bguberfain

It is good to se big companies like Microsoft launching LLMs. They have large amount of compute power and good scientists to create useful models.

Any plans to port to sglang or vLLM?

It seems to be something related the moving average calculation. So it is just a glitch on the chart.

This guy seems to be talking seriously.

Not to demerit the recording, but I felt more nostalgic for the last sentence of the article "Sometimes, the internet is good" than for the musics itself.

We all know it... but I think they were very bold in this warning about using your private messages to train public models. _Your messages with AIs will be used to improve AI at Meta. Don't share information, including sensitive topics, about others or yourself that you don't want the AI to retain and use_

"A watchdog kernel thread monitors RAM and NVMe pressure and signals userspace before things get dangerous." - which kind of danger this type of solution can have?

We can finally search for playlists with a giving song! A basic feature that Spotify is missing!

So they used a LLM with knowledge cut in mid 2023 to evaluate 2023? Seems like a classic leakage problem.

From paper: "testing set: January 1, 2023, to December 31, 2023"

From the Llama 2 doc: "(...) some tuning data is more recent, up to July 2023."

I think that there may be another solution for this, that is the LLM write a valid code that calls the MCP's as functions. See it like a Python script, where each MCP is mapped to a function. A simple example:

  def process(param1, param2):
     my_data = mcp_get_data(param1)
     sorted_data = mcp_sort(my_data, by=param2)
     return sorted_data

https://blogs.windows.com/windowsdeveloper/2025/05/19/the-wi...

Not available in my country :(

Unfortunately, it uses Miniconda, which does not allow usage in companies with more than 200 employees. I think it conflicts with AGPL license. I created a PR to fix that.

Can you provide more information about this “bigger teacher” model?

Until GPT-4.5, GPT-4 32K was certainly the most heavy model available at OpenAI. I can imagine the dilemma between to keep it running or stop it to free GPU for training new models. This time, OpenAI was clear whether to continue serving it in the API long-term.

Any chance you could release the dataset to the public? I imagine NewsCatcher and Polymarket might not agree..

It remembers me Theano [0].

[0] https://en.wikipedia.org/wiki/Theano_(software)

I agree. One image of what it is doing would improve the comprehension of the algorithm.

The file I mentioned is just the begining... there is a folder full of .dll files, renamed to .pyd. I understand that this is the proprietary part, that limits usage for 30 minutes, but I think it is too closed for a MIT license.

Thanks for sharing this! But I have some doubts about hidden installation procedures. It imports all functions from one_click (from one_click import *), which points to a compiled file. It then runs functions like install_webui and install_extra_packages. At least suspicious.

"Nemotron-4-340B-Instruct is a chat model intended for use for the English language" - frustrating

From the code, it seems to send information to https://vizly-notebook-server.onrender.com/ when in "production". Not so local (src: https://github.com/squaredtechnologies/thread/blob/d450bbf2a...)

It says it has support for Llama 2, but in a deep page you can read that it is "coming soon". Anyway, good to see support for serverless inference and painless train of Llama 2 models!

What brings to my attention in this article is the section named "Cold Start", where it generates questions based on a provided context. I think it is a good way to cheaply generate an Q&A dataset that can later be used to finetune a model. But the problem is that it generates some questions and answers of bad quality. All generated examples have issues: - "What is the context discussing about?" - which context? - "The context does not provide information on what Ray Tune is." - Not an answer - "The context does not provide information on what external library integrations are." - same as before I could only think of manual review to remove these noise questions. Any ideas on how to improve this QA generation? I've tried it before, but with paltry results.

Yep. Just download the safetensor version.

I like the experience of using this tool, but did not work for paywalls. For WSJ, it generated a summary of its pricing offer.

It reminds me Theano

freedom from having to ask seemingly stupid questions With almost no risk of receiving a rude answer

200Tb is an impressive mark for a "future" product, but I was more impressed by the 100Tb offered "today": https://nimbusdata.com/products/exadrive/pricing/

Seeing the first page of these articles always reminds me of this joke: https://twitter.com/hardmaru/status/1483262080329392129