Full transcript
0:00Have you ever wondered which model you
0:01should be using for which use case?
0:04Well, stay tuned to the end of this
0:05video as I'm going to cover the 10
0:06models that I'm using and the use cases
0:09that I actually deploy for them to scale
0:11my business from 0 to 10 million.
0:14Now, one of the challenges that we have
0:16is which model shall I use, which LLM
0:19shall I use? There's thousands and
0:21thousands of different models out there
0:23and it's a nightmare knowing which one
0:25to use because the way that they portray
0:26these models is incredibly complex. They
0:29talk about different context windows,
0:31different token amounts, different
0:33reasoning levels, different tools that
0:34these um these LLMs or models come
0:37equipped with. It's a headache. So,
0:39let's break down exactly what I'm using
0:42and how I'm doing it. So, we're going to
0:44start with the company. So, first of
0:45all, OpenAI. Now, for a long time, I
0:47actually stopped using OpenAI because
0:49they pretty crap. Um but recently,
0:51they've started to produce some really
0:53good models. So, starting with the first
0:56one, which is 40. here. I'm using it
0:58particularly for just image generation.
1:00Uh and it's incredibly good. I used to
1:02use uh midjourney but I've transferred
1:04across to uh OpenAI just because it's
1:06very easy. You can just use it from
1:08within the app. You don't have to be on
1:09Discord or anything like that as well.
1:11So that is my image generation which
1:13we're using for blogs. We're using for
1:15content on LinkedIn. We're using it for
1:16a whole range of um use cases. The other
1:21is 4.1. This is one of their new
1:23reasoning models that they've just
1:24launched and it's basically a cheaper
1:26version of their 03 which is like a
1:28super reasoning model. And what is 41
1:32doing for me? Well, it controls most of
1:34my agents. So you would have seen from
1:36the other videos any agents that I have
1:39within NA10 within
1:41any custom agents that I build via code
1:43or within triggery I usually have 4.1
1:46being the controller of that agent
1:48because they can query tools very well
1:50and it has a good level of reasoning.
1:54Then we have uh 4.5. Now this is an
1:56expensive model. So I don't use this all
1:58the time if I'm totally honest. I only
2:00use it if I really want the copy of that
2:02email to be good. And here I'm doing
2:04more copywriting and potentially some
2:06posts uh for LinkedIn as well. So it can
2:09act as like my content generator agent
2:12when I use it use it, but it is
2:14occasionally. The other is 03. And 03 is
2:16probably one of my favorite models out
2:19there right now. Hands down. It's just
2:21incredible. And the challenge with it is
2:24it's a little bit expensive, but the
2:25reasoning on it and its thought process
2:28and the answers that it give you is the
2:30best that you will ever find anywhere.
2:32So here I use it for anything that I
2:34want more of a complex answer to. So the
2:37will be like a deep research task if
2:39it's researching something for me. I use
2:42it as a business coach within like the
2:43project sections within OpenAI. So it's
2:45like my mentor, it's my coach, it's my
2:47advisor because it knows everything and
2:49it's connected to all of the different
2:50business uh data points that I've
2:52connected it to. Um, and I'm also using
2:55it for PRDS, which is basically when I'm
2:57building out a new um, part of the
3:00platform where I'm building out like a
3:01micro SAS. I use it to basically help
3:03build out what that product needs to
3:05look like, the tools that I need to use
3:07because it can go and run that research
3:08incredibly well. And then last but not
3:10least within the OpenAI family, it's
3:12Codeex, which is their very latest model
3:14that they just launched, which
3:15effectively is their equivalent to a
3:17cursor challenger. uh and codeex I'm
3:20using to write a lot of tests for all of
3:22these different SAS applications that we
3:24have uh being developed right now.
3:27Claude or anthropic I'm using um three
3:30different models here 3.5 3.7 and the
3:323.7 pro max. Now 3.5 for me is more of
3:36like a data tagging. So, if I'm using
3:39stuff inside of Clay or Freckle or with
3:42inside of NA10 and I just want to tag
3:44whether this person is XY Z or I just
3:46want to run some light research, 3.5 is
3:48the one to go for. It's very very
3:50efficient, very very effective, follows
3:53the tasks and is pretty cheap. Uh, in
3:56all things considered, 3.7 is my
3:59allrounder when I'm looking at doing
4:01most things. Uh, and I typically start
4:04with 3.7 and then from 3.7 I branch away
4:07to other models to check to see if
4:08they're any better. If they're not, I
4:10stick with 3.7 for pretty much most
4:12things. Uh, maybe recently Gemini 2.5 is
4:16starting to kind of encroach in that
4:18territory of being like my allrounder
4:20model. and then the 3.7 um kind of max
4:24thinking. I use this inside a cursor for
4:26basically building out any uh
4:28applications and code. From a Google
4:31perspect perspective, we have Gemini 2.5
4:34and Gemini 2.5 Pro Max. So the Promax
4:37again is just used for coding inside of
4:38cursor and the 2.5 at the moment is used
4:41for um executing on certain plans and
4:43helping me write a plan uh for something
4:46as well. So, a lot of the time you
4:48you'll find that I start with V3 to do
4:50some research and then I take it across
4:52to Gemini 2.5 to kind of start
4:54formulating a plan on how I would attack
4:56a certain area um using Gemini because
4:58it's great at doing that type of thing.
5:01But like I said, it's starting to become
5:02my allrounder. It's an incredibly
5:04effective and powerful model. And
5:06listen, it's free depending on uh you
5:08know what type of Google subscription
5:10you have as well. So, you can really
5:11take it to town. And there are some
5:12great open-source models on that as
5:14well. Now, like I said, I am using
5:17things like Plexity. I'm using things
5:19like DeepSk. I'm using um various
5:21different models on across hugging face
5:23and things like that. But if I had to
5:25break it down to 10 models, those are
5:27the 10 models that I use on a day-to-day
5:29basis without fail.