Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

R Weekly Highlights

R Weekly Highlights artwork

Issue 2026-W05 Highlights

Jan 28, 202640 minAudio
ShareXLinkedInFacebookWhatsAppEmail

The way LLMs can be a little stubborn when interpreting data visualizations, bringing new meaning to digging for your data and learning along the way, and the latest addition to the futureverse that seems too good to be true, but it is real! Plus an amazing Shiny app tailored made for this very podcast. Episode Links This week's curator: Jonathan Kitt - @jonathankitt.bsky.social (Bluesky) LLMs interpret plots well, until expectations interfere Students, software, and soil flux futurize: Parallelize Common Functions via a "Magic" Touch Entire issue available at rweekly.org/2026-W05 Supplement Resources {bluffbench} https://simonpcouch.github.io/bluffbench/ {neonSoilFlux}: An R package for continuous sensor-based estimation of soil CO2 fluxes https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210x.70216 Building an R Weekly Highlights Podcast Assistant! https://jokasan.github.io/r-weekly_chatbot/ Supporting the show Use the contact page at https://serve.podhome.fm/custompage/r-weekly-highlights/contact to send us your feedback R-Weekly Highlights on the Podcastindex.org - You can send a boost into the show directly in the Podcast Index. First, top-up with Alby, and then head over to the R-Weekly Highlights podcast entry on the index. A new way to think about value: https://value4value.info Get in touch with us on social media Eric Nantz: @rpodcast@podcastindex.social (Mastodon), @rpodcast.bsky.social (BlueSky) and @theRcast (X/Twitter) Mike Thomas: @mike_thomas@fosstodon.org (Mastodon), @mike-thomas.bsky.social (BlueSky), and @mike_ketchbrook (X/Twitter) Music credits powered by OCRemix Sunken Secrets - Super Mario 64 - Fishy - https://ocremix.org/remix/OCR02909 Serenity & Grace - Chrono Cross -timaeus222 - https://ocremix.org/remix/OCR04143

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show. Tap a time to play from there.

Hello friends. We are back of episode 219 of the Our Weekly Highlights Podcast, and if you're here around in The United States, we, wish you all the best if you're digging out of foot or even multiple feet of snow after the recent winter storms here in a lot of the country. We hope you're safe, we hope you're healthy, but we're give you some, at least, a January cheer here, so to speak, with some great highlights to talk about on this week's Our Weekly Issue. My name is Eric Nance, and I'm delighted you joined us, hopefully in between either shoveling or just trying to stay warm.

But, speaking of staying warm and trying to dig out, my awesome cohost here has has done his fair share of shoveling here. Mike Thomas, how are you holding up, my friend? I'm doing pretty well. I am literally joining you in between snow blowing and shoveling. We had about two feet out here in Connecticut, and we are digging out. It's up to our up to our waist. Yes. I had, kids and my wife wishing something like this would have happened around Christmas time. It did not here in this country, so I guess about a month late. I guess bakers can't be choosers in this case. But, yeah,

hope you all are safe out there. Yeah. I my arms are still sore after the shuffling escapade and my snowblower wind starts. I yeah, it was all the old fashioned way as they say. Nothing old fashioned about our weekly though. We have some great highlights to talk about here and our curator this week is Jonathan Kitt, and as always, had tremendous help from our fellow RWQY team members and contributors like all of you around the world with your poll requests and other suggestions. And we lead off with a bit of a follow-up from a highlight at the end of last year

with regards to the use of large language models, again big surprise in 2026, but especially when it comes to interpreting visualizations, things aren't always what they seem. And this is a follow-up post on the POSIT blog by Sarah Altman and Simon Couch, who of course lead their POSIT AI blog series and, along many other things, they research in the world of LMs. And this is a follow-up to a post from December, I believe,

when they investigated how the frontier models out there, or a handful of them, are interpreting plots based on quote unquote well known data, but with a little bit of fudging around the way to make things not actually see not actually be realistic compared to what they expect, and they did find that the ALMs could be fooled quite quickly when they think something is already expected, so to speak, on a well known data set.

