Sign in

troy_phd

@troyhernandez.bsky.social
253 followers 398 following 118 posts

AI founder, Dad, #Rstats

PostsRepliesMedia
Reposted by troy_phd
R Consortium @rconsortium.bsky.social · 16/10/2026
The full R+AI 2026 schedule is up! Nov 10-11. 100% online. Keynotes, talks, lightning talks, a workshop, and a tutorial. A taste: statistical thinking when AI writes the code, agents in R, private AI for sensitive data, and shipping R models to production. rconsortium.github.io/RplusAI_webs...
R+AI Nov 10-11, 2026
0196 19
troy_phd @troyhernandez.bsky.social · 01/10/2026
I tried this idea in the #rstats saber/corteza packages in the spring but it messed up caching and definitely didn’t keep the context below 10k tokens. Will have to dig in to the details.
130 3
Reposted by troy_phd
Henrik Bengtsson @henrikbengtsson.bsky.social · 30/09/2026
UPDATE: The CRAN submission system and the win-builder are up and running again. Packages are now flowing onto CRAN again. My R.matlab submission that got stuck in the inspection queue yday was re-checked automatically today and I just got an email say package "is on its way to CRAN" #RStats
081 8
troy_phd @troyhernandez.bsky.social · 29/09/2026
bsky.app/profile/troy...
010 1
Reposted by troy_phd
Henrik Bengtsson @henrikbengtsson.bsky.social · 29/09/2026
Heads up. Don't submit to CRAN right now. The CRAN submission system is currently broken, because Windows checks won't complete. If you submit your package will end up in the manual "inspection" queue. You'll see the same on win-builder R devel and R 4.6.1. The R 4.5.3 still works. #RStats
0117 11
troy_phd @troyhernandez.bsky.social · 29/09/2026
And finally, a jazz countdown derived from JazzWeek dot com youtube.com/playlist?lis...
youtube.com
Zab Sazzler's Week in Jazz - September 28, 2026 - YouTube
001 1
troy_phd @troyhernandez.bsky.social · 29/09/2026
We have an Adult Alternative AirPlay (AAA, derived from stations like 93.1 XRT in Chicago, WFUV 90.7 in NYC, KFOG 104.5 in SF) countdown: youtube.com/playlist?lis...
youtube.com
Lil' Casey's AAA Countdown - September 26, 2026 - YouTube
100 1
troy_phd @troyhernandez.bsky.social · 29/09/2026
We’ve got 3 countdowns every week… for now. Here’s the classic Billboard Hot 100 countdown; ie pop music youtube.com/playlist?lis...
youtube.com
Lil' Casey's Top 40 Countdown - September 26, 2026 - YouTube
100 1
troy_phd @troyhernandez.bsky.social · 29/09/2026
If you’re a fan of Top 40 lists, check out my startup’s Top 40 countdown playlists on YouTube! They’re generated entirely in #rstats using our AI packages; diffuseR, chatterbox, and whisper! Links below. Thanks for the diffuseR shoutout @rworks.bsky.social!
132 3
Reposted by troy_phd
Dirk Eddelbuettel @eddelbuettel.com · 27/09/2026
r^4 post #060: Using bubblewrap for R sandboxing Quick note on agentic R use via bubblewrap dirk.eddelbuettel.com/blog/2026/09... #rstats
Screenshot of r^4 blog post #060 at https://dirk.eddelbuettel.com/blog/2026/09/27#060_bubblewrap_for_R discussing R use inside the 'bubblewrap' harness helper
093 9
Reposted by troy_phd
R Consortium @rconsortium.bsky.social · 25/09/2026
"Sustaining R: Investing in the People, Infrastructure, and Community Behind R" Register for Oct 6 R Core / R Foundation panel discussion Simon Urbanek, Kurt Hornik & Heather Turner. Hosted by R Consortium. r-consortium.org/webinars/sus... #RStats
"Sustaining R: Investing in the People, Infrastructure, and Community Behind R" - Register for Oct 6 R Core / R Foundation panel discussion
0116 11
troy_phd @troyhernandez.bsky.social · 22/09/2026
Update: corteza scored 100% on ARC-AGI-3's public set with GPT-6 Astra! And due to how they calculate their metrics, I'm happy to report that our 100% is the best 100%. I made a dashboard with details: cornball.ai/arc-agi-3/ #rstats
cornball.ai
ARC-AGI-3 comparison
032 3
Reposted by troy_phd
Giles @gdeejay.bsky.social · 19/09/2026
#AI *improving* #rstats ranking on the index wasn't on my 2026 bingo card. 😲
R has improved on the tiobe index by 8 ranking points since the same month last year.
2133 13
troy_phd @troyhernandez.bsky.social · 15/09/2026
Yesterday I conflated Astra’s public and semi-private ARC-AGI-3 results. Additionally, we ran corteza overnight with Claude Opus 5 at xhigh reasoning, this time completing all 25 games (183 levels), scoring 98.6% again. The blog post is updated with the correction and new results. #rstats #AI
Results table comparing harnesses on the same 25 public ARC-AGI-3 games. Scores measure completion and action efficiency, not just the percentage of levels solved.

