look man … that’s your fault for thinking you should ever service a bottom bracket

shit …

I need a book to listen to looking for something that is not like but would be liked by someone who liked books like the restraint of beasts three to see the king after the war catcher in the rye starship troopers gangsters of capitalism Norwegian wood Kafka on the shore expanse and dune etc

potato tacos with a beef stock au jus

Well don't I feel stupid

This snapper was massive

if you don’t love this I don’t know if we can be acquaintances

So, a bit of just info. The “diner” scenes that were shot for American splendor is the same diner I helped write yehuda moon before I went to work as a bike messenger

oh god, we’re doomed. fairly sure this is the parents asking the kids how to work the TV remote of this era

gravel racing is a bit odd, I’m not sure what to make of it

I’m going to go ride an 88.7 mile gravel race tomorrow … I’ll let you know if I spot any gravel

Beautiful

@rchung.bsky.social thinking about what lessons from the epidemiology world can be applied to testing questions using the OpenGC dataset pmc.ncbi.nlm.nih.gov/articles/PMC...

imagine being this bad at being wealthy

I’m trying real hard to figure out whether this biting irony within irony within irony … or just not self aware ?

So why don’t we just make a bunch of dynamical systems physiologically model agents and make them have them complete scripts based on goal, motivation, stop signs, hills, etc. To generate labeled data to train a foundational model. Like here, here’s your vocabulary then train on real unlabeled data?

I'm looking at tokenization to get on ideas from what it does well and what it does poorly. I'm liking the sound of motif encoding, which encodes runs of data with shared temporal "shapes" into tokens. It still however, when you look at the tokens, makes the data less human interpretable.

we should probably try this

don’t throw rocks in glass boxes ? www.sciencedirect.com/science/arti...

Intensity or total work?

Posting this over here as well: Getting coded up for a big durability study, anyone want in on it?

And last one for tonight, tired is tired. So prior FPC1 doesn't really bias a decrement in sprint or endurance. Prior FPC1 mainly brings down FPC1.

Whether the prior segment had a sprint or endurance bias doesn't seem to have much effect on change in FPC1. That is going hard whether shorter higher intensity or longer lower intensity doesn't seem to matter on overall fatigue.

Piloting this out on a small slice of my data (58 rides), what I can already show is that change in FPC1 (gain function) over the course of a ride is correlated with FPC1 of the prior segment. Makes sense, the harder you go, the more tired you get.

For the minute I'm going to pick back up the "phased-MMP" or what the kids call durability now. The first thing to do is all the background work to wrangle the data into different formats for analysis. Here I'm splitting each ride longer than an hour into thirds.

I'm putting the second by second FPCA to the side for a moment. I have an issue where FPC1 is a little insensitive to short maximal sprints in the middle of very low cruising pace. You can bring the sensitivity in by adding FPC2, but then that makes it over sensitive to suddenly chopping the power

And my FPC scores plotted into the FPC modes of variation plots:

I only have a couple of files in intervals.icu so my PD curve is a bit wonky, but here is my PD curve pulled in on top of the study data:

And if you have a garmin or intervals.icu account hit connect with SweatStack to pull in your own data to interact with the study results.

The app pulls in the abstract:

Amazing!: fpca.paperplayground.app From @aartgoossens.bsky.social a Paper Playground paperplayground.app implementation of the FPCA model. It's a great concept to interactively showcase research results. cc @drphilipskiba.bsky.social

Getting this figured out. I get a little nervous about needing to make algorithmic decision to get it to work right. This is what I'm looking for though that FPC1 is a really dominating gain function, with a small bias adjustment for sprint vs endurance and very small one for middle distance

Looking at "top" rolling powers and limiting it to just the top 20 for each time window FPC3 is the gain term, so that's reassuring that the inflection points are not a simple function of the FPC number

3 day weekend, get up, ride a little, code all day, you know it’ll be nice

farming in the desert …

and last one, good night

normalized

normalized

normalized

Last step to really unlock the context is the scale the FPC scores to 1.

I just did this ride tonight. Warm up then used a hill as a bit of a launch, settled in then longer climb for a hard steady effort with a kick just before failure 2/3 of the way up, tried to sprint over the top but didn't have much, recovered, a final short climb didn't have much, recovered sprint

warm up and one long irregularly paced effort.

Make sense of the mess. Warm up, irregularly paced work interval, irregularly paced recovery, a crescendo decrescendo work interval, and recovery.

It makes sense of the mess. (just ignore the fact of just how low my CMP is these days)

I think this method to contextualize power output is fairly convincing. For example take the sequence of intervals. FPC1 (blue) shows you that the block of intervals sustains RMP greater than CMP and FPC2 (red) indicates alternation between acute and chronic strain.

I think I like this. I need to look at some more rides but I think it works. Is it useful? TBD. But for now I need to go out and make some more data.

Flipping FPC2 (red) so that positive is sprint bias and negative is endurance bias makes it a little more intuitive. Now we can see that when we start an interval we are sprint bias then as we hold it longer it becomes endurance bias if we can hold it long enough: