pwython

My comment on GLM-5 five months ago:

"How many pelican riding bicycle SVGs were there before this test existed? What if the training data is being polluted with all these wonky results..."

https://news.ycombinator.com/item?id=46974853

I set up an old Stream Deck to do the same thing. I stopped using it after a few days. This design looks great though, status lights are a nice touch. YouTube vibe coders will love it, traditional devs will keep MacGyvering their own toys.

I thought they were Turkish.

Looking forward to playing with Krea 2, I use Z-Image Turbo daily -- it has replaced my stock photo subscriptions, for realism and illustrations.

May I ask how much did the training cost you?

Here's a much larger list (1,057 sites): https://pastes.io/rcPg2RLC

Not yet, I have to decouple some of the specific/personal logic. When I do push it, it'll be here: https://github.com/alisorcorp

I had Claude whip up a local solution for me using Gemma 4 26b-a4b on my Mac and a Raspberry Pi Zero 2 W. It can do web search (valyu), file reading -- I have business & personal context stored in many markdown files -- weather, Apple reminders, and has cross-session memory. Streamline but capable agent that has been pieced together over a few weeks after using Karpathy's LLM wiki pattern, with bits of Hermes logic. Orpheus-TTS streams the spoken reply back with the first word usually landing in half a second. Voice input is openWakeWord for the wake word plus faster-whisper for speech-to-text, all on-device. I can run it straight on the Mac but I use it with the pi satellite and a cheap USB speakerphone (ConfCall MS13B). You can barge in by just talking over it rather than having to say the wake word again. Pretty handy Google Home replacement.

Same thing architects do if drawing lines gets automated: architecture.

Would you trust living in a high rise designed by AI?

Designing a system that survives production is the job.

I asked Fable to plan a security & performance audit of my website. It said it would check SSR & origin attack surface, CMS content injection, Strapi API surface, etc.

Just before asking for approval to run, it said one thing it wanted to "flag before running" was "Rate-limit and auth testing against prod will generate some 4xx noise in Railway logs and could trip the form rate limiter — harmless, but saying it now."

Ok fine, I said go for it, and it says:

"Running it. Quick recon first (prod URLs + the prior-findings baseline), then I'll fan out the audit tracks with adversarial verification."

Immediately after, I got the Fable warning about how it can't continue because of safety concerns, switching to Opus. In the end, Opus did a good job thanks to whatever Fable suggested doing. Things were fixed that Opus missed in a security/performance audit just the week prior. But what surprised me is that it used 55 agents. Burned 80% of my 5-hour window in 15 minutes (5x Max plan). I've never had Opus do that before on these audits.

First one I clicked on is 404: Programming Massively Parallel Processors: A Hands-on Approach (3rd Edition) https://www.cambridge.org/core/books/programming-in-parallel...

This actually worked well when I asked Gemini to generate a random color, character by character. I was getting Indigo/Electric Indigo a lot if I just asked for a random color on new sessions.

Amazon is investing $5 billion in Anthropic today, with up to an additional $20 billion in the future. This builds on the $8 billion Amazon has previously invested.

Today’s agreement will quickly expand our available capacity, delivering meaningful compute in the next three months and nearly 1GW in total before the end of the year.

They need a bunch of compute, now.

https://www.anthropic.com/news/anthropic-amazon-compute

pi-coding-agent (pi.dev) is also great. I've been using it with Gemma 4 and Qwen 3.6.

Big brands don't pay the big bucks to buy placements on run-of-network channels (ie. small random channels).

Not Flock defenders, just people explaining how this is not a CCPA violation. I could set up 100 cameras around town (with property owners permission) and record cars driving by, birds, etc all day. Then I could sell access to that footage to whoever I want. If they want to scrape license plates that's up to the customer and their problem. Or if they want to track birds, cool, that could be in the frame too.

At first I thought you were talking about an actual rotating fan, which would be an awesome addition to this. Just a small PC fan running at a very low RPM built into the side in a circular cutout, with that worn metal patina look.

then mentions that future readers "may even be an artificial intelligence rather than a human, how wonderful!"

My first thought seeing this post was, I need to find more literature like this, fine-tune a model with that + Logic Pro documentation, then give it an MCP to control Logic Pro and see if it can be my music production assistant.

I've been migrating a few Wordpress sites from Wordpress to Astro + Strapi recently, working in 'hybrid mode' so the entire site is static except for post previews in Strapi (only that one route is SSR).

Editing content in Strapi, once customized with CKEditor and such, is Wordpressy enough for the human Editors familiar with WP.

So far I'm loving the stack.

Genuine question: If I were to fine-tune a model with 10 years of business data in a competitive space, would you feel safe with cloud training?

My biggest gripe with Apple Music is shuffling. When I shuffle my entire library of thousands of songs, I'm hearing the same ~50 songs over and over UNTIL I add a new song to my library. Then suddenly I'm hearing songs I haven't heard in years. How does one screw up randomization that bad, and how has it not been fixed over the past several years?

Is this launched? Looks cool, but you should add a privacy policy.

Ah yes, that's right. I was looking at the M5 model last month wondering why there was no 64GB option.

Well that's. Just. Great. I bought a 64GB M4 Max MBP last month. I'm past the 14-day return window. I figured the M5 was near, but assumed M5 Max would come a bit later. Not sure where I came up with that.

Coincidentally, I was asking Claude today if something existed that could identify the key, chord progression, tempo, etc from a playlist of my favorite songs to see if there was any pattern that stood out so I could find similar songs with that vibe. Like a more music theory approach to discovering new songs versus the "people who liked this song, also liked these songs" way.

Even more coincidental, earlier today my wife was saying we should take our "Skylight Calendar" screen device that is hardwired into our wall with us when we move. I said I could just make a DIY one... and then I open HN and see the top post: https://news.ycombinator.com/item?id=47113728

Spooky.

Oh by the way, all of the "Open on Bandcamp" links I clicked were 404 pages.

Did you just "it's not x, it's y" me?

See: Rick Rubin.

"Rick Rubin says he barely plays any instruments and has no technical ability. He just knows what he likes and dislikes and is decisive about it."

https://www.cbsnews.com/news/rick-rubin-anderson-cooper-60-m...

https://en.wikipedia.org/wiki/Rick_Rubin_production_discogra...

How many pelican riding bicycle SVGs were there before this test existed? What if the training data is being polluted with all these wonky results...

My "actual job" is a designer, not a career engineer, so for me code has always been how I ship. AI makes that separation clearer now. I just recently wrote about this.[0]

But I think the cognitive debt framing is useful: reading and approving code is not the same as building the mental model you get from writing, probing, and breaking things yourself. So the win (more time on problem solving) only holds if you're still intentionally doing enough of the concrete work to stay anchored in the system.

That said, if you're someone like me, I don't always need to fully master everything, but I do need to stay close enough to reality that I'm not shipping guesses.

[0] https://alisor.substack.com/p/i-never-really-wrote-code-now-...

1 frame of Bob Ross = 1,800s

For those that have homebrewed a base model, does your output have the same AI-isms like overusing em dashes? If so/not, what dataset did you use?