Two corteza rows are highlighted in pale purple. Both use Claude Opus 5 and score 98.6%. At max reasoning effort, corteza completed 182 of 183 levels and 24 of 25 games using 7,597 actions. At xhigh, it completed all 183 levels and all 25 games using 9,154 actions.

Prime Agent’s 3 Claude Opus 5 xhigh runs scored 95.5%, 95.2%, and 94.99%. Respectively, they completed 179, 178, and 183 levels; won 24, 24, and 25 games; and used 13,850, 11,369, and 12,439 actions.

GPT-6 Astra with ARC’s provider adapter scored 100% at max and xhigh effort, and 99.7% at high. With ARC’s standard harness, Astra max scored 68.3%, Claude Opus 5 high scored 40.7%, and GPT-5 Sol max scored 13.3%. Level, game-completion, and action counts are blank for these ARC reference runs.
020 2
troy_phd @troyhernandez.bsky.social · 14/09/2026
ARC's analysis of GPT-6 Astra: 62.7% with their standard harness, 98.55% at max effort with their provider adapter harness (99.95% at high). corteza, our #rstats harness, at max effort with Claude Opus 5: 98.58%. Same 25 games. cornball.ai/posts/cortez...
cornball.ai
98.6% on ARC-AGI-3 in R
corteza, an R-native RLM harness, scored 98.58% on ARC-AGI-3. Across 25 games the model wrote 894 R helpers and kept calling them, 53 of them from before its context was compacted.
093 9
troy_phd @troyhernandez.bsky.social · 13/09/2026
janssonr is on CRAN: strict JSON for #rstats Duplicate keys, big ints, and NUL escapes are refused with classed errors. mx.client 0.2.1 is up too! Verify an encrypted @matrix.org bot from your matrix client and the R console: 7 emojis on 2 screens. cornball.ai/posts/jansso...
031 3
Reposted by troy_phd
Dirk Eddelbuettel @eddelbuettel.com · 04/09/2026
r^4 post #059: r2u, GitHub Actions, a Tragedy of the Commons, and a Fix About #r2u Timeouts at GitHub Actions, Analysis and a Fix dirk.eddelbuettel.com/blog/2026/09... A 'free' resource may get abused, defensive measure are taken, those hurt subsequent unrelated use, and how we cope. #rstats
Screenshot one of two describing blogpost #59 in the r^4 series at https://dirk.eddelbuettel.com/blog/2026/09/04#059_r2u_gha_tragedy_of_the_commons_a_fixScreenshot two of two describing blogpost #59 in the r^4 series at https://dirk.eddelbuettel.com/blog/2026/09/04#059_r2u_gha_tragedy_of_the_commons_a_fix
083 8
Reposted by troy_phd
Dirk Eddelbuettel @eddelbuettel.com · 22/08/2026
RProtoBuf 0.4.28 on CRAN: Small Updates An R interface to Google Protocol Buffers data serialization dirk.eddelbuettel.com/blog/2026/08... #rcpp #rstats /cc @troyhernandez.bsky.social
Screenshot of blog post at https://dirk.eddelbuettel.com/blog/2026/08/22#rprotobuf_0.4.28 describing release 0.4.28 of CRAN package RProtoBuf
052 5
Reposted by troy_phd
Dirk Eddelbuettel @eddelbuettel.com · 19/08/2026
PSA: If your Ubuntu-using CI jobs on GitHub Actions hang mysteriously today, consider adding sudo sed -i -e 's/azure\.//' /etc/apt/apt-mirrors.txt (or similarly for .list or .sources files). It has been added to the r-ci setup for #Rstats, and addresses the issue for ubuntu-latest.
031 3
troy_phd @troyhernandez.bsky.social · 14/08/2026
For more details: cornball.ai/posts/diffus... and github.com/cornball-ai/...
cornball.ai
diffuseR is on CRAN
install.packages("diffuseR") That works now! diffuseR 0.2.2 went up on 2026-08-08, its first CRAN release. When I introduced diffuseR in June 2025, it did two models and I was upfront about the shortc...
000 0
troy_phd @troyhernandez.bsky.social · 14/08/2026
diffuseR is on CRAN! AI images and video, locally in #rstats install.packages("diffuseR") Go make something weird!
172 7
troy_phd @troyhernandez.bsky.social · 13/08/2026
Oh, I know! I've got a crazy project with over 700!
010 1
troy_phd @troyhernandez.bsky.social · 11/08/2026
I wouldn’t agree that it’s disgusting. It’s certainly verbose! And definitely more safe than my prior approach, which was, “If they do something that silly, that’s on them!”
140 4
troy_phd @troyhernandez.bsky.social · 09/08/2026
I’ve recently started getting a bunch of bot followers here on Bsky. Feels like something is afoot.
010 1
troy_phd @troyhernandez.bsky.social · 06/08/2026
I'm happy to announce our new subtitles package! Go caption something! #rstats
0225 22
troy_phd @troyhernandez.bsky.social · 05/08/2026
Saw some great talks last year! Looking forward to it
000 0
troy_phd @troyhernandez.bsky.social · 23/07/2026
