sunshadow

Too late, we all went to cursor.

I can make this in couple of weeks, but I need an incentive other than just having fun

What made you use caddy in front of tiny-http? I guess tiny-http should be still fine on its own. genuinely curious

They made 11B$ last year. It has incredible amount of ads. If you haven't noticed, then that means they did a great job. (tip: look for the custom logo pins in the map. Its printing money)

According to the license, this repo is not that different than me just extracting the source code from the binaries. Pass.

There is no difference between this and a well known bot prevention mechanism, from the scraper perspective.

I don't see any difference than browserless?

If you're looking for open-source function calling: Checkout Functionary: https://github.com/MeetKai/functionary/

It's MIT licensed. Supports both CPU and GPU

Yes, Functionary: https://github.com/MeetKai/functionary/ MIT Licensed

Its actively maintained as open-source. You can even run it with llama.cpp

Original project: https://github.com/xenova/transformers.js

Unless you start a war against scrapers, you don't need to worry about that as I'll always find a way to scrape your site as long as its valuable to 'me'. Even if it requires Real browser + OCR :)

Crul looks nice, though, you cannot imagine how many startups that I've seen failed doing a very similar thing as Crul. Wouldn't rely on it. The problem is complex: Humans generating messy pages

You don't use XPath&CSS selectors at all (Except if you dont have choice). You rely on more generic stuff, e.g, "the button that has 'Sign in' on it":

    await page.getByRole('button', { name: 'Sign in' }).click();

See playwright locators: https://playwright.dev/docs/locators

These days, I'm not even using Go for scraping that much, as the webpage changes makes me crazy and JS code evaluation is a lifesaver, so I moved to Typescript+Playwright. (Crawlee framework is cool, while not strictly necessary).

Its been 8+ years since i started scraping. I even wrote a popular Go web scraping framework previously: (https://github.com/geziyor/geziyor).

My favorite stack as of 2023: TypeScript+Playwright+Crawlee(Optional) If you're serious in scraping, you should learn javascript, thus, playwright should be good.

Note: There are niche cases where lower-level language would be required (C++, Go etc), but probably only <%5

It would be nice to see it open source. As im concerned about the privacy policy as its a really valuable information

Give it 1 year

But its too expensive to become practical with the OpenAI API. Also, demo is cool until you see the real-world webpages, then you'll realize that this only works less than %50 of webpages.

You're good until this is cheaper than your salary.

I don't understand why people still spend time on jailbreaks of the proprietary models, while they can easily use uncensored open-source models these days. I feel like its kind of waste of time.

Unfortunately its gonna be abused like hell so you'll end up shutting it down or requiring API keys for each one of them :(

I would love to see the code to verify this is not a confirmation bias.

Do you have any explanations on why this performed better than Llama 2?

Mixture of Experts is actually some sort of ensembling

Unfortunately the license makes this somewhat useless. Hope they realize that and change it.

Checkout: https://github.com/facebookresearch/llama/blob/4d92db8a1db6c...

What are the benefits over https://github.com/microsoft/guidance/ ?

I think you need a small video and much more clear and simple marketing material. For example, these don't mean anything for me: "Boxcars, CableReady, and StimulusReflex"

I find it amusing that you felt you had to write "written organically" to claim authorship.

I've asked "are you using palm 3": It said: I am using the Palm 3 model. Palm 3 is a large language model...

Don't believe it :) Also, In the technical report, It mentions multiple languages, I've asked in Turkish which was supposed to be supported, but wasn't able to answer.

Even if its PaLM 2, its hard to trust to the model itself.

What is the page number you're referring to? If its 9, then I believe its talking about optimal numbers per token, not the real numbers that the model is trained on.