That reminds me of the toilet paper calculator during peak covid.

He put up a site in about half an hour and it ended up getting coverage on TV which led to ~10 million visitors and it making $5,000 a day in ads for a while. The core logic of the app is 6 lines of JavaScript to take a few inputs and do basic math on them.

I ended up chatting with the creator on how he built and hosted it back in 2020: https://runninginproduction.com/podcast/35-determine-what-yo...

That’s a mistake I see too many people make. They try to solve a problem and build a product from afar, using their imagination or trying to copy an existing product without having personal relationships with people having that problem (or without having it themselves). I guess it’s possible to create a product first and then find people who need it and iterate from there, but I’ve found that having personal relationships and then making something in response to a common problem they have is the only way to start and grow. That’s what I’ve found “talk to users” actually means.

Typst has been pretty amazing, and at my organization, we’re very happy with it. We needed to generate over 1.5 million PDFs every night and experimented with various solutions—from Puppeteer for HTML to PDF conversions, to pdflatex and lualatex. Typst has been several orders of magnitude faster and has a lighter resource footprint. Also, templating the PDFs in LaTeX wasn’t a pleasant developer experience, but with Typst templates, it has been quite intuitive.

We’ve written more about this large-scale PDF generation stack in our blog here: https://zerodha.tech/blog/1-5-million-pdfs-in-25-minutes

Reflecting on this, I kind of feel like the opposite has actually been true for the last 10+ years?

Early in my career, I saw many examples of homegrown AB test frameworks, event tracking frameworks, alerting tools, monitoring tools, etc. Now it seems like everyone just buys LaunchDarkly, Segment, DataDog, etc. Other than admin panels, I have a hard time thinking of any truly "internal" tools at my last few companies (FAANG was the exception, but they are big enough that I think it genuinely makes sense for them to build their own tools for things like applicant tracking, meeting booking, etc).

Maybe I'm just forgetting something - do other ppl have examples "internal tools"?

I integrated the function calling feature into my personal project and wrote a blog post about it here:

https://letscooktime.com/Blog/ai,/machine/learning,/chatgpt,...

Hopefully this saves you some time!

You can use sips together with iconutil to generate a complete .icns file for your app from a single 1024 by 1024 PNG without any third party software:

    mkdir MyIcon.iconset
    cp Icon1024.png MyIcon.iconset/icon_512x512@2x.png
    sips -z 16 16     Icon1024.png --out MyIcon.iconset/icon_16x16.png
    sips -z 32 32     Icon1024.png --out MyIcon.iconset/icon_16x16@2x.png
    sips -z 32 32     Icon1024.png --out MyIcon.iconset/icon_32x32.png
    sips -z 64 64     Icon1024.png --out MyIcon.iconset/icon_32x32@2x.png
    sips -z 128 128   Icon1024.png --out MyIcon.iconset/icon_128x128.png
    sips -z 256 256   Icon1024.png --out MyIcon.iconset/icon_128x128@2x.png
    sips -z 256 256   Icon1024.png --out MyIcon.iconset/icon_256x256.png
    sips -z 512 512   Icon1024.png --out MyIcon.iconset/icon_256x256@2x.png
    sips -z 512 512   Icon1024.png --out MyIcon.iconset/icon_512x512.png
    iconutil -c icns MyIcon.iconset

As a bonus, generate .ico with ffmpeg:

    ffmpeg -i MyIcon.iconset/icon_256x256.png icon.ico

Incidentally, does anyone know enough about the way sips scales PNGs to confirm that it makes sense to create the 16px version straight from 1024px, as opposed to basing it off 32px (and all the way up)? I.e., is it better to downscale in fewer steps (as currently) or in smaller steps?

> My script read through each of the products we had responses for, called OpenAI's embedding api and loaded it into Pinecone - with a reference to the Supabase response entry.

OpenAI and the Pinecone database are not really needed for this task. A simple SBERT encoding of the product texts, followed by storing the vectors in a dense numpy array or faiss index would be more than sufficient. Especially if one is operating in batch mode, the locality and simplicity can’t be beat and you can easily scale to 100k-1M texts in your corpus on commodity hardware/VPS (though NVME disk will see a nice performance gain over regular SSD)

>It took me a couple years to realize the smart people are just playing the game, the unsuspecting losers are "playing it straight" and getting endlessly frustrated.

This essay gets linked to a lot on here, but you might be interested in Rao's "Sociopaths/clueless/losers" taxonomy: https://www.ribbonfarm.com/2009/10/07/the-gervais-principle-...

I found that, as a person who (sometimes) creates things, my attitude towards other people who create things has significantly shifted.

When I was younger I would just see the fault -- the typos, the narcissism, the pointlessness, the already-been-done-better-by-others -- of people's work. But after trying a bit, it sometimes (for the same reasons the OP talks about) creating anything new ends up being really hard in part because while you struggle everyone is shouting "lame!" from the sidelines (and doing nothing). You see this a lot on this site in particular -- the "middlebrow dismissal" HN is famous for.

So now, even when I see something that I genuinely think is lame, I think of this problem and keep my negativity to myself. I might give useful criticism but I try to keep the balance in mind. Even if you publish something you've made that is kinda lame, I still think it is awesome that you are taking the effort to finish something. The best way to get better is by making multiple attempts.