WEBVTT
00:00:00.000 --> 00:00:01.600
Welcome to the Copilot Briefing.
00:00:01.679 --> 00:00:05.679
I'm Steve Corey, a Microsoft MVP and Copilot and Agents.
00:00:05.919 --> 00:00:12.000
This podcast is for business leaders navigating Microsoft Copilot, AI agents, and the future of work.
00:00:12.160 --> 00:00:25.199
In these early episodes, I'm sharing a handful of foundational conversations originally published on my YouTube channel because they provide important context for the decisions organizations are making about Microsoft AI today.
00:00:25.359 --> 00:00:30.960
Make sure to stay until the end because I'm going to have some special takeaways just for the podcast audience.
00:00:31.120 --> 00:00:32.000
Let's get started.
00:00:32.240 --> 00:00:36.079
Running AI is expensive, like really expensive.
00:00:36.320 --> 00:00:44.000
Microsoft's spending a fortune to power Copilot, and a lot of that cost eventually lands up on your budget.
00:00:44.240 --> 00:00:51.679
That's really obvious for something like Copilot Co-Work with that recent pay as you go announcement.
00:00:51.920 --> 00:00:57.280
But a huge shift is happening right now that could change the future of your AI, Bill.
00:00:57.920 --> 00:01:09.120
You see, Microsoft is deploying their own family of in-house AI models called MAI or My, as I've heard them uh pronounce it.
00:01:09.359 --> 00:01:20.000
They've been building these as a really a strategic answer to what the kind of wild economics of running AI at a global scale.
00:01:20.239 --> 00:01:30.560
In this video, I'm going to break down what these My models are and more importantly, what Microsoft's strategy means for the way you're going to be paying for Copilot.
00:01:30.719 --> 00:01:32.000
Let's get into it.
00:01:36.799 --> 00:01:42.159
So let's start with the headache that every business leader and IT department is dealing with.
00:01:42.480 --> 00:01:45.040
Powerful AI just isn't cheap.
00:01:45.280 --> 00:01:53.280
You're probably familiar with the$30 per user per month skew for M365 Copilot, you know, that price tag.
00:01:53.519 --> 00:02:01.120
And for many, this was already a really big investment, like hundreds of thousands of dollars in a lot of cases or more.
00:02:01.599 --> 00:02:03.519
But that's no longer the whole story.
00:02:03.680 --> 00:02:16.080
You see, as Microsoft has introduced a lot of these more powerful autonomous features like copilot co-work, and we're starting to see uh autopilots and Microsoft Scout, things like that.
00:02:16.319 --> 00:02:24.719
They're also rolling out more of that consumption-based pricing model on top of your co-pilot license.
00:02:25.280 --> 00:02:30.800
These advanced capabilities are they're usually just billed using co-pilot credits.
00:02:30.879 --> 00:02:37.039
That's kind of their standard unit of consumption, just to kind of keep things even across the board.
00:02:37.120 --> 00:02:39.199
But this is where things get tricky.
00:02:39.520 --> 00:03:01.120
Because as your team uses AI for more and more uh like things like automating workflows, running complex analyses, uh, building complex uh agents, the the variable cost becomes a source of major anxiety uh when it comes to budgeting, because the more you use these tools, the more you're gonna pay.
00:03:01.360 --> 00:03:13.439
And predicting what that bill is actually going to be has been nearly impossible because the model uh could be different from one month to the next, the prompts could be different, like all these different variables.
00:03:14.240 --> 00:03:25.199
So the Microsoft My Models uh is potentially a really a really good solution for this, for trying to clamp down on those costs.
00:03:25.439 --> 00:03:28.479
Uh MAI stands for Microsoft AI.
00:03:28.800 --> 00:03:35.439
And Microsoft announced a family of new specialized models that it built entirely in-house.
00:03:35.599 --> 00:03:49.360
The lineup includes models from advanced reasoning, coding, uh, and even some of the top-performing models are out there, are my models for things like voice transcription and image generation.
00:03:49.759 --> 00:03:51.919
But why why build your own their own?
00:03:52.000 --> 00:03:53.120
Why reinvent the wheel?
00:03:53.199 --> 00:04:02.080
You've already got the big uh powerhouses like OpenAI and Anthropic and and others dominating the market, just dominating.
00:04:02.240 --> 00:04:06.400
And it I would argue Anthropic is miles ahead of OpenAI.
00:04:06.879 --> 00:04:23.120
And most people want to just use anthropic models unless they're limited because of uh government um organizations or or maybe EU-related uh or EU companies and some of the just regulatory requirements, really.
00:04:23.279 --> 00:04:30.079
If you don't have that in the way, you've generally been wanting to use anthropic because it's the most complex and the most expensive.
00:04:30.160 --> 00:04:32.959
Uh, and if you didn't have to foot that bill, you wanted anthropic.
00:04:33.279 --> 00:04:37.920
But if you talk to an AI architect, they're gonna tell you a very different story.
00:04:38.000 --> 00:04:43.920
You don't just simply want to slam everything into anthropic models because they are the most expensive.
00:04:44.160 --> 00:04:47.120
So the strategy is actually really smart with the my models.
00:04:47.360 --> 00:04:55.360
First, it's about um becoming more self-sufficient and less dependent on these third-party partners.
00:04:55.600 --> 00:04:59.519
But the second reason is the one that matters the most to you.
00:04:59.759 --> 00:05:03.759
It's lowering the operational costs of running AI workloads.
00:05:03.839 --> 00:05:07.600
And if you're getting value from this video so far, then hit that thumbs up button.
00:05:07.680 --> 00:05:14.240
And if you learned one new thing or gotten excited about one new thing, then consider subscribing to this channel.
00:05:14.399 --> 00:05:18.959
It's just a free way you could support me so I can continue to make videos just like this for you.
00:05:19.199 --> 00:05:23.839
So here's the clever part Microsoft's been using a multi-model orchestration strategy.
00:05:24.000 --> 00:05:30.000
Think of it this way: you don't hire a brain surgeon to put a band-aid on a paper cut, right?
00:05:30.160 --> 00:05:31.199
That's just insane.
00:05:31.360 --> 00:05:36.319
That's a bad use of their time and a huge waste of your own money.
00:05:36.480 --> 00:05:46.399
In the same way, Microsoft realized that it was using the most powerful and expensive frontier models for simple everyday tasks.
00:05:46.560 --> 00:05:50.399
I've talked about this for weeks now, potentially months.
00:05:50.639 --> 00:05:56.079
People were using cowork for things that they should have been using M365 Copilot for.
00:05:56.399 --> 00:06:01.360
You know, just because it was there and it was, it felt free anyway.
00:06:01.600 --> 00:06:07.199
So they just stopped using Copilot and started using co-work for every little random, simple question.
00:06:07.360 --> 00:06:11.040
I was guilty of it because I just wanted to play around with the latest and greatest stuff.
00:06:11.199 --> 00:06:12.079
But back to this.
00:06:12.319 --> 00:06:19.519
So a lot, what happens now with this model routing, and it's not a it's not something new Microsoft created.
00:06:19.680 --> 00:06:22.399
A lot of companies are using a model routing.
00:06:22.480 --> 00:06:24.879
OpenAI might have been the first one I had heard about.
00:06:25.040 --> 00:06:36.720
Now, what what's really happening is a lot of the high volume, low effort requests are automatically routed to the cheaper, highly efficient in-house my models.
00:06:37.120 --> 00:06:51.839
And then the biggest, most expensive models are reserved only for the really complex queries that need that level of breakdown and analysis and deep reasoning, that level of firepower.
00:06:51.920 --> 00:06:56.319
And that's really the core of this cost-saving mechanism.
00:06:56.560 --> 00:07:01.040
So here's the breakdown of what the what models they've actually released so far.
00:07:01.279 --> 00:07:04.000
First, there is the My Thinking One model.
00:07:04.160 --> 00:07:15.519
It's the that's if you want to call it their flagship model, like how Anthropics got, you know, Opus and OpenAI has got what I think they're at 5.6 now.
00:07:16.000 --> 00:07:19.600
Um, it's their their really high-end model.
00:07:20.319 --> 00:07:30.639
Uh, the purpose of it is for complex reasoning, multi-step planning, math, coding assistance, which coding assistance is interesting because they have a different model specifically just for that as well.
00:07:30.720 --> 00:07:32.720
Uh, but also long document analysis.
00:07:32.800 --> 00:07:38.879
So this is like general purpose, like lots of stuff, but it's not just your basic chat model.
00:07:39.360 --> 00:07:44.639
It is, it's like they don't have like the basic model, they just have the deep reasoning one.
00:07:44.800 --> 00:07:53.199
Uh, but I think that's gonna work out because this is still gonna be way cheaper than the third-party models that also do the deep reasoning.
00:07:53.439 --> 00:07:56.879
Then you've got the my code one flash model.
00:07:57.120 --> 00:07:59.680
This is their purpose-built coding model.
00:07:59.920 --> 00:08:03.439
You already see it in things like GitHub Copilot.
00:08:03.600 --> 00:08:10.639
Uh, so it'll handle code generation, refactoring, agentic software development, the GitHub workflows, things, things like that.
00:08:10.800 --> 00:08:15.199
It's uh it it's it's definitely already showing up, and a lot of people are already using it.
00:08:15.279 --> 00:08:23.360
And I personally want to start using this thing more to really see how it stacks up with Kodaks and uh and Anthropic.
00:08:23.519 --> 00:08:28.879
Then there's the My Image 2.5 model, at least that's the current version of all these.
00:08:29.040 --> 00:08:34.240
Is uh um and this one is for this is their image generation and editing model.
00:08:34.559 --> 00:08:45.440
At some point in the recent past, I remember Copilot was switching from the DOLI 3 image generator to the, I think it was the native OpenAI image generator.
00:08:45.519 --> 00:08:55.039
And now we're starting to see the My Image Generator um is probably a replacement for using the uh the open AI model.
00:08:55.440 --> 00:09:06.399
And the the quality of the images inside Copilot is definitely way better than what we used to get a couple of years ago, maybe even last year.
00:09:06.559 --> 00:09:08.399
Uh, it's just miles ahead.
00:09:08.639 --> 00:09:14.080
I can't quite tell which one uh when it's using the my image versus the open AI.
00:09:14.240 --> 00:09:25.039
By now, it may be using the my image model all the time now, but this is more for the designing artistic type assets and marketing assets and things like that.
00:09:25.200 --> 00:09:30.639
Then we can start to get into some of the really interesting, more technical use cases.
00:09:30.799 --> 00:09:36.399
There is the My Voice 2 model, and this is for speech synthesis.
00:09:36.559 --> 00:09:41.919
So it's text-to-speech, it's voice generation, multilingual voice experiences.
00:09:42.480 --> 00:10:00.480
And if you think about the Teams feature, it may be Teams Premium, actually, where they can do uh you can get real-time voice uh real-time language translation inside meetings that actually might be using this model.
00:10:00.720 --> 00:10:04.960
Um, just as I'm thinking on the fly about this.
00:10:05.759 --> 00:10:11.440
So uh the the speech synthesis and and and text-to-speech and all that, very, very cool stuff.
00:10:11.600 --> 00:10:15.200
And then I think lastly, we've got the My Transcribe 1.5.
00:10:15.279 --> 00:10:18.639
So this is the opposite of the voice model.
00:10:18.799 --> 00:10:20.720
This is the speech to text.
00:10:20.879 --> 00:10:28.720
So uh transcription, meaning transcription, audio processing, and similarly, multilingual speech recognition.
00:10:29.279 --> 00:10:34.879
So those are the models as of the date of recording that Microsoft's put out and kind of what they do.
00:10:35.039 --> 00:10:40.399
They kind of run the gamut of the different AI workloads that you would come across.
00:10:40.639 --> 00:10:49.200
And these things are tuned enough without making you choose between, like, you know, 10 different models or 20 different models or whatever.
00:10:49.440 --> 00:10:51.840
But what does this mean for your cost?
00:10:52.080 --> 00:10:59.519
So let's talk about what this means for how you'll manage to and pay for your copilot workloads.
00:10:59.759 --> 00:11:05.519
First, the underlying cost for Microsoft to run a simple AI task is way cheaper.
00:11:05.679 --> 00:11:22.559
Now, while you're not going to see some line item discount on your copilot bill or your AI bill, what this efficiency is, it's really about making it more sustainable and a more potentially more flexible pricing model in the future.
00:11:22.720 --> 00:11:30.720
Well, in other words, it makes it harder for Microsoft to justify an increase in the cost of M365 copilot licenses.
00:11:30.879 --> 00:11:38.720
Now, that really is good news for us because so far they haven't ever raised that price of copilot.
00:11:38.879 --> 00:11:43.600
However, the models have definitely got more expensive in the recent past.
00:11:43.679 --> 00:11:51.840
And they continue to roll out all of those more expensive models to the copilot folks at the same price tag.
00:11:52.159 --> 00:11:56.320
So this is how they're going to be able to continue to do that in the future.
00:11:56.480 --> 00:12:05.440
If it's an expensive process, sure, maybe route it to Blaud or something like that, or or maybe, maybe even the my thinking model.
00:12:05.759 --> 00:12:13.519
But if it's a cheaper task, if it is like where is the so-and-so document ad, it doesn't need to go to some expensive model.
00:12:13.759 --> 00:12:17.840
The cheapest thing out there is probably going to be just fine for it.
00:12:18.080 --> 00:12:23.840
Second, the the my models are the engine that makes pay as you go really viable.
00:12:24.000 --> 00:12:36.159
So the copilot credit system, which kind of powers features like co-pilot uh co-work now, they all calculate costs based off of several factors, including which model is being used.
00:12:36.320 --> 00:12:48.559
So if the task is routed to a more efficient my model, it consumes fewer resources, copilot credits, which in turn means your consumption is going to be less.
00:12:48.720 --> 00:12:52.720
You consume fewer of those purchased copilot credits.
00:12:52.960 --> 00:13:03.840
So this should make the pay as you go system way more manageable and predictable because all across the board, your costs should be coming down.
00:13:04.080 --> 00:13:12.320
Now, it's never going to be completely predictable, but you'll get more value if you're using those P3 copilot credit packs.
00:13:12.559 --> 00:13:17.759
And if you're doing just straight pay as you go, at least it'll be a lower cost.
00:13:17.840 --> 00:13:19.759
It's probably easier to fit in your budget.
00:13:19.919 --> 00:13:35.360
Third, and this is a big deal, it gives you a new liver for cost control, especially if you're building custom solutions with AI, because the developers can now strategically design the automated workflows to be way more cost effective.
00:13:35.519 --> 00:13:42.799
For routine tasks, you could have your custom agent using the efficient MyModels to reduce all of your consumption.
00:13:43.039 --> 00:13:51.679
Then you could reserve those expensive, eye-powered models for only the critical steps that really need something like Anthropics Opus.
00:13:52.159 --> 00:13:53.360
So, what's the bottom line?
00:13:53.519 --> 00:13:55.600
Microsoft's MyModels are here.
00:13:55.840 --> 00:14:00.639
They are a direct response to the unsustainable costs of running AI at scale.
00:14:00.799 --> 00:14:02.720
At least that is my opinion of it.
00:14:02.879 --> 00:14:11.759
And this change towards more efficient in-house models is what makes the affordable enterprise AI thing a kind of a long-term reality.
00:14:11.840 --> 00:14:26.240
It's a major shift towards granular and control and a controllable future of your copilot costs and less dependency on external vendors like OpenAI, Anthropic, and whoever else.
00:14:32.320 --> 00:14:39.440
Remember, not everything needs the to run the most expensive model just to just to run, you know?
00:14:39.600 --> 00:14:52.320
You don't just throw everything at the most expensive thing because your profit margin, your, your, your, uh, what the feasibility of this solution starts to go way down if the return if the ROI isn't there.
00:14:52.480 --> 00:14:56.559
Not everything needs to run on the most expensive model.
00:14:56.960 --> 00:15:02.799
And in many cases, customers are throwing really simple tasks at really expensive models.
00:15:02.960 --> 00:15:05.600
And the model doesn't care, it's just gonna do it.
00:15:05.679 --> 00:15:10.080
And you're gonna get charged accordingly, and you could be saving a ton of money.
00:15:10.240 --> 00:15:13.679
So let me know what you think uh about those, the these new My models.
00:15:13.759 --> 00:15:20.000
If you even knew about them existing, if you've tried them, I want to hear about your experiences down in the comments below.
00:15:20.159 --> 00:15:26.240
So here are some of the key takeaways for the Microsoft AI models that are coming out.
00:15:26.480 --> 00:15:30.799
AI is clearly becoming cheaper just across the board.
00:15:31.039 --> 00:15:38.000
As hardware advances and uh better processes come out, cheaper models are coming out as well.
00:15:38.399 --> 00:15:58.080
And being able to continually reevaluate which models are being used, leveraging things like the model router, uh, also known as just the auto selector and copilot, lets copilot route your users' prompts to the correct model and more importantly, the most efficient model.
00:15:58.240 --> 00:16:06.240
This is how you'll be able to lower those costs over time and continue to drive the return on investment higher and higher.
00:16:06.480 --> 00:16:09.279
Thanks for listening, and I'll see you on the next show.
00:00:00.000 --> 00:00:01.600
Welcome to the Copilot Briefing.
00:00:01.679 --> 00:00:05.679
I'm Steve Corey, a Microsoft MVP and Copilot and Agents.
00:00:05.919 --> 00:00:12.000
This podcast is for business leaders navigating Microsoft Copilot, AI agents, and the future of work.
00:00:12.160 --> 00:00:25.199
In these early episodes, I'm sharing a handful of foundational conversations originally published on my YouTube channel because they provide important context for the decisions organizations are making about Microsoft AI today.
00:00:25.359 --> 00:00:30.960
Make sure to stay until the end because I'm going to have some special takeaways just for the podcast audience.
00:00:31.120 --> 00:00:32.000
Let's get started.
00:00:32.240 --> 00:00:36.079
Running AI is expensive, like really expensive.
00:00:36.320 --> 00:00:44.000
Microsoft's spending a fortune to power Copilot, and a lot of that cost eventually lands up on your budget.
00:00:44.240 --> 00:00:51.679
That's really obvious for something like Copilot Co-Work with that recent pay as you go announcement.
00:00:51.920 --> 00:00:57.280
But a huge shift is happening right now that could change the future of your AI, Bill.
00:00:57.920 --> 00:01:09.120
You see, Microsoft is deploying their own family of in-house AI models called MAI or My, as I've heard them uh pronounce it.
00:01:09.359 --> 00:01:20.000
They've been building these as a really a strategic answer to what the kind of wild economics of running AI at a global scale.
00:01:20.239 --> 00:01:30.560
In this video, I'm going to break down what these My models are and more importantly, what Microsoft's strategy means for the way you're going to be paying for Copilot.
00:01:30.719 --> 00:01:32.000
Let's get into it.
00:01:36.799 --> 00:01:42.159
So let's start with the headache that every business leader and IT department is dealing with.
00:01:42.480 --> 00:01:45.040
Powerful AI just isn't cheap.
00:01:45.280 --> 00:01:53.280
You're probably familiar with the$30 per user per month skew for M365 Copilot, you know, that price tag.
00:01:53.519 --> 00:02:01.120
And for many, this was already a really big investment, like hundreds of thousands of dollars in a lot of cases or more.
00:02:01.599 --> 00:02:03.519
But that's no longer the whole story.
00:02:03.680 --> 00:02:16.080
You see, as Microsoft has introduced a lot of these more powerful autonomous features like copilot co-work, and we're starting to see uh autopilots and Microsoft Scout, things like that.
00:02:16.319 --> 00:02:24.719
They're also rolling out more of that consumption-based pricing model on top of your co-pilot license.
00:02:25.280 --> 00:02:30.800
These advanced capabilities are they're usually just billed using co-pilot credits.
00:02:30.879 --> 00:02:37.039
That's kind of their standard unit of consumption, just to kind of keep things even across the board.
00:02:37.120 --> 00:02:39.199
But this is where things get tricky.
00:02:39.520 --> 00:03:01.120
Because as your team uses AI for more and more uh like things like automating workflows, running complex analyses, uh, building complex uh agents, the the variable cost becomes a source of major anxiety uh when it comes to budgeting, because the more you use these tools, the more you're gonna pay.
00:03:01.360 --> 00:03:13.439
And predicting what that bill is actually going to be has been nearly impossible because the model uh could be different from one month to the next, the prompts could be different, like all these different variables.
00:03:14.240 --> 00:03:25.199
So the Microsoft My Models uh is potentially a really a really good solution for this, for trying to clamp down on those costs.
00:03:25.439 --> 00:03:28.479
Uh MAI stands for Microsoft AI.
00:03:28.800 --> 00:03:35.439
And Microsoft announced a family of new specialized models that it built entirely in-house.
00:03:35.599 --> 00:03:49.360
The lineup includes models from advanced reasoning, coding, uh, and even some of the top-performing models are out there, are my models for things like voice transcription and image generation.
00:03:49.759 --> 00:03:51.919
But why why build your own their own?
00:03:52.000 --> 00:03:53.120
Why reinvent the wheel?
00:03:53.199 --> 00:04:02.080
You've already got the big uh powerhouses like OpenAI and Anthropic and and others dominating the market, just dominating.
00:04:02.240 --> 00:04:06.400
And it I would argue Anthropic is miles ahead of OpenAI.
00:04:06.879 --> 00:04:23.120
And most people want to just use anthropic models unless they're limited because of uh government um organizations or or maybe EU-related uh or EU companies and some of the just regulatory requirements, really.
00:04:23.279 --> 00:04:30.079
If you don't have that in the way, you've generally been wanting to use anthropic because it's the most complex and the most expensive.
00:04:30.160 --> 00:04:32.959
Uh, and if you didn't have to foot that bill, you wanted anthropic.
00:04:33.279 --> 00:04:37.920
But if you talk to an AI architect, they're gonna tell you a very different story.
00:04:38.000 --> 00:04:43.920
You don't just simply want to slam everything into anthropic models because they are the most expensive.
00:04:44.160 --> 00:04:47.120
So the strategy is actually really smart with the my models.
00:04:47.360 --> 00:04:55.360
First, it's about um becoming more self-sufficient and less dependent on these third-party partners.
00:04:55.600 --> 00:04:59.519
But the second reason is the one that matters the most to you.
00:04:59.759 --> 00:05:03.759
It's lowering the operational costs of running AI workloads.
00:05:03.839 --> 00:05:07.600
And if you're getting value from this video so far, then hit that thumbs up button.
00:05:07.680 --> 00:05:14.240
And if you learned one new thing or gotten excited about one new thing, then consider subscribing to this channel.
00:05:14.399 --> 00:05:18.959
It's just a free way you could support me so I can continue to make videos just like this for you.
00:05:19.199 --> 00:05:23.839
So here's the clever part Microsoft's been using a multi-model orchestration strategy.
00:05:24.000 --> 00:05:30.000
Think of it this way: you don't hire a brain surgeon to put a band-aid on a paper cut, right?
00:05:30.160 --> 00:05:31.199
That's just insane.
00:05:31.360 --> 00:05:36.319
That's a bad use of their time and a huge waste of your own money.
00:05:36.480 --> 00:05:46.399
In the same way, Microsoft realized that it was using the most powerful and expensive frontier models for simple everyday tasks.
00:05:46.560 --> 00:05:50.399
I've talked about this for weeks now, potentially months.
00:05:50.639 --> 00:05:56.079
People were using cowork for things that they should have been using M365 Copilot for.
00:05:56.399 --> 00:06:01.360
You know, just because it was there and it was, it felt free anyway.
00:06:01.600 --> 00:06:07.199
So they just stopped using Copilot and started using co-work for every little random, simple question.
00:06:07.360 --> 00:06:11.040
I was guilty of it because I just wanted to play around with the latest and greatest stuff.
00:06:11.199 --> 00:06:12.079
But back to this.
00:06:12.319 --> 00:06:19.519
So a lot, what happens now with this model routing, and it's not a it's not something new Microsoft created.
00:06:19.680 --> 00:06:22.399
A lot of companies are using a model routing.
00:06:22.480 --> 00:06:24.879
OpenAI might have been the first one I had heard about.
00:06:25.040 --> 00:06:36.720
Now, what what's really happening is a lot of the high volume, low effort requests are automatically routed to the cheaper, highly efficient in-house my models.
00:06:37.120 --> 00:06:51.839
And then the biggest, most expensive models are reserved only for the really complex queries that need that level of breakdown and analysis and deep reasoning, that level of firepower.
00:06:51.920 --> 00:06:56.319
And that's really the core of this cost-saving mechanism.
00:06:56.560 --> 00:07:01.040
So here's the breakdown of what the what models they've actually released so far.
00:07:01.279 --> 00:07:04.000
First, there is the My Thinking One model.
00:07:04.160 --> 00:07:15.519
It's the that's if you want to call it their flagship model, like how Anthropics got, you know, Opus and OpenAI has got what I think they're at 5.6 now.
00:07:16.000 --> 00:07:19.600
Um, it's their their really high-end model.
00:07:20.319 --> 00:07:30.639
Uh, the purpose of it is for complex reasoning, multi-step planning, math, coding assistance, which coding assistance is interesting because they have a different model specifically just for that as well.
00:07:30.720 --> 00:07:32.720
Uh, but also long document analysis.
00:07:32.800 --> 00:07:38.879
So this is like general purpose, like lots of stuff, but it's not just your basic chat model.
00:07:39.360 --> 00:07:44.639
It is, it's like they don't have like the basic model, they just have the deep reasoning one.
00:07:44.800 --> 00:07:53.199
Uh, but I think that's gonna work out because this is still gonna be way cheaper than the third-party models that also do the deep reasoning.
