black holes. You can have a relatively
[01:28] simple agent to come up with a few web
[01:31] search terms or web search queries, then
[01:34] you can hardcode in that you call web
[01:36] search engine, fetch some web pages, and
[01:39] then use that to write an essay. And
[01:41] this would be an example of a less
[01:42] autonomous agent with a fully
[01:44] deterministic sequence of steps, and
[01:47] this will work okay. In terms of
[01:49] notational convention, throughout this
[01:51] course I’ll use the red color, as you
[01:53] see here on the left, to denote the user
[01:55] input, such as a user query in this
[01:57] case, or in later examples, maybe the
[01:59] input document into an agentic workflow.
[02:02] The gray boxes denote calls to an LLM,
[02:05] and the green boxes, like the web search
[02:09] and the web fetch boxes that you see
[02:11] here, indicate steps where other
[02:13] software is being used to carry out an
[02:15] action, such as a web search API call,
[02:17] or executing code to fetch the contents
[02:19] of a website. Then, an agent can be more
[02:22] autonomous, where, given a request to
[02:24] write an essay about black holes,
[02:26] perhaps you let the LLM decide does it
[02:28] want to do a web search, or does it want
[02:30] to search recent news sources, or does
[02:32] it want to search for recent research
[02:34] papers on the website archive? Based on
[02:36] that, maybe in this example, the LLM,
[02:39] not the human engineer, but the LLM
[02:41] chooses in this case to call web search
[02:43] engine, and then after that, you may let
[02:45] the LLM decide how many web pages does
[02:47] it want to fetch? Or if it fetch the
[02:50] PDF, does it need to call a function, or
[02:53] also call the tool to convert the PDF to
[02:55] text? And in this case, maybe it fetches
[02:57] its top few web pages, then it can write
[03:00] an essay, decide whether or not to
[03:02] reflect and improve, and maybe even go
[03:04] back to fetch more web pages, and then
[03:06] it finally produce an output. And so,
[03:08] even for this example of a research
[03:11] agents, we can see that some agents can
[03:14] be less autonomous with a linear
[03:17] sequence of steps to be executed
[03:18] determined by a programmer, and some
[03:20] could be more autonomous, where you
[03:22] trust the LLM to make more decisions,
[03:24] and the exact sequence of steps that
[03:26] happens may be even determined by the
[03:27] LLM, rather than in advance by the
[03:29] programmer. So, for less autonomous
[03:32] systems, you will usually have all the
[03:34] steps predetermined in advance, and any
[03:36] functions it calls, like web search, and
[03:38] we’ll call that tool use, as you learn
[03:40] in the third module in this course,
[03:42] might be hard-coded by the human
[03:44] engineer, by you and me, and most of the
[03:46] autonomy is in what text the LLM
[03:48] generates. At the end of the spectrum
[03:50] would be highly autonomous agents, where
[03:52] the agent makes many decisions
[03:54] autonomously, including, for example,
[03:56] deciding what is the sequence of steps
[03:58] it will carry out in order to write the
[04:00] essay. And there’s some highly
[04:02] autonomous agents that can even write
[04:04] new functions, or we sometimes say
[04:06] create new tools that it can then
[04:08] execute. And somewhere in between are
[04:10] semi-autonomous agents, where they can
[04:12] make some decisions, choose tools, but
[04:14] the tools are usually more predefined.
[04:16] As you look at different examples in
[04:18] this course, you’ll learn how to build
[04:19] applications anywhere on this spectrum
[04:21] of less to more highly autonomous. And
[04:24] you find that there are tons of
[04:25] applications in the less autonomous end
[04:27] of the spectrum, they’re very valuable
[04:28] being built for tons of businesses
[04:30] today.
[04:31] And at the same time, there are also
[04:33] applications being worked on at the more
[04:35] highly autonomous end of the spectrum,
[04:37] but those are usually less easily
[04:39] controllable, a little bit more
[04:41] unpredictable, um and also a lot of
[04:43] active research as well to figure out
[04:45] how to build these more highly
[04:47] autonomous agents.
[04:49] And with that, let’s go on to the next
[04:51] video to dive deeper into this and to
[04:54] hear about some of the benefits of using
[04:56] agents and why they allow us to do
[04:58] things that just were not possible with
[05:00] earlier generations of these
[05:02] applications.

關係圖譜