The Right Model for the Task – Ep.567

Kurt Buhler and Eugene Meidinger’s August article on picking an AI model is just over a month old, and Tommy Puglia and Mike Carlo spend the hour testing whether it still holds. They walk purpose, capability, effort, and cost through real Fabric work, then open the FabCon Barcelona posts on Fabric apps and semantic views.

News & Announcements

• Connecting apps, databases, and AI on one foundation — Tommy reads this as the FabCon and SQLCon Barcelona highlight post, published the morning of the show, from a venue he puts at roughly half the size of FabCon Atlanta. The four areas he walks are Fabric apps on enterprise data, a database hub in preview for SQL Server, Azure SQL, PostgreSQL, Cosmos DB, and Fabric SQL, database agents for SQL and PostgreSQL, and a stronger base for AI through MCP, skills, and OneLake. Mike’s pick is Fabric apps: warehouse and lakehouse connectors, a control that can keep an app inside the organization, and TypeScript backend functions with a secret store, which he says moves the app from a novelty to something that can call APIs and Azure Foundry. He is less sure why he would open Fabric to manage databases that do not live there, and both of them treat database-agent skills as still gray. Mike then argues the success metric shifts from monthly active users to monthly active agents. Tommy counters with monthly active tokens. Mike’s objection is that a token target pushes everyone onto the most expensive model, and that tokens spent on Anthropic never show up in Fabric.

• Bringing governed analytics into the flow of work — Tommy reads the claim that Power BI is moving past dashboards: an app-building experience in Desktop, powered by Fabric apps, starts from a trusted semantic model and uses natural language to generate, preview, and publish a data application that can accept inputs, write data, keep shared state, and support operational workflows, with Fabric apps included for Power BI Pro. Mike hears an admission that the report framework has rails, and that an agent builds faster against a Fabric app because websites are what models have already seen. He calls Desktop the wrong place to start, because it is Windows-only and because Kurt has told him he has not touched Desktop in months. Tommy wants the argument on the path, not the shell: natural language to visuals goes through Fabric apps, not through the Power BI framework. Under that, Tommy spots a YAML file in OneLake that is not the semantic model itself. Mike recognizes the Open Semantic Interchange shape, including a total-revenue calculation written in both T-SQL and DAX, and says Microsoft now has a converter from semantic models into the spec. He wants those files to reach the OneLake catalog. They park the governance and managed-self-service consequences for a later show.

Main Discussion

Topic: How to pick the AI model for agentic development

The main article is How to pick the right AI model for agentic development, co-written by Kurt Buhler and Eugene Meidinger and published August 25, 2026. Tommy frames it as model choice for the agentic cycle, not as picking a Power BI semantic model. The piece says there is no single best model, names six dimensions (purpose, capability, configuration and provider, speed, cost, and behavior), and treats effort as the reasoning-token budget. Model choice sits beside context, prompt, and tools. The published page also walks hosting, including local open-weight models, and a pattern of using one model to do the work and a stronger one to review it.

• Match the model to the job, then watch the cost of getting there. Mike’s current habit is to let a system pick the model. Planning and architecture want a higher-reasoning model. Writing code, once the instructions are tight, usually does not. A cheap model that needs four or five correction turns can cost more of his time than a stronger model that lands the result once. He points at a GitHub Copilot preview that takes one prompt and routes it, and he compares the front of that router to the classifier he and Kurt have been calling Jev. Tommy’s read of the lower-effort models, Haiku, Sonnet, older GPTs, and Luna, is that they fit work that is already defined, such as organizing files or searching a tenant, and that they hallucinate more once the task gets harder.

• A hard task is one that has to choose among options. Mike’s definition is reasoning: architecture, a reply that has to account for what the other person said, a report design. Moving files, searching, and writing a function with a stated input and output are the low end. He wants the code-writing agent to follow a clear task, and a second agent to write the test, so neither one has to be the strongest model in the building. His analogy is an org chart. Architects, code writers, and people who move things map onto different models, and a better harness can beat a stronger model.

• Plan on a strong model. Execute the recipe on a cheaper one. Tommy does not want one model to both plan a medallion architecture and build it. Planning, including a Power BI report, belongs on a higher model. He likes plan mode in Cursor and VS Code: requirements go in, a few clarifying questions come back, and a step-by-step plan with validation and a definition of done can be saved and rerun. Once the steps exist, boiling water and adding tomatoes does not require a chef. He will run that execution below Opus.

• They do not score the same Fabric scenarios the same way. Tommy’s game is low, medium, or high, aiming at the smallest model that still works. An inventory of tenant artifacts is low for Mike if it is only a search, and high if the agent is writing the documentation. Relationships on a ten-table model are medium for both. For measures, Mike wants a high model to decide what should exist, a medium or low model to write the DAX, and a high model again to build tests and review pages. Tommy had marked both the planning and the writing as high, because he does not want a loop of “the context is wrong.” Notebooks that transform lakehouse data are medium for both, and lower if the agent has an example connection to copy. A full medallion build is high for the plan, through the Fabric MCP, and medium for the notebooks and pipelines.

• Start from the question the visual has to answer. Mike had been talking with Alper at Microsoft, who told him Power BI never did enough to say why a visual exists. With agents, Mike wants the questions that keep someone up at night first, then a handful of visuals, then the model that can feed them. Budget, actual hours, and revenue have to be real numbers on a page before the model work starts. Tommy’s design phase on Power BI Desktop bridge has been medium, because more effort overcomplicates the page. Mike would spend high there.

• Examples lower the model you need, and hosting is still an open choice. Mike’s notebook agents miss variable libraries and the connection pattern he wants unless he points them at a sample. Copying a pattern lets him drop the model. Closing the article, he says the hosting question is the one organizations still have to answer: frontier-lab models, or open-weight models when the goal is cost. The techniques in the August piece still apply. The harnesses around them have multiplied.

Looking Forward

Read the Tabular Editor article, then run one of Tommy’s scenarios on the smallest model that can finish it and write down how many corrections it took. Mike is also changing the close of the show: this episode ends with a new song, and he asked the comments to say whether the treat should stay.

Episode Transcript

0:02 measures. Drop the beat now. H feel the crowd. Explicit measures. Drop it loud. Good morning and welcome back to the explicit measures podcast with Tommy and Mike. Hello everyone and good morning Tommy. Good morning Mike. How you doing? I am doing well. I’m just clipping along, having a good time. Things are

0:33 we’re we’re coming off a week here, a couple of weeks, two weeks of us talk talking with Kurt, and I’ve got nothing but good feed feedback from the community about how the conversation was going. People really enjoyed what Kurt was saying. They loved his thought and thoughtfulness on on the podcast. So, just want to again, thank you for Kurt for jumping on the podcast. It was very wellreceived., and because of the overwhelming feedback that we got, we’re gonna have to get Kurt back on. He’ll have to be part of a staple,, getting him back in the mix again. Anyways,

1:04 we’ve we’ve been talking a long time about having Kurt on internally, you and I. And I’m just so glad the schedule worked out for two weeks because, man, did we jive together. I I enjoyed all the conversation. So, good stuff. It’s a little nerve-wracking as someone who runs like a podcast show of like bringing on new people like you don’t like you know them like we know we know Kurt outside of yeah the podcast we know him as a a friend we’ve been conferences with him I’ve talked with him but sometimes you

1:34 don’t know if like will you gel well on camera will you be able to talk about things well there’s a little bit of dynamics that have to happen that make it interesting and fun for the audience because we want to be partly very we want to be part informational, but we also want to be part fun. Like we don’t we don’t want to be a boring podcast here. So hopefully you’re enjoying that. If you have other individuals that you would like to have on the podcast, let’s go back to the audience here. Feel free to let us know in the comments

2:06 if you’d like to have someone on the show. We’ll check those. We’ll we’ll make note and see if we can get them on Can you make me a promise, Mike? Depends. So, before we ever end the podcast for good, whenever that unfortunate day comes, not and is ain’t coming soon., can we have a rune on? Oh, you want a rune on the podcast? I want a rune, man. I That’s a hard pull, man. I I know. I know. That’s why I asked the

2:37 way I did. So, that’s my on my bucket list of the podcast life. I’ll have to float an idea past him and see what he thinks. See if he’ll pick up on the idea. If we did do this, I’d say it’d have to be something magical around like a good number, episode number. If you had it, probably a thousand. That’s what I was thinking. I was like, we’re we can’t do it until it’s like a nice big good. Can you come on for 712? So, yeah.

