Stop Hallucinations! Unlock AI Secrets with Context

Description

Chapters

Transcript

0:00 How do I know
0:02 when to use Chat GPT, Claude, Gemini, or
0:06 another AI tool?
0:09 Aren't they basically the same?
0:12 And do they have strengths?
0:15 Well,
0:16 okay, that's a great question.
0:20 Um
0:25 if you're just getting started with AI,
0:29 it really doesn't matter which one you
0:31 use.
0:33 Um
0:34 because part of the part of the skill of
0:37 of
0:38 learning to use a large language model,
0:40 so Chat GPT is the most famous one,
0:43 Anthropic is the sort of darling
0:45 recently. Um Gemini from Google's really
0:47 good. Um what was the other one that was
0:49 mentioned here? Oh, another AI tool. Um
0:53 they all basically do the same thing.
0:55 You put in a prompt, they generate an
0:57 answer.
0:59 If you're if you're new to these things,
1:01 one of the things that you learn with a
1:04 large language model is that it's
1:06 actually very different than a Google
1:08 search. What Google has trained us over
1:09 20 years
1:11 is to
1:12 sort of treat
1:14 these interactions like a vending
1:15 machine. I'm going to put in a short
1:16 little prompt and it's going to give me,
1:18 you know, a bunch of links and then I'm
1:19 going to go deal with them.
1:21 Okay, that's perfect. That's great.
1:23 Um
1:27 with large language models like Chat
1:29 GPT, they need a lot of what's called
1:32 context. They need a lot of like
1:36 information about what you're talking
1:38 about.
1:40 The challenge with large language models
1:43 is that they've been trained on like all
1:45 the crap,
1:46 right? All the crap on the internet. Um
1:49 in fact, it's literally what what the
1:51 what these uh frontier model companies
1:53 do is they have robots that go out and
1:56 just scour web pages and grab all that
1:59 information and they do this magic
2:00 mathematical thing that gets all of the
2:03 information and it it encodes it in a
2:05 certain way
2:07 that it makes it instantly searchable.
2:09 It's actually quite remarkable
2:11 that um you can just
2:14 query the knowledge of everything.
2:16 The problem with that is if you've got
2:18 all of the knowledge out there, if you
2:20 ask a simple question like tell me about
2:23 this particular math formula, for
2:25 example, I don't query about math
2:28 problems because I'm shitty at math and
2:30 I don't care about it, but some people
2:31 like that.
2:33 But if you query about a specific math
2:34 problem with no context,
2:37 it will not only get the answers to the
2:39 the the right answers to the problem, it
2:41 will also bring in the wrong answers to
2:43 the problem and it'll bring in
2:44 conjectures about the problem, some of
2:46 which are great, some of which are
2:47 horrible.
2:49 And it sort of amalgamates them all
2:50 together and will give you
2:53 some answer that will look very right um
2:56 and often it is very wrong, right?
2:58 Because
3:00 um it's what's called hallucinations. It
3:01 just hallucinates. It it has
3:04 put things in there that it just made up
3:06 essentially or it it pulled from, you
3:08 know, different different avenues.
3:11 Um if you give it more context like,
3:14 "Hey, I want to talk about this math
3:15 problem from this
3:17 professor and, you know, I I want you to
3:20 act as a mathematics genius and things
3:24 like that." It will narrow
3:26 your question into a into a more narrow
3:28 thing. So, the general answer is if
3:30 you're if you're just starting with
3:32 large language models, it doesn't matter
3:34 which one you choose.
3:36 Think of the conversation more like a
3:39 conversation than a vending machine. So,
3:41 if it gives you an answer that you don't
3:43 like, tell it that you don't like it.
3:44 Tell it why you don't like it. Tell it
3:46 who you are, what you're trying to
3:47 accomplish.
3:48 And the answer will get better and
3:49 better and better. In In
3:51 I heard today in office
3:52 that uh a study was just done
3:55 Watch the full replay at
3:57 community.thesalon.ai.