Karsten Konrad 🇪🇺
@zabong69.bsky.social
Data Scientist and nerd. PhD in automated theorem proving just before AI became a subfield of Linear Algebra. Machine Learning, statistics, data visualisation, biology, recreational maths, computation, old pocket calculators, and video games
Pinned
As a former researcher on automated reasoning systems and NLP machine learning algorithms, I am very grateful to the Bsky community for clarifying my various misunderstandings about reasoning and the capabilities of NLP machine learning algorithms.
Reposted by Karsten Konrad 🇪🇺
Es ist mir scheißegal dass Erdbeeren Nüsse sind
September 19, 2026 at 11:37 AM
Es ist mir scheißegal dass Erdbeeren Nüsse sind
Very interesting thread.
Thinking is, for an anendophasic person like me, pushing around vectors, directions, in my head. I point my thought toward "cat" and then i virtually see, feel, hear it, have fuzzy ideas about Fritz the Cat and the feel of fur, maybe a story about a nightly visit from a neighbors cat, and so forth.
September 19, 2026 at 6:18 AM
Very interesting thread.
Reposted by Karsten Konrad 🇪🇺
Wonder what their handles are, because they're definitely on here
September 18, 2026 at 8:56 PM
Wonder what their handles are, because they're definitely on here
Reposted by Karsten Konrad 🇪🇺
I'm not sure I would make it in a cult, because I socialized a bit more than normal this last week and now I need a minimum of a full week with no human contact.
September 19, 2026 at 3:06 AM
I'm not sure I would make it in a cult, because I socialized a bit more than normal this last week and now I need a minimum of a full week with no human contact.
Reposted by Karsten Konrad 🇪🇺
TIL about Islamic mortgages, which appear to work like standard mortgages, except *technically* no interest is involved.
September 19, 2026 at 4:26 AM
TIL about Islamic mortgages, which appear to work like standard mortgages, except *technically* no interest is involved.
Reposted by Karsten Konrad 🇪🇺
Being a computer scientist who refuses to find anything about LLMs interesting right now is a bit like being a geneticist who refuses to find anything interesting about the recently opened Jurassic Park
September 18, 2026 at 7:19 PM
Being a computer scientist who refuses to find anything about LLMs interesting right now is a bit like being a geneticist who refuses to find anything interesting about the recently opened Jurassic Park
And so it begins <meme>
Jev can't reason. So I built an interpreter that reasons for it. It runs a logic language where the facts and rules are plain English.
September 18, 2026 at 7:32 PM
And so it begins <meme>
If there ever was a sensible rule, this is it.
if you EVER get called for a meeting and they say “don’t bring a union rep”
bring a union rep
bring a union rep
'The University and College Union (UCU)’s Edinburgh branch said staff had been called into “informal meetings at less than 24 hours' notice” and told not to bring union representatives.'
September 18, 2026 at 3:12 PM
If there ever was a sensible rule, this is it.
Reposted by Karsten Konrad 🇪🇺
I had a student ask the other day why we don't multiply vectors in the expected way (component-wise). So, <a,b,c>⊙<d,e,f>=<ad,be,cf>. My gut response was that we could, but it isn't as useful as the dot product or the cross product. But I decided to think about it and see if I could come up with
1/6
1/6
September 18, 2026 at 1:52 PM
I had a student ask the other day why we don't multiply vectors in the expected way (component-wise). So, <a,b,c>⊙<d,e,f>=<ad,be,cf>. My gut response was that we could, but it isn't as useful as the dot product or the cross product. But I decided to think about it and see if I could come up with
1/6
1/6
Test driving it now. Full Qwen still a bit too slow for my taste, but this might do the trick.
it's so nice to have a capable local model i can recommend to people in my life who only have normal laptops, super valuable drop. check out the new prism-ml quantization of Qwen3.8-27b. 6GB weights!
blog: prismml.com/news/bonsai-...
webgpu (browser-native!) demo: huggingface.co/spaces/webml...
blog: prismml.com/news/bonsai-...
webgpu (browser-native!) demo: huggingface.co/spaces/webml...
September 18, 2026 at 8:53 AM
Test driving it now. Full Qwen still a bit too slow for my taste, but this might do the trick.
Reposted by Karsten Konrad 🇪🇺
Well, that didn’t take long. And its probably like this - Jev is an LLM based on diffusion, trained to the max on synthetic classification tasks. Without reasoning and tools, what we have is a fast cheap and often dumb predictor. For which I’d still like to see real use cases.
Most convincing one so far from @mmastrac.bsky.social ! x.com/mmastrac/sta...
September 18, 2026 at 5:57 AM
Well, that didn’t take long. And its probably like this - Jev is an LLM based on diffusion, trained to the max on synthetic classification tasks. Without reasoning and tools, what we have is a fast cheap and often dumb predictor. For which I’d still like to see real use cases.
Whenever I said „learning is compression“, I didn’t mean it quite this literally.
galaxy brain: a LLM is ffmpeg
arxiv.org/abs/2306.04050
arxiv.org/abs/2306.04050
September 18, 2026 at 5:17 AM
Whenever I said „learning is compression“, I didn’t mean it quite this literally.
Reposted by Karsten Konrad 🇪🇺
As a discipline, mathematics needs to adapt to AI with superhuman math abilities, which is why I’m introducing Move2Prove, the world’s first proof-based workout plan with heavy sets, light reps, and a cooldown that's left as an exercise for the reader!