3:07 Yeah. Not not good. Yeah. We’ve already passed 420. That can’t be can’t be that one. Yeah, 500 was would have been good, but 500 would have been a good idea. We missed that one. Now we got to wait another 500. That’s okay. That’s right. We’ll we’ll make it do. All right. That being said, main topic for today is an article from tabular editor written by none other than Kurt Uler. So, so even though we had picked this article a while ago and

3:37 and Eugene is well there. So, Eugene is also on this one too. So, this is a co-written article on how to pick the right AI model for agentic development. This is not picking the right model for PowerBI. We’re talking about picking the right AI model for your agentic development cycle. Now, I’m also going to caution you here. This was written on August 25th. Okay? We’ll have to see how this lay this this like landscape has changed

4:07 because there’s already better models already. Some of them are cheaper, some of them are more capable. So, I’ll be I’m going to roll out the article and see how timeless this will be. Will this actually be a timeless article or is it more of a one-off? So, we’ll see. How sad is that that an article from two months ago like I don’t know if it holds up still. It’s already legacy. Well, that’s that’s actually in the world of AI now a month or two

4:38 months of time that’s like an a very long time and there could be a brand new pattern or we didn’t have Jev back then like we didn’t have other models in at our disposal like a lot can change right now. It feels like things are very fluid. I’m having a hard time keeping up. and so yeah I’ll be our main article today. Okay, that being said, let’s move over to Notion and our news articles today. So, Tommy, you’ve got a couple articles in our Notion notebook. What’s the news? What’s going on? Well, we got actually some big news.

5:09 There was the SQL Con Fabcon Europe 2026 that just happened in Barcelona and there’s some just starting, right? It’s like this week. Yeah. So, I think today so all these articles from six hours ago. So, Oh, yes. This is hot off the press. So, , if you are there, enjoy the conference. I hope you’re having a blast., but it’s now, what time is it in Barcelona, Spain? It’s like, , I think it’s five hours ahead. So, yeah, they probably just did the keynote.

5:40 I wish they had broadcasted that because that would have been fun to watch the Yeah. see the keynote and what they’re what they’re talking about over there. This is I love going to these conferences. This is the first one that I’m going to miss. I didn’t get any opportunity to volunteer at the community area. I’ll volunteer anywhere else. Just get me in. And they were very near very tight on volunteering and things this year. The venue was quite small. I believe Yeah. I think the numbers I think I heard was like at FabCon there was like

6:10 roughly around like 8, 000 people and they’re going to make that even bigger. So that’s going to be much like a Wow. This one I think was like 4, 000 5, 000 somewhere right around there. so it’s almost half the size of what FabCon Atlanta was. which which then means they got to be really selective around who they let in and I guess they’re trying to really keep it European centric stuff not bring a bunch of people in the US which yeah fair enough for me I’ll catch it next time hopefully. So I they did come out with some pretty good stuff. So there’s the main

6:40 obviously if you never really read the blog anytime there’s a major conference there’s always that main article that talks about the highlights. So we’re going to dive into that one. And then there’s one about fabric analytics and data governance that I wanted to bring to your attention as well. So the first one is connecting app databases and AI on one foundation new innovations at SQL con FabCon 2026. And Mike, I like the article already and I don’t know if you’re able to take a look because I

7:10 feel like this really drives well with our conversations and where we’re feeling like the trend is right the the high the bird’s eye view of where data needs to go and where especially fabric needs to go. This to me is in that right direction. So there’s three main areas where we’re seeing or actually four main areas that we’re seeing updates and where Microsoft’s really pushing. Again, that’s where you can read between the tea leaves of what Microsoft’s focusing on fabric apps building AI applications

7:41 directly on enterprise data. a lot of this is things that have already been out the connectors of fabric warehouses and lakehouses Postgress SQL support for transactional workloads TypeScript backend functions that we already know for back business logic and integrations and finally they’re going to introduce upcoming integrated application storage based on one leg so files going to be introduced to this so if you want to upload a receipt or really I think any any type of attachment they didn’t go

8:11 into too much detail So let me run through them and I want to get your thoughts. That’s the first one. Second is the database hub. Microsoft introduced the database hub. It’s in preview and it’s like the one lake hub or the data hub that used to be there. That’s a centralized management experience for SQL servers for Azure SQL, Postgress SQL, Cosmos and Fabric SQL databases. It’s also available as we’ll go into in VS code and obviously in the fabric UI which goes to database agents is a third major announcement.

8:42 This is available for database for SQL and Postgress SQL and Azure. These agents can monitor performance and workload health, help identify and prioritize issues, assist with investigation and troubleshooting and operate within a governance controls approvals auditing role based. Fourth is stronger foundation for AI u emphasizing their database agents already and again that’s focused very much on the database side for MCPs and skills one lake and operational business data and there are

9:13 a few SQL platform enhancements such as SQL server on Azure local disconnected operations Azure SQL hypers scale and just a few other vector indexing that’s generally available so Mike we’ve got fabric apps database hub dat database agents and stronger foundation for AI. What is tickling your fancy? Well, I really like I’m going to say of these three of the four things that were announced here, I believe the one that is going to be most revolutionary to

9:43 people in their day-to-day work or things are going to want to like it’s really changing what we do. It’s the fabric apps piece., by far is the announcement there. But it’s not only just fabric apps, right? So, there’s a couple things that I think are very relevant here that that have been landed. the features of fabric apps, they’re really rounding this out to become a really proper fullervice application development experience. You’ve already been able to connect to a semantic model and you’ve already been able to connect to a SQL server. They’re now extending more data connectors for

10:14 the the the app, right? One is a fabric warehouse. Again, that’s just writing SQL. Makes no makes makes sense to me, right? Lakeous, that’s really important to me, right? So the lakehouse gives me a lot more flexibility for storing files or having data being saved,, hey, upload this thing here, right?, right now we have a some some customers have a a way of making a portal and customers need to show up and upload files to something, right? Well, now you can build an a file

10:46 uploading application, a fabric app, where people can just upload whatever they want and then it goes right into the app. The app then saves it down to Lakehouse. Boom. Now we’re in fabric. Now it’s easier for me to manipulate that file with notebooks and other things. So this is huge to have some of these better data connectors. One thing I’m gonna I’m going to cover out here. They’re talking about Postgress SQL or SQL support and one thing I don’t understand is and while it’s good I’m happy to have that.