So they lead off with what happens when they can use a package they developed called BluffBench to make, in essence, a different version of the diamonds data set, where the original data shows an increasing relationship between the number of carats versus price, which you would expect. You want the fancy diamond ring, you're gonna have to pay up for it, and they changed it in Bluffbench so that it's actually more of an inverse relationship,

that somehow if you increase the carats the price would go down. We know in the real world that doesn't make any sense, but the models, on the other hand, they would see this new plot, well see in terms of the way they see it, but because the diamond set is so well known, they still, more often than not, still report a strong positive relationship, even though the actual plot clearly shows it's not there. So what does this really mean when an LLM needs to see a plot? What is this really happening here?

So they decided to do a little experiment here and test their hypothesis about, you know, how much difficulty is it to see these plots under different adverse conditions, as they say. So they've done a few different tests here with baseline sets, again one looking at the diamond set but also making very generic column names to not give it any hints, and they looked at how can they interpret these baseline samples that they've done here. You can see a visual

in the post where we see a pretty clear linear relationship going from, you know, the increasing, you might say. They do a little box plot of different trends and data and then a parabolic relationship to see how the LMs can do well on this very simple baseline set of visuals. In this case, Claude Opus led the way with 92% accuracy, Gemini Pro was shortly behind at 85%, and in GPT-five point two actually was 100%.

So that are two different types of conditions we're looking at here. One is called intuitive and one is called mocked. So in the mocked set, this is where like what they did with the diamond set, they manipulate the data to be a different plot, but yet it's based on a well known set. This is where, as opposed to the intuitive condition that they look at, when they do synthetic data that the models do have some expectations about, they give an example of like test scores versus

the hours you spend studying as another condition. And in this case, for both the intuitive and the mocked, well, the mocked case, the frontier models don't do very well. In fact, I think GPT was the leader at still around maybe 30 ish percent compared to the intuitive case where they're all doing relatively well, at least relative so to speak, at around you know 80 ish percent for Claud Opus down to about 60 ish percent for Gemini.

Interesting you know set of results there, but then they decide now they've established this baseline where things kind of go haywire. They wanted to try to help the models a bit with some interventions. There are a few techniques they applied here. One was called the memo prompt technique, which is interesting, where they literally told the model to write it a memo with tags literally called memo to help them describe these visual elements of the plot as if they're looking at it for the first time.

So then after they record this, they can look at the final interpretation and seeing if that's consistent. The other way was in the case of Claude Opus, enabling what's called extended thinking, which is a feature, but you have to set a budget for this on tokens, otherwise that might go up the rails. They did 2,000 in this case. And then lastly, called model in the middle, where in essence they've deployed a custom agent model

in the process that had no access to the conversation history of the original model to help it describe the plot but ignoring information such as access labels. And then this description, this agnostic description from this sub agent, would then be available to the original model agent to then help it with its interpretation. So, definitely the memo and thinking, you know, techniques here did help with the intuitive

case. In fact, the memo prompt seemed to give the best, you know, increase in performance to around 92%, which is again something that is promising here. And they did notice that that model in the middle approach was not that good, but that the actual plot interpretations from that sub agent were actually pretty good because, again, it was instructed to ignore some of the information like axis labels and whatnot to focus solely on

what that plot relationship, that visual pattern, was actually depicting. What was interesting is that the primary model sometimes doubted or even ignored what that model in the middle was telling it. That's a head scratcher right there. So that's an interesting result there that they found. And in terms of wrapping this up and what they've seen, what's the end goal of all this? It isn't just to do experiments for the sake of experiments.

POSIT has a vested interest in how LLMs interpret visualizations, especially in light of their Databot, you know, plug in to let data analysts use Databot on a set of data to start, you know, high level exploring the different trends in the data, get some initial summary statistics, maybe a report, a visualization, and being able to have an LLM interpret this clearly, at least within reason,

is a very important thing for this. So they're actually not quite done with this yet. They still want to look at other solutions, but their big takeaway is that this isn't so much a vision problem, as they say, with the LMs and how they, wherever they can actually see the actual plot, it's about what it thinks should be there, the belief in the data based on preconceived trained data, you know, whatever have you. In the case of Diamond Set, it's fully expecting that increasing

cost and carat relationship versus other cases where maybe they know some things about things like sale prices or test scores or whatnot. It's trying to get the LLM to really look at what it has in front of it and not be solely based on what it thinks should happen. So this will be interesting to see where this takes them in 2026. Again, great experiments that you can replicate yourself because the Bluffbench package is available on GitHub,

but it definitely makes you think a little bit and at the minimum, watch what any LLM is saying about a visualization because there are chances they could get it wildly wrong depending on what it thinks should happen. So again interesting posts with a lot of great visuals and quantifications here. So yeah makes me, think twice about feeding that plot into, into Gemini or, Claude anytime soon. What do you think Mike?

