00;00;00;00 - 00;00;24;21
Unknown
All right. Welcome to another amazing episode of Energy Bytes. Got rad. Dad. John. Coffee. How's it going? I'm Bobby Nealon and super excited to have John Henderson in from Scottsdale, Arizona, not Denver. Like I keep asking him every time I'm in Denver. But, head of Ford deployed engineering at Spot fire. Former advisors. Correct. And, yeah, I think we've been trying to get this done for a couple of years also.
00;00;24;21 - 00;00;43;01
Unknown
So really glad to get you here. It's great to be here. Really appreciate the time and doing this. This is a great timing with, you know, ATC and being in town. So happy to be here. Good weather week to be in Houston. Whether we're like I left Arizona, we had tornado warnings in Arizona. We don't get tornado warnings.
00;00;43;04 - 00;01;01;21
Unknown
So that checks out when the rest of the world is melting down. It's beautiful in Houston. And then when the rest of the world is beautiful, it's awful in Houston. Typically. Yeah, that's the that's how that works. But, yeah, I mean, just I think it's a really interesting time to have you on to obviously, like, you know, with everything with AI.
00;01;01;22 - 00;01;17;12
Unknown
But you know what, a couple of weeks ago had Claude dashboards announced and all these things. So just I'm sure you're getting it from every angle. Yeah. And, just you're not running on PHP or are you? Yeah. It's relevant. There's only a few vulnerabilities there. Yeah. That's wild. I mean, like, they're like, I think it's all about NSA.
00;01;17;19 - 00;01;38;03
Unknown
Like, literally like, turn it off. Like, if you can't create it now, like, and crazy, like, I'm wondering, like Facebook, right? They, they have their own version of PHP now. Right. But it's I wonder, is it a fork of a corrupted. Yeah. Version. God. Oh, man, I'm so glad I'm not in upset. There's some tech debt. Their security.
00;01;38;06 - 00;01;52;25
Unknown
I mean, you say that, but, like, there's going to be so much like space. Like you know what you're doing there. That is gonna like that is a huge frontier because a like you have 12 year old was talking to somebody about this earlier today. You have 12 year olds with uncensored models that can do whatever the hell they want.
00;01;52;26 - 00;02;13;29
Unknown
It's like we were ding dong ditching people and throwing, you know, putting people when we were 12. But yeah, it's like now God knows what they can do with with this stuff. And Michael Cortez's son was doing. Yeah. Like, yes, yes. I mean, Collins kid is making money off of rope books, games. He's making Robux games, and they're making money off of him.
00;02;13;29 - 00;02;35;07
Unknown
He's built like burning, like 2 billion tokens a week. Yeah, well, but, like, now he's having this, like, existential crisis of, like, he's like, why would I go to school if I'm making money doing this right now? It's, Yeah, it's a wild world. Sky's the limit. Instead of playing video games, they're like recreating and modding them. That is like, I'm so excited about that.
00;02;35;08 - 00;03;07;12
Unknown
Like, yes, I do want the Street Fighter characters and Mario Kart or, you know, Sonic. And now like a division within Microsoft that's actually just focused on making this happen and seeing if they can build a community. I love how it has, like this all spun off of PlayStation and everybody going diskless and then Microsoft, everyone like, is freaking out with all the GTA stuff and then those leaks and then Microsoft's just like, screw it, we're going to lean into this and like, yeah, I signed up for the little Microsoft Labs thing because I was like, hell yeah, I like I would love to be able to do this.
00;03;07;15 - 00;03;32;21
Unknown
Yeah. Anyway, hopefully no one out there is dealing with the PHP hack or railway. Railway also got hacked yesterday. I heard about that too. That's fine boy apps right. Like railways for deploying apps. Yeah, it's like a virtual virtual machines. And it's all discovered with AI, right? Probably done with AI to. Yeah, probably. Guess. Yeah. Or someone guiding go through PHP and find every little hole you can find.
00;03;32;24 - 00;03;47;12
Unknown
I mean but actually it should also they should also be the people on the other side that have the the patch like, you know. Right. I mean that's what it's going to be. It's just agents fighting agents like it's kind of crazy to think about, but it's almost, almost like people doing pen testing, right? It's like, go find the holes now.
00;03;47;12 - 00;04;07;15
Unknown
We'll go patch them. Right. All the bounty hunting is done by AI agents now. Yeah, yeah, I, I started a Aramis agent to start doing some of that, and I never finished it, but it's like, that would be dope. Just pointed at a bounty and let it go. Run through your process and see what the hell happens with some open source models.
00;04;07;17 - 00;04;13;18
Unknown
So that's that's kind of stuff. You're doing that spot fire, right?
00;04;13;21 - 00;04;36;11
Unknown
Every day? No, I mean, we get to build some really cool things, like we're, we're in for deployed engineering, so we're, we're the tip of the spear for the organization. So the things that that engineering can't get around to yet, we figure out where the gaps are for the product, for oil and gas clients and high tech manufacturing clients, what do they want from a workflow of visualization actions, agents skills.
00;04;36;18 - 00;04;56;12
Unknown
And then we build that. And then the things that get sticky with the customer base end up going into the roadmap longer term. Okay. So it's it's a it's cool to kind of step into the organization, you know, after being acquired and then being able to now kind of shape where, where the product's going. Yeah. And it's based on customer feedback and, and needs and wants from the industry.
00;04;56;12 - 00;05;22;21
Unknown
So that's one of the biggest things I think a lot of people have people in the weeds of it see how useful it can be, but the average person, I don't think, recognizes how much faster you can have that client to feature in the product loop completed. Nowadays, when you have a model like that supplemented with, you know, AI coding agents and stuff like that.
00;05;22;23 - 00;05;45;05
Unknown
But, you know, it's like, historically speaking, the cost of, you know, customizing some kind of enterprise software would be pretty significant unless it was, you know, a giant feature that everybody wanted. Whereas the future of this looks like people just customizing their stuff as much as they want. Like, it's pretty fascinating to think about from that perspective.
00;05;45;07 - 00;06;16;13
Unknown
So one thing I think might be helpful for listeners, especially if they're not familiar with you or like, so, let's talk about what you were doing at Enter Tell or what Anatole was doing, because I think it's important, like we even, you know, I moderated that, at the forum a couple of years ago. Where was we kind of showing the extensibility of spot fire, but how, you know, tell what you guys are doing within spot fire, you know, but then also, we had Brock from wise Rock and how they were kind of working outside of able to leverage it from the outside.
00;06;16;15 - 00;06;32;23
Unknown
But can I talk about what you guys were doing there and now how that parlays into this? Because because you were you've always been extending the, bounds of what you can do within Spotify, right? Yeah. We were I mean, we were building and gas workflows within Spot Fire, and we chose Spot Fires, a platform for two.
00;06;32;24 - 00;06;51;09
Unknown
Really, really two main reasons. So it already had massive market penetration within and gas. So it was easy to step into a company and say, hey, this is our solution and spot fire. I will bet money it looks different than your solution. In spot Fire it performs better. But it's also something we could iterate on quickly rather than, you know, a two week cycle.
00;06;51;11 - 00;07;09;07
Unknown
Time to change colors on a web SaaS like we had. We could change things on a demo. So you get questions and you can like feedback those questions to them in a demo in real time. You just couldn't have that flexibility with any other tools. So it was a it was a wonderful platform to build in. You have the full, you know, Python and R implementations.
00;07;09;07 - 00;07;27;26
Unknown
And so you can really build whatever you want the text areas or websites effectively. So like you can, you can make it look and feel however you want it to look. So yeah, I mean between what you guys are doing and then obviously you had mean literally queue engineering. Like they never even had their own web app where they built it all in spot fire.