11:17 I don’t think there’s a Postgress database item in fabric. You can’t make one. Okay. So, this is why it boggles my mind a little bit. They’re talking about it a lot, but there’s really no item in fabric that does this. So, I’m going to come back to that one, but okay, interesting there., the control access policy makes sense, right? You may build an app, you may want to say only people in my organization can have this. Do not share it outside externally. It does not go on the internet. Like, you got to have this. This is this is the control plane of what we needed before we really can

11:48 start taking off with fabric apps. The one that I’m most excited is backend functions. This is huge. Extend your application with TypeScript functions and run basic logic, process data, and connect to external sources with a secret store protecting your secret credentials. This is massive. Without this feature, your apps can just talk to a semantic model

12:15 and it’s they’re somewhat effective. This opens up a whole new world. This allows you to add chat bots and connect to Azure Foundry. This all tells you to run APIs. You want to build a management portal for your fabric tenant and what you want to do there. All of that now becomes part of the backend functions. And this is huge in my opinion. This takes making a fabric app from just a novelty item to like, okay, we’re we’re here to play ball. This is the real deal. So anyways, that’s just

12:47 the one around the fabric. I think fabric apps for me is the most impactful thing. Data hub is interesting. I’m not me personally, I’m not managing a bunch of databases., but I will say the part of this that I’m a little bit annoyed about is they say it’s the centralized control plane. Now I need to dig in this feature a bit more because I’m not exactly sure what this means. It says managing SQL Server, Azure SQL, Postgress SQL, Cosmos DB, NSQL database, and fabric. It sounds to me like this is going to bring data from other sources,

13:19 Azurebased services, and put them in one single place. And I really instead of like so fabric has Cosmos DB in it. Fabric has SQL databases in it. Fabric does not have Postgress, right? Right. So, I kind I would like them to say I want to like yes, I understand this is managing this, but why would I go to fabric to manage all my databases? It doesn’t I guess I’m just not quite

13:50 clear on why I would need to go into fabric to manage the databases. It makes more sense to me to like I’m going to fabric to manage the databases that I put in fabric. There could be a lot of those. Like that does make sense to me, but does it really make sense to have it for everything? I I don’t know. I think it it well for fabric apps too like the the different applications and different systems that organizations use it’s dare I say it’s an all or nothing right like if you can only connect to some of your data but not the other essential part

14:20 and obviously they wouldn’t be just saying Postgress without it actually being something that is a common situation that they’re finding over and over right I think I think this is maybe someone I mean it makes sense that you can essentially manage all this stuff because you may like You might have some databases that are in fabric. You might have some that are not and you want to see the own place. That makes sense to me. Am I going to go to fabric to see all my other databases that are non Azure if I don’t have any databases in fabric? I I don’t think I would. So I think this is more of a story around

14:51 look, you’re going to put databases in fabric. You’re going to have things you need to monitor. Oh, and there might be some other things as well. You may want to also see them so you have a central place. So I think that is why this exists, right? Again, in my day-to-day work and how I work with customers, not as relevant as for what I need, but there will be organizations I think really going to like. Oh, I I think that it’s going to be a huge win. I want to touch on something that they briefly ran through and it’s only really just two sentences in the entire article that I think is going to

15:23 begin to pick up. So they they’re talking specifically about database agents, this new database agents. And it says here, Mike, make sure I’m reading this right because if this is true, then they’re going to expand this to the the other agents, the data agents. And this is where the I think real core of where I think the biggest enhancement is by exposing the database agents capabilities or the data agents capabilities through MCPN skills and leveraging fabric’s connection to

15:54 operational data through one lake. They help trusted business information more accessible. So, it sounds like that they’re going to allow database agents to use your own skills and like your own installed or created MCPs. Am I reading that right here? this is Yeah, I believe there’s a pattern they’re following by doing this. So, I believe when you create a data agent today and a data agent can talk to lakehouses or semantic models,

16:25 warehouses maybe, I’m not sure, lakehouses. when you go to a data agent today that data agent by default gets its own MCP server. So when you submit a you can when you connect to it you say here’s the MCP location of the data agent. So when you’re needing to ask questions you ask questions you talk to the MCP server. The agent that agent and MCP server are wrapped into one and then it does what it needs to do. it it builds

16:56 you a CU for AI usage and then the come back out, right? That’s what it’s doing. It’s it’s not an the agent itself is the MCP server. So, it’s like some large language model and an API call all bolted together into one thing. Okay. Yeah, that’s not I guess that’s not how I read. It’s like can you install my own skills? That’s what I want to know. Can because there’s obviously and we already know this with data agents. It uses the skills for fabric and database agents are going to also use skills for fabric for

17:26 databases. But can I have my own? This is one area that I’m not quite sure what happens here. Tommy, does does the agent like so when you build the database agent, would you go there? Would you upload your skills into the database agent and then it would have access to those skills when you talk to the MPC server? That’s one that’s a little bit more gray to me. it. I’m guessing it’s going to look very similar, Tommy, to what they’ve done with the current data agents, which you can add skills to or you can give instructions to. It’s just not very refined yet at this point.

17:57 So, anyways, yeah, I’m not I’m not sure where to go with that one. that part of there’s another point here I wanted to make around Oh, yeah. Agents Tommy, I’m going to propose a weird idea here. Have we, this is maybe a general question for you, but it might be more relevant here on this one. Have we just made a new metric? I know that Microsoft looks at fabric and PowerBI usage as a function of monthly

18:27 active users, MAU. So, I know that’s a huge metric. So, a lot of my lens on like what is Microsoft going to make next? Well, if it doesn’t make more monthly active users, odds are it’s probably not going to get built, right? Cuz that’s really the name of the game for them is to how do we get more people to engage with the platform that we already have? That’s the game. I think that metric in the future is not going to be as strong as you think. Okay. So, there’s something else that

18:57 you’re thinking of. There’s a new metric that I believe applications and I think this is in general if you’re a software developer this is the metric you should actually care about now moving forward I believe you should have a monthly active agents I don’t know about that I think you need a monthly active agents so think of it right behind every agent there’s some request from a human there’s some demand that’s being met there’s something like I think about the apps and things that I today.

19:29 Most of my agents that I have, I’ve given them identities. They have things they can go ask and go talk and connect to. And for me, I was just looking I just counting up with my friend. I had he sent me a picture of like here’s all the here’s all the things I’m using. And I sent him a picture. I’ve got 22 agents that can run on various tasks when I need them to be spec specific for various things that I’m working on. Let if we if we extrapolate that let’s just say I’m Michael’s an outlier here right

19:59 Tommy how many agents would you say you have in a single harness or across across like just looking at with the things that you’ve built Tommy and you’ve had you’ve named various things that are agents and they’re doing things on your behalf how many agents would you oh I’m around the 18 I think right now okay to yeah but let me let me evolve think about this evolve your metric Yeah. Well, every single human is going to be able to build 10 to 18 additional

20:29 agents. They’re going to do various things. I’m not going to want to go in and write a notebook. I’m not going to want to go in and and manipulate things. I’m not going to want to go run a query against a SQL database. I’m going to want to use a data agent. I’m going. So I think the measure of success for application development and software now in the future will actually be what is your number of agents per month? I monthly active agents. Let me evolve it. No, I was going to be mats

20:59 monthly active tokens or monthly average tokens. You I don’t think you can measure that. Why not? How why couldn’t you measure tokens? Of course you can. You can measure usage. So think about it. I don’t want to put No, no, no. Hold on. in fabric. I do not care about adding more tokens to like I’m not I don’t want to pay for any tokens with my fabric capacity. Does Microsoft care? But there’s no way for them if I’m using Anthropics MCP server or anthrop like Microsoft’s never going to know. That’s information they’re never going to see.

