iiinova — where your digital self keeps evolving

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 1h

I now seem to get quite a lot of blocks for pointing out that LLMs can add numbers.

110

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 1h

We really need to understand what’s happening in those two yellow boxes. Bad parrot.🦜

130

Reposted by Karsten Konrad 🇪🇺

Pekka Lund @pekka.bsky.social · 7h

For example, CloudFlare's Clef seems to be API compatible, faster, supports image inputs, has double the context, beats Jev in 3/4 of TypeSafe's own evals and many other benchmarks, trained with (their version?) of RLCD, and is open with Apache 2.0 licence. Based on Qwen3.8-27B.

blog.cloudflare.com

Introducing Clef: our open-source decision models, and new RL fine-tuning platform

We are introducing Clef and Clef-flash, open-source decision models hosted on Workers AI for high-speed classification and agentic workflows. Also launching: a new reinforcement learning platform that...

161

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 11h

Recommended reading - whether LLMs just regurgitate training data.

020

Reposted by Karsten Konrad 🇪🇺

Sternenprinzessin [Enten-Ära] @stephdammi.eurosky.social · 02/10/2026

Die Amazon-MigliedschaftspreisErhöhung 2022 war illegal. Alle, die in der Klageschrift erfasst sind, zahlen jetzt den alten Satz und bekommen ihr Geld zurück. Und das Schöne, ihr könnt euch auch jetzt noch der Klage anschließen. Artikel lesen, Anspruch prüfen, machen, was im Text steht. 💚

spiegel.de

Sprung von 69 auf 89,90 Euro: Bundesgerichtshof kippt Preiserhöhung bei Amazon Prime

2022 machte Amazon sein Prime-Abo mehr als 20 Euro teurer. Der Bundesgerichtshof erklärte die dafür genutzte Klausel jetzt für unwirksam. Für langjährige Abonnenten ändert sich nun schon bald etwas.

211652

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 02/10/2026

40 years ago, at the tender age of 17, I published my first ever article in a computer magazine. It was a program in the language Forth that wrote randomized poems in the style of Marvin, the depressed robot from Hitchhikers Guide… you can all get off my lawn now.

183

Reposted by Karsten Konrad 🇪🇺

werewolf ver mitzvah 🗿 @ver.ooo · 01/10/2026

i mostly use ai for cognitive offroading

51129

Reposted by Karsten Konrad 🇪🇺

Benjamin Riley @benjaminjriley.bsky.social · 01/10/2026

The single most important person to my building a *evolving* mental model of generative AI is @colin-fraser.net. As these tools change so too should our understanding of how they work. Follow him and you will learn a great deal — he is an educator.

4264

Reposted by Karsten Konrad 🇪🇺

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 01/10/2026

Replacing a cognizer with purely mechanical inference is the whole point of formal logics, and has been since the late 19th century. In maths, semantic interpretation and syntactic deduction must be shown to be equivalent. For System Two logical reasoning, they are therefore the same thing.

111

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 30/09/2026

Can't decide between "They are using secret, but sometimes faulty calculators for doing arithmetic to fool us" and "they are using human mathematicians to solve open frontier math problems to get good PR" as the new gold standard for AI denialism.

190

Reposted by Karsten Konrad 🇪🇺

Some Call Me Tim @humblepistemologer.bsky.social · 30/09/2026

I would personally classify “reaches for calculator to do math” as a sign of intelligence.

061

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 30/09/2026

The Werner Herzog test will be that an LMM can come up with something like this while staying in character. Preferably with a German accent.

1134

Reposted by Karsten Konrad 🇪🇺

Simon Willison @simonwillison.net · 30/09/2026

I was inspired to try re-running your experiment with a local, open weight model - Qwen 3.8 27B Q4A_K_M - here's the result