00;07;28;01 - 00;07;46;05
Unknown
And I think what Zach with forecasts started off, you know, within there as well as just like, I mean definitely way ahead of its time. And you know, it was always frustrating to me because relative to, you know, even Tableau or especially power BI, it was like, man, like this, this tool is is great, but not enough people know about it, especially outside of our industry.
00;07;46;06 - 00;08;08;10
Unknown
Very true. Yeah. And they're not you know, they were not ever really brought up in the same conversations as power behind Tableau and some of the like the true like by tools. Yeah. Like Spotify doesn't consider itself a BI tool. They're not a dashboarding tool. They're a visual analytics engine. So being able to interact with the data more effectively, you can build algorithms in the back end to do whatever functionality you want.
00;08;08;10 - 00;08;25;26
Unknown
So that's how they like, they separate themselves. But it's always been very kind of carved into will and gas. And in high tech manufacturing, like where you actually have engineers and subject matter experts that need to pull the pull the trigger on, you know, 7 or 8 figure decisions. They're not they're not just relegating that off to an agent.
00;08;25;26 - 00;08;46;03
Unknown
It's still very, very hands on. Yeah, that's something I've talked about people because I mean, again, I think there, you know, is some maybe momentum against, you know, even just like tools in general, people like I can build my own, which we can totally get into that whole thing. But I will say, like even some of the engineers that have been very big, oh, man, I'm building this in Claude.
00;08;46;03 - 00;09;01;25
Unknown
And I got this tool and I, you know, made a dashboard or Streamlit or whatever super fast. Like they still say, like when I need to pull data in and, like, investigate the data and, like, slice and dice it like nothing beats, like they want to put it in a spot fire. They know how to chop it up, you know, move things around.
00;09;01;25 - 00;09;18;23
Unknown
It's like it's really hard to do that, you know, just talking to the data. You know, people people are still using Excel, even though there's a lot of better tools than Excel at their purpose built. But you can build a rocket ship in Excel if you want to build a. Well, I mean, that's what painfully slow that what Elon did on the way back from Russia.
00;09;18;24 - 00;09;38;03
Unknown
Like I think you got jumped into Excel and like ran all the calques and it's like actually this myself. Yeah. That's how we got flying cars to run in Excel to run differential. Yeah yeah yeah. If I never have to write any VBA ever again, I'll be happy. Man, there I took one of like my sophomore year in college.
00;09;38;03 - 00;09;56;12
Unknown
We had to take this VBA for Excel class, and we had to write differential equations in VBA to run in Excel. And I was like, what are we doing? Like, we had Matlab, we had all these other tools that were built to do math and were forcing this into an Excel worksheet. That's great. Little did I know, right?
00;09;56;17 - 00;10;12;14
Unknown
I go right into an industry that is that's just how we do things. And as a former Excel jockey myself, like, I still like I still know my way around. Like I was trained to not touch the mouse like it's all keyboards. So all investment banker trained and, you know, I don't think that tool is ever going away.
00;10;12;14 - 00;10;31;26
Unknown
And you can you could probably open source. I mean, I know there's open source versions of Excel and things like that, but that was right after Adobe took to Office Suite, I did it. I think it's but there's always been, there's always been like, yeah, but I think these are like even more direct. But it gets to like the like the Claude question.
00;10;31;27 - 00;10;49;13
Unknown
Yes, you can go and build your own thing, but that's, that's always even at Inner tell. So we built engineering and financial modeling tools and you know decline isn't that hard. But it has rabbit holes that you can easily go down. So it's like you talk to an engineer like, oh, I have my own DC. Like okay, great.
00;10;49;14 - 00;11;17;12
Unknown
Well what's your cost function like? What's a cost function? And he's just start talking about like the thing like outlier analysis where you, you know, your peak detections, all these things that have a huge impact on accuracy. And we competed. And when we're at interval the SB like their auto forecast bake off and we crushed it. And so like there's there's clearly huge deltas between you know industry solution A and industry solution B.
00;11;17;14 - 00;11;36;21
Unknown
So you know I think that's why we got acquired as we had a really good solution with great penetration. And it worked better than something, you know, that you can go and ask Claude to do. No. 100%. Yeah, I think I think so, I mean, like, I mean, in this world, where do you kind of hearing boots on the ground?
00;11;36;21 - 00;11;52;24
Unknown
Obviously, you guys are the tip of the spear. Like, you know, you're going into places. I mean, I'm sure you've you or your PhDs have people being like, well, I built this cool thing here, but like, I think the big thing too is like, okay, that's great. And that works on your laptop. Yeah. Now like now production into production.
00;11;52;25 - 00;12;15;00
Unknown
Yeah. Yeah, that's my favorite. Well, there's, there's always like two elements like one is, is governance and it trusts like I'm my like an inner tell. My biggest bane was getting like getting into IT procurement and having to pass that hurdle and just being approved. Because if it's involved it's going to take longer. You know, there's all sorts of hurdles and hoops you have to jump through.
00;12;15;02 - 00;12;31;06
Unknown
So, you know, with with spot fire, you have a tool that's already governed. They've already said, this is fine. This is this works. You can I mean, even like spot for copilot and spot for AI is part of, you know, 15.1. You can bring your own model, whatever model has been approved by it, you can have that and talk to it in a chat interface.
00;12;31;06 - 00;12;54;02
Unknown
And it'll build your visualizations within spot bar. So like it's already governed, you don't have to, you know, get anything else approved. And so everybody's also rowing in the same direction because they're all building on the same platform rather than in each individual direction. All right. So let's dive into that. Like so you've got like a chat interface and like so we can you can talk to that.
00;12;54;02 - 00;13;15;07
Unknown
And then you don't have to do any of the clicking and dragging, dropping expression like it. And obviously it has access to the data model that's behind it. You know, assuming the data is, in the canvas. Whatever. Yeah, it's it's very like it's very, very similar to you just opening up cloud code and doing things internally. You can build skills, you can build agents that have, you know, very like codified workflows, repeatable workflows.
00;13;15;07 - 00;13;33;14
Unknown
You can just call that and say, I have a well agent backslash. Well agent will scan your data sets. If you're connected a Snowflake or Databricks, it'll search all of your schemas and find other relevant data that's tied to it. So that's like the, the like the other element. You know, if I'm working in cloud, I'm generally connected to it.
00;13;33;14 - 00;13;51;09
Unknown
Repo I might have CSV, but the data is stagnant. Within spa fire you have markings and things that that additional contextual layer, it gets passed into the agent. It's just much, much easier to interface with. So rather than having to say, hey, select this field, I can just select it on the map and then the agent already knows what I've selected and interfaced with.
00;13;51;11 - 00;14;13;05
Unknown
That's my favorite thing that A been using and client facing stuff is. It's like, oh yeah, well have you tried mapping and Claude yet? Because it's not great. It's not very good or there for a while. It used to make cartoon maps for you when you asked it for a map. Power. I still makes cartoon maps. Well, yeah, they have real data on them.
00;14;13;08 - 00;14;33;24
Unknown
These are just like cartographer style maps, but the like the ability to go and interrogate the. The answer, I think, is also kind of where, you know, the subject matter expert, the petroleum engineer is never going away. They're still going to be making those decisions especially, you know, seven, eight figure decisions, like we're not handing off HSC decisions to an agent, at least not yet.
00;14;33;26 - 00;14;51;28
Unknown
We got to wait for that insurance policy to get rid of I guess despite everyone's. So you I mean you have you have the agent that's still working. It's in a familiar interface and you're able to then go and interrogate it and make sure it's giving you the right answer and use your your subject matter expertise to really kind of dive into things.
