Macuyiko

From the article:

"However, the most astonishing thing about this break is that the GPT–6 Astra did it entirely on its own. Carter Leffer only directed GPT–6 Astra to see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page."

I mean... I'm all for collaboration but I think this case is pretty clear, no?

Sarcasm is the lowest form of wit, but...

For what it's worth the original article is in Chinese: https://www.qbitai.com/2026/09/483600.html

The article linked here seems to be a (bad) auto-translated version.

It's not uncommon for images to be incorporated in an article like this, as they typically get shared around WeChat and correspond better with the 'culture'.

What I and my team have been doing in papers is just to refer to the OpenRouter slugs. It upsets reviewers because they will complain it "is not sufficiently clear to a wider audience" but I do agree it's the cleanest approach. Also goes to prove how much power OpenRouter has actually...

This reminded me of coming across LeoCAD (https://www.leocad.org/ - first release in 1997!) as a kid and playing around with it for endless nights, wondering how close it was to whatever tool Lego was using in-house.

Looking at screenshots of older versions gave a burst of nostalgia.

There is - to the standard Quake .map format Neverball uses.

Reminded me strongly of the paper "Deep Learning is Not So Mysterious or Different" from a year ago: https://arxiv.org/abs/2503.02113

Things such as AirLLM, or good old llama.cpp.

The input attribution part is interesting though, but I do wonder to which extent that is just assigning some sort of SHAP values to the input tokens, in which case it should be pretty portable to any kind of model.

What I typically end up doing is just recalc the slug and see if it matches the provided one. If it doesn't redirect to the most up to date slug matching the id. Though who knows if those old SEO patterns still matter these days...

My personal feel (completely subjective) is that during RLHF humans are incredibly sensitive to this pattern, especially when talking about personal or emotional issues. Any reply in the form of "it's not you, it's them" is such a dopamine hit that the LLMs started applying it for everything else.

Oh I see what you mean now, indeed:

    Score: 7
    ~~~~~~
    ~····~
    ~·~~·~
    .#..#.
    ......
    ..#...
    .#H#..
    ..#...

However, I think that you do not need 'time' based variables in the form of

    reachable(x,y,t) = reachable(nx,ny,t-1)

Enforcing connectivity through single-commodity flows is IMO better to enforce flood fill (also introduces additional variables but is typically easier to solve with CP heuristics):

    Score: 2
    ~~~~~~
    ~....~
    ~.~~.~
    ......
    ......
    ..##..
    .#H·#.
    ..##..

Cool puzzle!

Good point. I don't think the puzzles do this and if they would, I would run a pre-solve pass over the puzzle first to flood fill such horseless pockets up with water, no?

Yes. CP SAT crunches through it in no time, but of course larger grids would quickly make it take much longer.

See

https://gist.github.com/Macuyiko/86299dc120478fdff529cab386f...

Late, but reading all of the replies, and speaking from my own observation using Claude, Codex, as well as (non-CLI) Gemini, Kimi, Qwen, and Deepseek...

It's fun how we are so quick to assign meaning to the way these models act. This is of course due to training, RLHF, available tool calls, system prompt (all mostly invisible) and the way we prompt them.

I've been wondering about a new kind of benchmark how one would be able to extract these more intangible tendencies from models rather than well-controlled "how good at coding is it" style environments. This is mainly the reason why I pay less and less attention to benchmark scores.

For what it's worth: I still best converse with Claude when doing code. Its reasoning sounds like me, and it finds a good middle ground between conservative and crazy, being explorative and daring (even although it too often exclaims "I see the issue now!"). If Anthropic would lift the usage rates I would use it as my primary. The CLI tool is also better. E.g. Codex with 5.1 gets stuck in powershell scripts whilst Claude realizes it can use python to do heavy lifting, but I think that might be largely due to being mainly on Windows (still, Claude does work best, realizing quickly what environment it lives in rather than trying Unix commands or powershell invocations that don't work because my powershell is outdated).

Qwen is great in an IDE for quick auto-complete tasks, especially given that you can run it locally, but even the VSCode copilot is good enough for that. Kimi is promising for long running agentic tasks but that is something I've barely explored and just started playing with. Gemini is fantastic as a research assistant. Especially Gemini 3 Pro points out clear and to the point jargon without fear of the user being stupid, which the other commercial models are too often hesitant to do.