21:29 So I understand the token. Yeah. I don’t I but I don’t think I don’t think tokens is a good metric because if you measure if your if your requirement for making more tokens is the rule everyone can make more tokens they’ll just put everything on opus whatever whatever the best thing is and they’ll rip it and so everyone will just use the most expensive model not caring about efficiency so I think you do not want to measure the amount of tokens if anything you want to minimize the amount

22:00 of tokens being Yeah, everyone does. Every every system does. So, so having a metric around monthly active tokens, that’s not a metric that goes up. , you the the number counting the number of agents that attach to your platform, your software, your app. That I think is a much better use. It’s a it’s a proxy for people and more and more people are going to act interact directly with the agent and the agents will be doing work on behalf of the person. So, I think monthly active agents is going to be your next big

22:30 metric that software companies are going to need to care about. Well, I know they’re already beginning to integrate that with Google Analytics because since half the website visits now or a good portion are now actually coming from AI that as much as it’s a great segue to the article, Mike, I need we need to go over this other article bringing govern analytics to the flow of fabric analytics. And there’s really one main area that I’m curious on your opinion because I don’t know how how I feel about this. This is the section about PowerBI is evolving beyond

23:01 dashboards. And again, this Microsoft introduced vision of vagentic app era. So Mike, they’re introducing some bold statements here. I I that’s the the nicest the easiest way I can say that. So let me just read some from this and I’ll want to get your thoughts here. For more than a decade, PowerBI, which is crazy, PowerBI has helped tens of millions of users turn data into insights. Copilot has made the analysis report creation more conversational. We are now introducing a new app building

23:31 experience in PowerBI desktop powered by apps in fabric. Users can start from a trusted semantic model and use natural language to generate, preview, and publish purpose-built data applications of Microsoft fabric. So they’re saying this is office semantic model. The application can accept inputs, write back data data, preserve shared state and support operational workflows. To make this more broadly accessible, fabric apps will be get included with PowerBI Pro, giving existing users

24:01 create a natural path and report to purpose-built data operations. But that’s not it. Everything above depends on the semantic layer. We’re sharing an early look at semantic views, a new capability that allows users to find governed business metrics and semantics directly where data lives in one lake. We’re cementing our move towards a more open model where business definitions can be shared across workloads, tools, and AI

24:28 experiences across fabric. So Mike to Carlo, we got some is this is pretty major to me at least of what it sounds like. I feel like there’s some very bold statements here and this has been Microsoft’s holy grail with PowerBI honestly since the beginning. How can we use natural language to easily create a viewer insights for users to make that selfservice side of it? And it sounds

24:58 like they’re getting awfully close or they think they’ve now in a sense found the holy grail. So they have a nice little gift that shows or a little video that’s showing it actually build, but it’s not a PowerBI port, Mike. It’s a fabric app. Yeah. No, I I think the majority of So I think what you’re seeing here is an admission of PowerBI is a framework. It has rails around what it can and cannot do.

25:28 When you give an agent a fabric app, it’s able to build much better, more faster. It’s there’s a lot more data that’s trained on just building a website, right? That that’s way more trained on models. The models are way more capable there without a lot of like additional like skills and other things you need to add to it. So, I think giving your co-pilot and even desktop some ability to build fabric apps is good. I’m going to continue to argue here. Desktop is not the place to be doing

25:58 this. Yeah, because that they do have it. This is in desktop that this is happening. Yeah, I that’s a in my book that’s a miss. You should have started with the web., like the reason the reason I say this is because you can only use desktop on Windows and and people like Kurt has told me he hasn’t touched desktop in months now and we’re not going to be touching desktop more and more. So like I think we’re moving away

26:29 from desktop because our agents are more capable on the files themselves under the hood and not with% of where reports are being built right now. Give or take a percentage point here, Mike. I and I know how you feel about the web. I think everyone who’s listened to a podcast episode knows how you feel about the web. And again, no, it’s not that. It’s less it’s less about that. It’s more about desktop needs to be gone. It needs to go the by way of the power apps, honestly. focus on the topic. I love the desktop

27:00 one, but for the sake of the argument, Mike, let’s say this is in the web. Let’s just imagine and that imagine it probably will be there anyways like it’s not a big it will be it will be this ability. Let’s focus on here Mike is pretty interesting. We’ll say that their focus on the way the best way and this is to echo Kurt too. The best way to do national’s language to building your data is not through the PowerBI framework because it is so

27:32 intricate by design, but a fabric app can be HTML or TypeScript or whatever code it needs to be on the back end. and they’re just going all in that way for the ne the best path to take a semantic data and to build visuals on the fly with natural language is through fabric apps not through PowerBI. interesting that they’re going that way. I think this is the natural ev I think this is the natural evolution like let’s let’s just not even

28:02 let’s ignore desktop reports all the things right let’s just say what’s happening in the industry right now in the industry right now people are talking to an AI and by the act of talking to the AI it’s doing a much better job of getting the results I want give you an example Tommy the other day I had to do some manual analysis with Excel and literally pulling numbers and moving them between stuff Tommy me. I was angry by the time I was done with this process. I was I was livid because

28:32 I was it took me like two hours to do something simple that I know my agent could like and this is this is the point Tommy like once you’ve seen the speed like I have been I had I had a really hot heated conversation with a friend of mine who’s also very AIcentric as well and I was like I have zero patience for any application that does not have a good AI built into the application anymore. Right, Tommy, you’re using notion.

29:02 You’re talking to your agents. Your agents are doing these amazing document writing things. You can give it some direction. You set a process around your agents. Now, anything you use that doesn’t have an agent attached to it that can just automatically write, edit, click buttons, all the like the agent should be able to do everything the person does. Period. And for example, I was in an app that I was in QuickBooks, zero help from any agent. the worst app agent experience I have experienced. Right. And Tommy, that was a mess, an

29:33 atrocious mess. And I’m going to Excel and I’m trying to copy a file between things and I have a PowerBI report that has some data in it. So, I’m trying to put all these numbers together into the same pile and I’m going to myself, this is so painful. And I know my agent could have done this in a heartbeat had it had access to the right things. I could ask for the things I wanted and it would produce the output that I want. So anymore I’m like I I at the end of the two hours I I wasted my time moving around data and things that I know that an agent should be able to do. And this

30:04 this is going to become the demand from the audience. I like prompting and communicating with an agent better than having to click the buttons and spend hours of time doing something cuz everything I do now is comp. I build full apps. Full apps like that do things in a weekend. This is wild, Tommy. Like, so that’s the expectation I had with everything now with with this natural language and this ability and I,, this is going to be an easy experience for users. We

30:35 have to put in our parking lot because a conversation needs to be had on the governance side, right? The adoption side because it’s going to be very easy to do. And I think it’s going for the most part because it’s using a fabric app as the backend rather than a PowerBI definition. and how they, you know, since build visuals, I think we’re going to see this become something that can really explode that we need to be careful about too in terms of this goes into managed self-service. This may be the new realm of managed self-service

31:06 for people which we know has a lot of governance around itself. Mike, before we get to the article, I need you to look at the image underneath the semantic layer to find once it reuse. I’m actually I have it on my screen right now. Good. Good. Because I you might have noticed something about because I know I this is me mentioning the business metrics and definitions can be defined once in one lake and the same definition could be used in PowerBI reports and applications. They have an image next to it. The image is a file that’s in one lake but Mike that file is YAML.