00;14;51;29 - 00;15;13;28
Unknown
Yeah. No, I think that's one of the like to me, that's a big thing that we're going to start seeing across lots of software, right? Is like, hey, here's here's our foundation. If you want the colors different, here you go. You know, like there's a chat agent right there that can configure the entire thing for you because it makes it so much easier.
00;15;14;00 - 00;15;39;08
Unknown
So like not just the UI, but okay, let's write some scripts. Let's write the code. Let's figure out the data model like you're already doing those things anyway. Why not do it where the data lives and where you're going to visualize it to begin with? And so I think it's again, you know, right now most things are still going through like some kind of loop, right, of client, you know, request submission, whatever.
00;15;39;11 - 00;16;02;24
Unknown
Somebody pulls that and fixes it and then deploys it. But the future is like that's just all happening in basically real time. And so it's it's going to be I think it it brings so much more power to the engineer. Right. Because I would say the only gripe I could come up with for a spot fire would just be the learning curve if you were going from zero.
00;16;02;25 - 00;16;15;01
Unknown
Right? Like if you had never seen or use a tool like that, there's a little bit of a just with. Also, yeah, it's a steep learning curve for sure, but something like this where it's like, hey, I need you to do this, and then I can do it, and then I can tell you exactly how it did it.
00;16;15;01 - 00;16;35;29
Unknown
If you want to, you know, just train you the whole whole way through. Seems like such so much more powerful than not having to learn how to write an expression. Yes, learning a new expression language. I mean, those those are trouble from, you know, when I was going away from Excel, I was going to say, because you're coming, if you're coming from another one, like, yeah, Excel or Power Query or something else, it's in its effects.
00;16;35;29 - 00;17;01;04
Unknown
And yeah, probably makes it harder. They're all all different. And yeah, not having to learn that and go through, you know, all the, the heaps of documentation. It's just it's so much easier. So like, it's I think AI it's giving, you know, tools that have been around that have industry adoption that are, you know, familiar. It's giving everybody a little bit of a head start so you don't have to learn how to, you know, code in Python.
00;17;01;04 - 00;17;31;22
Unknown
I can just ask Claude or whatever my agent is builds that data function. You put it in and boom, you're already in production. Yeah. And I think to your point, like that governed way of connecting to the data, whether it's through the spot, assuming you can use it with Spotify server or direct connections or however they can act, but like having those governed ways to connect to the data because like I'm hearing from people, oh, I'm downloading CSV of data and shoving it into it's like, oh God, please don't like, like, you're doing the same thing that we.
00;17;31;25 - 00;17;50;26
Unknown
Yeah, that's the problem. Don't start with the story is connected to this way that's approved. And you you might get a great like react visualization out of Claude feeding it your CSV, but it's in that chart is interactive, but your data is still stagnant. And so you really haven't. I mean, it looks cool, it's pretty, but it hasn't really elevated your workflow at all.
00;17;50;29 - 00;18;15;19
Unknown
And it's not repeatable. Right? Yeah. Now put that in prod. Yeah, exactly. Okay. It works on my machine. Yeah. What about even just AI? For me spot fires had mods forever to write and because it is a website you could actually build your own, you know, awesome interactive visualizations if you wanted that were not spot fire native. So I mean like, what have you seen as far as like being on the build mods or users building their own mods?
00;18;15;22 - 00;18;37;13
Unknown
It's it's fantastic. So like we have, we have action mods. So what used to be like an iron Python script becomes a button that's repeatable action and that's authored and scripted in the same way that your visualization mods are. But the mods are, I mean, you can build literally anything. It's a sandbox. You can bring in any, any library you want, or you can have quad in any library at once.
00;18;37;14 - 00;18;51;07
Unknown
And it's it's easy to deploy. And again, it gets back to like that trusted permissioned tool. But it's sandbox. You can't get in trouble even if you're probably using PHP or. Yeah.
00;18;51;10 - 00;19;11;21
Unknown
Man, that's I'm so glad. Just so glad I don't have to worry about that. I think you hit a little bit early, but, like, let's step back in your career. So how did you even get into all this? I'm like, you started at the Excel level and. Yeah. Did you start in energy or in tech? So I started in risk management.
00;19;11;28 - 00;19;35;00
Unknown
So my, my first job out of college was with a risk management firm in DC. It was founded by a bunch of CIA guys. Fred who? My co-founder in hurtle, you know, he's a, And so they had kind of a unique proposition. There's like people capital technology and then they're they're additional information. So being able to leverage alternative information.
00;19;35;01 - 00;19;54;10
Unknown
You know, provide information to hedge funds and things that are trading around events, but worked a lot in the energy space. And then also financial institutions that were investing in energy space, private equity. And I just found the industry fascinating. Like, I could I could easily quantify how good a company is based on, you know, how many barrels came out of the ground.
00;19;54;10 - 00;20;23;26
Unknown
And it was it just really resonated. So, after a few years there, I joined a private equity backed, burst reserve portfolio company and in a corporate development finance role, were trained by investment bankers, got the whole Excel Excel modeling, PowerPoint creation backlog. And then I followed my my CFO, then to his own startup and kind of got the entrepreneurial kick connected with Fred in 2018 and built up Inner Tell.
00;20;23;28 - 00;20;42;25
Unknown
And then Inner Tell was built really just out of, you know, my like, what tools did I want to have when I was in a corporate development finance role where I had to work with engineering, but, you know, not having to rerun everything through engineering when we wanted to change assumptions. So we built the full asset evaluation model within within Spot Fire.
00;20;42;28 - 00;21;04;04
Unknown
What was the craziest, like correlation or, or stat that you found from your risk assessment days or in your finance days? Oh, these things are always fascinating to me. It's, because they're either like, super deep and niche or they're the most obvious, like dumb, simple, stupid answers most of the time. Right? Like, most of the time.
00;21;04;04 - 00;21;23;20
Unknown
Yeah. It's it's like Occam's razor. It's it's the simplest explanation tends to, you know, make the most sense. I don't, I don't really don't have, I don't have a good correlation for, for the risk management work I did. A lot of it was, you know, pre transactional due diligence or, you know, investigations, things like that.
00;21;23;20 - 00;21;49;08
Unknown
But what about the events? Events were kind of crazy. So especially around just information related to what era like what decade are we. Are we like 2007 to 2014 of social media? Exactly. Starting of Facebook and all that. I think I joined Facebook in 2007, something like that. Dating myself now. Yeah, that was 2004. I think it started.
00;21;49;08 - 00;22;07;21
Unknown
I remember like hearing about people when I was in college. It was mine was 2005 because I, you still had to have a college, an edu to join. And the first college I was when I had it was before they opened it up to the parents. And the first college acceptance that I got, I got an email, so I signed up with that.
00;22;07;21 - 00;22;31;08
Unknown
But that wasn't the end of going to anyway. It works. It worked. Right? No. We did some really, really cool stuff like event based trading. So working with pulling in social or like internet data, every bit of data you could possibly think of, like subtracting offshore drilling rigs and then figuring out. So I will name the company.
00;22;31;08 - 00;23;00;04
Unknown
But there was a company that was drilling offshore, offshore Africa, and they were doing exploratory wells and their public density. And then we were tracking all the AI traffic. So everything that was kind of moving towards that, that drilling rig. And then we found out that there was a production platform and that was getting brought in. And all the crews essentially needed to do, you know, poppy tests and all sorts of things, like they actually, you know, hit paydirt.
00;23;00;07 - 00;23;21;12
Unknown
We tracked all that traffic and figured out that their exploratory, well, most likely hit something. And so we put on a trade or our customer put on a trade, killed it. Yeah. And ended up they announced, you know, announced a massive, massive finding a couple of weeks later. That's awesome. So it was finding informational advantages in places and most people wouldn't look.
00;23;21;13 - 00;23;46;16
Unknown
Yeah. Are you using satellite data at that point or satellite data? Before it was good to I was gonna say that was early days and we were doing we were doing like, like before a lot of the main data brokers now, we were doing some really cool things, like putting a thermal imaging cameras on coal plants and gas turbines and seeing like which, which ones were fired up, putting like acoustic centers on pipelines and, you know, renting that that equipment out.
00;23;46;16 - 00;24;04;24
Unknown
And now, you know, now this is all automated. You can go and get a subscription for just about any of this. So we were building out, you know, informational edges and in that universe, which was pretty cool. Yeah. No, I always find it like the oil and gas data side of things is fascinating. Right? Because it's like historically it's like, okay.
00;24;04;25 - 00;24;31;20
Unknown
Yeah. You know, drilling rigs and rig count and permits and like these, you know, just kind of static, laggy type things. But then over the years it's like, yeah, like Todd this company was doing they were doing a, you know, satellite imagery of all the frack yards and counting frack pumps and then same thing on the locations and seeing how many pumps were on location and looking at maintenance and how many pumps stayed on the yard like crazy stuff.
00;24;31;20 - 00;24;50;05
Unknown
Right. Like, and I don't feel like clearing pads or talking about. Yeah, like and you, you still can see it like that's, that's still like. And then I think in various or another, another entity was having cell phones tracked and they were just buying that data from the cell phone brokers and figuring out, okay, well, this this cell phone belongs to somebody.
00;24;50;07 - 00;25;09;02
Unknown
Halliburton, this ones that Schlumberger. And then they could actually track the frack crews through cell phones. Yeah. In addition to the satellite imagery. Yeah. Yeah I won't bring it up because Colin fan but but yeah, I do remember them talking about it but then didn't like AT&T or someone changed their AI or something or it was iPhone.
00;25;09;02 - 00;25;34;11
Unknown
There's their privacy. They changed their privacy in one of the like iPhone releases and it it everything out. Yeah. Androids will still sell. You know whatever you want. They believe in capitalism. Exactly. Whether you like it or not. Android, the bot of mobile. So I guess thinking about back to the work you were doing, you know, you know, I guess before any of this.
00;25;34;11 - 00;25;59;11
Unknown
But back then, like, what do you what do you have now that you wish you had back then, like what would have been like? And, that's like I feel like it would be so overwhelming to try and be an analyst now because there is just so much more information, like, you have to synthesize so much, you have to distill it, you have to have some really good filtering of like, this is just junk noise, but whatever.
00;25;59;13 - 00;26;18;28
Unknown
Like I can't fathom what that's like today. Yeah. Being being on the on the buy side now or even sell side because you have to track everything just as well, is a very overwhelming field, I would imagine. Hopefully AI makes some of that, you know, synthesizing of data a little bit easier, but it's there's still so many data sources.
00;26;19;00 - 00;26;35;05
Unknown
But I can see companies hiring less analysts not to because we can do more, you know. So it's like it's not any less stressful. You're more efficient, but you have more efficient with more responsibility and probably less biased because, you know, every every data source you don't look at or you don't wait equally, you know, that's bias you're introducing into that.
00;26;35;06 - 00;26;39;23
Unknown
Yeah, that risk or trade or whatever.
00;26;39;25 - 00;27;05;21
Unknown
So let's head back to Spotify. Because honestly, I've used it here and there just in a more traditional way, you know, for different clients. But like sounds like there's been large leaps as far as, like so like, what does it look like, me soup to nuts. Like building out like a spot fire to an out with some of the AI stuff involved, like, so when when I was at dinner, tell we never really had access to the server, the enterprise version of it.
00;27;05;23 - 00;27;25;03
Unknown
So we missed out on a lot of the really cool features and things that, that they were building. And a lot of that's changed considerably in the last year. Yeah. So, you know, we have we have a lot of stuff in the works now and then what what's released this month in 15.1 is really, really cool on the AI front.
00;27;25;06 - 00;27;41;12
Unknown
But we never, you know, we never took took advantage of the alerting or the, you know, automation, the job reporting, the things that you could, you know, push to a server to be done. We just never really played around with we didn't have access to it. Now that that we do, I'm like, wow, this is really cool. Yeah, I wish I could have used this, you know, three years ago.
00;27;41;14 - 00;27;55;21
Unknown
Yeah. Because you guys had where even about compatibility, right. You know, like you building a building 12 so that it would work and, you know, whatever everyone else was using or that was a big thing. It's why our, you know, we wrote a lot of our code and they're like, well, why did you choose R versus Python?
00;27;55;24 - 00;28;16;20
Unknown
Because Python didn't exist when, you know, yeah. The Python data functions were not nearly as ubiquitous or that was a new feature. And you know that that we predated Python as a, as a feature within Spot Fire. So we had you know, we evolved as the tools became better and we built on those tools. And if I did it differently today, would I maybe I don't know.
00;28;16;22 - 00;28;37;07
Unknown
Yeah. But yeah I mean it's just yeah. People again that's a whole other thing too. I went into a product and you need to be able to sell to multiple people. It's like, yeah, I could build this into this one. But if they have, you know, they're it doesn't let them upgrade except every three LTS releases, which is a I think that's a that's a real struggle is getting it to continue on with, you know, every new LTS release.
00;28;37;07 - 00;28;51;23
Unknown
And it's internally it's a, it's a big thing where we're always trying to upgrade people to the newest, newest version that there, you know, it groups will accept because that's where all the new features live. So everything our engineering team has been working on for the last year, you know, you have to be on the latest version to be able to utilize it.
00;28;51;23 - 00;29;12;20
Unknown
So, there's a there's a lot of organizational I don't I don't want to say damage, but you know, optimization that could be done if they just pushed out, you know, and got everybody on the latest versions of not just our tool but probably a lot of things. Yeah. Okay. But so let's say like 15.1 I mean this has been announced or it's yeah, it's this this this week.
00;29;12;21 - 00;29;32;05
Unknown
Okay. And I'm assuming you've played around with I mean like, I mean, what I mean, what does it look like? I jump in and I connect to whatever table. Or do I just need to establish a connection to the whether it's information links or the the data connection with the tables and like and it can go sounds like I can go search that and it can it can search the data schemas that you have connected to.
00;29;32;06 - 00;29;50;02
Unknown
So if you're in Databricks or Snowflake, we also launched an 15.0 push compute. So you don't have to do all the compute locally. If I wanted to do a calculate a column or push a data function to Snowflake or Databricks, it can now happen in the cloud. And then we can retrieve relevant portions or aggregated portions of that.
00;29;50;03 - 00;30;14;02
Unknown
So you can still, you know, you can handle billion row data sets or multi-billion row data sets still with ease in the platform, which is really, really cool. But the AI component, yeah, it's aware of everything you've brought in. It has context. You can you can also publish knowledge basis to it. So it's aware of, you know, not only your internal documentation knows about spot fire and knows all the data schemas, it knows what you're connected to.
00;30;14;03 - 00;30;23;03
Unknown
And then it can build and add new data connections. If it's if it's seen that before, really, really, really powerful tool. Yeah.
00;30;23;05 - 00;30;47;12
Unknown
What about I mean and then as far as like skills or different things I mean do you all have like a repository because you guys have had like the, like the data science or data science thing with different data functions that have been available before, but now like are there different skills or, you know, like around like certain domains that you can load that, that you guys have available that people can build on or it's kind of be bring your own or both or it's a little bit of both.
00;30;47;14 - 00;31;10;01
Unknown
And then that's the insight agents is one thing my group specifically builds. So we'll build around workflows. We'll build, you know, skills and things that tackle, you know, repeatable workflows within multiple industries that we're targeting. So it's a little bit of both, but it's really easy to do. And now you can prompt your own own skill. So if you wanted to create something you can you can just ask ask your agent to do that.
