iiinova — where your digital self keeps evolving

to show my "intellectual weakness". If you could elaborate what this correction of my "state of the art" contributes now to the discussion I am happy to go deeper on this. Also, I do this discussion because I want to find something that falsifies my latest impression of this technology.

000

I don't work for free. I want to discuss a complicated topic. But if I have to explain and defend every word this leads to nothing or should we discuss now the well known saying "state of the art". I think this did not contribute anything meaningful, sorry. It was not meant to offend you or

100

I agree in all points! That's what makes this so interesting for me. If LLMs show intelligence, intelligence is something we can discover and describe. Also, I have to deal with them. They disrupted my field completely (document processing, business process automation and software development).

010

We see now the first steps of these LLMs to generalize and that is where it starts to get interesting in terms of intelligence. For me again the more intersting view is what this tells us about intelligence - even more than what it tells us about artificial neural net models.

000

I was opposed the last decade to call anything intelligent produced by a nerual net, because it could not generalize and abstract anything. Same with deterministic software - it doesn't process anything in a unexpected or even novel way (at least if you do it right).

100

Speech is highly connected to intelligence to a way that it is hard to separate both topics from each other. I think that is what we see in these vast neural nets that are trained on all kinds of data but mainly text. It is not only generic text it is ALL text we have available.

100

Yeah, 'state of the art' is a saying and not literal. If you find something more intelligent than our brain, let me know.

100

That's why the benchmark is median human performance - I think you pointed it our already :) I mean, these tests are probably also not the absolute proof. They are working on version 4 now since 3 was solved.

010

These are fundamental questions and a reason why it is so hard to understand and measure. To state only organic brains can produce intelligence is not a scientific view unless you have a source? Intelligence is discussed and has more nuances. That our brains are state of the art is undisputed.

210

ARC Prize - What is ARC-AGI?

They can be checked but here it is only a review not an approval. We talk also about the first signs of intelligence and not about solving all real world problems. They put some effort in designing these tests: "Core Knowledge Priors" arcprize.org/arc-agi#:~:t...

arcprize.org

The only AI benchmark that measures AGI progress.

200

Brain in a vat - Wikipedia

That is a wrong assumption. Our brains only have sensors and same can be done for LLMs. A LLM detects if it processes a picture a video or data from a sensor. en.wikipedia.org/wiki/Brain_i...

en.wikipedia.org

200

Demiurg @d3miurg.bsky.social · 11h

If Trump gets POTUS twice my dog is over qualified for any job.

000

Demiurg @d3miurg.bsky.social · 11h

Humans can't as well. Truth is not defined and we are dependent on the source. Current AI can absolutely weight the credibility of a source. The ARC AGI tests are exactly designed for this, though. It is not relevant if a source is true to solve these levels. They are solved or not.

200

Reposted by Demiurg

Steve Klabnik @steveklabnik.com · 13/10/2026

very excited to see this out of A\ this is our thesis at @ersc.io! But we are building a git compatibility layer: ersc.io/blog/what-co... enterprise adoption of new source control is a big ask, so making a smooth bridge from the past to the future is a big part of our offer

ersc.io

What comes after git

Our thoughts about the future of version control

712414

Reposted by Demiurg

Armin Ronacher @mitsuhiko.at · 13/10/2026

If you want to know what agents do these days: this took around 30-60 minutes of active prompting max. It’s a mobile iOS app that runs Python via WASM, ships ruff and ty, has a debugger and iCloud integration including a UI library inspired by solid and turtle. And a REPL.

2464

Demiurg @d3miurg.bsky.social · 13/10/2026

It is even exhausting. I get lectured by people 'how LLMs acshually work' just to find out they only did some chats with the gpt interface and have a lot of opinion...

000

Thank you for clarifying this for me.

000

Simon Willison @simonwillison.net · 12/10/2026

And since this sentence would have made no sense in the world of Turing, we need better tests.

100

Reposted by Demiurg

Shared some notes on Python 3.15, using it with uv, and features from Python 3.11 I'm excited to use now that 3.10 is no longer the minimum supported version for my own projects simonwillison.net/2026/Oct/12/...

simonwillison.net

Hello Python 3.15, goodbye 3.10, and now I can use 3.11

Python 3.15 is here! Here’s the essential What’s new in Python 3.15 guide, this year curated by 3.14 and 3.15 release manager Hugo van Kemenade. As someone who builds open …

3664

I provided objective tests and my experience seems to be different than yours. But I did only study chemistry and philosophy and train models for about a decade. So, what do I know.

