Vladimir Salnikov

Genuinely curious how it works, like chatGPT responses are not instant (and I assume it take some time to read the response and try to understand it if the question is complex)

Do people just, sit in silence for like couple of minutes or how it works ?

For the sake of my mental health I will leave social media until Danish election season is finished

Had an experience reading news/election campaigns while being immigrant in sweden during election season and it was not fun (to put in mildly)

How Europeans view average American city :

True, I have actually missed the exact point when labs stoped providing logits because of "distillation anxiety", somewhere in early 2024 ?

Oh yeah, somehow forgot about LLM-as-judge even though it was explicitly mentioned

And I agree that you can not really call it distillation in a sense how most people image it

Also realistically

Chinese labs most likely aren't directly training on Claude's outputs (because not 150k, not even 13 million is enough for something like a frontier-level performance lol)

My prediction - most likely they're using Opus outputs as seeds for synthetic data

Ok now I want to distill Opus just out of spite

Ah, didn't know that

Could it because Z.AI have IPOd and they could take more legal action against accusation ?

Nah they should stop pretending this is some "national security" issue

They have cut access to xAI and OpenAI for their coding tools also. It is pretty obvious that Anthropic is afraid of the competition at this point

techcrunch.com/2025/08/02/a...

And like in computer vision models you often end up just randomly guessing towards the end just to get a better score

Be honest, you like it because it is happening in Malmö this year 😂😂

I think your students will definitely appreciate;) tbh I see how it becoming increasingly more important both in research but also industry

Thank you for the link !

Yeah special course might be interesting, also I have still this annoying idea of adding VLM benchmarks to EuroEval

But I am afraid that has to wait at least till Summer because I am totally busy with current studies/research projects 😞

I must admit, that at this point I would strongly benefit from the university course that focuses just on model evaluation (both CV and LLMs)

Like I am taking many courses on how to train stuff, but imho it is as important to be able to evaluate what you have trained

Ahahaha that's a good one

Sweden announced new "AI strategy"

I think it is quite cool that more and more Nordic countries are interested in training open Language Models

news.cision.com/knut-och-ali...

I actually messed up and download a transparent .png but decided to leave it like that because it looks sick 🔥

I think we are slowly moving towards reinventing backpropagation from first principles 😁😁

I mean I think it won't be a problem considering usually target model in speculative decoding can just reject tokens, the issue that it might hurt the performance speedup

Wow, that actually might work pretty well in case where language/framework is fixed

Fast and Cost-effective Speculative Edge-Cloud Decoding with Early Exits Large Language Models (LLMs) enable various applications on edge devices such as smartphones, wearables, and embodied robots. However, their deployment often depends on expensive cloud-based APIs, cre...

arxiv.org/abs/2505.21594

Well quick Googling says that we can

I wonder what if doing speculative decoding this way :

Take smaller/distilled/heavily quantized model and run it locally, then generate "proposal" tokens for larger target model that runs on the cloud

Thanks for the reminder

Finally joined IDA students because of this 😁

a cat is sitting in a cardboard box and looking at the camera . ALT: a cat is sitting in a cardboard box and looking at the camera .

When you have stopped being GPU poor and realized there is whole another PyTorch to learn

It would be nice to see "cost" axis, I think this might be actual moat of open weights/open source models

Also I wanted to thank @dorialexander.bsky.social for an inspiration to work on synthetic data, which will be the focus of my work

I am very grateful to @kennethenevoldsen.bsky.social for the support during the hiring process (and of course mentorship outside of it)

Life update:

In March, I’ll be starting as a Student Developer on the Danish Foundation Models project at Aarhus University

I am excited to work with amazing people and contribute to Danish open source 🇩🇰