00;31;10;02 - 00;31;44;11
Unknown
Yeah. So it's getting easier to work with rather than, you know, there was a for some some of the agent work. You had a you had boilerplate code you really had to write in order to get it working. The skills kind of eliminate that. So please use skills. More people. Everyone just PSA. It just take two seconds, have Claude walk you through it for anything that you're doing a lot and repetitive type stuff like it is so useful once you understand IT skills and then mods which are recently those are those are always fun to mess around with and have like the context window.
00;31;44;11 - 00;31;57;00
Unknown
As you're you're building up that context, have it warn you when it needs compaction. And yeah, lots are cool if you haven't played around with it. Where do you,
00;31;57;02 - 00;32;31;06
Unknown
So far everything is happening within Spot Fire. Are there any plans or existing, kind of external connectors that y'all are either looking at or currently building to where it's like from coworker? I could basically have it build my dashboard in spot fire and then ideally pull that dashboard back up. Or would I have to I could have it build it, but then I'd have to go into spot fire to see it as that.
00;32;31;08 - 00;32;52;13
Unknown
It's so we've we've been doing a lot of work on MCP and just connecting out to other data sources. But also we've had, you know, a lot of those discussions of how, you know, how can we manipulate this tool with the coding interface without having to open the tool? So we've we've worked on that. I mean, if you've ever just pulled open a P, it's, you know, it's essentially zipped XML that can be parsed.
00;32;52;13 - 00;33;15;29
Unknown
So you can do it. Now there's just not really a I guess a supported way. Yeah. Per se. Well, I think that's the thing the claudin and GPT have to I think they're starting to build the ability for like you to customize an artifact to like your thing. I don't know what that actually looks like. And in execution, if it's like, yeah, we'll build it.
00;33;15;29 - 00;33;37;21
Unknown
Just give us all your code and it's fine. We'll train on. Yeah, we definitely won't train on it and build a skill that does it every time after that. But yeah, I, I think we were talking about this before we kind of started, but I think that's an area where, you know, there's there's a couple camps on UI with what the future of UI looks like.
00;33;37;24 - 00;34;02;13
Unknown
And just as an engineer, I, I don't see a world where the tools that we use don't have plots, maps, charts, tables. Like it's got to have that. Right. And so like that's where I see from an enterprise perspective, at least in our industry, the biggest limitations of these things. It's like, yeah, sure, you can, you know, draw me a plot where you chart.
00;34;02;15 - 00;34;20;12
Unknown
That's great. But that's not going to be days versus depth. And you know, like it's not going to be these customized things that we need or like a las you know, like a log. It doesn't have the ability to do that very well, at least natively. And so I don't know how you know what that looks like.
00;34;20;14 - 00;34;40;11
Unknown
Obviously, I think I'm curious what your thoughts are, because yes, the chat is great when it's needed and the chat can pull things or generate little things, but it's not the same as like if there was a native integration across all of these things where it's like, hey, here is your spot fire dashboard. Like, that'd be awesome, but I don't know if that's ever a reality.
00;34;40;14 - 00;35;03;11
Unknown
Yeah, it's a it's kind of it's an interesting dynamic like how, you know, right now we're building within the, the interface either with Copilot or Spotify AI that's nest, you know nestled inside of the tool. Like like to your point where you have to have the visualization layer, you have to have the maps, you need to be able to interrogate the data.
00;35;03;11 - 00;35;23;23
Unknown
As long as the subject matter experts still involved in that decision making. I don't like we're not going to move to an agent interface. You still need a UI, and I'm a big UI UX nerd, that's what most of the time I was talking at spot Fire conferences was about interface, and I think it's one of the also the things that differentiated us at that beautiful spot fire.
00;35;23;24 - 00;35;52;01
Unknown
Like, yeah, like we took a lot of like care and polishing that, and making it look different. It was also part of our branding and our differentiation. Logical too. Yeah. It's not just beautiful, but it's logical. Are you sad that you may not touch CSS anymore? I get to work with. I now get to work with our UX and UI teams at Spa Fire, and instead of having to change all that CSS, I can, I can make recommendations on things that I would change from an interface, which is cool.
00;35;52;06 - 00;36;19;06
Unknown
Being able to shape the product that I've worked in for a decade plus now is it's a really rewarding experience. I think there were people like you, and I think we had Andrew Munoz, like some of y'all probably knew the product better than a lot of people. Yeah. Internally. Spotify agreed. Yeah. So. Well, even just like. And you really I mean, I want to give y'all accolades on it because I remember going from spot fire to Power BI trying to build something.
00;36;19;08 - 00;36;38;15
Unknown
And I was like, how the fuck do I do this? I cannot figure this out. And it's just a completely different way of doing it. And I was like, this is so stupid compared to how I do this and spot fire. And it's like, oh, okay. It's like it's the little nuanced things that you wouldn't see if you're just kind of scanning, but when you're using it, you're like, oh, this makes so much sense.
00;36;38;16 - 00;36;56;15
Unknown
You take a lot for granted. Like if you if you ever go in the the drop the aggregation dropdown for a spot fire expression or on a visualization, it's you don't just have average count. Some like P10, P50, like interquartile range, you have all sorts of statistical algorithms that are pre-built into the tool that you don't have to.
00;36;56;16 - 00;37;16;14
Unknown
Right, you know, custom Dax logic or anything. Yeah. I mean, like, I mean, you even you have all the clustering and regression models and stuff like in there, I can and it would also give you all your, you know, p statistics and r squared. And I know you didn't have to use R to do that. You could but calculated Swanson's mean and power BI versus do it and spot fire.
00;37;16;14 - 00;37;38;15
Unknown
And it's it's I had no business running regressions. We were working together but we were and they were valid. And it's like, I'm glad I had a math major to make sure. I think I think it's funny though, too, because, I mean, I think with cloud dashboards coming out, it kind of like flipped it because like, I think for a while ahead of that, everyone was, oh, I mean, people don't people aren't going to use dashboards anymore.
00;37;38;15 - 00;37;51;02
Unknown
They're just going to, you know, things are going to be sent to their slack or teams and they're going to get, you know, they don't they won't need. And then all of a sudden cloud dashboards and be like, oh man, this is amazing. I can make dashboards. Like it's like, so you always wanted the dashboard. You're not going away.
00;37;51;05 - 00;38;06;14
Unknown
Yeah. Well it's like well worth it on here. That was talking about like they've tried to replace Excel and like we built an Excel plug in. We never thought we would build that. Yeah. Like those are the like you think dashboards are going to go away and then everybody still wants a dashboard to be able to view their KPIs and things.
00;38;06;14 - 00;38;23;01
Unknown
And I think, you know, they were invented for a reason because humans and like, that's the easiest way for us to process time series data most of the time. Right. Like it has a function. It didn't just appear. And so like the fact that we're just like, no, we're going to throw that out. It's just crazy to me.
00;38;23;02 - 00;38;48;17
Unknown
Right? Like but of course it's because, hey, GPT hadn't built it yet. So no, everything's everything's chat until we finally get around to building the dashboard. Still need to visualize it. Yeah, we still gotta have it. Yeah. I think one of the I'm not going to name names, but, former colleague of mine sent me this a few months back, but it was it was kind of what you're saying, like how, you know, a Spotify or DHCP is just like, kind of like zipped up XML.
00;38;48;17 - 00;39;05;23
Unknown
But he was like, I use cloud the other day to effectively reappoint databases within a spot fire file. I couldn't open anymore because the source didn't exist, but because it basically deconstructed the DHCP, modified the XML, rebuilt it, and he's like, it actually worked. Like, I probably know who you're talking about. Yeah.
