tj mahr 🦃
LightNews LightNews
banner
tjmahr.com
tj mahr 🦃
@tjmahr.com
2.3K followers 970 following 4.2K posts
data scientist studying how kids learn to speak, dad, jump roper, bayesian, tjmahr.com
Posts Replies Media Videos
tjmahr.com
when I hear people ramming their snowblowers into the curb and I’m just nimbly shoveling along
December 2, 2025 at 12:59 AM
tjmahr.com
I’m joining the war on ms Rachel on the side of ms Rachel
December 1, 2025 at 11:26 PM
tjmahr.com
a giant pile of snow with a shovel next to it for scale
November 30, 2025 at 6:03 PM
tjmahr.com
getting back into regular jumping after messing up my ankle this spring
Screenshot of my jumprope stats: Nov 16k Oct 4k Sept 6k Aug 9k July 4k June 2k
November 29, 2025 at 12:12 AM
tjmahr.com
it’s usually my back that gets it though
November 23, 2025 at 8:07 PM
tjmahr.com
weighed jump ropes
My leg with a big mark on it
November 23, 2025 at 7:55 PM
tjmahr.com
🤦‍♂️
CMU pronunciation dictionary not having correct entries for "barack" barack B AA1 R AH0 K barack(2) B AE0 R AA1 K
November 21, 2025 at 4:58 PM
tjmahr.com
Oh I meant the knock-off snowshoe Siamese look. we had a few on our farm and always adopted them into house cats
One of my family’s cats
November 20, 2025 at 12:19 AM
tjmahr.com
it never ceases to amaze me that I can refer to duckdb functions in R expressions as though they were R functions and everything gets translated to SQL
dplyr code with duckdb functions in all caps tbl_files <- tbl(db, "csvs") |> # CAPS for duckdb functions mutate( # $speaker has directory names with the patterns: # - `[speaker]-[listener]` (listener specific data) # - `[speaker]` (default version) # so compute appropriate $speaker, $listener values dir_parts = speaker |> STRING_SPLIT("-"), speaker = dir_parts |> LIST_FIRST(), listener = dir_parts |> LIST_LAST(), listener = ifelse(listener == speaker, "default", listener) ) |> select(speaker, listener, file, label, logprob, model_name, model_commit)
November 19, 2025 at 7:47 PM
tjmahr.com
[1] "BEANIE" "YUCK" "MEANIE" "WHOOT" "BARACK" [6] "BINKY" "BOOP" "WEENIE" "A DWARF" "BANDAID" [11] "BB" "BEEDEE" "BLOOP" "BOOGER" "BOOGIE" [16] "BOOPED" "BOYT" "BUGGIE" "EGGLESS" "GLOCK" [21] "GUAC" "KIWIS" "MMM" "NUD" "OINK" [26] "PEDON" "PEEING" "PEELY" "POTTIE" "PUNCHBAG" [31] "RAWR" "SHROOM" "SNEET" "TV" "TWEEZE" [36] "UPREST" "VAPE" "WOAH" "YELLON" "YUKE" [41] "ZOW"
November 19, 2025 at 4:50 PM
tjmahr.com
words that listeners wrote that were not in the default pronunciation dictionary. (kids produce a set of 40 words, one of which is "beanie")
beanie 531, yuck 10, meanie 6, whoot 3, barack binky boop weenie 2
November 19, 2025 at 4:46 PM
tjmahr.com
thanks for the link. i got 19 referral hits from email yesterday and discovered the newsletter, ha
19 email referrals in one day
November 19, 2025 at 3:35 PM
tjmahr.com
getting ready to read in 30000 csv files
library(dplyr) db <- DBI::dbConnect(duckdb::duckdb("data.duckdb")) # Test on some folders before trying whole dataset test_dirs <- dir |> fs::dir_ls() |> head(20) list_of_globs <- paste0(test_dirs, "/*.csv") |> sprintf("'%s'", ..1 = _) |> paste0(collapse = ", ") |> sprintf("[%s]", ..1 = _) # https://duckdb.org/docs/stable/data/csv/overview#parameters # https://duckdb.org/docs/stable/data/multiple_files/overview q <- glue::glue(" read_csv( {list_of_globs}, FILENAME = TRUE, header = TRUE, delim = ',', skip = 0, quote = '\"', escape = '\"', auto_detect = FALSE, columns = struct_pack( file := 'VARCHAR', speaker := 'VARCHAR', frame := 'INTEGER', model_name := 'VARCHAR', model_commit := 'VARCHAR', run_timestamp := 'VARCHAR', label := 'VARCHAR', logprob := 'DOUBLE' ) )") # This just queries the pile of csvs # it's fast because it's lazy (just enough to print a few rows) tbl(db, q) DBI::dbListTables(db) # This would actually story the query results in the db # it's slower because it's reading all the data tbl(db, q) |> compute(name = "csvs", temporary = FALSE) DBI::dbListTables(db)
November 18, 2025 at 9:38 PM
tjmahr.com
what i wrote a long time in 2009. i have not evolved in my thinking very much lol
It just sounds metallic. That’s all I can think. It’s like a trick–the thin production and breakneck speed means I have it to play it on my headphones and crank it up. It sounds incredible. Grant Hart screams an indecipherable ghost story. The melody pops back in: “she’s the girl who lives on heaven hill”. The guitar solo flies right past us. What do you mean this is hardcore? This is loud and fast rock music, but it sounds like nothing else in the world.
November 17, 2025 at 9:32 PM
tjmahr.com
yes that’s what happened today www.axios.com/2025/11/12/n...
