iiinova — where your digital self keeps evolving

Training metrics of the mimo-v2.6-pro and mimo-v2.6-flash reinforcement-learning runs, live from the trainer's logs.

for sure ]. i think even though this is for RL, cost display here is quite interesting mimo.xiaomi.com/rl/#overview !

mimo.xiaomi.com

mimo-v2.6 RL

Agents here mean basically codex harness runs (and till recently that harness is very bloated and CPU/memory hungry)

the paper looks like a rewrite of a lean formalization campaign. All of the recent proof "papers" look like they are reconstruction of lean proofs. Lot of scripts being run for sure and 10K agents will all consume lot of resources even without lean

you are forgetting the CPU and memory cost, 10K agents meaning millions of CPU hours compiling lean slop, lot of VMs

Gabor Csardi @gaborcsardi.org · 14/09/2026

Reposted by SMT

New beta of #Rstats rig: LOTS of EXCITING stuff: - new: web site: rig.r-lib.org - new: user-mode R installation: rig.r-lib.org/admin-vs-use... - new: manage R packages (CRAN + git): rig.r-lib.org/pkg-guide.html - new: manage R projects and virtual environments: rig.r-lib.org/proj-guide.h...

rig.r-lib.org

rig – the R Installation Manager – rig

34919

SMT @pp0196.bsky.social · 12/09/2026

CanardAbsurd: Durable #RStats workflows with DuckDB state and Quack remote SQL (Slop) Sloppy attempt to solve some recurring issues around durability (and resumption) especially for (frontend or api) exposed (costly) workflow coordination. repo: github.com/RGenomicsETL...

Heng Li @lh3lh3.bsky.social · 23/03/2026

Reposted by SMT

Long reads carry multiple small vars and SVs and their phasing. LongcallD is the only caller that tightly integrates germline/mosaic small/structural vars/MEIs and their phasing in a single C program. One command line to get competitive small variant calls and better SVs. Led by Yan Gao.

04522

Reposted by SMT

Miles McBain @milesmcbain.bsky.social · 09/09/2026

Big difference is obviously macros. But they worked damned hard to keep “the code is just data” philosophy that underpins macros.

011

SMT @pp0196.bsky.social · 07/09/2026

the lisp point is very funny given the author www.stat.auckland.ac.nz/~ihaka/downl... By the way, are people privy to the R-like rewrite he talks about in that talk ?

SMT @pp0196.bsky.social · 06/09/2026

but "that is" bioinformatics day job :D

SMT @pp0196.bsky.social · 05/09/2026

httpuv could support the required changes, but i guess this is not a priority at posit github.com/rstudio/http.... alternative would be to build a http framework on top of nanonext (or a rust crate including the rust binding to nng because why not)

github.com

Support for SSE · Issue #404 · rstudio/httpuv

Adding support for Server Sent events would be a nicety to have. I can imagine a websocket like API that only allows sending. I'm aware that SSE is pretty limited in HTTP/1.1 but people have reques...

SMT @pp0196.bsky.social · 05/09/2026

nanonext http server is the way to go indeed since httpuv(hence plumber-2-) is not planning to get server sent events anytime soon. Of course there are other http frameworks in R and one could have some vibeslop one these days

220

Reposted by SMT

Vini @vinimdocarmo.com · 24/03/2025

Well, it's now backed by S3 and I've made the code public for anyone interested: github.com/vinimdocarmo... 🦆

021

Reposted by SMT

Henrik Bengtsson @henrikbengtsson.bsky.social · 21/08/2026

The Futureverse Resource Specification proposal: a syntax for resource declarations, monitoring & scheduling; fits <- map(xs, model_fit) |> resources(\(x) memory(4*object.size(x))) |> futurize() www.jottr.org/2026/08/11/r... #RStats #parallel #hpc

081

Reposted by SMT

DuckDB @duckdb.org · 17/08/2026

Quack, quack! A DuckDB v2.0 preview is here! 🦆 In this post, #DuckDB co-creators Mark Raasveldt + @hannes.muehleisen.org preview the headline features of DuckDB v2.0 (expected to land this autumn): • DuckDB as a server • Triggers • async I/O • + more Dive into the deets: duckdb.org/2026/08/17/d...

2559

SMT @pp0196.bsky.social · 17/08/2026

C Dev move enough to break the weird java inspired naming conventions on some corners of base R :D

000

SMT @pp0196.bsky.social · 16/08/2026

There is indeed overlap in the theses of composeq composeq.dev and the working hypothesis of the duckhts project github.com/RGenomicsETL... i.e adding custom kernels only when duckdb sql would underperform and get composition for free

composeq.dev

CompoSeq

Composable foundations for genomic data.