00;39;05;26 - 00;39;24;19
Unknown
But I mean, it's just like that. I mean like, yeah, it is possible. I mean, this is like, will we support it? Probably not, but it is possible. And that's, you know, it's one way you can interface with it. But, you know, our value. I think the value of Spotify, like we were not a dashboarding tool. We're a visual analytics tool.
00;39;24;19 - 00;39;45;23
Unknown
We're we're not like, I don't think we're threatened by the cloud dashboard because that's more I think that's encroaching on the power BI space. And being business intelligence, we're not a BI tool. We're not streaming real time drilling or frack data. Yeah, exactly. And then you know that that's what you want to dashboard for is for that that real time data just streaming through and looking for alerts and flags and things.
00;39;45;23 - 00;40;06;04
Unknown
But, our value is being able to interrogate the data, being able to do like visual analytics and data science from the platform and, you know, natively, intuitively, and then also call on the agent where I need to bring in additional data, change things that I don't know how to or that I, you know, I don't want to learn how to build a data function, ask the agent to do that.
00;40;06;06 - 00;40;23;27
Unknown
And then rather than having it in a cloud environment, it's now in a production environment you can send send it and save it to the library. Yeah. We're not. So that's I think like that's the, the unique edge that we have. And that's where I think a lot of our moat is. Yeah. No, that makes a ton of sense.
00;40;24;04 - 00;40;44;29
Unknown
And do you have anything like, in there as far as, like. All right, so you've got this say even like a daily report with exception reporting or whatever it is, or like pump by exception type thing, like where say, then you can run it on your, your job, you know, as a job, automation job. And then we're like, it can actually process like the data came in.
00;40;44;29 - 00;41;00;08
Unknown
They and provide a, an output like using LM to like kind of look at it and send it, summarize it and send it out. Like I'm just thinking like I think a big problem then that people have with all the AI is then the orchestration side of it too, right? Where it's like, you know, how do I do this every morning?
00;41;00;14 - 00;41;29;26
Unknown
Or, you know, like without me having to go ask it or, you know, I want to build it once and then it should just work forever. It runs. And every morning I get this. Yeah, yeah, yeah. Putting it into production is always the, always the sticking point. But like, is that a workflow like that that I'm talking about like possible where like it can then provide like an analysis of, you know, maybe like you've created a workflow that a visual that looks the same, but it can go through and look at, say, the, the downtime reasons or codes or comments then like kind of provide a summary back, say to the production manager and
00;41;29;26 - 00;41;45;11
Unknown
stuff like that. You can have it interpret the data, you can throw it into an automation job and not run and then do all that, open it and do everything that you would, you know, you would do as you open that, that tool and go through things. So yeah, a lot of it, it is it's repeatable. It's and it's things that, you know, you can throw into a library.
00;41;45;11 - 00;42;14;06
Unknown
Everybody has access to it. Yeah. That's awesome. So yeah, give me some, give me some ideas that I didn't have. Yeah. It opened some doors. Yeah. What's the what's what's something you're excited about that y'all are currently working on? Oh, I can't I can't say too much. We're I mean, the, the cool thing about my group is that we're, you know, we get to build the things that that we want to build, that clients say, hey, be great if it does this.
00;42;14;13 - 00;42;35;01
Unknown
So we have a lot of on the, you know, industrial engineering side and being able to pull in new data sources, building connectors to everything, doing a lot of work on the midstream side, but also in high tech manufacturing. There's a few things that I'm I'm really excited. What's in our our roadmap for for the next year?
00;42;35;02 - 00;42;53;10
Unknown
I can't go into too much detail, but, I think a lot of the, the, the benefit of being in our role is we get to figure out what's, you know, we hear customer A wants this, we build it and we figure out is it used? If not, we, you know, abandon it. If it is, it gives into the product roadmap and it becomes a feature in the next LTS version or.
00;42;53;12 - 00;43;11;06
Unknown
Yeah. And I guess with that, I mean, we'll say with the AI behind office, I mean, hopefully every major dev team is using AI to accelerate their doing, but like, what have you guys seen even internally, the ability to like take a, an idea to, you know, model times or so much shorter like to build a new visualization mod.
00;43;11;08 - 00;43;39;03
Unknown
You know things that used to might take weeks can take days. Yeah. So we're we're burning through our backlog of ideas faster than we ever had. Have you broken your GitHub yet? I have not, probably one of my engineers has. The minute you start, the minute GitHub starts becoming a blocker for you is when I feel like you're hitting the stride of this, of this stuff because you're just like, oh man, everyone's waiting on GitHub.
00;43;39;06 - 00;43;58;07
Unknown
Who was it? Somebody I can't remember. Who was somebody rebuilt, get their own version of git and just recently open sourced it because they went from like 2000 commits a week to 30,000 a day. Yeah. And it's just not literally. No, it's not at all. It wasn't built for that. And so it's like, yeah, it's it's crazy.
00;43;58;10 - 00;44;22;18
Unknown
And they're all very well commented pushes. Yeah. Oh yeah. Well documented very lengthy documentation. That's my favorite. Yeah I didn't write shit for documentation for myself before. No. Like it's so it's so hard. I mean it hasn't been. No I love the idea of I'll write the documentation and then write the code. Then I'll write code. Yeah, but sounded like a good idea that that was probably disclosed in our due diligence.
00;44;22;20 - 00;44;28;18
Unknown
Like, our code could probably use some more comments. Yeah, yeah.
00;44;28;21 - 00;44;46;08
Unknown
That's the hardest part of a startup though, is like when you're when you're running, you're just dead sprinting documentation is always like, oh, it's the last thing I just doing. But now, now it's read my code base. Tell me what I did, give me the documentation. I don't I don't want to read the code, much less no. Yeah.
00;44;46;10 - 00;45;11;22
Unknown
The documentation wrote in there, but, so yeah, I think I think that's it. Man. This has been been awesome. I'm glad to see the, like, full cycle of spot fire coming back around. I also yeah, I'm right there with you that, like, interfaces aren't going away. Like they were created for a reason and they're still going to be around as long as humans are making decisions.
00;45;11;25 - 00;45;32;04
Unknown
And so, yeah, I'm glad to see all, you know, a lot of the problems that a lot of incumbent companies in our industry have is that they push back against new technology or incredibly slow to adopt it, even though their clients are screaming for them to adopt it. And so it's really nice to see companies that are doing that.
00;45;32;10 - 00;45;52;03
Unknown
Yeah. And I think, I mean, I'll I'll be honest. I mean, I've been kind of, I mean, forever I was being in the spot fire and risk man, I don't know. But at the same time, I think most of the companies that I'm interacting with are still on version 14, you know, that, you know, like and I think it's an important point, hopefully people listening and we'll talk to their IT folks about this.
00;45;52;03 - 00;46;10;15
Unknown
But it's like if you want, you know, you can't complain that, you know, tool XYZ is not, you know, made for this modern era. If you're not willing to upgrade, if you're not in the modern era, you're willing to just like for versions behind you only go, you know, get a quad enterprise account, just let people go wild with the newest model that comes out yesterday.
00;46;10;15 - 00;46;28;19
Unknown
But just plug in your app, but you're not willing to upgrade. Don't worry about your data. Yeah. Need to say a few the API, but you're not willing to upgrade. You know, tools like Spot Fire and then you're going to hear internalize them for that. You know, like that's a I mean, I think that's a fantastic point is that, you know, even the difference that I've seen between like 14, six and 15 is immense.
00;46;28;19 - 00;46;50;29
Unknown
And the, the, the additions we've done in copilot and now like that, that spot for AI's natively in the tool. A lot of the features we built in Industry Pro that are just dialed in for our industry and our workflows that we're doing on a day to day basis, like it's you just don't have access to it because a lot of the things are built with, you know, limitations on 14 8 or 14 six.
00;46;50;29 - 00;47;10;01
Unknown
And so just there's a huge leapfrog from from our last LTS version of where we are now in the next version. And what we have planned. It's, it's, I mean, we're we're definitely keeping up pace with, with how advances have happened with AI and, you know, where our tool needs to be to, to still be relevant and powerful for customers.
00;47;10;03 - 00;47;32;03
Unknown
Yeah. And and then how does like a customer get access to your team. Like, I mean, like, is this something through Spotify. They have to like request or it just comes out like certain projects come up and they're like, hey, John, your team needs to take a look at this or I, I've my mandate that I've personally kind of just helped try and push in is that I want our team to be more customer facing, customer forward.
00;47;32;03 - 00;48;01;17
Unknown
I think my team wants to also be more customer facing. So, you know, well, we sat down with a major last week after an AI conference and talked to them like I didn't know what copilot was or AI. We were you know, they were also on 14 or on 13. I know this this happened. So, no, I, I think our group's provided to anybody that really wants to, you know, if you have, if you have a gap or something that you really wish the product had, I'd love to talk to you and figure out where where are we missing certain things?
00;48;01;17 - 00;48;19;05
Unknown
What visualizations, mods, workflows, whatever. Where can we improve our tool in the short term? We build out a proof of concept. You use it, you test it, you abuse it and figure out where, you know, where do we need to kind of shape things. So sure it's a yeah, reach out. Okay. It's yeah, we're pretty open on that.
00;48;19;10 - 00;48;35;26
Unknown
So we definitely work a lot with our customer success teams and figuring out which names need more help. But, yeah, we're not we're not really picky in that in that regard. Okay. Yeah. I think there's a shockingly obvious like hot take in there. It's like you don't get to complain about the software if you're not on the newest version of the software.
00;48;35;27 - 00;48;54;18
Unknown
Yeah. Especially again because how fast things move out, like, you know, a couple years ago. Sure. I mean, a lot of people liked 12 better than 14 or like everyone loved ten or was it ten, which was the one where they made the they made some big UI changes. Was it seven and a half to ten? Ten? Yeah, ten was like their spot fire ex yeah yeah yeah that's what it was.
00;48;54;20 - 00;49;11;12
Unknown
Yeah it wasn't 1010 x. There are people not liking that even just because the the no one likes new UI. No one like no. Remember office the new office came out and they're like, what the hell happened? It's like, oh yeah. It's just more logically laid out and it's easier to navigate. But you've never done it before. Remembered that version, right?
00;49;11;14 - 00;49;42;27
Unknown
Yeah. Yeah. We're we're not we're not easy, easy creatures to change. Yeah. Anytime you get any kind of new, new interface, it's always difficult to adapt. Yeah. So yeah. Speed around. Yeah. What is your current daily driver. Which model are you currently kind of daily driving. Anything with a 5.5 behind it. So been splitting out workflows across to really whatever, whatever model makes the most sense from a cost standpoint and effectiveness.
00;49;43;00 - 00;50;12;20
Unknown
So I'll generally do my planning and like opus five five and then have haiku or sonnet to some of the grunt work. Yeah, no, that's exactly how you should do it. Don't, don't. No reason to burn those tokens. Yeah. Nobody's token maxing anymore. I well turn off our this week. We're we're hackathon this week internally. Yeah it's a yeah it's or have you messed with any of the harnesses or anything like that.
00;50;12;23 - 00;50;31;00
Unknown
No I've, I've, I've wanted to because I think harnesses are the next. Like it has a huge impact on quality of model. And I actually I kind of want to hear your thoughts on how do you, how would you rank model versus harness versus skill versus a like if you had to rank those, what do you think the most important.
00;50;31;02 - 00;50;51;03
Unknown
No. So perfect example. I've been working on some classification models for us and I got it. The one I'm talking about is on our land side, and I don't remember the specific stat, but I got it to, like, I always test, you know, most of our clients are on Azure, so most of the stuff we're doing is through ChatGPT.
00;50;51;06 - 00;51;13;07
Unknown
And so, but it's interesting with the new models they don't have, there isn't a thinking or you can turn thinking on and off, but there isn't a temperature parameter. And you know, what that was is used for historically speaking. And the older models is to basically turn off or up its creativity or its ability to use synonyms, is how I describe it to people.
00;51;13;08 - 00;51;37;20
Unknown
Right? It's a good way to do it. And so when you're doing data extraction or you're doing classification, you don't. You want repeatability, you want the opposite of creativity or synonyms. And so with the new models though there is that you can't very that knob is gone. So I did this benchmark of four one many 41545 for mini and Luna and Terra.
00;51;37;22 - 00;51;56;20
Unknown
And it's funny because for one mini for what that specific task was the best one and also happens to be the fastest and the cheapest out of those three. And so it's like it but that was just for a very specific extraction problem, right. Or a very specific classification problem. And so it depends on the layer that you're working in.
00;51;56;23 - 00;52;21;07
Unknown
So like again if you're doing like high level big things, fable, opus or soul, whatever makes a lot of sense. But once you get down to the execution, you actually don't want the agent to have thinking, really, you want it just to go do what it's told. Exactly. And so there's elements that going back to the classification piece, I was able to go through and get our classifier up to like whatever it was, 98%.
00;52;21;12 - 00;52;43;03
Unknown
And then I completely swapped the model out to a couple different open source models as well as Claude reran it and I got the exact same stats. And so when while I'm doing this, Claude is like, this is exactly what you want, because this shows that the repeatability is outside of the model now. Right. And like that's and so it's yes, the model is this foundational piece.
00;52;43;03 - 00;53;06;17
Unknown
But then you build all these rules and more deterministic processes that extend it so that it is repeatable because these things can, can vary. And so, you know, I think the, the harness and the things around the controls essentially in the knowledge that it has inherently is the foundation before you ever even get to the model. Right?
00;53;06;19 - 00;53;33;15
Unknown
Especially now, the models are all so damn good. Yeah. That, you know, it's it's not the it's not the high level logic or knowledge. It's the, the industry stuff that you're putting in the harness or that are built into the skills or the rules or whatever. You know, your configuration looks like it's all the other stuff that really gets it up to that 99, 90, you know, the into those percentiles where you're really like, oh, wow, like, this is great.
00;53;33;18 - 00;53;58;11
Unknown
I don't have to go run this through another iteration of the agent because it knows exactly what it was supposed to be doing, because it had all the right tools and stuff out of the box. It's a so yeah, that's my long winded approach on that. But ultimately the harness, I think, is one of the most important pieces because IT governance, it's the governance of the model of what the model is doing, the knowledge that it has access to or shouldn't use and things like that.
00;53;58;13 - 00;54;22;24
Unknown
So that's a really big one. You know, the skills to me are super useful if you're doing repeatable processes, because that's what it is. Right? And then, you know, you can string you can nest skills with inside other skills. And now that's like a workflow. Right. And so those are good for really deterministic stuff. But from that high level of like where do you really fine tune that.
00;54;22;26 - 00;54;46;13
Unknown
It's absolutely in the harness outside of going fully fine tuned model stuff, which more and more it's seeming like most people are leaning into the harness side because it's much cheaper, faster and easier, and you can get comparable results without having to go completely off the deep end. And it's been millions of dollars fine tuning something, but it's also getting easier and cheaper to fine tune stuff, so who knows.
00;54;46;18 - 00;55;08;14
Unknown
I'm fascinated in like the open source model world because it's it's crazy what they've been able to do in the last six months. How close they are to Frontier intelligence with not just that, but then these models are so good, and then they're distilling them and quantizing them. And it's like, now I can run a 30 billion parameter model on my MacBook, and that's crazy.
