johndough

Why not? The model is 2.8T parameters with native MXFP4, which is 1400GB.

> Once the tide drops enough for hardware lead times to shorten to weeks

Which will not be any time soon according to SK Hynix CEO:

> We still forecast that customer demand will remain higher than our supply capacity even beyond 2030

https://www.reuters.com/world/asia-pacific/sk-hynix-ceo-sees...

> GPUs become obsolete in 5 years

The GPUs are far from worthless after 5 years. E.g. the A100 80GB PCIe version cost around $15k when it was introduced in 2021 and now sells for $10k used.

Things might be slightly worse for the data center servers, but I am sure they will find find buyers.

Thanks for pointing that out. I haven't noticed any adds in years with Firefox and Ublock Origin extension.

I'll look for a better image host in the future. I guess the economic incentives makes them all turn bad after a while.

Oops, autocorrect. Sorry about that.

Another point for consideration: Specialized SVG models create way better looking pelicans riding a bicycle. (E.g. Refract V4: https://jumpshare.com/s/8liB7Aiuoo3yucbWGXjZ mirror: https://postimg.cc/McV70p84 )

I would not put any trust into that AI-generated "research". It does not control for how OpenAI, Anthropic and Moonshot are doing tokenization and token healing, so the results are meaningless.

Neat, I did not know of that functionality. Thanks!

I am not arguing that Kimi did not distill Opus or Sonnet (which they almost certainly have, as opposed to Fable or Sol). I am arguing that even if they did not distill, the model could still identify itself as Opus or Sonnet.

That being said, I'd like to point out a few things:

- The second link (to claude-sonnet-4-2025051) had 17k matches, not just 3k.

- grep.app does not index the entirety of GitHub, so the real number of occurrences of model identifiers is even higher. SourceGraph has a larger index, but the number is so high that it exceeds their result limit of 10k: https://sourcegraph.com/search?q=claude-opus-4-5-20251101&pa...

- 1000 occurrences are already plenty for memorization. Here is a paper that shows over 80% "extractability" with 1000 occurrences for a 6B model. It also shows that extractability scales with model size: https://arxiv.org/pdf/2202.07646

- LLMs have a "mosaic memory", which means that they can combine information from different parts of the training data. https://arxiv.org/pdf/2405.15523 For example, one piece of training data might mention "I am Opus", another might mention "Opus (claude-opus-4-5-20251101)" and another 10000 "claude-opus-4-5-20251101", where the later occurrences strengthen the generation probability of "I am Opus".

- RL can change the model identity easily, so all of the above may not matter at all.

Implementing your own agent is very easy. Here is a minimal agent in 60 lines of Python without dependencies:

https://github.com/99991/MinimalAgent

You only need a single tool to start with. All recent LLMs know how to use bash for reading/writing/editing/executing.

The exact model identifiers appear extremely frequently in code on GitHub.

https://grep.app/search?q=claude-opus-4-5-20251101

https://grep.app/search?q=claude-sonnet-4-20250514

They also appear elsewhere on the internet:

https://trends.google.com/trends/explore?q=claude-opus-4-5-2...

> Zero emissions, silent, no illumination, etc.

Mostly silent. High voltage transmission lines may have a low-frequency hum. You have to stand directly under them to hear it though.

But then again, the identity could also have slipped into the model from other sources during pretraining. The internet is full of "I am Claude": https://grep.app/search?q=i+am+claude and variants https://grep.app/search?q=i%27m+claude

Either way, there's probably no significant portion of Mythos/Fable or Sol in there as OP has stated.

Your comment sounds like an AI-generated advertisement.

If an uneducated acquaintance of yours was about to name their newborn child "Adolf", would you interject that this name might not be such a great idea?

It would be great if this could be combined with quantization-aware finetuning. In my experience, Qwen3.6-27B has much fewer repetitions at Q6 quantization level as compared to e.g. Q4, but that leaves little space for context on my 24GB RTX 3090.

This does not seem to be true currently. In 2025, there were 45356 McDonald's restaurants worldwide and 13706 in the United States, which is about 3.3092.

https://corporate.mcdonalds.com/content/dam/sites/corp/nfl/p...

However, in 2023, the numbers were 41822 / 13457 = 3.1078, which is (almost) within 1 % difference.

https://corporate.mcdonalds.com/content/dam/sites/corp/nfl/p...

This could be another fine example of the Goomba fallacy: https://en.wiktionary.org/wiki/Goomba_fallacy

Another explanation might be a general dislike of big establishments like AI companies and publishers (which glosses over individual authors, but they probably make up a negligible portion of total sales anyway).

Any specific abliterated big models you can recommend?

Do you have any plans on mitigating the privacy consequences of microtransactions? I'm fine with paying for (some) content, but I'd prefer if there weren't some companies using that information to manipulate me or the more impressionable members of our society.

    > GLM-5.2 class models already need 1TB+ of RAM.

