Subscribe
Share
Share
Embed
Chris Gmyr and TJ Miller talk about the long-awaited Turkish coffee finally landing, Unitree versus PiDog robot dogs, and a tour of TJ's forum comment analysis system at Luma. They also compare notes on AI as enhancement rather than replacement, multi-level synthesis patterns for both work reviews and product features, and the role of LLM-as-judge evals in shipping production AI.
Ask a question, suggest a topic
Podcast
- https://slightlycaffeinated.fm
- https://bsky.app/profile/slightlycaffeinated.fm
- https://x.com/slightlycaffpod
Chris
TJ
References Mentioned
Creators and Guests
What is Slightly Caffeinated?
Join Chris Gmyr and TJ Miller as they dive into the world of PHP, Laravel, and all things programming, while also sharing insights on family life and other musings.
Chris Gmyr (00:00)
Hey, welcome back to the slightly caffeinated podcast. I'm Chris Gmyr Yo TJ, so what's new in your world this week?
TJ MIller (00:03)
I'm TJ Miller.
⁓ man, ugh.
Lots and lots of housework. Still still on that grind. You know, I got I got my son like super stoked finally to be mowing the lawn again because he's like income stream. I upped his dollar amount for it. And so he's like now happy to do it. So I've got at least that off of my plate. But like I still I've got like
All the landscaping around the house needs to be redone. I've like never done it. So like we've got like a raised bed around the front and then along the side of the house, there's like a Bed that's like level with the house like level with the ground and everything and I tore out all the dead bushes shortly after we moved in and then did nothing with it So it's just been like now weeds So I've been tearing all of that up
Gotta get new weed barrier and like mulch and everything. blowing my back out doing yard work, that's all I've been doing. Yeah, that's.
Chris Gmyr (01:07)
Yep.
TJ MIller (01:13)
That's honestly it. How about you,
Chris Gmyr (01:17)
Yeah,
TJ MIller (01:18)
Ooh, I take that back. did, I totally forgot. I went to Comic-Con. I did, that was last week. Yep, all right, so I didn't do anything. It has been all housework.
Chris Gmyr (01:19)
I mean, yeah.
Yes, you did. You mentioned that last week.
Yeah, so with me, just a lot of the same stuff. I'm just plugging away at work, doing a bunch of AI stuff, working on a rig, which I mentioned yesterday, the local testing and seeding, DevEx tool, CLI AI thing, which is going really well. I'm going to be doing a demo of that later today. So moving right along with that.
TJ MIller (01:51)
Yeah.
Chris Gmyr (01:59)
And winding down on Scouts for the year. We have our last den meeting tonight. It's just going to be like a little playdate at a park just to kind of wind down and kind of be done for the year. Then we have like a graduation and like a picnic and a barbecue like over the summer, but nothing actually planned. So a nice little break until we start planning for next year when the school session starts.
Yeah, just do a little bit of yard work as well. My wife is going out of town next weekend or next Thursday through Sunday. So I'll be home with the kids, taking a couple days off, and she's going to be going to Puerto Rico for a girls trip for her sister's birthday. So yeah, it'll be fun. So yeah, just doing some prep for that.
TJ MIller (02:45)
How sick!
Chris Gmyr (02:51)
Yeah, just regular family stuff. Getting into the summer mode. It's been getting a little hotter here. So it's been like low 90s during the day. And the kids have been going to the pool after school. Yeah, I'm just kind of doing that.
TJ MIller (03:05)
Yeah, that's, yeah, it's starting to get somewhat consistently warm. It's been really weird. had like two or three days where it was like in the eighties, like, and humid and warm. And then it snapped back into the fifties. Like I woke up this morning and it was like 40 outside. It's just, yeah, it's like, it's been all over the place. So it's been, you know.
Chris Gmyr (03:18)
Mm-hmm.
wow, that's a change.