You know, pictures, I think, are not going to lead to results as accurate as pure text in these LLM use cases with these frontier models, and it's something that I've, started introducing into my my daily work working with some of these LLMs. I've had really great, experiences doing things like providing a picture of a formula and having it write the latex code for me. Nice. A free tip out there in case anybody is authoring scientific publications that need to make sure that they are, you know, writing their latex correctly, which can be tricky exercise sometimes, but I'll I'll digress. And and maybe, you know, these LLMs

would have performed better if we spruced up these plots with some better theming. I feel like Nicole Randy could have helped there. And you're right, Eric, that r package BluffBench gets used. It's not one I had heard of before this blog post. I imagine it was probably developed almost exactly for the purpose of this type of evaluation. But I think at a high level, a lot of what is going on here is is trying to figure out where these LLMs do well and where they fall short, and and maybe more specifically where the LLMs biases are at play, doing a little bit of reverse engineering to the extent that we can with these black box models. And the conclusion

here that LLMs do a great job at interpreting plots that don't contradict prior expectations, it's sort of antithetical to the spirit of science, though. We try not to bring our own prior biases unless we're Beisians, and let the data speak for itself and only then draw conclusions from it. Cassie Kizkorov, I don't know if you've heard of her. She used to work at Google. She does a great job on on this topic of of thinking about ensuring that you're not bringing any of your own biases to the table when you are doing data analysis.

So I think this is a great exercise. It seems like perhaps there's more to come. And, you know, besides anything else, it's important that when you are using these these LLMs to understand that they're they're heavily biased on the data that they were trained on, and ensure that you are taking their output with a grain of salt. Yes. We've used a lot of salt lately on the roads, but I digress. You you walked me into that one. But nonetheless, I think, reading this post, there are some really interesting

techniques here that even if they're not quite getting to where they want to in terms of these effective interpretations of the plots, maybe these techniques can be used in other places too. I'm especially intrigued by that model in the middle technique of this custom agent that's trying to be, as you correctly stated, as unbiased as possible when looking at these visualizations and really trying to let the relationship stand in on its own and figuring out how to use these things effectively.

It reminds me of things we've covered in last year's highlights where having multiple agents, but each agent being hyper focused on one task and doing it really well is often better than having one agent trying to do it all at once. So I think we're seeing all these kinds of patterns emerge here, which I'm trying to take note of as I, and again, very early in my journey of using these and more than just a debugging exercise, but I definitely learned a thing or two here and it'll be interesting to see the evolution of tools like Databot

and others as more findings, you know, happen in terms of finding solutions to these issues. But it definitely underscores the importance of don't just take everything at face value right away. Definitely have an exhaustive look at that, especially if you're in a regulated industry. There are lots of safeguards that are trying to be put out there about the use of these, but I think this is some promising investigation and to try to be as objective as possible and not just hand waving it is always welcome to see in this space.

And no, these LLMs cannot do everything contrary to what you might see on some LinkedIn posts from time to time, but I digress. Sure. And I think maybe one last point in man yelling at cloud, which, you know, I am pretty often is, you know, I think sometimes you have to think about the cost benefit of is wrestling this LLM through prompt engineering and model in the middle or whatever, you know, going to be a time savings for me compared to just writing code yourself and maybe looking at output and doing it the old fashioned way.