31:41 This file to define your star schema or define your semantic model is not connected to your semantic model. It’s in a lakehouse. At least from what I’m understanding here. or from what I’m seeing here. And Mike, this is OSI. This is what Snowflake has been developing. And I hope to goodness they’re using that same definition and that same schema because that’s what it looks like here to me. Yes. Tommy, so I can I can 100% I I’m looking at this one. I know that I know

32:12 exactly the image you’re talking about and this image is in fact what I would call the OSI specification. It’s got the same terms and again to be very clear here they listen to us. I’m very pro on well I’ve been saying Microsoft you better stink and participate in this world this this OSI language and world of things and so two things have happened right this is being announced which is awesome by the way and because there’s also when you think about your tables that

32:42 you have in lighouses you don’t get a column description you don’t get relationships between things you don’t get the dialect of language that’s that’s on this one you don’t get to define an expression. And if you look in here, you can notice here inside this one, they actually have two new dialects that are in here. They’re talking about an So, this is look at the example here, Tommy. I’m going to go a little bit deep here for a hot second. I I know what you’re going to Yeah, continue. In the image they show you, they say total revenue. Total revenue is the

33:14 calculation. It’s an aggregation. It’s a sum. And if you read the language here, they have two different dialects. This is so cool to me. This means they’re just saying the term total revenue can be defined in TSQL and in DAX and here’s exactly how you write it on this table to tell the systems either this would what I would argue here this is going to be a consumption for the SQL database it’s going to be consumption for the SQL data warehouse you can say DAX and TSQL

33:45 this is huge because what this means is I can define a common metric in some generic place aka here with the lakey semantic views and that then can be used wherever it makes sense anywhere in any model or in any SQL database too huge and so this is what this is and they’re calling it semantic views which is okay I’m okay with the name this is initial reaction well what this is this is the enterprise

34:16 semantics this is what this is this is this is the whole company right doesn’t matter what table doesn’t matter where it comes from. And the other thing that I need to present here as well on the OSI the OC this is the the open standard that snowflake is participating with data bricks. A lot of partners have have shown up. I’m actively participating in the weekly community meetings around the OC spec. So I’m in there meeting the people. I’m seeing what they’re building. I’m actively providing feedback and changing to the spec that’s that’s there inside

34:47 the the the system. So this is Microsoft is participating. They have their own converter now. They can convert semantic models into the OSI spec. So that’s something that that is now available to us. Huge. Again, these are all these are big movements here. So yeah, I am very excited about this part. This is something that’s greatly needed., and I think there’s going to be, now again, I don’t love the fact that this is all just YAML and stuff on top of things, but the fact that they’re at least starting down this route and

35:17 giving you flat small files that are going to do this, this is a major win. And so I think this means I would have to argue here the stuff that we store here in these files will have to make it to the one lake catalog and and this is where you’ll have like better definitions and linkings of things across the entire semantics layer of everything that you’ve got. So there’s a lot more challenging things that are under the hood on this one. We’ll have to talk about this and unpack this a bit later. I have some very strong opinions around this one as I’ve been working on this one for months now

35:48 and really trying to hash out what this looks like. You may also see a workload appear in the near future. I was thinking I was thinking the same thing. Yeah, something like that. I think there’s gonna be a lot of people creating workloads for this because Mike, what I love about what we do is people always tend to ask us from time to time. Are you ever running out of topics like No, no. there are so many implications of all this. But Mike, we got 20 minutes. So, let’s jump into our main topic today. and Mike. So, I think this does fit with

36:18 some of our discussion here, which has been throwing agents. So, we’re like, yes, we’re not talking about the actual topic at hand here, but I do think Tommy, some of the things that we’re talking about, touching in these news articles, like agentic experiences, the first one, especially how we want to be agentic first, right? I want to talk to my agent to have my agent do things. I’d rather edit my semantic model using the MCP modeling server versus doing it by hand. Now

36:44 there’s a lot of really good improvements that have been happening that make this easier. So now when you talk to these different models or different systems, right? How do you know which large language model to pick? And I think that’s where this article is going to touch on. Go ahead, Tommy. Give us a So what I would like to do since Kurt wrote this is disagree with everything. No. since he can’t say anything. Well, the one thing I do want to do is a lot if you missed it, we had four episodes with Kurt Buler, who is again just he’s the

37:14 chief innovation officer or head of innovation and he’s just all around awesome dude and also smart dude. A lot of our conversations the last two weeks have been very much touched on a lot of this. So, a few things I want to do today, Mike, is I want to make sure we don’t cross over too much the last week. Try to take a few different angles here. But first, let’s give the lay of the land here about this article. And the article is from Kurt and Eugene. How to pick the right AI model for agentic development. This is on the Tableau editor blog. There are a few things here that we’ll

37:45 touch on in terms of different models are optimized for different jobs. There is no single best model selection because it’s based on the task requirements and always using the artist model like oh fable for everything. And there are six major dimensions that Kurt and Eugene talk about that distinguish a model. Purpose, capability, the configuration, and the provider, speed, cost, and behavior. And what and beyond model choice, you should also choose an appropriate effort level, the reasoning

38:16 budget for the amount of tokens that it’s allowed to use. And one of the big elements here is model choice and this is in the article. Model choice is a core building block of effect effective agent agentic development. The others are context prompt and tools. So knowing that there’s effort levels, there’s you know those six main areas and I would like to go through with you Mike on those six areas. if you we can start there. But before we do that, what in

38:46 the article stuck out to you both from the article itself, but also the conversations we had with Kurt last the previous two weeks, your own experience in choosing models. Tommy, a lot of a lot more of what I’m doing now is I’m trying to use systems that are going to be automatically picking models based on what what factors there. So

39:17 they how do you pick models? Okay, so let me say this way. When you’re picking models you’re trying to pick models that are going to get the job done, right? So this is this is like not this is not a just semantic model pieces. This is just like generally like what you’re trying to solve when you have issues or challenges you’re trying to resolve. For example, there are various tasks that happen along your workflow, right? If I

39:48 want a planning, I’m thinking, I need a higherend reasoning model to help me out. Right? If I am writing code, if I have good instructions, I don’t need as smart of a model to write the code. Most models are pretty good at writing code, but having appropriate tests and things after the code is written is also incredibly important here. your your experience Tommy with using a particular model only comes after you have spent some time in Fable versus

40:19 Luna versus GPT. And so your personal experiences with you pick a model, you run some task through it and then you get like a feeling around how good it is, right? That’s very unquantifiable. Yes, it’s it’s very difficult. Like you can say, I feel like I get better results out of,, the fable model. Better Yeah. result. Exactly. Right. And then usually what happens over time is you’re like, “Wow, I used that model and wow, that really burned through my credits really fast.” Right. So then you

40:49 get this this gauge of like, okay, well that was a really good model, but it was very expensive to use. It was a really I could get what I want done, but it was like cost me more, right? So Microsoft has come up with co GitHub copilot. It’s in preview right now. another thing called hydrofusion. in you give the agent one prompt and the agent decides there’s a little bit of a router in here and this sounds to me like they’ve added Jev like the front part of this is like Jev is in here like hey here’s the

41:20 prompt which model should I use right so it but that’s probably not what’s going on Microsoft or other companies have oodles of data around what tasks what model how much cost was it per task per month model, right? There’s so much data being generated by just talking to the agents while you’re doing a process. So, the real measure here should be you’re trying to get a prompt in, you’re trying to minimize the tokens used, but

41:52 you’re trying to maximize the output or the delivery of that result. Right? For example, I could give a a very nonsmart model, low-end model, not a lot of reasoning. I can just say here build me this website, right? And it may get it halfway right the first time and then I got to talk to it again. Okay, well now fix this, fix this, fix this. Then it fixes something, right? And then and then it comes back and I have to fix it again. So we may be talking four or five turns on something that comes back from the AI agent. Great. Super cheap.