TJ MIller (03:30)
with the weather being all up and down. We've had some like really, with that temperature change, we've had really big storm fronts move in. So like all my joints are extra sore on top of, know, tearing these flower beds apart. you know, I talked about how I used chat GPT to help pick the color for my living room, right? And like visualize it. So I did the same thing with the landscaping.
Chris Gmyr (03:51)
Yep.
TJ MIller (03:55)
fantastic. I'm like, I've got basically a game plan together of what's going where, what to purchase. And then I was like, yeah, can you go find these at a local nursery or at Lowe's or something. And then they told me where to find all the different plants and everything. I told them wanted low maintenance, Michigan native stuff. So that was a really, really cool use case, man.
Chris Gmyr (03:56)
Nice.
TJ MIller (04:19)
I'm digging it for more of that. And my backdrop, I'm kind of doing the same thing. like, look, I've got all this stuff. I need a better way to organize it. I used to have skateboards behind me. I'm actively using them, but also they keep falling off the walls and breaking everything. So I'm like, how should I rearrange stuff? So it's like, use shelves. So I'm now building shelves to put up. that's been.
Chris Gmyr (04:44)
you
TJ MIller (04:45)
a super fun use case. Yeah.
Chris Gmyr (04:48)
Nice. Yeah, I
gotta try that out too, because my wife, we have a bunch of landscaping that needs to be updated, trimmed back, replaced with some native stuff. We have too many like, ornamental, like invasive stuff that the last homeowner and just the area in general just plants by default because it looks nice or it's like deer proof because they don't like eating them.
TJ MIller (05:02)
Mmm.
Chris Gmyr (05:11)
because they're not good for them. So it's like, we want to get rid of all this stuff. And she has a bunch of ideas, but I think being able to visualize it in a bunch of different ways. So we have the survey that we can put in there and be like, here's the top down, be able to gauge the area and measurements and stuff like that. Then we can go out and take actual pictures and then upload those and see if we can do a bunch of stuff with that. That'd be pretty cool.
TJ MIller (05:35)
Mm-hmm. It was,
yeah, because I was just going to do like a bunch of like low stuff, like low to the ground plants. But it actually like suggested putting in like some small like a Japanese maple or something, like a small little tree here or there. And it looked really nice. So I'm always like weird about putting trees too close to the house. I'm scared of roots. I've been bit by that too many times here. But like those those little trees, I don't think it'll be too bad.
Chris Gmyr (05:56)
Yeah.
Yeah, yeah, totally. We have huge trees here. We have like big oaks like all around the house and in the back and just, yeah, it's crazy. Crazy big trees.
TJ MIller (06:09)
Yeah, have, yeah,
we have like a huge tree in our front yard and ⁓ it actually reached under the house and like got roots and like crippled our sewer systems. like it's, I'm always like concerned with that, but then I guess the little trees shouldn't be too bad.
Chris Gmyr (06:26)
Yeah, not too bad at all. Nice. So moving on a little bit, did you get your Turkish coffee?
TJ MIller (06:30)
Yeah.
I did. I did. It was like right over by, it was right over by like towards my sisters and I was in that area and yeah, was like, I'll just zip over there and grab it and I did and it was great.
It was great to finally get it, however, didn't hit the way I wanted it to. So I was like, satisfied, but I'm now wanting to go find a coffee, coffee shop. They do coffee, but it's also a little lunch joint.
Chris Gmyr (07:15)
Mm-hmm.
TJ MIller (07:19)
I kind of want to go find a better one now. But I'm not going to make... Oh no, I'm definitely not making a pilgrimage out of this one. But I did finally get the Turkish coffee from the place that I planned on getting it from. So... Mission accomplished. Yeah. So... Cool, man. I've got this like weird...
Chris Gmyr (07:22)
There you go. That's your next homework.
Nice. Mission accomplished.
TJ MIller (07:41)
obsession of late. I don't really know where it came from, but I have been super into these robot dogs. The Boston Dynamic ones, they kind of like.