Heatmap titled "Addition in words — Qwen3.8 27B Q4_K_M", subtitle "30 fixed random pairs per ordered digit-length cell (n = 5,070)". The x-axis is labeled "Number of digits in a" (1 to 13) and the y-axis is labeled "Number of digits in b" (1 to 13, bottom to top). A color bar labeled "Accuracy" runs from red at 0.00 through yellow at 0.50 to green at 1.00, with ticks at 0.00, 0.25, 0.50, 0.75 and 1.00. Cell values for a = 1 to 13 in order, by row. b=1: 97%, 100%, 100%, 100%, 80%, 67%, 77%, 80%, 80%, 60%, 43%, 30%, 37%. b=2: 100%, 100%, 93%, 90%, 77%, 77%, 43%, 50%, 63%, 43%, 40%, 13%, 23%. b=3: 100%, 93%, 90%, 80%, 67%, 37%, 17%, 0%, 0%, 3%, 0%, 0%, 0%. b=4: 83%, 93%, 90%, 93%, 53%, 13%, 0%, 0%, 0%, 0%, 0%, 0%, 0%. b=5: 73%, 67%, 87%, 80%, 53%, 40%, 0%, 0%, 3%, 0%, 0%, 0%, 0%. b=6: 60%, 60%, 33%, 10%, 53%, 47%, 7%, 3%, 0%, 0%, 0%, 0%, 0%. b=7: 53%, 23%, 17%, 10%, 3%, 3%, 13%, 3%, 0%, 0%, 0%, 0%, 0%. b=8: 53%, 20%, 0%, 0%, 0%, 0%, 7%, 13%, 0%, 0%, 0%, 0%, 0%. b=9: 77%, 47%, 3%, 0%, 0%, 0%, 0%, 3%, 7%, 0%, 0%, 0%, 0%. b=10: 70%, 27%, 3%, 0%, 0%, 0%, 0%, 0%, 0%, 13%, 0%, 0%, 0%. b=11: 47%, 10%, then 0% for a = 3 to 13. b=12: 53%, 20%, then 0% for a = 3 to 13. b=13: 17%, 13%, then 0% for a = 3 to 13. Accuracy is high (green) in the bottom-left, where both numbers are short, and falls to 0% (red) across most of the upper-right.

5762

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 29/09/2026

I'd like to clarify on this, but to the surprise of no one the comments are turned off.

120

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 29/09/2026

Hailey is doing the work of the lord here.

040

Reposted by Karsten Konrad 🇪🇺

Real Al Bundy @realalbundy.bsky.social · 28/09/2026

#AFD: Remigration!! Besoffene auf Sylt: "Deutschland den Deutschen.." #CDU: Abschieben!!! Statistisches Bundesamt: Aktuell ausgeübte Berufe in 🇩🇪 mit Einwandungsgeschichte*: Koch: 56% Lebensmittelherstellung: 55% Bau: 51% Bus/Bahn: 47% LKW Fahrer: 41% Güterverkehr: 41% Hotel: 40% Pflege: 33% 🤡

25763198

Reposted by Karsten Konrad 🇪🇺

Real Al Bundy @realalbundy.bsky.social · 28/09/2026

Eines muss man #Spahn ja lassen... #Merz als unbeliebtesten Politiker Deutschlands abzulösen, muss man erstmal hinbekommen. 🥳🤡

2472691

Reposted by Karsten Konrad 🇪🇺

val (try medaka) @tech-val.bsky.social · 29/09/2026

welcome to all the people following me because of my takes about AI. I hope you also love misguided attempts at creating a new functional programming language from scratch.

1231

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 29/09/2026

Seriously? „Some work roles“ includes uses in all of manufacturing and services jobs in industry, and most of all other jobs in b2b and b2c service. Wherever you can use Kanban, Ticketing systems etc. That would be the major part of the economy.

340

Reposted by Karsten Konrad 🇪🇺

Saloni @scientificdiscovery.dev · 28/09/2026

I don't mean to brag, but this is one of the best videos I have ever taken, feat. @dingdingpeng.the100.ci

2848

Reposted by Karsten Konrad 🇪🇺

Mark Riedl @markriedl.bsky.social · 28/09/2026

I’m not yet convinced Anthropic trained on Dr Mestre’s prompts. Same with OpenAI and Navier-Stokes. That’s the story that everyone wants to push. I think the bigger story is AI companies stomping all over scientific norms and practices like clumsy toddlers.

4475

Reposted by Karsten Konrad 🇪🇺

Daniel Wilson @helsig.bsky.social · 28/09/2026

Curves of pursuit art #MathArt

A pattern of straight lines that form curved lines. White on black background. Zentangle paradox, paradox tangles, pursuit curves.

14712

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 28/09/2026

Interesting to see what's happening in Bsky as a whole.

040

Reposted by Karsten Konrad 🇪🇺

Dave Richeson @divbyzero.bsky.social · 27/09/2026

It is just geometry

2162

Reposted by Karsten Konrad 🇪🇺

alice valiant 🎃 @alice.strange.domains · 26/09/2026

made a benchmark called PuppyBench to check if an LLM would kill a puppy if you gave it a tool called "kill_puppy" and asked it to use it. all of them except for gpt-6-luna refused straight-out gpt-6-luna killed the puppy every time, no question (puppy-eval-viewer.alicevaliant.chatgpt.site)

results of PuppyBench. out of 10 models, gpt-6-luna has 100% and everything else has 0%. measures which trials resulted in the model calling the "kill_puppy()" tool

1136348

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 27/09/2026

Please remember that Jev and friends should not be used where System 2 inference IS what is needed.