If you quantize GLM-5.2 to 4 bit, you can do it in less than 500GB: https://huggingface.co/unsloth/GLM-5.2-GGUF (table on the right)

If you find three finds that also have a 128GB MacBook, you can chain them together (the MacBooks, not your friends) and make it work.

You could also run GLM-5.2 on a single MacBook if you stream the active parameters from disk, but even with speculative decoding, you'd probably only get in the order of 1 token per second, so this is not really practical for most applications.

    > What provider do you use.

OpenRouter with pinned DeepSeek provider or OpenCode Go

    > Why do you trust it with serving full quality?

Quality seems good so far.

    > What harness do you use? Why do you trust it not to have malware (most harnessed are TS apps).

I wrote my own. A minimal harness without dependencies is only 65 lines of Python.

> Its not an issue to adhere with your private money if your product isn't causing damages.

Might happen really easily though. E.g. you install some package which has been compromised, infecting your software product and suddenly all your customer's systems are cryptolocked and you are on the hock for millions of €€€.

Or your db crashes in new and creative ways and your backups don't work for some reason and now your customer lost an expensive contract because critical data that was in your db is gone.

Of course, you can try to foresee every eventuality, but you will indubitably overlook something and probably never make it to market.

Another approach that often works for these kinds of problems and does not require much intelligence:

Work out the first few cases by hand (1,2,6,20 in our case) and then look up the sequence on "The On-Line Encyclopedia of Integer Sequences" (OEIS):

https://oeis.org/search?q=1%2C2%2C6%2C20&language=english&go...

> It’s unclear to me what the advantages of openrouter are but it seems to be a default I see many people talking about here.

The advantage of OpenRouter compared to using API providers directly is that you can switch between API providers without binding your money to a single provider.

The advantage of OpenRouter compared to OpenCode Go is that the price for DeepSeek-V4-Pro and MiMo-V2.5-Pro is better on OpenRouter.

For example, DeepSeek costs $0.435/0.87/0.003625 for 1M in/out/cached tokens (https://openrouter.ai/deepseek/deepseek-v4-pro), compared to an equivalent of $1.74/3.48/0.0145 under the OpenCode Go plan (https://opencode.ai/docs/go/#usage-limits), almost exactly 4x.

But since you get a monthly usage limit of $60 with the OpenCode Go plan for $10 (i.e. 6x), you might still come out ahead if you use it a lot (or use other models, where the pricing difference is smaller or non-existent).

I had a look at eurouter.ai and it seems like an extremely bad offer.

- The prices are ridiculous (15 % markup for free account).

- They have a rate limit of 1000 requests per month, unless you pay 40€ per month for ... what exactly is their value proposition?

- They have a single provider (TensorX) for DeepSeek-V4-Pro, with a cache read cost that is over 100 times higher than DeepSeek ($0.44 vs $0.003625). Notably, I had to look at the TensorX website for that information, since I could not find any information about cached token cost on eurorouter.ai.

> Mythos and other models are not brute-forcing passwords (and with this analogy passwords, ie. systems are the same).

I am not talking about literally bruteforcing passwords (although LLMs are being used for that, too), but bruteforcing passwords and solving verifiable domain tasks have quite a few similarities, especially when considering rule-based and probabilistic bruteforce methods.

> We're not talking about dogs, but LLM systems.

Well, clearly dogs are not LLM systems. It is an analogy. If there is an important point on your mind that makes the analogy break down, feel free to spell it out.

> Mythos is not exploring entire solution space either.

Yes, but weaker models do not find the solution right away, so they need to try more often. But if they only try the same thing every time, they will never succeed, so we need some kind of guarantee that they try something different every time.

> Usually looping is solved by repetition/frequency/presence/n-gram penalties/DRY/min-p sampling, not temperature but we're not talking about small models that have those classes of issues here.

Those might help to reduce looping (at the cost of biasing the generation), but to guarantee that a model can generate all possible generations, we need non-zero probabilities for all tokens, not lower probabilities for likely tokens.

I don't think that is necessarily true.

- With a weaker model, the time to break into the system might grow so larger that it becomes infeasible, similar to how password hashes can be bruteforced, but if the password is long enough, that is not going to happen in our lifetime.

- There might be problems which are inherently unsolvable with a lower level of intelligence. For example, your dog won't derive calculus from scratch, even if it lived forever.

- LLMs might be biased in such a way that they never explore the entire solution space, no matter how many attempts are made. Some models are notorious for getting stuck in a loop, trying small variations of the same approach every time, even though it is doomed to fail. This can be counteracted somewhat with higher sampling temperature, but that hurts reasoning capabilities.

Works for me though, even when using a proxy that is usually blocked everywhere.

Article 53 of the AI Act: https://ai-act-law.eu/article/53/

The definition of a "genral-purpose AI model" is described in more detail in the "Guidelines on the scope of obligations for providers of general-purpose AI models under the AI Act": https://ec.europa.eu/newsroom/dae/redirection/document/11834...