Hey buddy, Man, Sunday on Hacker News had some interesting stuff pop up. Lemme give you the quick rundown on a few things that got people talking.

LLMs and Pelican Bikes

First off, there was this post by Simon Willison about the last six months of AI language models, using pictures of pelicans on bicycles to show how things have changed. Kinda weird way to show it, but the point was how these image AIs are getting better, or maybe just different. People in the comments were saying how even though it's cool, asking for something open-ended like "pelican on a bicycle" can still make the images look kinda samey or just... not quite right. Someone else brought up that they heard OpenAI might be losing a ton of money on this stuff, like Uber but on a massive scale, and their paid subscriptions aren't even covering the costs. Also, a few folks mentioned that for some specific things, local AI models are still better because the big ones refuse to generate certain copyrighted or "spicy" things.

Here's the link: https://simonwillison.net/2025/Jun/6/six-months-in-llms/

Building Computers for Bossy Leaders

Then there was this article titled "Building supercomputers for autocrats probably isn't good for democracy." Pretty heavy topic. It's basically asking if companies helping countries with not-so-great governments build powerful computers is a bad idea for the rest of the world and democracy. The comments got into a big debate about what counts as "disinformation" and whether expressing certain opinions should be allowed everywhere. Some people pointed out that maybe Western countries aren't exactly innocent either and do similar things, just maybe less obviously. There was also some historical chat comparing it to past fights over ideas, like Galileo and heliocentrism.

Check it out: https://helentoner.substack.com/p/supercomputers-for-autocrats

Cloudflare's AI-Coded Security Stuff

Cloudflare apparently let AI code one of their security libraries, specifically for OAuth (that login stuff you see everywhere). Someone took a look at it. The big worry in the comments seems to be that AI coding could create a bunch of "mid-level technical tornadoes," especially with less experienced engineers or consultants. People are wondering if AIs are actually "competent professionals" or just good at mimicking. One funny comment mentioned feeling weirdly polite to the AI, even though they know it doesn't care. Another thought was that it would be a bad sign if AI code relied heavily on comments to explain itself, showing it might not really understand what it's doing.

Read about it here: https://neilmadden.blog/2025/06/06/a-look-at-cloudflares-ai-coded-oauth-library/

Why Android Can't Use Some Ethernet Adapters

Okay, this one's a bit techy, but interesting. Someone wrote about why Android devices sometimes can't use those USB-C Ethernet adapters that use something called CDC Ethernet. Turns out, it was a crazy debugging journey that led back to a single, overlooked piece of code – like a tiny typo or a simple rule (a "regex") that was way too strict, only recognizing network names like "eth0", "eth1", etc., which is apparently super old-school and dumb. People in the comments were amazed how a small thing could break a whole class of devices and compared it to other frustrating debugging experiences. It sounds like a classic case of a tiny detail causing a huge headache.

The deep dive is here: https://jordemort.dev/blog/why-android-cant-use-cdc-ethernet/

Remember <Blink> and <Marquee>?

Haha, this post was a throwback to the early days of the web, talking about those old HTML tags like <blink> (makes text flash) and <marquee> (makes text scroll). The article was from 2020 but popped up again. People were reminiscing about how annoying they were but also how they were part of that wild west era of the internet. The comments talked about why they were actually bad ideas (like not working well on different screen sizes, or you couldn't even link to the flashing/scrolling text). There was also a debate about whether killing Flash was a mistake (some said yes, others said no way, it was proprietary and bad for the open web). Fun bit of web history.

Take a trip down memory lane: https://danq.me/2020/11/11/blink-and-marquee/

Ask HN: How to Get Good at CUDA?

Someone asked how to learn CUDA (that's NVIDIA's tech for programming their graphics cards for heavy-duty computing) to a professional level. The advice in the comments was pretty practical. A lot of people said the only real way is to get actual professional experience. But to even get that, you need to be good enough to land the job. Some suggested focusing on the tools around CUDA (like the compiler and performance analysis tools) if you're in it for the money, more than just the math side. If you know C++, apparently you can jump in easier. People also recommended specific YouTube series and examples like "prefix scan" as good starting points.

See the advice here: https://news.ycombinator.com/item?id=44216123

Why People Don't Like DNS over HTTPS (DoH)

Finally, a post asking "Why not use DNS over HTTPS (DoH)?" got a ton of comments. DoH is supposed to be more private because your internet provider can't easily see what websites you're looking up. But the article and comments pointed out the downsides. A big one is that it sends all your lookups to one provider (like Google or Cloudflare), which just moves the "peeper" problem instead of solving it – now that one company sees everything. Some comments were convinced DoH was pushed by Google specifically because things like Pi-hole (which blocks ads by messing with DNS) were hurting their ad money. There was a lot of back and forth about whether the privacy benefits are worth the centralization and control issues.

Read the discussion: https://www.bsdhowto.ch/doh.html

Okay, that's the main stuff. Catch you later!