Again, it would be fun to have some unbiased method to uncover some of those underlying persona's.

On the homepage it says "Sinmple" above "Export SQL", fyi

A coin measurer is still my goto explanation. Especially with most models having an inset for the coin to rest on / fit in. The hole itself is then just to quickly/easily get the coin out again with your finger.

With so many different coin sizes and types in the empire, I think this makes most sense.

Wikipedia also mentions this:

> Several dodecahedra were found in coin hoards, suggesting either that their owners considered them valuable objects, or that their use was connected with coins — as, for example, for easily checking coins fit a certain diameter and were not clipped.

I've noticed that puzzles that can be solved with CP-SAT's presolver so that the SAT search does not even need to be invoked basically adhere to this (no backtracking, known rules), e.g.:

    #Variables: 121 (91 primary variables)
      - 121 Booleans in [0,1]
    #kLinear1: 200 (#enforced: 200)
    #kLinear2: 1
    #kLinear3: 2
    #kLinearN: 30 (#terms: 355)

    Presolve summary:
      - 1 affine relations were detected.
      - rule 'affine: new relation' was applied 1 time.
      - rule 'at_most_one: empty or all false' was applied 148 times.
      - rule 'at_most_one: removed literals' was applied 148 times.
      - rule 'at_most_one: satisfied' was applied 36 times.
      - rule 'deductions: 200 stored' was applied 1 time.
      - rule 'exactly_one: removed literals' was applied 2 times.
      - rule 'exactly_one: satisfied' was applied 31 times.
      - rule 'linear: empty' was applied 1 time.
      - rule 'linear: fixed or dup variables' was applied 12 times.
      - rule 'linear: positive equal one' was applied 31 times.
      - rule 'linear: reduced variable domains' was applied 1 time.
      - rule 'linear: remapped using affine relations' was applied 4 times.
      - rule 'presolve: 120 unused variables removed.' was applied 1 time.
      - rule 'presolve: iteration' was applied 2 times.

    Presolved satisfaction model '': (model_fingerprint: 0xa5b85c5e198ed849)
    #Variables: 0 (0 primary variables)

    The solution hint is complete and is feasible.

    #1       0.00s main
      a    a    a    a    a    a    a    a    a    a   *A* 
      a    a    a    b    b    b    b   *B*   a    a    a  
      a    a   *C*   b    d    d    d    b    b    a    a  
      a    c    c    d    d   *E*   d    d    b    b    a  
      a    c    d   *D*   d    e    d    d    d    b    a  
      a    f    d    d    d    e    e    e    d   *G*   a  
      a   *F*   d    d    d    d    d    d    d    g    a  
      a    f    f    d    d    d    d    d   *H*   g    a  
     *I*   i    f    f    d    d    d    h    h    a    a  
      i    i    i    f   *J*   j    j    j    a    a    a  
      i    i    i    i    i    k   *K*   j    a    a    a

Together with validating that there is only 1 solution you would probably be able to make the search for good boards a more guided than random creation.

All of the above is true, but between solving quicker, and admitting we gave context:

I do agree with you that an LLM should not always start from scratch.

In a way it is like an animal which we have given the ultimate human instinct.

What has nature given us? Homo Erectus is 2 million years ago.

A weird world we live in.

What is context.

Weirdly it has gotten so far that I have embedded this into my workflow and will often prompt:

> "Good work so far, now I want to take it to another step (somewhat related but feeling it too hard): <short description>. Do you think we can do it in this conversation or is it better to start fresh? If so, prepare an initial prompt for your next fresh instantiation."

Sometimes the model says that it might be better to start fresh, and prepares a good summary prompt (including a final 'see you later'), whereas in other cases it assures me it can continue.

I have a lot of notebooks with "initial prompts to explore forward". But given the sycophancy going on as well as one-step RL (sigh) post-training [1], it indeed seems AI platforms would like to keep the conversation going.

[1] RL in post-training has little to do with real RL and just uses one shot preference mechanisms with an RL inspired training loop. There is very little work in terms of long-term preferences slash conversations, as that would increase requirements exponentially.