021

SMT @pp0196.bsky.social · 16/08/2026

There are some design choices here, for example I stick to only the duckdb stable C API to make the extension easily maintained and portable across duckdb versions. Moreover can be used with any duckdb client not just the R package duckdb.org/community_ex... Therefore it would speak C Arrow Data

duckdb.org

duckhts

DuckDB Community Extensions DuckDB extension for reading HTS file formats via htslib

120

SMT @pp0196.bsky.social · 16/08/2026

I did not do a benchmark against these crates, only against tools I ported (e.g mosdepth). I knew about exon and it was an inspiration but the design is a pure C duckdb extensions on top of htslib, cgranges and other C libraries

SMT @pp0196.bsky.social · 15/08/2026

This is great work, happy to see DuckDB usage in this space. Related self plug, i made a HTS file formats reader extension rgenomicsetl.r-universe.dev/Rduckhts that allow fast imports directly into duckb and avoid RSamtools and al. There are community readers for single cell formats

rgenomicsetl.r-universe.dev

Rduckhts: 'DuckDB' High Throughput Sequencing File Formats Reader Extension

Bundles the 'duckhts' 'DuckDB' extension for reading High Throughput Sequencing file formats with 'DuckDB'. The 'DuckDB' C extension API <https://duckdb.org/docs/stable/clients/c/api> and its 'htslib'...

111

Reposted by SMT

Eddie Ruiz @ed2uiz.bsky.social · 14/08/2026

Can databases scale scientific analyses beyond RAM? I’m excited to share dbverse, a composable database framework built to do just that. Across benchmarks, dbverse delivers ~1000x speedups and scales #spatialomics analyses to millions of cells. Preprint: www.biorxiv.org/content/10.6... 1/🧵

biorxiv.org

dbverse scales spatial omics analysis with embedded analytical databases

Spatial omics datasets are increasing in size and complexity, exceeding the memory of standard computers and thereby limiting data analysis. Here we present dbverse, a framework for larger-than memory...

242

Reposted by SMT

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

i must say that Claude models generate less of this stuff compared to GPT models but it is significantly worse in both model families

000

the 1K thing is not a joke, i actually grep a repo that was basically a port of a previously handcoded one, half of the code is made of these "helpers"

210

GitHub - sounkou-bioinfo/s7contract: 'Go'-like interfaces and 'Rust'-like traits with 'S7'

my safety nudge is "use S7 properties and field validators" i even made a package for organizing this thing github.com/sounkou-bioi... :D but i get SuppressWarning(as.numeric(x)) - that one is reward hacking since i develop with warning as errors- .read_tsv_safe_file_1, .read_tsv_safe_file_2

github.com

'Go'-like interfaces and 'Rust'-like traits with 'S7' - sounkou-bioinfo/s7contract

120

thanks ! i will check what can be transferred

i use extensively the subrepo pattern (some .sync folder) and made some (project scoped) skills around some coding patterns, but hard to pull against the attractor bassin toward overly defensive code is baked in by RL. I should make new attempts to specialized the skills a bit more

Tim Dunn @timd.one · 22/07/2026

No matter how much one tries, LLMs cannot stop generating disgusting #Rstats code (1K .something functions checking if string is empty or similar) and i think this is getting worse. No wonder people do not want to read code anymore or prefer using slop language they don't know to read

240

Reposted by SMT

CRISPR off-target search can find the right locus while hiding plausible alignments within it. I wrote about how Sassy v0.2.5 uses recursive backtracking to report every reasonable alignment—more than 9× as many in under 30 seconds. blog.fulcrumgenomics.com/p/why-crispr...

blog.fulcrumgenomics.com

Why CRISPR Off-Target Search Should Report Multiple Alignments Per Locus

How Sassy enumerates every reasonable alignment without sacrificing runtime

054

SMT @pp0196.bsky.social · 28/07/2026

#CRAN Release of version 1.5.1-0.1.3 of Rduckhts: #RStats Bindings to the #htslib based #DuckDB extension for HTS file formats This comes with a BigWig reader via libBigWig used with (WIP)DuckVEP: the fastest Ensembl VEP-compatible consequence predictor in the West written in #C #Bioinformatics

031

SMT @pp0196.bsky.social · 25/07/2026

#RStats is this SSL issue on cran.r-project.org/submit.html restricted to me or others are facing the same issue ?

SMT @pp0196.bsky.social · 18/07/2026

had something in Rtinycc too, all these warnings are in the upstream libraries i wrap. Rduckhts being flagged is odd to me since there is no linking to R involved, but i guess this is because the configure logs are scanned.

040

SMT @pp0196.bsky.social · 18/07/2026

i don t have the list, just a cryptic mention here www.stats.ox.ac.uk/pub/bdr/disc...

030

SMT @pp0196.bsky.social · 18/07/2026

#Rstats ah looks like CRAN fedora machine's tool chain got updated, and lot of C code nonsense in the vendored libraries my packages wrap are not allowed anymore. Good though that there are 25 other packages concerned :D I have on month to fix it or packages will be rightfully boot out

SMT @pp0196.bsky.social · 14/07/2026

best way to send a LLM into a loop and burn all your tokens, in #RStats/C project, say go look at @coolbutuseless.bsky.social repos github.com/coolbutusele... for what we can use

github.com

coolbutuseless - Repositories

Cool, but useless. coolbutuseless has 255 repositories available. Follow their code on GitHub.

020

SMT @pp0196.bsky.social · 10/07/2026

CRAN release of version 1.4.0-0.1.0 of Rduckhts, the #RStats It was also nice to confirm that are our port of Brent Pedersen 's great mostdepth algorithm is likely the fastest for that application space in the West. But riker shows us there is space for memory bounding !