Latest / How I Work / Which AI model should I use for which task?
Transcript
- 0:00Speaker 1: If you've opened up an AI tool lately, instead at
- 0:03Speaker 1: a list of names like flash so one, it opens Fable, Mythos,
- 0:08Speaker 1: plus a pile of numbers that seem to go up
- 0:11Speaker 1: and down at random, You are not alone. The naming
- 0:15Speaker 1: within these platforms is genuinely confusing, and most of us
- 0:19Speaker 1: just pick one and hope for the best. Today, Neo
- 0:23Speaker 1: and I walk through the big four platforms Gemini, Chatch, FT, Cord,
- 0:26Speaker 1: and Copilot and sort out which model to actually use
- 0:30Speaker 1: and when. You will learn some tricks that cut through
- 0:34Speaker 1: all these confusing names and how to avoid the token
- 0:38Speaker 1: bills that have left some companies hundreds of thousands of
- 0:42Speaker 1: dollars out of pocket just for one employee using the
- 0:47Speaker 1: wrong models for the wrong tasks. By the end of
- 0:51Speaker 1: this episode, you will know exactly how much brain power
- 0:55Speaker 1: to throw at any given task for the AI tool
- 0:59Speaker 1: that you're using.
- 1:06Speaker 2: Welcome to how IAI with me, Doctor Amantha Imba and
- 1:10Speaker 2: Neo Applin, head of Inventium AI. Each episode we share
- 1:15Speaker 2: one practical way to use AI better at.
- 1:17Speaker 1: Work and in life. No fluff, no tech jargon, just
- 1:21Speaker 1: things you can use straight away. So Neo I feel
- 1:25Speaker 1: like in the news at the moment, there is a
- 1:27Speaker 1: lot of talk about new and different models launching, like
- 1:30Speaker 1: Mythos and Fable, and they've all got different names, and
- 1:33Speaker 1: then there's different numbers, and I feel like it can
- 1:36Speaker 1: be very confusing to the average person what is going
- 1:40Speaker 1: on with all these models you have?
- 1:42Speaker 3: These AI companies are horrendous when they're naming their models.
- 1:45Speaker 3: There's agents, there's GPTs, there's projects as models, there's fable.
- 1:49Speaker 3: There's too many different terms that mean very little to
- 1:52Speaker 3: the average person. But what it's about is how much
- 1:57Speaker 3: the model thinks and how deeply it thinks through your problem.
- 2:01Speaker 3: Is for the different models you've got, and so you've
- 2:03Speaker 3: got smaller light models that do maybe a little bit
- 2:05Speaker 3: of thinking, and you might have the heavier thinking models
- 2:08Speaker 3: that are smarter and will work through your problem more. So,
- 2:11Speaker 3: that's what it's about. It's about how much brain you
- 2:14Speaker 3: would like to put towards your problem. Now, the thinking
- 2:19Speaker 3: models will cost more for these companies to run. Therefore
- 2:22Speaker 3: they'll let you use it less each hour, each month,
- 2:25Speaker 3: each year, whatever, in a particular period of windows. And
- 2:29Speaker 3: so you'll find that if you use the deep thinking
- 2:32Speaker 3: big models, you'll get fewer uses each hour each day
- 2:36Speaker 3: than you would if you're using the small and light models.
- 2:39Speaker 3: So it's really about you're getting the right model for
- 2:41Speaker 3: the right task. And so if you've got something important,
- 2:44Speaker 3: use the thinking model. When you don't have anything particularly important,
- 2:46Speaker 3: use a smaller model.
- 2:48Speaker 1: So let's unpack the four main AI platforms in terms
- 2:54Speaker 1: of getting into Gemini, CHATCHABT, Claude and copile it to
- 2:59Speaker 1: understand and what are the models within each one and
- 3:02Speaker 1: what should we be using when? How does that sound
- 3:05Speaker 1: as a plan?
- 3:06Speaker 3: Love? It sounds great.
- 3:07Speaker 1: Let's start with Gemini, I think, and that's probably anecdotally
- 3:11Speaker 1: for us and the clients that we're working with with
- 3:14Speaker 1: inventium AI. We're not saying too many Gemini users. But
- 3:18Speaker 1: let's start there and cover off what should we be
- 3:20Speaker 1: using for what in Gemini.
- 3:22Speaker 3: Now, when I go into this, I'll tell you the names,
- 3:24Speaker 3: but also underneath the names when you've got there's a
- 3:27Speaker 3: model picker on each of these tools. Usually it's in
- 3:30Speaker 3: the prompting bar or around the prompting bar. When you
- 3:33Speaker 3: see the names, there's always a little bit of text
- 3:36Speaker 3: underneath it which gives you a clue as to what
- 3:38Speaker 3: it is. So, for example, here in Gemini on my
- 3:41Speaker 3: pick list today, I've got three point one flash light
- 3:44Speaker 3: and then underneath that it goes fastest dances. Underneath that
- 3:47Speaker 3: you've got three point five flash and it goes all
- 3:50Speaker 3: round help. And then the three point one pro says
- 3:53Speaker 3: advanced math and code. So it's giving you a clue
- 3:56Speaker 3: of which one to use when. So you can then
- 3:59Speaker 3: look at your go other numbers. Yeah, who gets but
- 4:01Speaker 3: the numbers. The numbers will always change, but it's a
- 4:03Speaker 3: little comments underneath that often mean more than the actual
- 4:07Speaker 3: name of it. So in Gemini they've got three point
- 4:10Speaker 3: one flash light, three point five flash and three point
- 4:13Speaker 3: one pro. Now the numbers can be confusing because you
- 4:16Speaker 3: look at it going three point one pro. Well, three
- 4:18Speaker 3: point five is better than three point one, so three
- 4:22Speaker 3: point five flash should be better than three point one pro.
- 4:25Speaker 3: It doesn't actually work that way. So that's why I'm
- 4:27Speaker 3: suggesting forget the numbers, go for the little text underneath it,
- 4:30Speaker 3: so that the pro one, the advanced method code, would
- 4:33Speaker 3: be the smartest model. And so when would you use that?
- 4:35Speaker 3: You'd use it on things that matter. So if you're
- 4:38Speaker 3: working out something like strategy, or you're doing problem solving
- 4:41Speaker 3: or Hey, I'm dealing with this really complex problem and
- 4:43Speaker 3: a market thing and a research thing, and how do
- 4:45Speaker 3: I put those things together? Then you'd use the pro
- 4:48Speaker 3: model here everyday use, I'd probably use flash And then
- 4:52Speaker 3: if you've got something just as quick fast light though,
- 4:54Speaker 3: summarize this paragraph into two words or whatever, I'd probably
- 4:57Speaker 3: use flash light for that.
- 4:59Speaker 1: And we'll get in to chat chapt next. But just
- 5:02Speaker 1: as a general question with the deeper versus quicker thinking,
- 5:08Speaker 1: something I wonder when let's say I'm working on refining
- 5:12Speaker 1: a really important piece of content, and I think to myself,
- 5:16Speaker 1: this is not a complex task like refining copy based
- 5:20Speaker 1: on the rules that I give it, but it's a
- 5:23Speaker 1: really important one to me. So in that kind of scenario,
- 5:27Speaker 1: which model should I be using.
