OP here. I read a lot of Wuxia novels, some of them are well translated, and I usually end up giving up on the poorly translated ones.

I tried a few models (Gemini-2.5, Gemini-3.0-pro, Claude Opus 4.5), and found that they all improved the quality quite noticeably, even without being provided the source. The trick was mostly asking them to be super conservative in the changes they made.

Here's a diff if you're curious: https://www.diffchecker.com/DKgmzLdM/

You might enjoy Asara's Personality Basins, it's on a similar vibe: https://near.blog/personality-basins/

It has a lot of good ideas that are probably true, although some are rediscovering things psychology had found out long ago.

Actually you know what can cause big personality change? A good plastic surgery/facial correction. I heard it's actually one of the few things that can do that. And a positive one, too; towards more confidence and extrovertedness (if you think it's 'positive'), a trait most psychologists considered fixed in any other cases. But I can't find the source right now, or if I even heard it. But it makes perfect sense to me. Like a good jaw realignment that fixed your whole face, blaw, suddenly you're this other person.

One of my all-time favorite essays, pairs very well with pg's Cities And Ambition (https://www.paulgraham.com/cities.html), as I wonder about what forms of transcendence various cities encourage you to aim towards.

Mmm interesting.

I'm not sure if it's possible in pure CSS, maybe it is? It relies on creating new SVG elements.

I think including the SVG in whatever website would usually be the simplest way to use it and it's super lightweight.

Feel free to take a look at the SVG generation function here [1] and lmk if you have any ideas to make it happen in pure CSS.

[1] https://github.com/venkr/gradient-gen/blob/main/src/EllipseG...

Hey HN!

Check out this fun tool I built to make cover images for my blog. The main goal: interesting looking random gradients, that all draw from a single color palette.

I stumbled upon this post from Justin Jay Wang [1], who designed for OpenAI - describing methods for random gradients, including the one that shipped on OpenAI’s homepage from 2020-2022.

Since the post included ~10 SVGs, I was able to reverse engineer the SVG parameters, and decided to put up a generator. You can see the exact deets in my Github README [2].

I also use it to generate cover images for my Substack [3] - if you want to see what using a bunch of images generated by it looks like.

[1] https://justinjay.wang/methods-for-random-gradients/

[2] https://github.com/venkr/gradient-gen

[3] https://venkii.substack.com/

Great work! Maybe you can add an option to use the 'heightmap' method described by Justin in his blog post. And maybe let users add their own palettes? :)

Thanks! I'll look into adding custom palettes & the heightmap method. Might not get to it tbh - but I'll take a look.

Looks awesome! It would be good to add a license to your repo!

Done - thanks :)

Really like the thesis implied by the title - kinda unfortunate the post tangents & doesn’t spend more time on that explicit claim.

I’ve briefly written about issues in the past: https://venki.dev/notes/google-gemini-is-bad

My first thought would be, re:

>It’s baffling that a tech giant like Google can’t seem to deliver the same level of reliability

It’s actually not that baffling to me! Shipping simple features gets harder and harder as scale, esp as the features must interact well with more and more existing features + get through bureaucracy.

For example, the regions bit you mention: Anthropic/DeepSeek can just pool all their regions together into a single API and mostly vibe it. But since Google already offers their regular compute by region, it’d be a little weird to not do that for LLMs only.

GCP probably just has a lot of complexity due to being enterprise-oriented, and unfortunately this carries over. Unencumbered startups get to start from a blank state though.

Any sense if applying later will be detrimental?

The main reason to delay: I've started writing relatively recently, and expect I might have more posts I'd showcase in my application by then.

My guess is not much? Because we are doing rolling applications, so we are somehow trying to judge how many good applicants in total we are going to get (classical secretary problem). Applying early means we might let you in with a lower bar if we end up getting a lot of great applications later and raise our bar. Applying later might be better if we realize we were overly conservative in the beginning and are disappointed in the later applications.

Thinking about it, my guess is we will probably let promising people who applied early know that they are on some kind of waitlist and extend an invite to them if we end up disappointed with the later applications, so if you are flexible, I think that makes early strictly easier. I don't expect the effect to be that large though.

Ooh very cool - I've been looking for something like this.

Top feature I'd care about before using this would be the ability to read the actual 10K inline and jump to sources, perhaps as an iframe or so.

Oh you can actually do that in this! the nodes of the map if you click them have references to the actual 10Ks which are readable as html in the space. Not only can you read the original text, but the html has been 'chapterized'; you can also read summaries of chapters. Further, you can query anything in chat - where if you select the document and ask, it retrieves the right passages to answer your questions too.

Kinda the opposite - Firestore documents don't inherently have a userId of the owner. Rather, they just have arbitrary fields.

Making all your documents have a `owner_uid` or `userId` field is just a convention they recommend, because it helps you write rules.

So rather: they have no default system for handling documents that can only be accessed by a given user, but rather you have to construct it using `firestore.rules`, and you end up with something oddly default-insecure.