rawrawrawrr

Non-blogspam link: https://qwenlm.github.io/blog/qwq-32b-preview/

discussion: https://news.ycombinator.com/item?id=42259184

I wish I knew that there existed a public quantum AI company as an easy short before they went out of business.

The model page says the prompt is:

The system prompt used for training this model is:

   You are a world-class AI system, capable of complex reasoning and reflection. Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.

from: https://huggingface.co/mattshumer/Reflection-70B

Nice! It would be a better benchmark to compare this prompt (w/ gpt-4o, claude) with whatever the original model was compared to.

Google Colab has this, I wouldn't be surprised if there was a Jupyter widget to implement something similar.

Edit: looks like Mercury (A jupyter extension) has them: https://runmercury.com/docs/input-widgets/

Title says "for everyone", but post says "Phind-405B is available now for all Phind Pro users". I guess everyone on earth has paid for Phind :)

Are you a VC? If they really didn't care about their investment exits, that would be crazy.

Can run on CPU (slower) or AMD GPUs.

Might have issues if you're from Texas or Illinois due to their local laws.

Video games, for one.

Parody is covered under fair use.

That's kinda hilarious, because I think the book was exactly wrong in its predictions. This can be evidenced by the continuous failures of his AI company, Numenta.

Ramen Lord is an awesome guy. The sheer devotion is something to marvel, as someone who is constantly distracted by Reddit/Hacker News/latest tech gizmo of the day.

Fun fact: in Japan, ramen is considered (Japanized) Chinese, but Japanese in the US. It's an interesting game of cultural telephone.

Babel is a decent sci-fi book, it's actually quite anti-capitalist. I'm not making a political statement.

Did you write your own VSTI, or what did you use?

It's not a big deal imo. I started on digital piano, moved over a few years to a real piano. The brain adapts quickly.

It depends. Right now once we hit 6-8 bit precision inference, H100s/A100s are not memory-bound, but compute-bound.

Git is to SVN as Sapling is to Git. It's really a great tool, makes version control a lot more productive imo.

If you're planning on using the brake pedal, then that's not one pedal driving.

> Chinese foreign ministry spokesman Wang Wenbin on Tuesday urged the Netherlands "to be impartial, respect market principles and the law, take practical actions to protect the common interests of both countries and their companies and maintain the stability of international supply chains".

This amuses me greatly. I had no idea China is explicitly pro-market capitalism.

Fun fact, a Nvidia RTX 4090 has ~1TBps memory bandwidth, for an apples to oranges comparison.

It's research, not meant for commercialization. The main point is in the process, not necessarily the output.

Doesn't seem like there's anything wrong with the methodology in the quote. It's perfectly fine to compare two things, even if one is not meant for the task.

$6 billion dollars raised in 2 months for their series C is blowing my mind. What does Anthropic have that OpenAI or other LLM startups don't? What other companies have raised that much in a single round?

I understand that this is an ad for airsequel, but as a musician and programmer I have no idea what this is. What kind of sheet music does it accept? Do I need to manually enter in all the fields?

This is fine, its a standard enterprise marketing technique. Watch a webinar, get a gift card.

The old design made it easy to copy paste some CSS to use Google's CDN to add fonts to a website. With the new design, I have no idea how to do this. Am I stupid?

Edit: Pressing Select [font name] + adds it to a "shopping cart", where it displays the code to copy it to a website. I am dumb.

I found it quite useful for summarizing the last X hours of chat messages.

To preempt this drama, Phind claims that they were inspired by WizardLM's technique of generating datasets using GPT-4, but didn't use their model or data.

> Not a bad context

A little understated, this is state of the art. GPT-4 only offers 32k.