iiinova — where your digital self keeps evolving

Downside of course is that integrated memory is very scant on the cloud. For my use case, the throughput on a T4's PCIE is just not enough, only allows us to offload a couple cores' worth of SIMD classification

021

The details are beyond me, but I think it compiles the offloaded code to LLVM bytecode and then from there into native GPU code, so it needs to be written GPU-like, whereas Vectorware is trying to avoid even that. Currently only works on x86 linux... but that's most of the cloud, so...

130

Brand spanking new. Manuel Drehwald announced it offhandedly at the scientific computing unconf meeting in Montreal last week, he merged it into night the night before.

140

That makes more sense, I was wondering what in the world HONK is. By the way, did you take a peek at the new gpu offloading on nightly? It's still very unstable, but promising. I managed to get JSON byte classification working on it yesterday.

110

Jason Kint @kint.bsky.social · 17/09/2026

I want to be happy for you but what in the world does this mean? This is the DevX equivalent of 'curtains for zoosha'. Did I miss some blog posts?

100

Reposted by Nicolas Roman Posner

Eye popping. Court just unsealed Summary Judgment motions in New York Times v OpenAI / Microsoft. Easy to see why the AI companies had so many redactions (pink). Their own people wrote the lede for NYT, the "largest theft of labor in human history"... "more and more substitutive..." /1

4253422317

Nicolas Roman Posner @nrposner.com · 18/09/2026

Position, velocity, acceleration, jerk?

Michael Fry @bigdirtyfry.bsky.social · 16/09/2026

Reposted by Nicolas Roman Posner

Your Grandad Cyclops talks about his school days

6246841587

Reposted by Nicolas Roman Posner

The Onion @theonion.com · 17/09/2026

“While I am proud to have served Tibetan Buddhism in my current body since 1940, I could not pass up the opportunity to inhabit one of pop’s most electrifying showmen.” — Dalai Lama

theonion.com

Dalai Lama Announces One-Year Residency Inside Body of Bruno Mars

DHARAMSHALA, INDIA—Touting it as an exciting limited engagement that would precede his next official reincarnation, His Holiness the Dalai Lama announced Wednesday that after his death he would partic...

20101364

Reposted by Nicolas Roman Posner

By Cory Doctorow (GPG 0xBF3D9110957E5F4C) @doctorow.pluralistic.net · 12/09/2026

Once you understand the corporate culture of AI hyperscalers is everyone cooking their brains by locking themselves in the bathroom, holding flashlights under their chins, and saying "Aaaaaaaaaay Eyeeeeeee" until they wet themselves in terror, things snap into focus: pluralistic.net/2023/06/04/a...

A cutaway view of a stone tower containing an elaborate water-powered medieval geared machine. Rising out of the machine is a rainbow-tinted, pixelated God with the robes, beard and all. To one side of this scene is a cluster of tiny people in Jesus-era robes, falling about themselves in religious ecstasy all over a stone staircase, atop of which stands a robed priest with his hands upraised. The background is a set of shining golden rays, emanating from pixel-God.

231183431

Reposted by Nicolas Roman Posner

Minor Mobius @minormobius.bsky.social · 14/09/2026

You can’t improve on a Henderson post but if you want more I offer henderhead.mino.mobi/wheel With sliders and a phase map and a twin wheel started a billionth of a second later. One important feature of chaotic systems is that small changes lead to big changes, visible toward the end of this

1807

Reposted by Nicolas Roman Posner

Bryan Cantrill @bcantrill.bsky.social · 10/09/2026

Slides from my keynote this morning at #RustConf on the history of Rust at @oxide.computer; video to come! speakerdeck.com/bcantrill/ru...

speakerdeck.com

Rust, In Sickness & In Health

Keynote given at RustConf 2026 on September 10th, 2026 in Montreal. Video to come!

410311

Nicolas Roman Posner @nrposner.com · 09/09/2026

A startup founder lamenting that his engineers refuse to work 80+ hour weeks like investment bankers (who get high salaries, large bonuses, and stable paths to promotion) despite giving them 0.1% equity! The same guy tried to warm up the room with misogynistic 'jokes'.

On my way to the airport now! Hope to see y’all there!