We're famous for those ones, right? But there's like a company that like makes them that you can buy at a consumer level, like unitree go to I I don't know why I just like i've gotten Like really hooked on on wanting one really really bad and like obviously i'm gonna shove iris in it or like give iris control of it because like Of course, that's what i'm gonna do with it
On one hand, I feel like they're really expensive. Like it's a lot of money. But on the other hand,
I think it seems reasonable for what you're getting on the flip side. So it just feels like a lot of money. But I think it's probably worth your bang for your buck. I think the basic one that you can get an SDK access for is like.
$2,700, think. Because I don't know. I've got to look at it a little bit closer, because there's weird stipulations around SDK access and availability. So I have to dig into that a little bit deeper. I think there's an open source community SDK. But if you want to use the official SDK, that whole package ends up being closer to $8,000.
Chris Gmyr (08:57)
Well.
TJ MIller (09:21)
So I don't know. And I actually, I think I figured out where I got this from. Like there was like a TikTok or something I came across somewhere where somebody was like showing off their like hacked version of it. He's like, I figured out how to hack it. like I've made it do all this like extra stuff. And I don't know, man. Like.
Chris Gmyr (09:38)
Mm.
TJ MIller (09:44)
I think it would be so cool to play with. I have zero practical use for it. I just want to like play and experiment and like, and learn. And I think my son would, well, I think my son would get a total kick out of it. But he also, as soon as I told him I was going to like give Iris control of it, he like flipped out and he's like, dad, you can't do that. Like, do you want Terminator? And I'm like, dude, Iris is so...
Like, I'm not even worried.
Chris Gmyr (10:13)
Yeah, I mean, it's it's a little weird, I don't know, it'd probably be cool of like maybe not being like the same entity as Iris. But, oh, you could have like Iris's companion being like a virtual dog and then like your. AI bot dog companion is like a subset of that or something like that. like Iris doesn't control.
TJ MIller (10:37)
yeah.
Chris Gmyr (10:40)
like the dog, it's just like her like virtual companion, also like seeds the data it needs to be like your like physical dog companion of like, TJ is having a rough day or whatever. go over there and sit down so he can like pet you or something. I don't know. Something crazy like that.
TJ MIller (10:57)
Yeah.
Yeah, no, I could totally see that. Or like, you're just slapping like an MCP server together and like, I don't know. At least be able to like get data feed in and like be able to like read from it. It's got like cameras and LiDAR and speakers and yeah. I think, and I think it would just be cool to like.
Of course I'm gonna make it super vaporwave and I think it would be just fun as hell to take places. And just be weird. I don't know. It'd be fun to walk around downtown Detroit with my robot dog.
Chris Gmyr (11:28)
you
Yeah.
of like a little.
Yeah, have like a little app on your phone to like have it sit or like do tricks or something or like speak or whatever.
TJ MIller (11:44)
yeah, they have a whole controller. Like
you can buy a controller for it, you know, and like have, just have that walking around with it. It's got like a little like display screen on it. So you can like walk it around corners and things. yeah, I just think it'd be so much fun. Like there's just a plethora of ideas. like, look, I can put a leash on it and throw my, my roller blades on and we can just go like ripping around the subdivision. Like that'd be super cool.
Chris Gmyr (12:08)
Could it like a pet set your other host of animals that you have at the Miller Zoo?
TJ MIller (12:15)
Dude, it totally could. Like, it totally could. ⁓ Yeah, no, I just think it would be a really fun time. And my son's super into STEM, and so I know it would be interesting to him. But the cool thing is,
Chris Gmyr (12:19)
Like rough, get off the couch.
TJ MIller (12:32)
looking at this and being like well there's no way I can afford it right now and I also don't know how I'm gonna like come up with.
that amount of money where I don't feel guilty spending it on a robot dog for funsies. There's a Raspberry Pi project called PiDog and I think that's closer to the like 200, $250 price point and then I think you have to buy a Pi on top of that. So not too bad and I guess it's like really pretty capable too.
Chris Gmyr (12:44)
Yeah.