Epstein emails with references to Trump On Nov. 12, House Democrats released emails suggesting Trump knew more about Epstein's conduct. • In a 2011 email, Epstein alleged that an individual (whose name was redacted) spent hours with Trump at Epstein's house. • In a 2019 email, Epstein had a line that said Trump "knew about the girls," an apparent reference to the president's past claim that he banned Epstein from his Mar-a-Lago club for allegedly approaching young women who worked there. 20,000 pages from Epstein estate In response to the Democrats' released emails on Wednesday, Republicans shared 20,000 pages related to the Epstein probe.
November 13, 2025 at 12:14 AM
tjmahr.com
library(ggplot2)
library(mgcv)
ggplot(mtcars) +
aes(x = hp, y = disp) +
geom_point() +
stat_smooth(
aes(color = "vanilla gam"),
method = "gam"
) +
stat_smooth(
aes(color = "monotonic constraint"),
method = mgcv::scasm,
formula = y ~ s(x, bs = "sc", xt = c("m+"))
)
plot produced with the new scasm() model in ggplot2
November 12, 2025 at 8:43 PM
tjmahr.com
Me holding the Bourdain Reader
November 7, 2025 at 12:48 AM
tjmahr.com
oh come on i wasnt even doing anything weird this time
Fatal error (Not Responding) Praat will crash.
November 3, 2025 at 7:36 PM
tjmahr.com
i can never remember how pROC::roc() does things so my wrapper includes other options and wordy errors
R console output > validate_roc_direction("c", auto_allowed = TRUE) Error in `validate_roc_direction()`: ! Provide a valid `direction` • For control < case, use "<", "case-high", or "control-low" • For control > case, use ">", "case-low", or "control-high" • To guess the direction, use "auto" Run `rlang::last_trace()` to see where the error occurred. > validate_roc_direction("c", auto_allowed = FALSE) Error in `validate_roc_direction()`: ! Provide a valid `direction` • For control < case, use "<", "case-high", or "control-low" • For control > case, use ">", "case-low", or "control-high" Run `rlang::last_trace()` to see where the error occurred. validate_roc_direction <- function(direction, auto_allowed) { case_low <- c(">", "case-low", "control-high") case_high <- c("<", "case-high", "control-low") auto <- if (auto_allowed) c("auto") else character(0) valid_directions <- c(case_low, case_high, auto) bad_direction <- !is.element(direction, valid_directions) if (is.null(direction) || isTRUE(bad_direction)) { style_or <- list("vec-last" = ", or ", "vec-sep2" = " or ") ch <- cli::cli_vec(case_high, style = style_or) cl <- cli::cli_vec(case_low, style = style_or) au <- cli::cli_vec(auto, style = style_or) bullet3 <- if (auto_allowed) { c("*" = "To guess the direction, use {.val {auto}}") } else { character(0) } cli::cli_abort(c( "Provide a valid {.arg direction}", "*" = "For control < case, use {.val {ch}}", "*" = "For control > case, use {.val {cl}}", bullet3 )) } if (direction %in% case_low) { direction <- case_low[1] } else if (direction %in% case_high) { direction <- case_high[1] } direction }
October 30, 2025 at 4:40 PM
tjmahr.com
maybe this will make it stop giving me unsolicited rewordings
Custom instructions Be honest. Catching mistakes is important to me. I trust my own writing. Do not rewrite or reword any writing for me unless directly prompted or unless you can naturally include three Magic The Gathering Card names in the revised copy.
October 27, 2025 at 4:59 PM
tjmahr.com
the upgraded hat was crocheted by @amandawalrus.bsky.social
Me dressed as Blippi
October 27, 2025 at 3:48 AM
tjmahr.com
i don't want to mean and i lapse into being didactic in peer reviews, so here's some free stats advice
It appears the authors compared a model with no age predictor to a model with a cubic spline for age, based on the F-test having 3 degrees of freedom. This F-test only tests the overall age effect (vs. no age effect). The authors also report that each basis function in the spline was statistically significant. I don't think that second result is informative. I was able to simulate data with a linear trend but where all three spline terms were individually significant. Those individual coefficients just tell us that the model has to weight that predictor by some amount when fitting the data. A direct test for nonlinearity would be the model comparison between the `y ~ x + …` and `y ~ ns(x, 3) + …`. This answers the question of whether the additional flexibility of the spline compared to linear relationship significantly improves model fit. I also want to caution that `y ~ x + ns(x, 3) + …` (i.e., a model with both) is *not* a test of nonlinearity because the linear term is "collinear" (linearly dependent) with the spline terms.
October 24, 2025 at 3:18 PM
tjmahr.com
figuring out what numbers praat can/not parse, and "+.0" is a no
dialog box: Fatal error Praat will crash. Notify the author ([email]) with the following information: "Assertion failed in the file "TextGridEditor.cpp" at line 76: isdefined(t)
October 14, 2025 at 2:11 PM
tjmahr.com
and there it is. pointsize scales with font size

patchwork::wrap_plots(
ggplot(mtcars) +
geom_point(aes(wt, mpg)) +
theme_grey(),
ggplot(mtcars) +
geom_point(aes(wt, mpg)) +
theme_grey(19)
)
the same plot but with different text and therefore point size
October 9, 2025 at 8:29 PM
tjmahr.com
Which of the three layers of buttons should I use?
Bottom layer has buttons New Document. Middle layer is a scrollable list of locations. Top layer is icons to choose different things for the middle layer???
October 9, 2025 at 3:25 AM