- 5:30Speaker 3: If it's more language and just summarize these things, then
- 5:34Speaker 3: I would be fine with a smaller model, one of
- 5:37Speaker 3: those just everyday ones, which in this case in Gemini,
- 5:41Speaker 3: the every day one is probably the three point five
- 5:43Speaker 3: flash or you can get away with a flash light
- 5:45Speaker 3: if it's just words and how do I put these
- 5:47Speaker 3: phrases together? Things like that, But if you want it
- 5:50Speaker 3: to actually analyze things and think through things, if it's
- 5:53Speaker 3: important to get it right. In your case, it's important
- 5:55Speaker 3: to get it right. I would definitely be using the
- 5:58Speaker 3: everyday one here, that's the three point five flash rather
- 6:01Speaker 3: than light because it's actually think through how it does
- 6:03Speaker 3: its word summary, those kind of things. But often we
- 6:07Speaker 3: need to do things that really matter. If it really matters,
- 6:09Speaker 3: I'm often doing PRO. But if it's really matter, usually
- 6:12Speaker 3: it's about thinking through a problem really matters, rather than
- 6:15Speaker 3: just summarize the sentence. Well, I need to summarize it, right, Yeah,
- 6:18Speaker 3: probably everyday model is probably fine for those things.
- 6:21Speaker 1: Okay, let's move on to chat GPT. What are the
- 6:24Speaker 1: different model options in chat GPT and what should we
- 6:28Speaker 1: be using when they've improved.
- 6:30Speaker 3: There's an instant then thinking, then PRO. That makes a
- 6:34Speaker 3: lot more sense to me. Instant is don't think, just
- 6:37Speaker 3: shoot me the answer straight away. And I love that
- 6:39Speaker 3: and that's exactly what it does. Instant effectively, don't think,
- 6:42Speaker 3: just do so use your own brain, go ahead thinking,
- 6:46Speaker 3: we'll actually think through it. That's what I'm generally on
- 6:49Speaker 3: all of the time. And then PRO, if it matters,
- 6:53Speaker 3: that's the one you'd pick, So do not. You get
- 6:55Speaker 3: quite a few pro uses, so a lot of people
- 6:57Speaker 3: are too scared to use them until they actually need them,
- 7:01Speaker 3: in which case you've actually got a whole bunch of
- 7:02Speaker 3: potential uses that you've missed out on using. So do
- 7:05Speaker 3: not be afraid to use the higher models.
- 7:08Speaker 1: So let's look at Claude now, which I know is
- 7:13Speaker 1: the AI tool that we use most often, and that
- 7:17Speaker 1: certainly we're saying a lot of our clients moving a
- 7:19Speaker 1: lot more towards Claude. And this is where certainly the
- 7:24Speaker 1: term fable you might have heard in relation to Claude.
- 7:27Speaker 1: What's going on there and what models should we be using?
- 7:30Speaker 3: Okay, so Claude have four models at the moment, well
- 7:34Speaker 3: almost four. So there starts at Haiku, that's their small
- 7:38Speaker 3: fast light model, then it goes to Sonnet, then to opus,
- 7:41Speaker 3: and then more recently there is fable. So instead of
- 7:44Speaker 3: giving you the words like chachipt does what's chechibit? He
- 7:48Speaker 3: got instant thinking and pro it kind of does what
- 7:50Speaker 3: it says in the box. I kind of like that
- 7:52Speaker 3: now with Claude, think of it as poems. A haiku
- 7:55Speaker 3: is a Japanese poem which only has a couple of words.
- 7:57Speaker 3: If it's only that three lights it's a tiny poem
- 8:00Speaker 3: is a larger poem, and Opus is quite a large poem,
- 8:02Speaker 3: and a Fable is obviously a huge poem. So it's
- 8:05Speaker 3: about poem sizes. That's how they've named their models. So
- 8:08Speaker 3: if you kind of get that in your mind, and.
- 8:09Speaker 1: Can I just say neo, I've never heard anyone describe
- 8:13Speaker 1: it like that. To me, that is quite revelatory, So
- 8:16Speaker 1: thank you. Continue.