TJ MIller (13:09)
Um, I've, I've been poking around at it I think that that might be a, uh, happy medium, right? It'd be like a fun project because it comes in, like, it's a kit. And so like we assemble it. So it'd be like a fun project for me to do with the kiddo. Like we can assemble it, we can program it. Like, and, uh, I think, I think the pie dog project is even like already open claw enabled.
or capable, or people have already figured that out. So that's a solved problem. So if I wanted to toss Iris in that, or have Iris control that, or experiment with Iris as pet by proxy, I think that's probably the way to go. And it's not small, but it's also definitely not gigantic dog size.
Chris Gmyr (14:00)
Yeah.
TJ MIller (14:01)
But I think that seems a lot more feasible and like a really fun project. But gosh, do I really want one of these unitrees?
Chris Gmyr (14:11)
Yeah, the only thing that came to mind was like starting a GoFundMe for like TJ's like RoboDog support animal or something.
TJ MIller (14:21)
Gosh, that's totally what I'm gonna do with it, right?
Chris Gmyr (14:24)
You
TJ MIller (14:26)
I mean, already have like my mini poodle follows me everywhere. like, so I'm gonna have like my little, my little like 11 pound poodle and my giant robot dog just following me around the house.
Chris Gmyr (14:36)
Could you make like a satchel or something for the robo dog for your poodle to like get on its back and run around?
TJ MIller (14:43)
I'm totally... Yes. Just like a super Star Wars style. yeah. my gosh, that'd be so good. That'd be so good. I think I saw somewhere too that like people make like covers for them. And so someone has like dinosaur covers for these Unitree dogs. So you can make it look like a...
Chris Gmyr (14:45)
Ha ha ha ha.
That'd be hilarious.
TJ MIller (15:07)
like Triceratops. Like this is even better. Yeah. Mm-hmm. Yeah. I'm, I don't know when, but probably this summer I'm gonna end up doing like the Pie Dog project. then, yeah, maybe I'll do a, funsies, go fun me or something. Like help me get a robot dog.
Chris Gmyr (15:09)
Oh, that's hilarious. That's awesome.
That's funny. Nice. Well yeah, keep us posted on that. Looks like a pretty cool company and project. But yeah, let us know.
TJ MIller (15:35)
Yeah.
Yeah, Unitree's
got all sorts of like robots, not just like the dog. They've got like bipedal robots and all sorts of stuff. I don't know. Yeah, well, we'll see where that goes. But moving on from that, still on AI, I could share with you a little bit about like the recent project I'm working on at Luma, because it's got, I don't know, it's pretty neat.
So we have part of our learning platform, right? We have eNuggets, which are like these learning units. And as part of an eNugget or freestanding, you can have like a little forum. So.
We have the forums, forums have comments, comments have replies. We go, I think, up to three levels deep in replies. So what we wanted to do at an admin level, so not for the people using the forum, it's like the forum admins, or like the organization that owns the forums admins. And
what we wanted to do is create a summary of like here's like the summary across all of the comments. And then we wanted to do some like synthesis analysis on it. So we wanted to grab like, I guess taking a step back at like the architecture, then we can like talk about each stage and like what pieces it breaks apart.
And the reason the architecture is the way that it is is because you know, you can have a forum have 2000 15,000 comments, right? Like who knows like it can have Infinite number of like comments associated with it. And so there's no way you're going to be able to fit all of those comments
inside of a context window. And even if you could, you're gonna get into context window degradation, right? You're putting too much information inside the context window and that causes problems on its own. So we take a two-pass approach. So first we go through and process comments.
Chris Gmyr (17:56)
Mm-hmm.
TJ MIller (18:08)
in thread aware batches of 100 comments. And we do like, this is what we call the classification phase. Inside this phase, we do a few things, which we can dive into detail about. But the second pass is then the synthesis, right? So it takes a compressed view of that first pass. We do like all the classifications and then like,
it uses those classifications to then do like synthesis and like get you a snapshot of like everything at that point in time. Right? does that all follow? All right. So for each comment in the batch and we do this as structured output. So we give it a batch of a hundred up to a hundred comments.