42:24 We eventually got to the solution, but it took a lot more of my time to then have it build the thing, right? So in that scenario, I am minimizing the token usage potentially, but I have maximized or I’ve I’ve increased my time and I’ve it took me longer to get the output that I wanted. So, if you think about these factors and you push around these factors, you’re trying to find the model that optimizes on these different characteristics. Does that make sense, Tommy? 100%. Because when you get to the lower

42:56 models or the ones that in a sense have less effort like Haiku, Sonnet, the older GPTs, Luna, they are Yeah. Yeah. Exactly. They are great if things are already well defined, you know, in terms of like search for something in my in my tenant, right? And that’s one of the examples that Kurt goes over. I have these 10 files. Can you just organize them into a folder? It’s really basic. It does not require a lot of Yes. planning beforehand and it is being conscious of that because it to

43:28 the time point of view you do tend to seem more if you’re trying to do something call a little more complex with those lower models a little more hallucination like we did it. Yep. And it’s very much like working with someone where like did you like I know you published the report but did you verify everything? Did you make sure everything’s aligned? It’s like well no the report’s done like okay let me take a look at this. So, and I think this is an important thing to your point though, Mike, unfortunately, you can read the the providers documentation or their

43:59 descriptions on the models. And I have to admit, most of the providers do a great job on their doc sites or the documentation to aptly explain each model and its purpose and when best to use it. Now, that being said, is everyone going to do that? And unfortunately there is a lot of you don’t have a usage tracker or you don’t you don’t have an accurate tracker right where it’s my feedback it’s what prompt did I use how many times did I have to refine it that we don’t have that data

44:29 for so there is a bit of feel to this that being said you have to start to me anytime a new doc model comes out what you’ve also found is there’s also recommendations how to prompt it when fable came out they have their own page in the book documentation that talks about the best way to prompt it and provide instructions. So that goes into it as well. I I don’t want to focus on the prompting, but understanding the model’s purpose and also how capable it is is like to me like the defunct part here. Mike,

45:02 I’ve been more and more in line with how you feel about how important is the model or provider and in terms of it’s not as important as the harness and it’s not as important as in sense the effort. So, I’m I’m an anthropic guy like that’s really been my tool choice for really most work. However, you can see if you’re using it in a in an improper place, it doesn’t matter which model you use. And so the

45:34 configuration the provider the provider would be like OpenAI or Enthropic. The model would be obviously which version of their model that they have. Sure. And Mike, when you think about the harder tasks actually can we define what a harder task is because I think people say more complex tasks, right? use a bigger model for more complex tasks. That’s the always the the saying. Well, can we define that for people? What does it mean to have a complex task?

46:04 I so I I would say a complex task is what where you require reasoning, right? Something where you would think about the alternatives and the different options, right? Architecture design. writing an email needs a bit more reasoning in the email like what you know what are they saying what do I think and feel about what they’re saying and then articulating out an answer back from that there’s more thinking that is required to get a feel and look for c certain things right planning out an architecture planning out a

46:36 design planning out a report those things require more reasoning because there’s many different options you could pick and so you’re reasoning through the best options or the optimized solution for customers or things. Those things are high-end reasoning. Lower end reasoning, moving files, searching through things, writing functions, writing code. While writing code can be quite complicated, you don’t really want to have the code writer agent reasoning about what it’s writing, you want to have very clearly defined tasks

47:06 for the agent to write code to do, right? Hey, I need a function that does this, right? I’m going to give you some text in and the output must be the number of characters that are there. Right? And then you can tell the agent to okay and once you’ve done that you give a different agent okay I’m going to give you a function that you put text in and numbers come out by counting the number of characters you need to test write a test that you can run against this function I’m going to give you right so now you have two adversarial reviews on things so both of

47:36 those two agents separately don’t have to be super smart but they both have to have like enough capability to get things done so again going back to Tommy like what’s easy or hard, right? A lot of those traditional activities that you would put people in front of, like think think about roles in your company, right? If you’re going to hire an architect or you’re going to go hire someone who knows more knowledge and background information, that’s the person that’s where you’re going to put the higherend models. Models are just like, let me say this again, models are

48:09 a very good analogy to people in an organization. It’s a very good analogy. You have people that are man you have agents that are managing other agents. You have agents that are experts in architectures and designs. You have other agents that are good at writing code. You have othera agents that are just good at just moving things around. Right? So the models and the harnesses, the patterns of those that this is why this is interesting right now, Tommy, because there’s so many moving parts. You can have a much less performant model, but your hardness is better.

48:41 Therefore, you get more out of it. Right. Right. I think it’s Yeah. And I think it’s important to note the planning side of this. I I want to emphasize that because you mentioned that where a lot of people consider the whole process would be done one model like I need to build an a medallion

48:56 approach in fabric using these X Y and Z data sources. So I need to plan this and execute it and they consider that all the same part of the same process. But one thing I want to emphasize here is planning should be done by a higher level model especially if it is more complex even the using PowerBI desktop bridge and the design of report. I think the the one of the things I love about cursor and I’m pretty sure VS code has this too is there’s something called a plan mode where all it’s going to do is bas gather

49:29 the requirements that you gave it in the context that it’s in and put together a very de and then ask you some questions to clarify some multiple choice usually like four to five questions and then write a detailed step-by-step plan that you can read in terms of what and then what are we going do what is the what are we pulling in what is the execution steps what is the data test that we’re going to do at the end to run it validation steps and then when what does done look

49:59 like and that plan can be executed at any time so like it there’s really a markdown file that you can save locally and relook at which I like to do a lot too as I’m trying things out I’ll create multiple plans on the same prompt just to make sure that I’m I’ve seen everything but that’s the planning side that is again taking the complex requirements and even if the requirements are not terribly complex if it’s going to be a longer running process I want a plan designed from that first there is that a building in like

50:31 cloud desktop too in cloud code there’s some great skills out there plugin I think one’s brainstorming that’s available by default in in cloud they might have bought the company or the person who built it but regardless that plans it but execution if you have a detailed requirement and testing you can have that run on you don’t have to run that even on opus you can run that on a lower model because you’ve already in a sense said here’s the step by step with what you have to do take your example

51:02 again like if you’re making a meal if I already have the step by steps of boil water add the tomatoes sauté and step by step I don’t need to be a chef to do that I just have to follow the directions basically and that’s really what those lower models can do. Yes, exact that’s that’s why we need to be talking about this because this is this is how this makes sense. So let’s let’s go down to the article a bit more. I’m going to keep unpacking this article. I think a lot of the things that we just talked on are touching on on things in the article

51:32 why this is so good for you to read. You need to read this article, right? How AI models are difference differencing from each other, right? Purpose, their capabilities, what configuration, cost, right? I was talking tokens earlier. Each model costs something different, right? Right. a Fable model is going to cost much more than a Luna model, right? So, routing the correct task to the right model is so important and and that’s really a big function of your model speed. One I didn’t really call out here was tokens per second. I think Tommy, we’re going to see some

52:03 blazing fast. So, Open AI is having I believe today is dev day. Yeah, open AI which will be interesting. Change everything as always. Last time they did dev day, it was underwhelming. They didn’t really announce anything that was like revolutionary. So, I think they’re trying to make or hype up this Dev Day a little bit more., I also know, have you heard of the company called Cerebrris, Tommy? No. I think I’ve talked to you about Cerebrris. S E R E B R A S. They just went public. They had their public IP