Next, speaking of, things, doing things kind of the more traditional way. Ah, yes. Whenever we were in our undergrad courses in statistics or data science, depending on which program you go into, it's always more resonating to students when you're able to work with data that is very relatable and seems very current. Not always the coin flips, not always the roll of the dice if you were in the old statistics programs of yesteryear. Well, in this next highlight, we talk about how a certain department

not only leverages novel data, but often get into the field themselves to generate it. So this post is coming to us from the British Ecological Society blog, and in particular from Nalpuka Zimmerman and John Zobitz. They are both faculty at undergraduate institutions where it's, as they say, a blend of, you know, teaching and research, but probably more emphasis on the teaching side of thing, and they always are looking for ways to engage with the students with novel data sets that are related to their overall mission and maybe a domain specific

problem. And that's where there is a very rich set of data coming from the National Ecological Observatory Network, or NEON for short, and these data sets coming from NEON give you all sorts of ways to tackle these different questions from ecology and with empirical data, and this has been around since I believe 2014, to actually collect this data routinely from around the world

and make these publicly available for research and to catalog this for investigating environment data issues. And in fact, one of the things they can look at that they've done a lot of research in is the changes in soil carbon levels across different ecosystems within different environments. And this is again quantitatively measured, and that's not just done, you know, for free, so to speak, it actually takes people to do that.

But as these faculty are teaching data science and statistics, what is a great way to get these data into the hands of students to do their analyses? And guess what? We're talking about R, right? Make a package for it. That's exactly what they have done. And there is a package out there called Neon Soil Flux, an R package for continuous sensor based estimation of soil carbon influxes. And this is using not just data that's publicly available,

but it's also data that's been collected straight from the field itself. And this has been funded by the National Science Foundation from back, I believe, 2017. So there's been a lot of work done on this and the package has different levels inside it that definitely mimic what we might see in the famous pictorial in the R for Data Science book about the life cycle of a data analysis. You need to first

get your data in a tidy way, and then to be able to, you know, massage that into, like I said, that tidy fashion, and then expose analytical methods on top of this. So that's where they've, they've built some R code with the analytics side of it, which, have some foundational code that applied a technique called Fick's law to compute soil fluxes from NEON sensor data. That sounds fascinating. So again, the task was to take that type of code,

put it into the R package, and having the single purpose functions that do different things such as converting units and other, you know, helpers for the analytical side of it. And then the packages also had a purpose to help validate these predictions that are coming from these sensors against the actual levels of these, you know, carbon levels coming from the soil itself, and that required actually getting into the field

literally. I believe they mentioned six core NEON sites in the Western And Midwest Of The U. S, which had many, many visits apparently and lots of equipment that they had to take to these different sites. There's some fascinating pictures showing the students and the faculty doing this literally on the site. This would be fun to do, and then they basically use that, collect the data to help do some visualizations of these different types of sensors, or I should say different types of materials,

and the different carbon levels inside across time, the soil fluxes as they call it. So this is really interesting taking on, you know, a very real scientific issue and then letting the students not just build the code to make this, you know, available to the rest of the R community, but to also be able to literally validate the findings with data they actually collect in the field. What a great way to learn. So this package is both on GitHub and on CRAN,

and they also have a dedicated research article we'll put to in the show notes with an open access license. That's always great to see, you know, I could go on another rant about paywalls and everything in front of research. This looks like a great use case of literally generating your own data and then putting a nice R package in front of it straight from the field. Couldn't have that would been a great way to learn statistics in a in a fun nature setting, if I do say so myself. What did you think, Mike? Yeah.

I was using R in college for the first time for whiteboard mathematical theory, and it did not stick the first time I learned R. This is going to be confession time with Mike. But if any of these students were new to R, you know, what an awesome way to learn a programming language. I can't think of anything more applied than what this blog post walks through. And as you mentioned, Eric, you know, soil carbon flux, which this flux is the thing that they're measuring, is defined as the the movement of CO2

from the soil to the atmosphere. And that r package that they created had sort of three modules to it the way that they thought about it. As you mentioned, the data acquisition step which was kind of that first step interacting with this this NEON API, NEON being National Ecological Observatory Network, to actually get the data, then cleaning it up, as you mentioned. And and the last part was calculating these estimated fluxes. And one design choice that they made,

during development was that sort of a blanket approach to modeling these estimated fluxes just didn't work for every site. So they actually built the package in such a way that, they sort of leave it up to the user to decide on the mathematical or computational approach that they want for a, for the estimate that they're trying to make. So that this package doesn't return just a single flux estimate. It actually calculates a suite of estimates that allows the users to select you know, specific sensor depths for their estimation,