And we get for each comment a structured output of sentiment. It's either positive, negative, neutral, mixed. We have concerning flag, which is like a balloon, like just true or false. Is this like, is this comment flagged for frustration, confusion, or negativity? Is the comment aberrant? Is it, know, just like off topic or mismatched in relativity?
you know, themes like.
in relation to like trucking and stuff, right? Safety awareness or training feedback. And then key quotes. like if there are like, I don't know, like any like key quotes that are like particularly insightful or ⁓ articulate, right? So that's, we do that in like thread aware batching.
And then what we get into is like the synthesis step. So we take a compressed view of all of those like first past results. We have like sentiment distribution, theme frequency and ranking. We list all the flagged comments, all the key quotes, and then like we add in all the forum metadata into the context.
So we end up with a pretty rich data set to then be able to provide back to the admins of like, yeah, here's like kind of your status across this whole conversation, which can be like really useful if you're trying to increase engagement in the forums,
you can get that insight into this forum's performing really well. Look at all of this engagement. The engagement's very positive. Or looking at this one, I'm like, oh, we maybe didn't approach this one very well because there's a lot of negative sentiment. maybe Joe is just always being an asshole in the forums. We need to pull him aside and talk to him because he's always coming up in these negative comments.
I think this is like such a, I don't know, I've been thinking a lot about.
I don't know, maybe ethical AI and like, I don't know, this is like super tangential, but like I've definitely been noticing externally to the development community, but absolutely with inside the development community, but maybe just not inside like my direct bubble, a growing anti-AI sentiment like.
in a pretty big way, right? we're, and I definitely share a lot of these concerns about data centers. We've got like a lot of uproar about that. A lot of like ethics and creativity, especially with generative artworks, which I definitely use generative, I use it for generating art, but I'm also not like ripping off other art styles.
Like I know that it's unethically trained on copyrighted material and everything, but I'm not trying to recreate this copyrighted material. When I'm doing stuff with AI artwork, for the most part, it's me trying to express my creativity in a way that I can't recreate on my own. And that, I think, is pretty gray area. But like...
It's caused me to... And then there's all the like, AI is gonna take my job, fears and fear mongering and massive misunderstandings and this AI bubble and like, so...
just kind of seeing that growing and like paying a little bit closer attention to that instead of like being my bubble where we're all like AI-pilled and like, you know, hacking and building like crazy. That's kind of caused me to think about my like approach with like building these like AI solutions and like, you know, thinking about things more of
assistive tools rather than like let's replace this job and like replace this the skill it's like no let's enhance that right it's it's additive it's enhancing it's you know a partner collaborative so i think this is like such a great one of those tools right where it all
exists fine without it, but like with it, gives such easy insight into making these types of discussions so much more useful and rich for the end user by like offering these insights to the admins that can then like act on act on this without having to like gather feedback. And so that also made me like really excited and like looking at the other things that I've built.
at Luma2 are also, very much like, they're tools to collaborate with and give more information to the user to then make a decision, like make a human to end then make that decision. We're not taking away decisions where like we're enriching data, we're enriching that experience, you know?
Chris Gmyr (24:00)
Yep, exactly. And like you said, it's a tool and enhancement to whatever the feature or site or application is, is not meant for a complete replacement. like when you were talking, this thought popped into my head of like, if we rewind way back to the beginning of like web dev, there was like basic HTML, right? Where everything was static. Everything was very much
like handmade push via FTP to the server. And when you wanted to change it, you would have to go in and make those manual edits and push that. Then little bits of like JavaScript and like Ajax was introduced. And it's like, well, now I can submit a form. And the whole page doesn't refresh. just like removes the form and says like, thank you for contacting me type of thing. Where like when I hover over a button, like it does some funky stuff or
you know, being able to play a video now and like all this stuff that started like sprinkling in these little enhancements to it. Compared to where we are now, where everything seems like it has to be like a full blown like SPA and JavaScript and interactivity like all over the place. And it's like, well, it doesn't really need to like we didn't need to replace the thing. We only needed to enhance the thing. And I think that's where.