000

I am very grateful I had a conversation with a person that obviously has such expertise in all these fields.

000

Sorry, you conclude it like this. Your apporoach was to lecture me on how well this technology is understood and that - despite my links - there is no proof that these models are capable. Have a nice day anyway.

000

Counter arguments so far were thin. I enjoy the dsicussion with you, though.

100

Thus we can not create the perfect 'illusion' of intelligence just to trick users. If it is something intelligent or not - we are about to find out. So far, I think we have strong independent signs that we found something.

200

How can we expect something we didn't know it will happen? It was just that I am quite certain that we do not fully understand the internals of these weights.

200

It was more easy for smaller models and e.g. optical transformer. We could generate heat maps to exactly see which part of an object triggered which layers and generalizations. Latest LLMs have outgrown this by far now.

100

For me the point is that we do not know what the weights express - no matter if it is open or closed. Sure, closed doesn't help at all and is for sure bad for the field but I think the lack of understanding is not restricted to closed models. It is the weights and what exactly they represent.

100

I understood you then wrong when you said: "We can see inside the box. We can see what goes in the box. We can see what comes out. We have to do the equivalent of closing it all up to run it but nothing magical changes when it's closed."

100

But going beyond these basic traces leaves us without a clue about the symmetries, how and why they emerge and what we can derive from this.

Screenshot of one part of the video where you see a matrix of Fourier symmetries only emerging when the so called "grokking" effect takes place.

100

The video explains it nicer than I could do - but when we follow the paths of tokens going through LLMs it seems they form Fourier symmetries. Only when these emerge the output starts to generalize. The video explains how this was done for basic arithmetic functions.

100

Sure! I would not advocate for them! I am only assuming here and ofc I do not have any internal information.

000

The most complex model we actually understand

The symmetries that are found are beyond our understanding: www.youtube.com/watch?v=D8GO...

youtube.com

YouTube video by Welch Labs

100

Not sure I agree on all. I do believe the expertise of Mr. Chollet on this topic. The paper I linked in the other thread shows the limits of mechincal i/o observability. It is very complex to trace every single token through the network and understand what each layer contributes.

100

They are the providers of frontier models in this case. OpenAI, Anthropic and probably to some degree Spacex monstrosities.

100

Sure, we can't say if it is just marketing or if they have something in the basement we probably do not want to find out. The latest and very fast advances in model capability are suspicious, though, and it might imply that they use techniques for improvement they do not fully understand themselves.

100

We only programmed the way the training data shapes the weights. We lack the full understaning of all processes leading to the outputs we currently have: arxiv.org/pdf/2501.16496 It also seems latest models reason more in their latent space. forklog.com/en/researche...

arxiv.org

100

They do test with different harnesses. These regulate how the state is kept for the model.

100

Arc AGI is games humans can easily understand and solve. The test is if LLMs can do the same. Up until very recently they were really bad at this but now, Astra solved it and more than that it solved it with a higher efficiency as the median human.

100

About the ARC Prize Foundation

Arc AGI is a non profit organization around François Chollet - he is a well know researcher and founder of Keras (and more). arcprize.org/about

arcprize.org

ARC Prize - Mission

000

I also think, this is exactly what they warn us about. If we let these models improve themselves, we might get in trouble fast, since we haven't understood the basics and they could get too good to really control what they do by simply brute forcing further down this path.

100

The way these models work lacks observability. It is there and we can measure it - it just is a very complex thing to do before CoT tokens are generated. But it might be we have found a scalable approach and that would mean that we can go further down the way to find even more intelligent models.

200

Solving ARC AGI 3 was the next level of capabilities for these models, which came way faster than expected. That it can do this now better then the median human might be the first glimpse of intelligence and that is very interesting about intelligence itself.

200

This raises the question, how to measure this? And we have the fundamental problem that intelligence itself is not a hard defined thing we can use to run these models against it. So, I guess ARC AGI is a good approach.

200

I think it is safe to claim that LLMs now imitate speech to a degree that we can not clearly state if it is a human or not, right? So, to be able to "talk like a human" is no longer the baseline for a test for intelligence - we are beyond that.

300

The most complex model we actually understand

It is very limited how far we can follwo this - unless you have another source? www.youtube.com/watch?v=D8GO...

youtube.com

YouTube video by Welch Labs

100