00:07:53.439 --> 00:07:56.879
Then you've got the my code one flash model.
00:07:57.120 --> 00:07:59.680
This is their purpose-built coding model.
00:07:59.920 --> 00:08:03.439
You already see it in things like GitHub Copilot.
00:08:03.600 --> 00:08:10.639
Uh, so it'll handle code generation, refactoring, agentic software development, the GitHub workflows, things, things like that.
00:08:10.800 --> 00:08:15.199
It's uh it it's it's definitely already showing up, and a lot of people are already using it.
00:08:15.279 --> 00:08:23.360
And I personally want to start using this thing more to really see how it stacks up with Kodaks and uh and Anthropic.
00:08:23.519 --> 00:08:28.879
Then there's the My Image 2.5 model, at least that's the current version of all these.
00:08:29.040 --> 00:08:34.240
Is uh um and this one is for this is their image generation and editing model.
00:08:34.559 --> 00:08:45.440
At some point in the recent past, I remember Copilot was switching from the DOLI 3 image generator to the, I think it was the native OpenAI image generator.
00:08:45.519 --> 00:08:55.039
And now we're starting to see the My Image Generator um is probably a replacement for using the uh the open AI model.
00:08:55.440 --> 00:09:06.399
And the the quality of the images inside Copilot is definitely way better than what we used to get a couple of years ago, maybe even last year.
00:09:06.559 --> 00:09:08.399
Uh, it's just miles ahead.
00:09:08.639 --> 00:09:14.080
I can't quite tell which one uh when it's using the my image versus the open AI.
00:09:14.240 --> 00:09:25.039
By now, it may be using the my image model all the time now, but this is more for the designing artistic type assets and marketing assets and things like that.
00:09:25.200 --> 00:09:30.639
Then we can start to get into some of the really interesting, more technical use cases.
00:09:30.799 --> 00:09:36.399
There is the My Voice 2 model, and this is for speech synthesis.
00:09:36.559 --> 00:09:41.919
So it's text-to-speech, it's voice generation, multilingual voice experiences.
00:09:42.480 --> 00:10:00.480
And if you think about the Teams feature, it may be Teams Premium, actually, where they can do uh you can get real-time voice uh real-time language translation inside meetings that actually might be using this model.
00:10:00.720 --> 00:10:04.960
Um, just as I'm thinking on the fly about this.
00:10:05.759 --> 00:10:11.440
So uh the the speech synthesis and and and text-to-speech and all that, very, very cool stuff.
00:10:11.600 --> 00:10:15.200
And then I think lastly, we've got the My Transcribe 1.5.
00:10:15.279 --> 00:10:18.639
So this is the opposite of the voice model.
00:10:18.799 --> 00:10:20.720
This is the speech to text.
00:10:20.879 --> 00:10:28.720
So uh transcription, meaning transcription, audio processing, and similarly, multilingual speech recognition.
00:10:29.279 --> 00:10:34.879
So those are the models as of the date of recording that Microsoft's put out and kind of what they do.
00:10:35.039 --> 00:10:40.399
They kind of run the gamut of the different AI workloads that you would come across.
00:10:40.639 --> 00:10:49.200
And these things are tuned enough without making you choose between, like, you know, 10 different models or 20 different models or whatever.
00:10:49.440 --> 00:10:51.840
But what does this mean for your cost?
00:10:52.080 --> 00:10:59.519
So let's talk about what this means for how you'll manage to and pay for your copilot workloads.
00:10:59.759 --> 00:11:05.519
First, the underlying cost for Microsoft to run a simple AI task is way cheaper.
00:11:05.679 --> 00:11:22.559
Now, while you're not going to see some line item discount on your copilot bill or your AI bill, what this efficiency is, it's really about making it more sustainable and a more potentially more flexible pricing model in the future.
00:11:22.720 --> 00:11:30.720
Well, in other words, it makes it harder for Microsoft to justify an increase in the cost of M365 copilot licenses.
00:11:30.879 --> 00:11:38.720
Now, that really is good news for us because so far they haven't ever raised that price of copilot.
00:11:38.879 --> 00:11:43.600
However, the models have definitely got more expensive in the recent past.
00:11:43.679 --> 00:11:51.840
And they continue to roll out all of those more expensive models to the copilot folks at the same price tag.
00:11:52.159 --> 00:11:56.320
So this is how they're going to be able to continue to do that in the future.