September 18, 2026 at 1:15 AM
As a discipline, mathematics needs to adapt to AI with superhuman math abilities, which is why I’m introducing Move2Prove, the world’s first proof-based workout plan with heavy sets, light reps, and a cooldown that's left as an exercise for the reader!
First impression: Jev is a one-shot LLM classifier that is fast and very cheap. May be confidentally wrong sometimes. Can't web search, or use tools. Possible applications seem to be a bit niche to me, but Bsky AI knowers, do your thing.
September 17, 2026 at 8:22 PM
First impression: Jev is a one-shot LLM classifier that is fast and very cheap. May be confidentally wrong sometimes. Can't web search, or use tools. Possible applications seem to be a bit niche to me, but Bsky AI knowers, do your thing.
Jev is a large language model, not a theorem prover. But it can guess whether a propositional formula is true, false or contingent. Randomly guessing correctly at least 9 times of 12 has a probability of less than 0,4%. Impressive.
September 17, 2026 at 7:53 PM
Jev is a large language model, not a theorem prover. But it can guess whether a propositional formula is true, false or contingent. Randomly guessing correctly at least 9 times of 12 has a probability of less than 0,4%. Impressive.
I am now part of the Jev party.
September 17, 2026 at 7:24 PM
I am now part of the Jev party.
Reposted by Karsten Konrad 🇪🇺
I had access to the new Claude Projects & was able to do some very complex work
Here, I asked it to go through all the images, videos and records about Umberto Eco's famous 33,000 book library & try to reconstruct it, including book locations, in 3D. Not perfect, but wow eco-library-map.netlify.app
Here, I asked it to go through all the images, videos and records about Umberto Eco's famous 33,000 book library & try to reconstruct it, including book locations, in 3D. Not perfect, but wow eco-library-map.netlify.app
September 17, 2026 at 6:22 PM
I had access to the new Claude Projects & was able to do some very complex work
Here, I asked it to go through all the images, videos and records about Umberto Eco's famous 33,000 book library & try to reconstruct it, including book locations, in 3D. Not perfect, but wow eco-library-map.netlify.app
Here, I asked it to go through all the images, videos and records about Umberto Eco's famous 33,000 book library & try to reconstruct it, including book locations, in 3D. Not perfect, but wow eco-library-map.netlify.app
Reposted by Karsten Konrad 🇪🇺
Reposted by Karsten Konrad 🇪🇺
*taps the "high enough dimensionality, interpolation *is* extrapolation" sign*
arxiv.org/abs/2110.09485
arxiv.org/abs/2110.09485
September 17, 2026 at 10:29 AM
*taps the "high enough dimensionality, interpolation *is* extrapolation" sign*
arxiv.org/abs/2110.09485
arxiv.org/abs/2110.09485
Reposted by Karsten Konrad 🇪🇺
Expect some record breaking speeds.
🇦🇺 🐊 A river home to crocodiles has been approved as the 2032 Brisbane Olympics rowing and canoeing venue by the sport's governing bodies after a "thorough and robust assessment".
➡️ u.afp.com/SjbB
➡️ u.afp.com/SjbB
September 17, 2026 at 11:00 AM
Expect some record breaking speeds.
I feel vindicated by using Pi. Claude Code is expensive.
HarnessTax: How Much Does the Harness Matter for Coding Agents? https://harnesstax.github.io/ (https://news.ycombinator.com/item?id=49733726)
HarnessTax: How Much Does the Harness Matter for Coding Agents?
What does a coding-agent harness actually add, and at what cost? It turns out your Claude models may not need Claude Code… We evaluate 21 model–harness pairs spanning seven models and three harnesses—Claude Code, Codex CLI, and Pi—on SWE-bench Lite and Terminal-Bench 2.0.
harnesstax.github.io
September 17, 2026 at 10:59 AM
I feel vindicated by using Pi. Claude Code is expensive.
I can easily imagine being radicalized into being more pro EU after meeting the UK negotiation team members in person.
Has anyone considered Mark Carney was radicalized into the worlds most pro EU man by being directly exposed to Brexit for years on end
'Canada is in the EU and we're not' would be an incredibly funny long term outcome of Brexit
September 17, 2026 at 10:49 AM
I can easily imagine being radicalized into being more pro EU after meeting the UK negotiation team members in person.
Reposted by Karsten Konrad 🇪🇺
I've written a blog post responding to the letter about maths and AI signed by 25 Fields medallists. As with the Leiden Declaration, I didn't sign it, but I agree with much of it and welcome its existence.
gowers.wordpress.com/2026/09/17/w...
gowers.wordpress.com/2026/09/17/w...
Why I didn’t sign the Fields medallists’ letter
When I was around 11 I heard for the first time about Fermat’s Last Theorem. I was immediately captivated by the problem statement, as well as by the accompanying story, and made a fairly ser…
gowers.wordpress.com
September 17, 2026 at 7:01 AM
I've written a blog post responding to the letter about maths and AI signed by 25 Fields medallists. As with the Leiden Declaration, I didn't sign it, but I agree with much of it and welcome its existence.
gowers.wordpress.com/2026/09/17/w...
gowers.wordpress.com/2026/09/17/w...