00;55;08;19 - 00;55;27;08
Unknown
Well, you can now and go. You can go to Apple and rent a, you know, a high end machine for $180 a month, which is less than your Cloud Pro subscription, and you can run a 70 billion parameter model locally, and your data never leaves your your silo. Well, have you heard I was talking to a guy at the coffee thing before this.
00;55;27;11 - 00;55;45;24
Unknown
Like about dark gloom. Something about like basically like if you're using Mac, like basically you can use other people's like Mac minis and. Yeah, like where it's like in vedere has something very similar called pear PAR. Yeah. But yeah, we I've actually been messing with it on our network. Yeah. He said if he didn't do anything bigger than we can do it on his MacBook air.
00;55;45;25 - 00;56;07;19
Unknown
Basically you can just run. You're running on other people's idle. Yeah, Mac machines or something like that. It's like distributed. Well, it's distributed in a sense. It's distributed in that if that machine has more compute, it can load the whole thing on that machine. But it can't take 330 gig instances and make 190 gig GPU. Unfortunately, they can't do that yet.
00;56;07;19 - 00;56;33;26
Unknown
But is that how they're doing? Like the protein folding on like PS5? They're doing crazy stuff. But even like, you know, being able to fold and quantize these giant models to run on a phone or a laptop is absolutely insane. Yeah. Everything you do locally now, I mean, even like what DB 2.0 came out today, I can like now it's got like async IO and like be able to run things like ten times faster.
00;56;33;26 - 00;57;08;03
Unknown
And yeah I love duck duck DBS. Amazing. Yeah it's so good. But you see like George Frazier from five Tran ran duck on his iPhone as a test. Like they gave the same performances like a Databricks medium cluster. Like like that's where I, I, I struggle with it. Not going that way though. Like because I mean, when you're in the big leagues, that shit is so expensive when you're talking about these billion, 500, 100 plus per parameter model, it's so expensive and no one has the compute.
00;57;08;06 - 00;57;28;24
Unknown
Right. And so it's like, okay, well then now everyone's quantizing this stuff so they can run it on laptops or gaming PCs. It's like shit, that kind of works. I mean, even my I've got a video, it's got an eight gig GPU, but before all the Ram shit popped off, the reason I bought that laptop because nothing was soldered in it.
00;57;28;24 - 00;57;55;14
Unknown
So I upgraded the the Ram to 96 gigs of Ram just to have it is my workhorse. And now they figured out what the mixture of expert models. You can offload the entire model into the Ram or the SSD, and then only load whatever you can fit on your GPU at that time to run. And so I can now run these huge models just because I have way too much Ram for this stupid machine, and you're still getting like a usable tokens per second.
00;57;55;16 - 00;58;17;05
Unknown
Very usable. That's amazing. It's crazy. And so yeah, the other day I was like, I wonder if I can just spend like 500 bucks on like a really badass SSD, like external SSD, and just do it this way in the future, but haven't gone down that rabbit hole yet. So hell of a speed around me. We slowed it down.
00;58;17;06 - 00;58;35;07
Unknown
Yeah, no, it was a good, good conversation. I have a we had if we go to Scottsdale where we we're going to go golf in the morning, and then where are we going to grab lunch and a beer. I mean, golf anywhere. Yeah. But I mean, anywhere you particularly like or. I mean, is is amazing.
00;58;35;08 - 00;58;38;08
Unknown
Yeah.
00;58;38;10 - 00;58;57;29
Unknown
And then, like, beer, they're like. So the. I love coming back to Houston because the food in this scene is a it's it's just impossible to beat. I tell people it competes with New York and L.A. like our food scene in Houston is awesome. Arizona is pretty good to, And so good beer. It depends on the vibe.
00;58;58;00 - 00;59;18;10
Unknown
Like, do you want the, the bachelorette party vibe or like. Yeah, or or something a little more sophisticated? Lots of good. But you're also seeing, like, chains like Ushi and stuff like that are opening up there. And, I guess what I wouldn't like. What do you prefer if, you know, I, like, look like low key stuff?
00;59;18;12 - 00;59;39;21
Unknown
There's a place called hush public House. It's just phenomenal food. It's real small bar. And the like. I don't like cauliflower, but they have these cauliflower appetizer that's just packed with flavor. It's amazing. So there's, there's little gems like that. And then you can just go crazy over the top brunches and I'm sure all sorts of stuff.
00;59;39;22 - 00;59;59;05
Unknown
I've heard it, I've heard it's awesome there. It's fun Town. Yeah, yeah, a lot of the spring training there to a lot of that. Yeah. So that starts picking up in February or. Yeah. March. Nice. I know y'all have everything. It's like it's a Bachelorette destination. They're the Astros. Just insane in and of itself and something for everybody.
00;59;59;06 - 01;00;17;00
Unknown
You really do. We have like some of the best spas and resorts in the world. Some of the best golf in the world. Like, it's a, it's a Mecca for so many different things. And the food, the food seems great. Shopping is amazing. It's it's a it's a fun place to visit, for sure. Nice weather is miserable.
01;00;17;02 - 01;00;34;29
Unknown
Several months a year. But, you know, I turn off my air conditioning two weeks ago, and it'll probably be off until the end of April, if not early, like mid-May. Yeah, the Houston mind cannot comprehend that. Yeah, I'm not a big humidity. I don't like humidity. And, luckily I landed here in a on a good week weather wise, but very much.
01;00;35;00 - 01;00;57;01
Unknown
Yeah. Yeah, our our weather in Scottsdale is dry and pretty tempered most of the year. Yeah. No, I feel like if you know, you've lived in Houston too long, when you travel west of Midland and your skin and body doesn't understand what dry dry is, and it's like I'll have a runny nose for two days. Just because it's dry like this is so bad.
01;00;57;08 - 01;01;15;20
Unknown
I finally remember this. I'm going to Colorado to get chapstick before. Oh yeah? Yeah, it's like I've never worn lotion in my entire life. I've never had that problem with go to Vegas for a long weekend. I come back on, my shit's cracked, and like, worst experience is traveling from Houston to, like, a ski town and then having like.
01;01;15;22 - 01;01;36;26
Unknown
Because, yeah, you get the runny nose, bloody nose, the chapped lips. And it's just it's the it's like ultra dry and you have massive sun exposure and. Yeah, that's that's never fun. This has been awesome. I know I feel like I took up most of the speed round time, so. Oh no, it's all good. We can, where can people find you?
01;01;36;27 - 01;02;01;14
Unknown
Yeah. How are they? Get in touch with you. Jonathan.henderson@spotify.com. Reach out on LinkedIn. I think I'm I'm there periodically. I'm mostly mostly just browsing rather than posting, but yeah, stalking working working with the word. Yeah. Well, I mean, I, I already served the pot with spot fire way back and like during Covid, but I think you guys really need to get out and market this the stuff you're doing when it, you know, when it's when it's publicly available.
01;02;01;14 - 01;02;19;01
Unknown
And people got to know man. Because yeah, that's one thing we're working on is just getting, you know, all the cool things we're, we're building, just getting some, some marketing out behind it. And so people are aware and then pushing people to upgrade can't, can't complain about what doesn't exist when you're not willing to. Yeah. To get on to that latest version 100%.
01;02;19;04 - 01;02;37;04
Unknown
Well awesome. Well, thanks for making time for us. I know you're yeah. Got a busy schedule here with ATC and everything, so thanks for having me. This is awesome. It's been great. Guys, if, Yeah. Like what? What you hear? Please give us a, like, a sub and, tune in next time. We appreciate it. Thanks, guys. Yeah, thanks.
01;02;37;07 - 01;02;37;17
Unknown
Thanks.