- 8:18Speaker 3: So, yeah, these are the poem sizes and effective. It's
- 8:21Speaker 3: how smart you want it to be and how much
- 8:23Speaker 3: thinking you want it to do as well. So, as
- 8:27Speaker 3: I said, we've got hik So and Opus. Don't worry
- 8:29Speaker 3: about the numbers, just go for the names here. The
- 8:31Speaker 3: numbers will always change. Now, just recently they had had
- 8:35Speaker 3: released Fable. If you've also heard of Mythos, Mythos and
- 8:39Speaker 3: Fable are pretty much the same model. This is the
- 8:41Speaker 3: way you saw in the news. It was like, oh,
- 8:42Speaker 3: we can't release it. It's too scary, it's too smart,
- 8:45Speaker 3: a little hack every system in the world and all
- 8:47Speaker 3: those kind of things. So Fable is Mythos. It's the
- 8:51Speaker 3: full blown model. But they've just said you can't talk
- 8:54Speaker 3: about certain things, so you can't talk about chemical stuff
- 8:57Speaker 3: because someone might create chemical warfware. You can't talk about
- 9:01Speaker 3: security stuff because you might try to hack something. And
- 9:04Speaker 3: so what they've tried to do is get Fable without
- 9:06Speaker 3: those problematic topics, and then they released it. They've released it,
- 9:12Speaker 3: and since then there's obviously politics involved, but the US
- 9:16Speaker 3: government has said, well, in a way, you could actually
- 9:19Speaker 3: get it to review your security on your code, and
- 9:22Speaker 3: so therefore, no, we're not happy with that, and because
- 9:26Speaker 3: of that there's been discussions, and then Fable has been
- 9:30Speaker 3: paused will put it that way, and so you can't
- 9:32Speaker 3: actually use Fable at the moment. So for most people,
- 9:36Speaker 3: do you need something which is as awesome as Fable, No,
- 9:39Speaker 3: I'd probably say no. If you're getting AI to build
- 9:44Speaker 3: your full software suite or whatever, Fable would be absolutely
- 9:47Speaker 3: amazing for that. Opus is great Fable from everything I'm
- 9:50Speaker 3: reading from people who've been using it, it's mind blowingly amazing.
- 9:54Speaker 3: But it's not available just yet. So what I'd be
- 9:56Speaker 3: doing is Haiku small fast light every day kind of small,
- 10:00Speaker 3: like summarize this kind of stuff. Haiku is great. I
- 10:03Speaker 3: work in Sonnet pretty much all the time, but then
- 10:06Speaker 3: I turn it up to Opus when I need to
- 10:09Speaker 3: because Opus is it does use more tokens, which we
- 10:11Speaker 3: talk about later. But Opus is very smart and does
- 10:14Speaker 3: an awesome job. But don't sit in the most thinkingest
- 10:18Speaker 3: model all the time because you're just going to waste
- 10:20Speaker 3: all your usage and then you won't have the great
- 10:22Speaker 3: usage ready for when you need to use it.
- 10:25Speaker 1: Okay, let's now talk about co pilot. What are the
- 10:29Speaker 1: different models.
- 10:30Speaker 3: There right now? I need to give you a little
- 10:32Speaker 3: bit of market knowledge here. So Microsoft have had invested
- 10:37Speaker 3: in chat ChiPT, so actually Copilot under the hood has
- 10:40Speaker 3: a chat chipitt Brain, not just that you can use
- 10:43Speaker 3: chat Chipt's models in Copilot, and they've also recently done
- 10:48Speaker 3: a deal with Anthropic so you can use Claude inside copilot.
- 10:54Speaker 3: So this is within the copilot model. Pickup, So in
- 10:56Speaker 3: copilot you've got auto and I like the little text
- 10:59Speaker 3: and decides how to think. It's got quick response. That's
- 11:02Speaker 3: it's a qui equivalent of haiku, so don't think, just
- 11:04Speaker 3: give me a response and then think deeper, which is
- 11:06Speaker 3: its equivalent of something like the thinking ish models. So
- 11:10Speaker 3: I'm generally on auto all the time and it will
- 11:12Speaker 3: decide how long to think. But if it's important I
- 11:14Speaker 3: will go think deeper because it will always do a
- 11:17Speaker 3: better job of thinking if I tell it to think
- 11:20Speaker 3: deeper than auto if it thinks so if it's important, strategy,
- 11:24Speaker 3: market research, all those kinds of things, or how do
- 11:26Speaker 3: I do this is really important, I'm going to get
- 11:27Speaker 3: it right. For a client, absolutely think deeper. But I
- 11:31Speaker 3: encourage you to do some tests and say, all right, oh,
- 11:33Speaker 3: now try the Opus model because you can use Opus
- 11:36Speaker 3: that's Claudes model inside Copilot, which is quite amazing. And
- 11:40Speaker 3: there's also a GPT from AI Open AI and so
- 11:44Speaker 3: you can actually use one of the GPT once you've
- 11:46Speaker 3: got the think deepers in quick responses and all the
- 11:48Speaker 3: same things, so you've got more choices in Copilot, which
- 11:51Speaker 3: is actually pretty cool.