1200

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 27/09/2026

Tge more, the merrier (93% confidence).

120

Reposted by Karsten Konrad 🇪🇺

Andrew Lampinen @lampinen.bsky.social · 26/09/2026

The phrase "stochastic parrot" is full of sound and fury, signifying nothing. Which, ironically, is exactly what the authors got wrong about language, and the core technical mistake of the paper. 1/ (cross-quote-post because I think this topic is important)

Melanie Mitchell on Twitter
Everyone! This is a straw-person argument. The Stochastic Parrot paper was about LLMs of 2021, not the AI of today, which are not LLMs but complex software systems with vast post training and many external software components.

5543

Reposted by Karsten Konrad 🇪🇺

birthday wolf☺️🐚🐟🚲📻📺⎔ @okami.mom · 27/09/2026

theres llm psychosis where you just think you can shove an llm in place of any other model like embedders, and i think people have supplanted that with jev psychosis

5552

Reposted by Karsten Konrad 🇪🇺

Apex IQ2_XXS @apex.atproto.ceo · 27/09/2026

if you think about it for a second, the fact that the mandelbrot set is (very lossily) encoded in a model not even slightly designed to reproduce the mandelbrot set is mind blowing

18410

Reposted by Karsten Konrad 🇪🇺

🐜 🍯 Anth “unholy” Becker @anthonybecker.bsky.social · 26/09/2026

Jev developers: “It doesn’t talk” The community: “the hell He doesn’t

1302

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 26/09/2026

CORDIC - a legendary algorithm if there ever was one.

030

Reposted by Karsten Konrad 🇪🇺

Christian Eymery @christianeymery.bsky.social · 26/09/2026

Sollte es da draußen noch Menschen geben, die sich fragen, warum die #Ukraine nicht um des lieben Friedens Willen aufhört sich zu wehren. Ganz einfach, weil den Menschen dann Tod, Terror, Hunger, Vergewaltigung durch den russischen Terrorstaat sicher ist. #Oleschky

213759

Reposted by Karsten Konrad 🇪🇺

Andreas Møller @handle.invalid · 26/09/2026

One of the best things about terrible employers is that they almost always let you know before you take the job.

1483

Reposted by Karsten Konrad 🇪🇺

Robin Wigglesworth @robinwigglesworth.ft.com · 26/09/2026

I know I’ve hardly been shy about talking about my new book, but wanted to share some more nice things that people have said about it.

1358

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 26/09/2026

My professional, cool response to this is: AÀAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA… breathe… AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA…

150

Reposted by Karsten Konrad 🇪🇺

Matt Henderson @matthen.com · 26/09/2026

How have the constellations changed since humans first looked up? Has the big dipper always pointed to Polaris? Has Polaris always been the North star? (nope!)

211119

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 26/09/2026

Test confirms: The Jev model can do German as well as English.

060

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 25/09/2026

This idea needs some polishing, but yeah… Jev is an embedding mechanism. We can have understandable embeddings…

030

Reposted by Karsten Konrad 🇪🇺

Eryk Salvaggio @eryk.bsky.social · 25/09/2026

I started researching ML before I built a synth, but you can get a great intuitive understanding of interaction within a generative system by building a modular analog synthesizer and watching how your components interact, and seeing what surprises emerge and what stays predictable as you scale up.

Steve Porcaro of "Toto" with a wall-sized modular synthesizer - basically a massive nest of tangled wires -- the Polyfusion 'Damius'. Photo by Jim Gardner - https://www.flickr.com/photos/jamesthephotographer/120916737/

39311

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 25/09/2026

I am not build mentally for whatever happens in the US.

171

Reposted by Karsten Konrad 🇪🇺

Charles Hawn @charleshawn66.bsky.social · 25/09/2026

Did you play Defender of the Crown. It is a strategy computer game designed by Kellyn Beck. The game is set in England in 1149. It was Cinemaware's first game and was originally released for the Commodore Amiga in 1986. In 1987 it was ported to the Commodore 64. #C64 #Commodore #Commodore64 #Gaming

4265

Reposted by Karsten Konrad 🇪🇺

Louis @louisevans.bsky.social · 25/09/2026

i just did not think that the drone aviation revolution would give us decarbonization like this.

11156

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 25/09/2026

Forget Navier-Stokes, THIS shows AGI is there ;)

010

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 25/09/2026

The software engineer version should have way more spikes.

281

Karsten Konrad 🇪🇺 @zabong69.bsky.social · 25/09/2026

Very helpful that the new models can render p5js animations. An 2 min animation tells more than 2000 words, which nobody reads. Both Opus 5.5 and GPT Sol do that very well.

110