A bit of a rant, but this is the kind of fact checking I wish the media and all our EU "trusted sources" would have jumped on instead of going for the most trivial and idiotic cases only a toddler (or a journalist) would get stumped by. (Example: recent posts on Tiktok 'claiming to be images from Pakistan but taken from Battlefield 3...' again. Who is impressed or even surprised by this kind of investigation?)

Much more interesting, but also with more effort required, so of course it never happens.

It would have a more beneficial societal effect, because it is this kind of article, neutrally written, deep investigation, that truly would make people capable to self-discover "maybe I should question a bit more things".

The model seems to be viewable here:

https://netron.app/?url=https://madebyoll.in/posts/world_emu...

From an age perspective (but the crowd here will not like that): before I trusted myself I could always find it back so I don't need to save it. Now I can't anymore, but I don't care so much.

I am not so sure, but indeed it is perhaps also a sad realization.

You compare this to "a human" but also admit there is a high variation.

And, I would say there are a lot humans being paid ~=$3400 per month. Not for a single task, true, but for honestly for no value creating task at all. Just for their time.

So what about we think in terms of output rather than time?

Some more interesting approaches in the same space:

- https://github.com/openai/evolution-strategies-starter

- https://cloud.google.com/blog/topics/developers-practitioner...

And perhaps most close:

- https://weightagnostic.github.io/

Which also showed that you can make NNs weight agnostic and just let the architecture evolve using a GA.

Even though these approaches are cool and NEAT even is somewhat easier to implement than getting started with RL (at least that is what based on so many AI Youtubers starting with NEAT first) they didn't ever seem to fully take off. Although knowing about metaheuristics is still a good tool to know IMO.

A few weeks ago I was planning to design a model I could send to a local 3d printer to replace a broken piece in the house for which I knew it would be impossible to find something that would fit exactly.

I looked around through a couple of open source/free offerings and all found them frustrating. Either the focus on easy of use was too limiting, the focus was too much on blob, clay-like modeling rather than strong parametric models (many online tools), or they were too pushy to make you pay, or the UI was not intuitive (FreeCAD).

OpenSCAD was the one which allowed me to get the model done, and I loved the code-first, parametric-first approach and way of thinking. But that said I also found POV-Ray enjoyable to play around with around the 2000s. Build123D looks interesting as well, thanks for recommending that.

I follow RL from the sides (I have dabbled with it myself), and have seen some of the cool videos the article also lists. I think one of the key points (and a bit of a personal nitpick) the article makes is this:

> Thus far, every attempt at training a Trackmania-playing program has trained the program on one map at a time. As a result, no matter how well the network did on one track, it would have to be retrained - probably significantly retrained

This is a crucial aspect when talking about RL. Most of the Trackmania AI attempts focuses on a track at a time, which is not really a problem since they want to, given an individual track, outperform the best human racers.

However, it is this nuance that a lot of more business oriented users don't get when being sold on some fancy new RL project. In the real world (think self-driving cars), we typically want agents to be way more able to generalize.

Most of the RL techniques we have do rather well in these kinds of constrained environments (in a sense they eventually start overfitting on the given environment), but making them behave well in more varied environments is way harder. A lot of beginner RL tutorials also fail to make this very explicit, and will e.g. show how to train an agent to find the exit in a maze without ever trying it on a newly generated maze :).

Very disheartening. HF is doing so much good in the AI community, much more than regulators understand at the moment.

Have a look at https://arxiv.org/pdf/2306.11695.pdf which also uses the norm of inputs based on calibration

Wow, this brought back memories. I could swear I wrote a blog post about this years ago but couldn't find it.

A quick search on the local file system revealed `vnccrawl/crawler.py` from 2016 [1] using what looks like a Shodan data dump and calling out to `vncviewer.exe`. I remember randomly logging into some instances and also seeing a lot of cool random systems, including a lot of them controlling industrial systems. Guess I never ended up writing that post.

One would think that on today's Internet it would take only a couple of seconds for those to get compromised, but obfuscation as security, perhaps?

[1]: A random tip from that file: Using a password of 12345678 gives access to way more 'weakly secure' instances.