- 11:54Speaker 1: Let's finish by talking about tokens because this is so
- 11:58Speaker 1: important because you know you might just be listening and going, well,
- 12:02Speaker 1: why wouldn't I just have the AI think deeply about everything?
- 12:07Speaker 1: I mean, what are the implications of that? And I'd say,
- 12:11Speaker 1: in if you're a solopreneur and you've just got your
- 12:14Speaker 1: own individual license, or maybe you're in a small business
- 12:18Speaker 1: where I think the pricing is a little bit different
- 12:20Speaker 1: depending on how things are structured in your company. It's
- 12:23Speaker 1: so different to if you are working in a large
- 12:26Speaker 1: corporate and you've got an enterprise plan because what we
- 12:30Speaker 1: know there is that often there's a per user charge,
- 12:34Speaker 1: but there is also a charge for tokens. So how
- 12:38Speaker 1: do we need to be thinking about this because quite frankly,
- 12:40Speaker 1: I know we've both heard some nightmare stories around token
- 12:43Speaker 1: usage going crazy and skyrocketing the prices that people are
- 12:49Speaker 1: paying for their AI tool.
- 12:51Speaker 3: Yeah, so it's all about II recommend use the right
- 12:56Speaker 3: model that you need for the task you're doing. So,
- 12:58Speaker 3: for example, I need to summarize twenty customer feedback things
- 13:02Speaker 3: into a couple of themes. An auto model, a regular
- 13:06Speaker 3: model is probably fine for that. Or you might need
- 13:08Speaker 3: to say, hey, those same twenty comments do they reveal
- 13:11Speaker 3: a product strategy risk or does that really impact what
- 13:14Speaker 3: I need to do for the next kind of sets
- 13:16Speaker 3: of investments that matters, in which case I probably use
- 13:19Speaker 3: a larger top tier, a thinking model and Opus model
- 13:22Speaker 3: for that. There's also a timeframe thing, So if I'm
- 13:25Speaker 3: using Opus, it's going to spend a lot of time
- 13:27Speaker 3: thinking through the thing. It will take me ages to
- 13:28Speaker 3: get that response back and if you're going to get
- 13:30Speaker 3: frustrated with it taking a lot of time, then you're
- 13:33Speaker 3: wasting a lot of tokens. Okay, but let's actually go
- 13:35Speaker 3: back to the token thing. So if you're in a
- 13:38Speaker 3: particular plan, and only some plans have this, So if
- 13:41Speaker 3: you're in say everyday claud plan, this is a thirty
- 13:44Speaker 3: bucks a month plan, then you will have a usage limit.
- 13:47Speaker 3: So you can actually go to your settings page and
- 13:50Speaker 3: under settings there's a thing called usage and you can
- 13:52Speaker 3: see that you've got a five hour windows, a little
- 13:54Speaker 3: bit of a bar how much you've used, and there's
- 13:56Speaker 3: a weekly limit window and there's a bit of a
- 13:58Speaker 3: bar there, and you can also top up and buy
- 14:01Speaker 3: more on top of that if you need. And what
- 14:02Speaker 3: they're basically saying is you can have a certain number
- 14:05Speaker 3: of uses of your different models over that five hour
- 14:09Speaker 3: window or one week window until we say you've used
- 14:12Speaker 3: it enough, you've overused what our models do now, Haiku
- 14:17Speaker 3: is about two to three times cheaper than Sonnet, and
- 14:20Speaker 3: son It is minimum two times cheaper than Opus, so
- 14:24Speaker 3: there's like a minimum six times differential in price on
- 14:28Speaker 3: these different models. So if you're using Opus for hey,
- 14:32Speaker 3: summarize this email, you are burning through your tokens and
- 14:36Speaker 3: you'll hit at the end of the five hour period
- 14:38Speaker 3: and AI is going to say tote sos. You're going
- 14:40Speaker 3: to have to come back in a couple of hours, right,
- 14:42Speaker 3: And that becomes a problem if you actually needed to
- 14:44Speaker 3: get some work done. So picking the right model means
- 14:46Speaker 3: you can actually get some work done, which is great.
- 14:49Speaker 3: Now if you're in a company plan, and some companies
- 14:52Speaker 3: have a plan where it's simply you can you've get
- 14:56Speaker 3: access to it. But each time you get AI to
- 14:59Speaker 3: run a query to get a prompt or together to think,
- 15:01Speaker 3: that goes kachin kachin, kachin kachin on your bank balance.
- 15:05Speaker 3: So if that's the case, then you might be costing
- 15:08Speaker 3: an awful lot of cash to your poor IT department.
- 15:12Speaker 3: And we hear stories where some companies have got like
- 15:15Speaker 3: a five hundred thousand dollar AI bill for that month
- 15:20Speaker 3: and things like that because people have just been using
- 15:22Speaker 3: these things on summarize this email or stuff they don't need,
- 15:25Speaker 3: or they're letting it go crazy. And so my encouragement
- 15:28Speaker 3: is pick the right model for the job. But also
- 15:30Speaker 3: if you do have, say Claude, check your usage limits
- 15:33Speaker 3: and see whether you're actually using it right, so you
- 15:36Speaker 3: can actually see whether you're over using any of these models.
- 15:39Speaker 1: And just for context, that five hundred thousand dollar AI
- 15:42Speaker 1: bill that you talk about, that was just from one
- 15:44Speaker 1: employee from memory that particular story that you're referring to,
- 15:49Speaker 1: and they.
- 15:49Speaker 3: Were also getting AI. This is APIs and chlor Cowork
- 15:54Speaker 3: and sorry clock code, I should say, and all those
- 15:56Speaker 3: kind of things. So they were going nuts, right, But
- 15:59Speaker 3: here's the they were going nuts but not watching what
- 16:02Speaker 3: they're doing, and that's coming to be a bigger problem
- 16:05Speaker 3: because like Claude has its own usage limits and a
- 16:08Speaker 3: lot of companies they're off this. You know, you can
- 16:11Speaker 3: be the five our limit you stop and instead it's
- 16:13Speaker 3: just a paper usage. There's another thing coming out in
- 16:16Speaker 3: a couple of months time, which is Microsoft Cowork, which
- 16:18Speaker 3: is the same as clor COK in Microsoft, they will
- 16:21Speaker 3: also have usage limits. So picking the right model is
- 16:23Speaker 3: good practice now for when all of these things go
- 16:26Speaker 3: to usage only.
- 16:28Speaker 1: Thank you so much, NEO. Hopefully you are now feeling
- 16:31Speaker 1: more confident and clear in terms of which model you
- 16:34Speaker 1: should be using for all the different tasks that you're
- 16:38Speaker 1: using AI for. Thank you so much for listening and
- 16:41Speaker 1: we will see you next week. How IAI was hosted
- 16:45Speaker 1: by me Amantha Imber and Neo Applan. A big thank
- 16:48Speaker 1: you to Martin Imber who does our sound editing, and
- 16:50Speaker 1: Jim Rubio for production support, and thank you to John
- 16:54Speaker 1: Kilby who composed the theme music