00:11:56.480 --> 00:12:05.440
If it's an expensive process, sure, maybe route it to Blaud or something like that, or or maybe, maybe even the my thinking model.
00:12:05.759 --> 00:12:13.519
But if it's a cheaper task, if it is like where is the so-and-so document ad, it doesn't need to go to some expensive model.
00:12:13.759 --> 00:12:17.840
The cheapest thing out there is probably going to be just fine for it.
00:12:18.080 --> 00:12:23.840
Second, the the my models are the engine that makes pay as you go really viable.
00:12:24.000 --> 00:12:36.159
So the copilot credit system, which kind of powers features like co-pilot uh co-work now, they all calculate costs based off of several factors, including which model is being used.
00:12:36.320 --> 00:12:48.559
So if the task is routed to a more efficient my model, it consumes fewer resources, copilot credits, which in turn means your consumption is going to be less.
00:12:48.720 --> 00:12:52.720
You consume fewer of those purchased copilot credits.
00:12:52.960 --> 00:13:03.840
So this should make the pay as you go system way more manageable and predictable because all across the board, your costs should be coming down.
00:13:04.080 --> 00:13:12.320
Now, it's never going to be completely predictable, but you'll get more value if you're using those P3 copilot credit packs.
00:13:12.559 --> 00:13:17.759
And if you're doing just straight pay as you go, at least it'll be a lower cost.
00:13:17.840 --> 00:13:19.759
It's probably easier to fit in your budget.
00:13:19.919 --> 00:13:35.360
Third, and this is a big deal, it gives you a new liver for cost control, especially if you're building custom solutions with AI, because the developers can now strategically design the automated workflows to be way more cost effective.
00:13:35.519 --> 00:13:42.799
For routine tasks, you could have your custom agent using the efficient MyModels to reduce all of your consumption.
00:13:43.039 --> 00:13:51.679
Then you could reserve those expensive, eye-powered models for only the critical steps that really need something like Anthropics Opus.
00:13:52.159 --> 00:13:53.360
So, what's the bottom line?
00:13:53.519 --> 00:13:55.600
Microsoft's MyModels are here.
00:13:55.840 --> 00:14:00.639
They are a direct response to the unsustainable costs of running AI at scale.
00:14:00.799 --> 00:14:02.720
At least that is my opinion of it.
00:14:02.879 --> 00:14:11.759
And this change towards more efficient in-house models is what makes the affordable enterprise AI thing a kind of a long-term reality.
00:14:11.840 --> 00:14:26.240
It's a major shift towards granular and control and a controllable future of your copilot costs and less dependency on external vendors like OpenAI, Anthropic, and whoever else.
00:14:32.320 --> 00:14:39.440
Remember, not everything needs the to run the most expensive model just to just to run, you know?
00:14:39.600 --> 00:14:52.320
You don't just throw everything at the most expensive thing because your profit margin, your, your, your, uh, what the feasibility of this solution starts to go way down if the return if the ROI isn't there.
00:14:52.480 --> 00:14:56.559
Not everything needs to run on the most expensive model.
00:14:56.960 --> 00:15:02.799
And in many cases, customers are throwing really simple tasks at really expensive models.
00:15:02.960 --> 00:15:05.600
And the model doesn't care, it's just gonna do it.
00:15:05.679 --> 00:15:10.080
And you're gonna get charged accordingly, and you could be saving a ton of money.
00:15:10.240 --> 00:15:13.679
So let me know what you think uh about those, the these new My models.
00:15:13.759 --> 00:15:20.000
If you even knew about them existing, if you've tried them, I want to hear about your experiences down in the comments below.
00:15:20.159 --> 00:15:26.240
So here are some of the key takeaways for the Microsoft AI models that are coming out.
00:15:26.480 --> 00:15:30.799
AI is clearly becoming cheaper just across the board.
00:15:31.039 --> 00:15:38.000
As hardware advances and uh better processes come out, cheaper models are coming out as well.
00:15:38.399 --> 00:15:58.080
And being able to continually reevaluate which models are being used, leveraging things like the model router, uh, also known as just the auto selector and copilot, lets copilot route your users' prompts to the correct model and more importantly, the most efficient model.
00:15:58.240 --> 00:16:06.240
This is how you'll be able to lower those costs over time and continue to drive the return on investment higher and higher.
00:16:06.480 --> 00:16:09.279
Thanks for listening, and I'll see you on the next show.