52:34 recently. I was try I was trying to get on the on the Cerebrus public IP because this companyy’s phenomenal. So it the company’s name is Cerebrous AI and you know Tommy when you buy a graphics a GPU chip from Nvidia right that’s go this little like maybe like inch and a half 2 in by 2 in square of a of a graphics chip. Okay the cerebrus system the thing that we’re talking I’m showing you here the wafer how when they make circuit board chips. Yeah. Yeah. This is an AI chip that is

53:05 the size of the entire wafer. One chip from this company costs a million dollars. Oh god. That’s how much these chips cost. Yes. Now, why do you why is this so important? Right. There are so many more transistors and things on this thing that the model can just screaming fast. So, when you’re when you’re working with a a model today, you’re probably getting like 25, 50, maybe 140 tokens per second to run a model. These models on lower end like not high-end models but like

53:35 medium to lowerend models like ones that are just good for code writing. These things can rip 1, 000 tokens per second. You can’t scroll like look the speed of this thing is insane. And I think I’m going to put my little hat on here. My little guest hat here is they’ve made a Cerebrus has got a commitment from OpenAI to land these chips. These chips I think will be on OpenAI models. And I think I think OpenAI is going to land some models on these chips and you’re going to get this like Luna fast. You’re

54:06 going to get like,, fable fast. They’re going to put really big models on this chip and this thing is just going to be screaming fast and we’re going to see the speed of models go up. Now, Tommy, this is going to I’m going to go off a little tangent here, right? If I’m an AI provider, if I’m like I’m looking at the market right now, what do we need to do, Tommy, in order to get more tokens to be used? Build a million-doll chip that does, 000 tokens per minute. Part of it, right? So, in order to get

54:39 more people to use tokens and more of them, you need to make the cost of prices, the price to write to build a token cheaper. You need to drop the you need to l literally drop through the floor the amount of time and cost it takes to run to get a token done. This is Javin’s paradox. We’ve talked about this in the podcast a long time ago when Seth was on the podcast, but Jav’s paradox is if you want to have more usage of something, you have to bring the price down. And the lower you bring the price, the

55:09 absolute more someone will consume. So if you can there’s a there’s a scaling function here somewhere that’s like if I bring the price down by half do I get more than double the usage maybe right this there’s exponential curve to this is why we have light pollution today right light is almost free to give when it was really difficult to make candles and drag oil around with you wherever you had to make light we didn’t have a lot of light there wasn’t a lot of light pollution you turn out lights at night because it was so difficult to get the

55:41 light to move where you’re going. Well, now it’s a flick of a switch. It’s a light bulb. It’s electricity. We can transport this stuff anywhere. So, light is now so cheap that we have pollution based on light. And I think this is the same way that the AI and the token generation stuff is going to go as well. We’re going to get so we are already so hooked on it. They’re actively trying to push the price of tokens down so low so that way everything like this is the majority of what people will be consuming in the future. It’ll just be

56:12 tokens on everything, dude. And I this goes back to what we talked about at the beginning, right? Your average month or you didn’t say the token one, but it’s the number of agents and what we’re actually using. Mike, I wanna I know we’re getting near the end here, which great conversation already, but I want to do a little game with you if you don’t mind, unless you add anything here. So, I do I do not. Okay. So, what I what I want to do here is take this in the realm in the environment of fabric and we’re going to play low, medium, high. Mhm.

56:43 And what we’re going to do is I’m going to give you a scenario in fabric that you would do and you give me the type of effort or model that you would use for it. Obviously, we’re talking very general here. If you’re listening, I don’t necessarily take this completely and just say, “Oh, they said high for everything.” So, I want you to say and try to take the lowest denominator here, right? Not your preferred, but what you think you get by here. So, I’m gonna run through you and I want to see how much we actually match up on these.

57:13 Make sense? Yeah, sure. All right. So, first one, I’m doing an inventory of my tenant. I want to know all the artifacts and have them documented. What size model am I doing that one? Yeah. Are you doing low, medium, high? I want that to be right. I’m gonna pick high. You’re gonna pick high for that? I’m going to pick a low for that. It should be a simple search. But wait, am I adding am I having it make

57:44 the documentation or am I just looking through documentation? It’s it’s just creating an inventory. It’s not necessarily going to document everything with the descriptions. Sorry, maybe I misunderstood the question. If I having it make the documentation, I’m using a high reasoning model. If I’m having it search through and comb through and recall data from the existing knowledge base, yeah, I’m probably going to say medium. I would say medium then or maybe low. But maybe I would give it a little bit more reason because I don’t want it to give me dumb answers. Starting from a standard semantic model and we’ll say there’s about 10 tables in

58:14 here. Create the relationships probably medium. That shouldn’t be too hard. That’s what I had as well. Create that same model. We’re bare bones. I’ve given ample requirements. I need to create measures. DAX measure development. DAX is a bit harder. I think I would do a mix of things. I would probably do a high-end model to

58:44 talk about, hey, I’m planning out some measures for this model. Yeah, let’s think about these, right? What measures should I should I be building, And then I’m probably using like a medium or a low, probably like a medium to actually write the measures, right? So the medium model is going to go read like best practices. It’s going to go understand the it’s going to go search for things. is going to make sure that it uses the right language and skills. But the high-end reasoning model, I want to make sure that I’m building the right things because measures, I think you don’t just slap measures in a model, Measures are designed to be paired with a visual like what aggregations do I

59:16 really care about. So in that part, that reasoning element should be more thought up. Yeah. Is a little bit more mundane. I had high on both because to your point you they may not get the DAX wrong in sense of like an error in the DAX but to to go back and forth with the model saying the context is not right here. I don’t want that refinement. So I went high high. Let me also give you one more here too. So I’m going to go high on the reasoning at the in the beginning. I’m probably going to go medium or low on the writing

59:47 of them. But I’m going to come back with another step in here that is important which I will put on high. I’m going to have it build some data quality tests. Hey, I want you to go build a report page or I want you to go run these queries. Grab some relevant dimensions to that measure. Make for me tables using that functions using those measures. So, what I’m finding I’m doing now more is because I because now I can make the agent manipulate I can have it help me generate the model. I can

60:18 have it help me write DAX in the model. But now I’m actually having the agents help me build tests and take this measure and combine it with different dimensions, put it on a page, let me review it. I want I still need to be the reviewer in the process, the human in the loop, but I’m making the agents build more for me so I can just see what they created and then the output of that creation. I think the test is it’s almost like documentation on the testing side a little bit is what I’m doing now more. No, 100% with DAX. You don’t want to go back to the chat going hey for us it

60:50 should be this it’s not this and then the amount of reasoning has to do because again it’s just context there so there is some like models if there are star ambiguousness to DAX yeah there well and there’s the model shape really determines what DAX things you can do right right and if you’re I think we think a lot of in terms of just very simple DAX

61:10 measures I’d almost sorry I’m I’m I’m continuing to unpack this idea particularly Tommy Right? Because there’s such an interconnected element of the model and then the report and the visuals that you put on there. Right? Since that’s so interconnected, I’m probably having an exercise around the model and building things for the model, but I’m also probably trying to build out a mocked up version of what visuals are on a page. Hey, I need bar charts. I’m going to try

61:40 like I want the fact that we have agents at our disposal. We can describe the kinds of questions that keep me up at night. Tom, you say this all the time. I love that., hey customer, what questions of your data keeps you up at night? Give me that information. Right? And then we take that information and say, okay, based on what you you told me, like look, I don’t know. I want to give you some of my consulting business metrics, right? I’ve got numbers of employees. I’ve got people working on various projects. Everyone