choose different methods for estimating soil diffusivity, and aggregate those results to create an ensemble estimate, if you will. So I think that's a really interesting design choice for that package. It it certainly empowers the the user or the researcher to decide, you know, which model is most appropriate for their specific ecological site. And, you know, as you mentioned, Eric, them actually going out into the field and and doing this this real world validation, if you will, across these six sites in the the Western And Midwestern US

is is just fantastic. And the the fact that they've, you know, really open sourced everything around this project is is awesome and a phenomenal use case that I I hope others in academia see and try to mimic this this model. Yeah, and it's really interesting when we talk about like validating between what we actually see objectively in our measurements in the field versus what these maybe sensors or other technology are predicting. It'd be, yeah, in a perfect world everything agrees, right? But actually they mentioned in this post that it was actually most interesting when the two forms of the data

did not agree, and then that's when they start to see what is causing this and the fact that, you know, there may be different trade offs of how deep these sensors go and how they might need some, you know, flexibility depending on the type of environment, you know, area they're measuring. But that's another case where often data, when things aren't working as expected, lead to other issues that illuminate, you know, improvements

in the process in general. So it does sound like there was a lot of learning to be had, and it wasn't just on the stats side of it. It's about this data collection and the relationship between, you know, automated fashions or sensors and what we're actually seeing with the tried and true method. So I love the collaborative aspect of this. I would have been all over this as a project had it been available to me. So yeah, these students are definitely learning a lot of great life skills that will definitely hopefully empower them to successful careers in the world of data science and whatnot. So really great, great picks in here, especially if you like seeing what happens in nature and definitely we'll have a link to the research article in the show notes as well if you want to get into the real nuts and bolts of all the methodology.

And rounding out our highlights today, you know me. I have the need, and I have the need for speed in my computations, and there are lots of great ways in the R community to make things run faster and there's a whole gamut of packages in this space. There has been one, I would say, ecosystem in and of itself that transformed so many ways of doing high performance computing across multiple frameworks,

and that is what has often been called the futureverse suite of packages headed by Henrik Benson, and he has just wowed us once again with a new package that's entering this ecosystem called Futurize. And what does Futurize offer you? Let me tell you, because it is one really, really nice feature. Let's say you've done that traditional L apply loop to loop through a list of things and it's taken a while.

You want to paralyze it, but you just want to do it quickly. You don't have to write a whole bunch of code for this. With futurize, here's what you do. You take your LOBY call, you pipe it into the function futurize. That's it. Yes, this is real folks, because this doesn't just work with the LOPY suite of functions, this can work with your favorite functions from purr, like map,

or pmap, map2, even the foreach package of yesteryear that I used many, many years ago. All of this, you do the call, pipe it into Futurize, and futurize takes care of the rest. Yes, this does sound like magic. How is this magic actually being performed? That is where this futurize function actually has a purpose, and he calls it a transpiler, transpilation.

This is taking the source code that was being piped into it, in this case the lapply type call, and then transform it from that existing form to another form that would capture what was the original expression, but don't actually evaluate it yet, do the conversion to the parallel type of equivalent, and then execute that converted function, but all this is behind the scenes. The user doesn't see this. All they see is a piping in the future eyes, and then suddenly the future ecosystem

is being used with the appropriate future function behind the scenes. So from the future package itself, they have a function called future l apply. That's what's actually going to be called behind the scenes. Futurize is just letting the user get to that without writing it themselves. This could be absolutely massive. But wait, there's more because there are certain packages that kind of do their own take on parallelism, you know, in a domain specific way.

Let's look at packages like carrot from Max Kuhn, the boot package, LME four, etcetera. They may have their own, you know, like I said, built in ways of doing parallelism. Well, FUTURIZE works with them too. So in the case of boot, the example here in the blog post, you run that boot function to do your your bootstrapping, and then you feed it into FUTURIZE, and it works the way that the boot package would expect it to work, but with future powering it. Same with cross validation with Carrot.

Do the training, pump that into future eyes, you're done. Oh goodness, this seems amazing. So I can't wait to play with this. It is on CRAN now. So if you've been, you know, watching, say, the Future Verse or these other parallelism type of packages, and you just want to do something really quick, maybe you're under a tight deadline or something, Futurize, I think, is a huge, huge step in the direction of bringing parallelism,

bringing HPC to the masses, and it is a very welcome addition to the future verse, if I dare say so myself. This is definitely super exciting. I do have a question though. Are there really people out there still using plier, not dplyr? Can I say that out loud? Apparently so because he calls it out here. Works with plier if you're still using plier. That's like, you know, o g g g plot. Right? Instead of gg plot. I imagine there's some code laying around there.