011

Good luck, and see you there!

Certainly it would be very pleasant (for me) if models converged on this and it became standard, but it's not clear where we can bridge from ought to is.

000

They're certainly using Rust for internal infrastructure (Predrag Gruevski gave a good talk a few months ago about his work at OpenAI) but it's not clear to me that Rust would seep into the habits of the models themselves. AFAIK, the Rust hires are for infra and dev tooling, not model training.

200

Though I am reminded of a GenAI workshop at my fiancee's workplace that encouraged them to 'use AI to code a game'... on locked-down corp laptops without access to any compiler or interpreter, so the LLM had to write Snake in HTML.

000

Probably also worth noting that lots of machines these days have a Python interpreter built in, Mac and Linux ship with it, while most do not have the Rust toolchain installed. If permissions and environment are sampled from genpop, Python is more likely to be present.

210

We can say much the same about the benefits of more verbose static typing. Assuming the median user isn't even reading the code these days, it doesn't matter either way, and even if it makes it easier for the LLM to write correct code (not sure if this is clear at all) that's still not its goal.

100

If most users don't request it up front and don't (know to) demand it afterwards, why would the LLM/harness bother with it? After all, it's not trying to produce the most optimum code possible, it's just trying to satisfy the user within reasonable effort bounds.

100

And certainly, most users aren't including 'and make it very performant!' in their explicit prompts, nor do they ask the LLM/agent to set up benchmarks and run them after each change to detect regressions, regardless of language used!

100

Regarding performance, I'm not aware if major LLM providers do any post-training to encourage the use of more performant tools or the generation of more performant code (or if the harnesses try to encourage this either), but I would assume not, or at least not as a top priority.

100

Well, I think of it in terms of training set, post-training (and harness construction) and 'did the user ask for it?'. There's a whole lot more Python out there than C, C++, or Rust, perhaps combined, so that's no mystery.

100

When Vectorized Arrays Aren't Enough: Array Optimization from Bytecode to Assembly- Nicolas R Posner

SciPy 2026 talk recordings dropped a couple days ago! Here's mine, 'When Vectorized Arrays Aren't Enough: Array Optimization from Bytecode to Assembly www.youtube.com/watch?v=xaEm...

youtube.com

YouTube video by SciPy

It is not necessary to remind everyone that A Textual Passage is indeed A Procession of An Individual Character, for instance. Yours truly.

000

and so on, considering that some of our guests are abundantly known and recorded in the Standard Social Register and find it tedious to have their full names declared at each introduction, such that these might be omitted.

100

``` Dear Module, May I have the distinct pleasure of introducing our Distinguished Guests, An Integer Number, first of his name, A Textual Passage, A Conjunction of An Integer Number and An Integer Number, henceforth A Numeric Tuple For the pleasure of tonight's dance? ```

100

Then, to gather our companions in a Module other than their own, simply introduce them by name, maintaining proper decorum by introducing the most senior (never 'primitive'!) Types first, strictly following their ancestral precedence, such as no child type should be declared before its parents:

100

To do the same for a public library API, be sure to close with `Yours faithfully` instead.

100

``` Dear Compiler, If it causes you no undue hardship, might we trouble you to introduce our young friend T into polite society at your convenience? Much obliged. ```

100

Thus, I submit for your consideration, a protocol to ensure the privacy of our companions is respected, while preserving civilized means of arranging lively gatherings: Within a given Module, we may declare, following the formal introduction of a Type or Diplomatic Proposal:

100

or if all our companions are in public at all times, which I must say would be regrettably taxing on anyone's faculties, let alone that of a Type which has yet to attend its first cotillion.

100

Indeed, it fails to specify whether the privacy of all our companions is assumed, in which case we have no means by which to host a convivial gathering in my good friend's Module,

100

Clearly, most humans prefer to write the former. But if all LLMs are effectively power users with encyclopedic understanding of all programming languages and ecosystems... why should they prefer the first set of abstractions to the second? I should expect the reverse.

100

because this makes the language easy to learn and write, at the cost of performance and, IME, clarity of intent. Rust also provides high-level abstractions, but they're aimed at verbosely specifying behavior, including memory access patterns, at some cost to speed of writing and readability.

100