62:11 needs to know how much work they that they have accomplished, how much work is left on the contract in totality, and how much revenue have we generated on a per month or per week or per day basis, right? What’s that what does that all look like? So, those are the things that I’m actively looking at over and over again. And so, now that I understand what I’m looking for, I need the agent to go build me a handful of visuals that meet those needs, right? And then I interact with those visuals and say wow I really should have like on this graph it needs to have like

62:42 a budget on this graph it needs to have a so you need you need to interact with the real outputs or the ideas because that all then informs okay where does the budget number come from where did the actual hours come from where does revenue get calculated those things are fundamental first and then you can walk into okay great now that we have more of a semblance of what the the intent right of the report. Now we can go build the model and I think I was

63:12 talking with Alper I don’t know if you heard Alper from Microsoft he’s one of the he’s now in research I believe but Alper was telling me he goes we had a big disservice when we built PowerBI and he he was pivotal in building almost of all the visuals and really but he goes we don’t do enough job defining what we’re trying to visualize we don’t do a good enough job defining why this visual exists like the intent but now with agents We have full capability to build on intent. Why? What

63:43 is the intention here? Start with that and from there everything rolls out. Sorry. I I No, no. I I love that because I can’t emphasize enough. I know we’re talking specifically models, but the way Mike We’ll save this for another episode. The way I build my instructions I’m very very proud of in terms especially when you’re doing a full development. So I’m going to run through three more and then we’ll we’ll get out of here. So speaking of design, your design phase of building a report, if you’re going to use the skills or PowerBI desktop bridge, what are you

64:14 using? Low, medium, or high? Probably high on bridge. Okay, I’m actually a little bit more complex. So, interestingly enough, I’ve been using medium because I’ve noticed more is not always better because it can it can over complicate the report because you really start adding. So I’ve been actually doing medium. All right. Notebook just we’ll say general creation of a notebook to connect to data in a lakehouse to try to do transformations. I could probably do medium on that one. I would do medium. Yep. Yeah. Because especially if you have I

64:44 mean you might be able to get away with low as well if you have other good examples inside notebooks, right? Examples of how to connect to database. One thing I found it faltering on when I use agents with notebooks is there are there are patterns of how you’d like to connect to a SQL database, a lakehouse, how you need to store things with variable libraries. Sometimes those details get missed and it does a weird thing when it builds connections to data sources. So in some situations I I need to like have a little bit a better

65:15 example. So if you give it more examples, it seems to deliver better results and then I don’t need as high of a engine modeling to to match things. One of the things I would point out here, especially working with notebooks, is you will you will want to find good examples of code that you’re looking to replicate. Agents are incredibly good at copying patterns. The pattern may not be the same, but it’s good at copying the pattern or mimicking the pattern. So having examples of, hey, go look at my sample code here. make a connection like

65:45 that on this other net new notebook. Much better practice there. Honestly, Mike, and I’m already already adding this as an episode. The more we go through this, the more again for those listening, the instructions and how what instructions you’re giving the model regardless of the model is going to dictate so much. And it’s hard to have a lot of these. So, the last one I have here is is going to be get building medallion architecture. So again, I know that’s super general, but give me the load.

66:16 , if I’m if I’m having it build the entire like, if it’s reasoning through with me to like what needs to go where and how do we do that, the build of it is much less complicated, right? It’s notebooks, it’s pipelines, it’s things like that, right? So I’m I’m definitely So one of the tools that I’m for sure going to be using will be the fabric MCP server, right? Oh, yeah. Yeah. Because that can that that you can use that to create a pipeline. can use that to create the notebook. So, I’m going to do a lot of planning up front with a high-end reasoning model, thinking through what this looks like, how to

66:46 build it., and then on the creation of the notebooks and the individual items, I’m probably using some medium stuff, dude. I love it. So, I think that’s everything, Mike. So, what a great conversation today. I think so as well. I think this was very useful and extremely helpful as well. So, anyways, really good article. I like where we’re going with this thing. A lot of things in here in this tabular edit article I think are very on point. I still think a lot of these things are even though this was written two months

67:17 ago, I still think this is very relevant. I I think this is extremely useful. What I will say though is there’s a lot of harnesses that are out there right now. Mhm. Right. There’s a lot of harnesses that are helping you build better things. and so I think there’s a lot of considerations here. the article was very like gen general that like these are techniques and patterns you can use then and you can still use them now they still apply one of the things I think that’s also very relevant as we close here Tommy is the last part around how do you host the model where does it go

67:50 and I think this is another big question for organizations is because you can get models from Frontier Labs you can also use open- source open weight models to also reduce costs So, not only we just talking about like Open AI and Anthropic, we could be talking about a lot of other firms that are building really good models that are again, how do which one to pick? That’s the problem here. So, anyways, really good article., I’m going to

68:22 apologize, Tommy, because I have a surprise for you at the end of here. So, okay. All right. Let’s do it. What is it? A new song. It’s a new song. ,, go ahead. Let’s do the full wrap here. Is there any final thoughts you have for this article before we we we finish out here? No, I think we’re I now you got me excited. I can’t think of anything else. So, you can find us in Apple. Do you want me to do the full rundown? Do the full rundown, Tommy. Let’s go. Oh, boy. You can find us in Apple, Spotify, wherever, get your podcast. Make sure to subscribe and leave a rating. It helps us out a ton. Do you have a question ID or topic that you

68:53 want us to talk about in a future episode? Head over to powerbi. tips/mpodcast. Leave your name and a great question. And finally, join us live every Tuesday and Thursday, a. m. Central on all of Powerad Tips social media channels. Excellent. And because at the end of our episodes, we’re we’re changing something up here. We’re going to see how this goes. Looking for feedback. Let us know in the comments below if you like our new ending. With that being said, we hope you enjoy this. Tomy’s like, “Oh no, what’s going to happen now?” All right, here we go. Tommy, this one’s for you. You’re going to enjoy this one.

69:24 We’re going to end you on a new note today. Well, also for those of you are at Fabric Conference, we hope you’re really enjoying it. And these are this is going to be a treat for anyone who sits through and listens to the entire episode. We will now be giving treats at the end of our podcast. Enjoy.

69:57 Monday memo says accelerate. Burn the tokens. Don’t you hesitate. [singing] Dashboard watching every prompt you send. Use too little and you’re at the end. Secret agents. building systems in the basement. under the radar. Just trying not to get the

70:27 paper. Tuesday m the script again. Token budget bleeding. Cut it then. Same old boss with a brand new fear. Yesterday’s hero is over here. Secret agents.

70:58 building systems in the basement. Secret agents under the radar. Just trying not to get the paper.

71:32 So we spin up tools they never bless shadow stack to survive the test by seat for claw when the gate gets closed by seat for gro when the path gets froze. More tokens, less tokens. Make up your mind. We’re solving work they leave behind. A sanction ain’t the dream with chose. It’s the only way the deadline go. Business users in a quiet war. Sanction path don’t open

72:03 anymore. If the company can’t decide the lane, secret agents keep us in the game. Secret agents, building systems in the basement. Secret agents under the radar. Just trying not to get the paper.

72:38 Heat. Heat. Explicit measures. Pump it up. Be it high. Tommy and Mike lighting up

73:08 the sky. Dance to the day. The laughs in the mix. Fabric and A. I get your feels. Explicit measures. Drop the beat now. Pus kings feel the crowd. Explicit

Thank You

Want to catch us live? Join every Tuesday and Thursday at 7:30 AM Central on YouTube and LinkedIn.

Got a question? Head to powerbi.tips/empodcast and submit your topic ideas.

Listen on Spotify, Apple Podcasts, or wherever you get your podcasts.