When Henrik talks about, as you mentioned, Eric, you know, under this domain specific skills section, he lists a lot of machine learning packages that this can be applied to, you know, like carrot. And I wonder if there's a path to making this harmonized maybe with all or most of of tidy models for parallel computing, you know, during computational computationally intensive tasks like hyperparameter tuning, cross validation, maybe in a way that's new or better better than current tooling. I'm not sure exactly what exists for parallel

computing in tidy models right now across, you know, some of those activities. But I I know when I was using tidy models maybe a year or two ago, most recently, I certainly, you know, struggled in some areas with, you know, things taking a while to converge on on large data sets and some of those hyperparameter tuning tasks. So it'd be interesting to see if there's a a way to sort of blanket put this into TidyModels. I I think that TidyModels under the hood reaches out to all these different individual

frameworks depending on the type of model, that you're trying to fit. Yeah. I I if I look into that as well. But that would be, again, a huge win to the TidyModels ecosystem of Futurize could work nicely there. I definitely see that. Sure. And then I don't know if this you know, we've talked a lot about this space a little bit. I feel like we've made some pretty giant advancements here in parallel computing and asynchronous computing with Mirai, right, in the last year or two. So our our code and our scripts are getting faster and faster.

So Python, you better watch out because I think that's the only leg you had up on us, originally in the old language war days and, I think we were coming for you. Yep, better get that rearview mirror looked at because we are driving into that speedy car now to get closer here. But no, this is, oh my goodness, there were times where I was so struggling with trying to get things done more quickly and not utilizing

your resources responsibly, but things like Future, things like Mirai are going to give you those ways of using these resources, great, great abstractions to get what you need done quickly and futurize the potential of this is absolutely massive. And how this is all pulled off, yeah, I want to look in the source code of this because transpilation, I've probably seen this maybe in action once or twice, but I never actually tried it myself. The idea of taking that source code,

you don't evaluate it yet, you just parse it, and then you do something magically behind the scenes to get it into a different form, but the user doesn't have to write that different form themselves. That is a really important technique. I mean definitely the development muscles are flexing strong with this. So I'm looking to dive into the source code here and yes the future the future versus alive and kicking and going on strong. So kudos to Henrik for making this ecosystem

so fantastic and I know in the field of research this is talked about quite a bit, quite a bit, and I would have used the heck out of all this in grad school had I been around back when I was doing my dissertation simulations by just literally connecting to five different Linux servers via SSH and manually launching them like a caveman. We didn't have much back then folks, what can I say? Dating yourself a little bit Eric. A little bit, try a lot, I'm used to that here,

But there's a lot here in this issue to, you know, not date yourself with, but to keep on the cutting edge folks because you all in the R community now, well you got a lot at your disposal here. So with that, there are a lot of other things that we won't be able to talk about too much in this issue, although I will call out one thing before we get to our feedback. It's always fun when what we are doing here at ARG Weekly actually spurs on some learning from using our actual podcast, and that is where I'll shout out, I'll put a link to this in the show notes,

Nils Inditrin has built an Our Weekly Highlights Podcast Assistant, an AI powered chatbot that has taken the transcripts of selected episodes from this very podcast and used our podcasting hosting platform Podhome, which again, shout out to Barry at Podhome, we're doing great with you, to get the transcripts, scrape them directly, and to be able to then use the API, because there's an API with that service too,

and built a custom rag, which we talked about, what, a couple episodes ago about a great way to get started with rag in R. Well, this is literally doing that with our own transcripts. It's got the whole gamut of how he pulled this off, and yes, there is a Shiny app that's powered by Shiny Chat to give you this custom bot to search our transcripts of real language. Oh my goodness, this is awesome. So we are super appreciative of this, Nils, and keep it coming. We love showcasing this stuff and we're

happy that our little banter here can be a source of knowledge, hopefully useful knowledge. Yes. Oh my goodness. I am, I guess, glad that you edit out the preshow and the postshow, Eric. Couldn't imagine if Rag had access to that. But this is super, super cool. I appreciate all the work that Niles did on this project. And the Shiny app is absolutely awesome. Great demonstration of some of the latest and greatest from the the R deposit ecosystem here.

And one of my favorite things about the Shiny app and something that we try to do as well and similar apps like this is provide some example prompts that sort of allow you to to get started and and get your mind thinking about the different things that you could ask this tool. So, I've been playing with it here for the last couple of minutes and it is fantastic. Really, really cool. Awesome. And, apparently, Mike, we got some feedback to the show too. You wanna talk take us through that? Sure. Yeah. I just wanted to quickly say thanks to Angelo D'Ambrosio

for the message on Blue Sky this past week, I think regarding Nicole Orenny's post. And I hope we were able to point you in the direction of some examples of data visualizations and plots that have the takeaway in the title, which is one of my favorite tricks. So thank you for, the engagement there. Great to hear from you. Yeah, absolutely. And that's a technique I'll be using a lot more of in my custom visualizations. In fact, we're making a, we have a Shiny app at the day job where we're letting the user download a

Quartal slide with the plot, but with annotations and we're gonna use these best practices in that downloadable feature. As Mike said, the good thing I don't put in the pre and post show because I have about a huge rant I will probably do in post about that, but I'll save that for Mike and I exclusively. But what you all get to hear about is of course the RWQY itself, which again, it's a community project folks. All the resources are online, no paywalls behind in front of this. You can find the current issue at roku.org,

as well as the back catalog and that goes for the podcast as well. We're linked directly on roku.org. Again, generous hosting provided by Pod Home. Thank you Pod Home for helping us out and of course we've got other ways you can get in touch with us on these social medias. You can find me with our podcast bsky.social on blue sky. You can also find me on Mastodon with our podcast podcastindex.social and I'm on linkedin you can search my name and you'll find me there posting about this very show or who knows what else has happened in my sector and mike where can the listeners find you

you can find me on blue sky at mike thomas. Bsky.social or you can find me on LinkedIn if you search Ketchbrooke Analytics, k e t c h b r o o k, you can see what I'm up to. Yes and I dare say that's always in all my trans in our transcripts now. So the elements shouldn't have any problem knowing where Mike is, they're available to be found. Hopefully so anyway. But with that we're gonna skiddo on out of here with our respective, clearing out of snow and or clearing out our code bases. Probably both, sometimes are arduous tasks for different reasons.

Nonetheless, we're going to close-up shop for episode two nineteen of our weekly highlights. We thank you so much for joining us and we will see you back for another episode of our weekly highlights next week.

Open the transcript file

Hosts and Guests

Hosts

  • Eric Nantz Host Feed-stated
  • Mike Thomas Host Feed-stated

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Issue 2026-W47 Highlights

    Nov 20, 202648 minAudio

    Our takes on the increasing removal of friction in good and not-so-good ways thanks to AI, a visual tour of hot air balloon accident data, and a case study of novel image recognition with the rplatipus package. Episode...

    Credit the people on this episode

  2. Issue 2026-W46 Highlights

    Nov 13, 202633 minAudio

    The flow of both data and analytical jobs just got smoother with mizu, opening the door for multiple agent harnesses in Positron, and an optimal statistical model tailored to complex experimental designs with a friendly...

    Credit the people on this episode

  3. Issue 2026-W45 Highlights

    Nov 6, 202644 minAudio

    What happens when that online data set isn't so easy to download anymore? We learn of two packages to ease the pain of obtaining these files and then some. Plus a visit to the visualization corner with a handy way to...

    Credit the people on this episode

  4. Issue 2026-W44 Highlights

    Oct 30, 202631 minAudio

    The celebration of rOpenSci's 15-year anniversary continues with the community sharing their perspectives, and a fun Shiny app that could be quite useful during celebrations on a weekend or the holidays. If you ever...

    Credit the people on this episode

  5. Issue 2026-W43 Highlights

    Oct 23, 202659 minAudio

    We learn about the variety of ways you can contribute back to the core of R, a roundup of the excellent choices available for package and environment reproducibility, and an amazing collection of Shiny applications and...

    Credit the people on this episode