I guess it’s not entirely pay per usage yet as I haven’t used this since last year. Time to actually cancel
000 0
troy_phd @troyhernandez.bsky.social · 22/07/2026
Anyone can cook. Now anyone can cut out a background in #rstats. rembg is on CRAN via ONNX and the onnxr package. png::writePNG(rembg("remy.jpg"), "remy_rembg.png") cran.r-project.org/web/packages...
0202 20
troy_phd @troyhernandez.bsky.social · 20/07/2026
I’ve been thinking of doing something automated via my corteza harness, but right now it’s very manual. I set them in the same directory and assign one the role of “reviewer, don’t write any code”. Make Claude do the work and commit. \copy the summary and tell codex: Review Claude’s work: [paste]
010 1
troy_phd @troyhernandez.bsky.social · 18/07/2026
I’ve found Claude to be good at big picture planning and long running, more auto pilot, sessions. GPT seems to be a much better reviewer and will one-shot small to medium tasks more deeply, but requires more manual oversight
110 1
Reposted by troy_phd
Dirk Eddelbuettel @eddelbuettel.com · 11/07/2026
Norm Matloff (1948-2026) @matloff.bsky.social I was made aware of the post below made by his daughter. We lost another friend and valued long-time contributor the #Rstats community, gone far too soon, and sure to be missed. His memory will be a blessing.
Screenshot of post on X in the account of Norm Matloff in which his daughter announces his passing. Gone too soon, and sure to be missed.
312937 129
troy_phd @troyhernandez.bsky.social · 26/06/2026
Now you can use your Claude or ChatGPT subscription to run the R agent harness, corteza. No API key, no per-token meter, and it's on CRAN! Thanks to @hadley.nz for the provider-native search idea, and to @pp0196.bsky.social, whose Codex OAuth PR became tinyoauth. cornball.ai/posts/claude...
cornball.ai
Claude and Codex subscriptions, now driving an R agent
A couple of months ago, corteza was an R session you could hand to an LLM, as long as you fed it an API key and watched the per-token meter tick. As of this week you can point it at your Claude Pro/Ma...
0173 17
troy_phd @troyhernandez.bsky.social · 25/06/2026
🎉 chatterbox TTS is now an R package! Thanks to Chris English for all the testing and @dfalbel.bsky.social for teaching me about garbage collection in torch! 🙏 #rstats
072 7
troy_phd @troyhernandez.bsky.social · 24/06/2026
New releases across the tinyverse dev toolchain: tinyrox, rformat, tinypkgr, and saber. People who aren't me are filing issues and PRs now! Big thanks to @eddelbuettel.com , @gmcd.bsky.social , and Christophe Regouby #rstats cornball.ai/posts/tinyve...
cornball.ai
Tinyverse Toolchain: Updates and Thanks
Back in February we released the tinyverse development toolchain: a handful of small, low-dependency packages for documenting, formatting, checking, and introspecting R packages. Since then they’ve be...
060 6
troy_phd @troyhernandez.bsky.social · 23/06/2026
whisper v0.4.0 now on CRAN 👄 Speech-to-text 2.5x faster with JIT compilation Added lightweight serve() using base::serverSocket() #rstats
040 4
troy_phd @troyhernandez.bsky.social · 23/06/2026
Very cool! CRAN bound?
120 2
troy_phd @troyhernandez.bsky.social · 23/06/2026
Talk to @matrix.org from R. mx.client 0.1.1 just landed on CRAN, completing the mx.* trio. Install it for config, rooms, sync, and markdown; add mx.crypto for Olm/Megolm E2EE. h/t @eddelbuettel.com for finally getting me onto Matrix. #rstats cornball.ai/posts/mx-cli...
cornball.ai
mx.client 0.1.1: use the Matrix protocol from R
mx.client 0.1.1: config persistence, room resolution, sync cursors, and Olm/Megolm E2EE orchestration. The opinionated layer on top of the mx.api/mx.crypto plumbing.
062 6
Reposted by troy_phd
Dirk Eddelbuettel @eddelbuettel.com · 21/06/2026
BBC article about Brian Ripley of the #RStats Core team given this week's excellent news concerning 2026 Rousseeuw Price for Statistics recognition for him and the core team. www.bbc.com/news/article...
bbc.com
Prof Brian Ripley on developing R to manage the world's data
Prof Brian Ripley has spent the past three decades working on a computer software that analyses data.
419383 193
Reposted by troy_phd
Dirk Eddelbuettel @eddelbuettel.com · 17/06/2026
Reproducing the Rousseeuw Prize citation ordering with a few lines of #Rstats code from an older blog post. And once again sincere congratulations for an award so well deserved.
R code and output:

> x <- readLines(pipe("svn log"))  # run in R's svn repo
> rx <- x[grep("^r\\d+ \\|",x)]
> who <- gsub(" ","",sapply(strsplit(rx,"\\|"),"[",2))
> twho <- table(who) 
> twho[rev(order(twho))[1:5]]
who
  ripley maechler   hornik       pd     luke 
   29782     9965     9092     2778     2253 
>
1173 17
Reposted by troy_phd
The R Foundation @r-foundation.bsky.social · 17/06/2026
The 2026 Rousseeuw Prize for Statistics has been awarded to the R Project. #RStats www.rousseeuwprize.org/2026
rousseeuwprize.org
The Rousseeuw Prize for Statistics
The Rousseeuw Prize for Statistics is a biennial prize to celebrate outstanding contributions to statistics research.
2280106 280
troy_phd @troyhernandez.bsky.social · 11/06/2026
Is there more to this story? I’d like to hear it!
100 1
troy_phd @troyhernandez.bsky.social · 11/06/2026
A fresh vignette is available here: github.com/cornball-ai/... And as you can see from this post, codex 5.5 is comparable to the new controversial Fable model bsky.app/profile/royf... The point is not “Claude bad, Codex good,” it's that R’s AI tooling should not have a single point of failure.
github.com
030 3
troy_phd @troyhernandez.bsky.social · 11/06/2026
I’m excited to announce that #codex is now available (typically through a chatGPT subscription) from R through corteza. Logging in is enabled by the dev version of the new tinyoauth package, already in the CRAN queue! A big thank you to @pp0196.bsky.social for the contribution.
121 2
troy_phd @troyhernandez.bsky.social · 11/06/2026
I strongly believe that R’s AI tooling shouldn't rely solely on one closed subscription path. That is certainly convenient for some users, but it cannot be the whole story. This is why I wrote the FOSS corteza package. It's an LLM agent harness written in #rstats, running in the R console.
120 2
troy_phd @troyhernandez.bsky.social · 11/06/2026
Like many, I experienced similarly unpredictable performance from Claude Code in the spring. It was hard to tell whether the problem was my code, my prompt, model drift, or something else. For serious engineering, silent degradation is worse than an explicit refusal.
110 1
troy_phd @troyhernandez.bsky.social · 10/06/2026
From @alphaxiv.org on X: > As believers of open research, we are disappointed to see Anthropic silently degrading Fable 5 for AI development... x.com/askalphaxiv/...
130 3
Reposted by troy_phd
SMT @pp0196.bsky.social · 07/06/2026
Rbebelm: #RStats Bindings for 'BebeLM' 'CPU' Inference For A 8B-1B Active model, not that bad A tiny R CLI "Agent" inspired by Corteza github.com/cornball-ai/... repo: github.com/sounkou-bioi... Upstream Rust crate : github.com/maximecb/beb... LFM2.5-8B-A1B blog: www.liquid.ai/blog/lfm2-5-...
021 2
troy_phd @troyhernandez.bsky.social · 07/06/2026
bsky.app/profile/troy...
000 0
troy_phd @troyhernandez.bsky.social · 07/06/2026
bsky.app/profile/troy...
000 0
troy_phd @troyhernandez.bsky.social · 07/06/2026
If you’re into #rstats and #obsidian vaults, try pensar and let me know what you think cran.r-project.org/package=pensar
cran.r-project.org
pensar: LLM Wiki Engine
Personal wiki engine with a large language model (LLM) as research assistant. Supports guided sessions through a 'Claude Code' &lt;<a href="https://github.com/anthropics/claude-code" target="_top">htt...
050 5