hopefully will kind of even out in the AI world is like we don't want to fully replace humans because they are valuable for what we bring to the table, but for things that can offload some of the more like tedious work or connect the dots that it would take like us too long to do. But like there's a limit to that and like an outer boundary of where we.
don't want to push the humans out. Like, I don't think we've figured that out yet. It'd be really interesting to see how that all evolves because, I feel like it should be like an enhancement to our jobs and an enabler for the tools that we use and information that we want to gather. But I don't know, trying to replace, you know, me or someone else, you know, in the field or these other workers like, I don't know, I don't.
That's not a great way to go.
TJ MIller (26:17)
No, I mean, and that's also just not where AI is at. And like, I think there's just...
I think that's a big part of the bubble, right? Is like, AI can do some stuff really well, but like, yeah, I can do some crazy stuff with AI, but that's only because I know what I know. And I can build things really fast with AI because I know how to build things. I've been building things with Laravel for 12, 13 years or something, right? Like,
Chris Gmyr (26:39)
Mm-hmm.
TJ MIller (26:51)
It's my human knowledge and experience that like empowers that stuff to happen. So, I don't know, I've always been a big proponent of like keeping human in the loop because...
that's important, like who's gonna know better really and like, I don't know. I'm gonna continue to build things that way and like, and keeping humans in the loop and like not, you know, automating that away and like, it's just, it's not.
It's not feasible. And also not ethical at this point, as far as I'm concerned either. So I don't know. So the system's really, I don't know, it was really cool. It was really interesting to build and figure out an architecture that could handle processing thousands of comments and still kind of come up with a
Chris Gmyr (27:27)
Yeah, yeah, totally.
TJ MIller (27:44)
high level summary of everything that still maintains a reasonable level of fidelity in the system. And so it was cool because we were building it as an API to integrate into the legacy system. And I'm like, dude, I need a UI to visualize this and debug this now before we integrate the API into the legacy system. And so.
quad slapped together like the perfect admin panel for it, like right off the rip. And I was like, this is like, this is sick. It's got everything I needed. And then what we did is we also, with all of this stuff that I build at Luma, for every AI feature I build or like a pipeline, I'm also building an evaluation system, like an LLM is judge evaluation system to go along with that.
and that's important to, it's important for iteration, but then it's also important for quality control as you make adjustments moving forward. So ⁓ it's like I ran an analysis, then ran the evaluation. Fed that evaluation back into Claude code with the analysis as context, and then had it iterate on everything until the evaluations were coming back consistent and clean.
Chris Gmyr (28:45)
Yep, 100%.
TJ MIller (29:03)
and then that evaluation then can like stand as your like gold standard moving forward, right? So we got clean evaluations consistently, great as we move, make improvements or adjustments moving forward, we know we can go back to that evaluation and that we had at one point in time.
everything working well. We upgrade a model or we want to maybe see if we can get away with Haiku instead of Sonnet. We can use that evaluation to evaluate the quality coming out. So I think those are super, super important if you're building production AI systems, like building evaluations as first-class citizens.
Chris Gmyr (29:30)
Mm-hmm.
Yeah, yeah, totally. Nice.
TJ MIller (29:49)
Yeah, yeah. And API turned out super clean. I was super stoked with how it all came together. Yeah.
