Iolaum

Care to substantiate that? Not that I disagree, just curious.

My (not thorougly informed) take is that they are are taking measures to have the capability of restriction but are also adding provisions so that this capability does not add friction to the current practice of enabling export of open weight models.

So to me it's like they are keeping their options open rather than leaning on one versus the other.

He didn't say cash, he said assets.

Also funny enough - AAPL is siting on 62B$ cash according to google search.

If only those who are intended to oversee the proper use of those powers could take offense at this misuse ...

It's not gonna happen but:

> Despite that complexity, much of the variation in performance is explained by just a single variable: model size.

please release model size :p

They will not come. You can only spend so much time speaking on the phone.

But if you get unlimited tokens, with scheduled/automated jobs and subagents you will burn a LOT of money.

If I was NSA I 'd create obscure websites that are easy to to parse for crawlers and have poisoned instructions there for ALL LLM's to pick.

Kinda like https://github.com/elder-plinius has been doing but covert and with actually harmful (for targets) instructions.

I used to be surprised in Sci Fi novels with mentions like programmer archaeologist. Makes much more sense now :p

Give me 5B $ in tokens and I 'll work something out :p

US wants to ban non us models from as many places as possible. This move enables them to get users in more places.

It may make it harder but does that make it a blocker in the age of AI?

In the age of RSI, is it really beyond our capabilities to evolve a particular LLM architecture, harness and inference engine towards a place where we can get that understanding?

I m afraid it's more a matter of prioritization. "Good enough" understanding brings as much money as "better" understanding.

Indeed, I m surprised how little attention qwen-3.8.-flash-next seems to be getting while it's so well fitted for 128gb unified machines.

Curious if GrapheneOS phones were also vulnerable to this.

Wild guess even if not in this particular case it's best to consider the possibility as non theoretical.

Companies buying those machines need years to get their money back. If they didn't believe they would make a good ROI they wouldn't buy as much.

You are forgetting how powerful aligned incentives are. If their interests align, they will find a way to reach an agreement.

> actually have to pay taxes

Interested in recommendations of politically stable countries where companies can't practice tax avoidance.

</s>

LLM's have already trained on similar enough code

Not your fault - and thanks for the gift link - but NYT still wanted me to login to read it.

Their prerogative, but no thanks.

Indeed - Even with llama.cpp you can store n-grams to disk and use it. I m using unsloth's llama fork atm but a new llama release is out that may be better - hadn't had time to test it yet.

It's not luck it's starting conditions. People have written books about the benefits of geography and natural resources.

One could argue that all this whole "Pacing the Frontier" bullshit is the industry freaking out regarding the danger of open models.

P.S. That's not to mean there aren't dangers regarding AI. I just don't trust the people making money from selling AI to manage those risks ethically instead of "protecting" us from those risks like pimps but with suits and good manners.

Qwen-3.8-flash-next @Q4

Interesting that he used an open source AI tool to do that: https://github.com/Autumn-27/ARTEX

and that he didn't know enough to cover his tracks and was easily identified.

Biggest reason for an OAI subscription instead of Ant imo.

Biggest loss is that Ant models look like they are genuinely better.

This may be a very good paper but when I saw Apple regarding the authors org I stopped.

With openAI's compute they likely can do the post-training for a decision model on a luna like model in the order of days. I woulnd't be surprised if the stuff before and after that training took more time.

I mean Orinth-1.5 35B has been out for a while and is considered a better model in that space. Not sure why they didn't start from there.

Once it's open weight people will be able to inspect and compare it's tokenizer, architecture etc and tell.

I am really curious why after Qwen-3.8-flash and deepseek-v4.1-flash newer open weight models (or proprietary but they don't tell) don't use n-grams. It looks (to me) like they are a cheep way to add more knowledge to the model.

What do I mean by cheap? You can rely on the SSD to retrieve the relevant tokens as no computation is needed meaning you can leverage storage (or cpu ram if you don't have unified memory) to serve part of the model which (to my understanding) is much cheaper to get than GPU RAM.

Anyone know what am I missing? Or is it that the pace of iteration for labs slow enough that they can't actually leverage it yet?

Don't know what they wanted to convey but I don't like that pure diesel cars are in the 51% bucket instead of the 49% one.