Sure, that’s a good point. Blackmail would indeed have been more likely to succeed, if the affair hadn’t been exposed at this time.

Newsworthy:

• Jensen Huang had an intimate relationship with an alleged Chinese spy

Irrelevant:

• The woman’s modelling career.
• Jensen’s wife and family. It would’ve been just as bad if they had an open relationship, or if he was single.

Just to clear up some potential misunderstanding… Ampere is the Nvidia GPU architecture, and your card is in the same category as RTX30xx. This can matter when it comes to hardware support for llama.cpp and other software.

Having tensor cores affects the speed your card does certain calculations, but AFAIK there are different types of tensor cores in GPUs and if you’re using quantized models (which you most likely will be, unless you’re training your own models) then it’s INT8 performance that matters.

Your amount of VRAM determines how large (which usually correlates to how “intelligent” the model is) model you can fully load on the GPU. If the entire model doesn’t fit in VRAM it can use CPU RAM swapping (partial offloading) which severely harms performance, though MoE models are less affected.

So both VRAM and tensor cores affect generation speed, but in completely different ways.

According to this page ChatGPT would tokenize “seventeen” as [“se”, “venteen”] but AFAIK tokens become multi-dimensional vectors which represents its meaning (all vectors are the same length and the numbers in the vector have nothing to do with the letters that spell it out, just so you don’t imagine something like a C++ string). So an LLM can’t answer how many of each letters there is in a token either, unless explicitly trained to do so. To be able to immediately answer this, I think they would need to include huge amounts of training data to cover all tokens and letter combinations.

But I would have expected that since the strawberry incident, all models would have been trained to do what Qwen tried to do here, and break the word down into letters. Because “s e v e n t e e n” tokenizes to [“s”, “e”, “v”, “e”, “n”, “t”, “e”, “e”, “n”] so then they only have to train the model to be able to convert any token into its individual letter, and to count occurrences of something in an array.

However, if you disallow thinking and force the LLM to answer immediately, then you’re back to the original problem and will likely get an incorrect answer.

I wonder if they didn’t read this section:

There are no entries in the certificate transparency logs for these intermediates in that period, indicating that CT-enforcing clients like Firefox, Chrome, Edge and Safari could not have been attacked.

or if they did read it, but don’t trust it.

Might be more relevant for LoRA training, right?

Would’ve been nice if the model was made to fit in 22GB VRAM, so one could fit it together with a graphical desktop environment on a 24GB GPU… But I guess we’re not NVidia’s intended user base.

Well, it’s just a guess. But also Gemma4 increased to 31B, so I think there’s a possibility that Qwen does the same.

27B dense was a good size for 3.8, but I’m guessing they’re going to add a few B parameters in order to make it noticeably better.

This one supposedly beats Jev at its own game:

https://huggingface.co/Cloudflare/clef (27B)

https://huggingface.co/Cloudflare/clef-flash (9B)

Requires a modded llama-server to run, but I find this project especially interesting. They have extracted the fine-tuning and decision making head from Clef (which is a modded Qwen3.8-27B), to make it possible to run both Clef and normal Qwen at the same time, while sharing the majority of the parameters.

https://huggingface.co/trevest/Qwen3.8-27B-Clef-LoRA-GGUF

It has indeed been really good, so much better than similarly sized models a few years ago. Would be interesting to know how far it is from what is possible to fit into a 27b model.