Chris Gmyr (29:58)
Yeah. And I like the,
⁓ the multi-level synthesis and that's what I have worked into, like a handful of my, workflows, I guess. so like I wrote about this in the blog about like, processing, you know, data from like social sites, this podcast, like the website, and grabbing content from like my vault and
everywhere else that I inject into it. it's kind of like a layer, like grab the raw data, do some larger synthesis on top of that, and then distill it down until actionable insights or next steps from there. I do the same thing for a daily and weekly and monthly review at work. every day at the end of the day, I do a review-daily skill that pulls in like,
GitHub activity, Jira, Confluence activity, any notes within my meetings that I have saved, all this stuff in it. It fills out my daily work journal. And then at the end of the week, it aggregates all those journals into a distilled version of a weekly review. And then the monthly review goes through the weekly reviews and distills that down into
themes, like signals, of like rising signals, like falling signals, things like that. Things that I can add to like a brag doc or like next promo cycle type of thing. But it's really important, like you pointed out, to do that leveling because it's like way too much if I just load it in like a month's worth of raw.
meeting logs and get logs and stuff like that. So you got to do like the smaller processes and get it into like a manageable space and amount of data. And if you need like additional nuance or to look into things like at least for like obsidian and notes, it's like you can link to everything below. So then you can ask the AI to go back and find the nuance of that, which, like you said, does the same thing for the threading and
TJ MIller (31:38)
Mm-hmm.
Chris Gmyr (32:04)
everything in the forum that you can link back to. So if you need to expand on anything, you just go resync this one thread in this one forum type of thing, and then you can get additional data from that. But yeah, this whole synthesizing of data and distilling it down into a more manageable working file and for context is something I use all the time.
I think it's a great pattern.
TJ MIller (32:28)
Yeah, I think you just have to be...
aware that that's a lossy, right? Like you, it's just, you have to be mindful of that. Like every time you distill it, you're losing a little bit. You're losing a little bit every time you do it. So like it's just, and that's not, that's the point. Like it's also the point, right? So you just have to be like smart.
Chris Gmyr (32:37)
There you
TJ MIller (32:56)
about how you're doing that and being aware of when and where it's happening. think that's where those evaluations come in really handy to you is like, we're missing these details, right? So it's not maybe just summarizing, but getting these structured outputs of multiple fields. You can still keep.
level of fidelity in there. Instead of like multiple paragraphs, you can now have like a small structured object representing like still pretty rich data that you're then using as your distilled data set later on. It's just something you have to like really consider as part of that process.
Chris Gmyr (33:32)
Yeah.
And maybe there'll be better technologies or options to do that, because at least on my side, it's all Markdown files. So it's just a lot of data. And it's hard to scan through those eventually. So maybe Vectorize Database or some other sort of tech layer on top of that could help with getting it a little bit more lossless. But like you said, that's kind of the point, too. But you've got to be careful of which
which layer you're like fuller analysis on because if it gets too lossy then you're going to miss some of those details in there or it could point you in a different direction then that's actually accurate.
TJ MIller (34:11)
And it's, really have to take a step back and think about what you're asking of that data, right? Like that first pass, we're coming out with five different data points that are really pretty rich. And like that's the data we want out, like specifically in the distillation process later on in the pipeline.
So like, just, you want to get careful and you want to be careful and be specific about like what you're asking of that data and like really kind of thinking about that upfront. And like, you can distill content different ways for different purposes too, right? Like you don't have to like this, I only distill this once and then use it. Like if you're going to maybe use it in a different pipeline, maybe you want to distill it in a different way and you're asking something different of that data.
Chris Gmyr (34:53)
Mm-hmm.
TJ MIller (35:05)
That's super valid as well.
Chris Gmyr (35:08)
Yep, exactly. Nice. Well, that sounds awesome.
TJ MIller (35:10)
So, yeah.
Chris Gmyr (35:11)
Sweet. Well, I know we're getting a little up on time. Do you want to wrap up for today? Sweet. So thanks for listening to this Lately Caffeinated podcast. Show notes and all the links and social channels are down below and also available at slightlycaffeinated.fm. If you have any questions for us or have a content suggestion, go to the Ask a Question page on our site and we'll feature it in an upcoming episode. Thank you all for listening and we'll catch you next week.
TJ MIller (35:15)
Yeah, we wrap up there.
See ya.