huydotnet

I recently subscribed to Claude, and was very unhappy about the usage limit of the $20 pro plan. Then I found a trick, since I got free Google AI Pro via my phone carrier, I use Opus 5.5 High for planning, and then dispatching agy to do works.

Most of the time 3.8 works fine, but it's a bit slow if compare to 3.7 Flash. If there's already a detailed plan, 3.7 can complete the task much faster. And the best thing about agy is the usage limit was very generous.

hey, thanks for the suggestion. unfortunately even with lichess's annotation, it's still not enough to identify hanging pieces or other board situation for the diagnostic, so we still need to run stockfish on each move again :D in fact i intentionally get just the raw move data from lichess during import.

Did not install the app but it looks cool. But I was wondering, how does the app can hijack macOS login screen like in the demo? is it even possible?

If you're looking for a carefully crafted/written work to explain internal combustion engines, look no further than this one https://ciechanow.ski/internal-combustion-engine/ (the Mechanical Watch article from the same author was featured on HN a while ago).

For anyone jumping into this thread hoping to see capacitor use for timing, there is this blog post about something like that:

https://notes-huy-rocks.translate.goog/posts/diy-pomodoro-ti...

(google translate link because the original post was in Vietnamese)

One don’t have to buy a new computer to run Qwen3.6 or Qwen3.5 (35B A3B), given that they can already run Qwen3 30B A3B.

In fact with a 64GB mac, you can run pretty much all of the latest Qwen models.

Also, anyone who has been following local LLM are well aware that the quality and performance has become way way better since Qwen3.5

I love that local LLMs are being discussed more often on HN recently. But for the post, I find it strange that the author claimed they were working with local models from day 1, but wrote a post that still links to Qwen2.5 and Qwen3 in mid June 2026.

I've been building the same thing for a while https://github.com/huytd/octocmd It has everything you need to throw away the mouse: keyboard tab switching, search and click, vim-style clicking, keyboard scrolling.

I'm gonna use this article to explain to my peers about LLM quantization!

They are different quantization types, you can read more here https://huggingface.co/docs/hub/gguf#quantization-types

you just answered your own question, "AI hobbyists who has 4090 at home". And they are pretty much targeted user of Unsloth since the start.

I've tried both. Each has pros and cons. Two things I don't like about superpowers is it writes all the codes into the implementation plan, at the plan step, then the subagents basically just rewrite these codes back to the files. And I have to ask Claude to create a progress.md file to track the progress if I want to work in multiple sessions. GSD pretty much solved these problems for me, but the down side of GSD is it takes too many turns to get something done.

Bangle is a good one https://banglejs.com/ Quite fun to work with. I got one from a conference, and my kid has been enjoying it a lot.

exactly what i think when reading the top of the article, maybe the author turned off vebose mode

Joke about train line aside, I think Railway fits right in the spot that Heroku left.

They have a nice UI, support deploy any kind of backend-involved apps as long as it can be built into a docker container. While many PaaS out there seems to prioritize frontend only apps.

And they have a free plan, so people can just quickly deploy some POC before decide if it's good to move on.

Anyone know if there is any other PaaS that come with a low cost starter plan like this (a side from paying for a VPS)?

Context: This is Railway the PaaS company, not your daily commute vehicle (which is good in general, still bad for many users, like me).

> Due to a miscommunication with the factory, the injection pins were moved inside the heatsink fins, causing the cylindrical extrusions below.

What happened after this? the factory have to replace the casting mold at their own expense or you have to pay for it?

Well, to whoever downvoted my comment: It's supported now!!!! https://lmstudio.ai/blog/claudecode

yup, I've been using llama.cpp for that on my PC, but on my Mac I found some cases where MLX models work best. haven't tried MLX with llama.cpp, so not sure how that will work out (or if it's even supported yet).

I was hoping for the /v1/messages endpoint to use with Claude Code without any extra proxies :(

I think this is one of the cases where strictly applying the guideline fails the reader, but yeah, I can see that this guideline make sense most of the (other) cases.

Unrelated to the conversation, but the post title was something like "Starlink roam 50GB is now 100GB and unlimited slow speed after that", then a minute later it's now "Roam 50GB is now Roam 100GB".

Was this change made by a mod or OP, and why would someone making that change? I do think the original title was more descriptive, and the new title was completely out of context, or it's imply that everyone is using Starlink and know what's Roam 50GB is.

very nice! would be nicer if it can be playable on mobile, i know where i'm gonna spend my time waiting for my wife at the mall now.

Thank! I've been trying with a conversation with 20, 30 comment threads, with about 5 replies per thread, so far so good.

I heard in Chrome, there's a gemini nano model built-in as well, maybe this is a good example to integrate it.

Thank you so much, that is a fair point! It's part of a series of mistakes I made, the product started out as a free to try and only showed the pricing after the user used up all their credits (I didn't even have a landing page back then). I'll update the landing page to make this clear!

I think the sustainable way is to put more and more backlinks out there, more blog posts, etc. I actually suffered from it too.

Many reasons: 1) lack of marketing, 2) I stopped working on it for a while, 3) because of #2, the app lacks new features to attract users.

Another one but turned out it was never really a big deal: some chatbots from frontier AI labs started to support those niche features (people still coming to my app for the flexibility of using multiple AI models).

I think the biggest problem was #2, life kept pulling me the other way.

I started my project in 2023 and posted here, made 20k that year. The traffic has been slowly decreasing during 2024, and last October, I was officially entering losing territory, where the cost of running it exceeded the total earnings (mostly due to free trials).

It's been a good journey. Thank you so much to whoever keeps running this thread!

While everyone training AI, this man train a rat. Are you gonna release the open weights (or the rat)?

Great project btw!

Hey, I'm the author of the post. Thank you so much for the kind feedback!

Speaking about total time/cost, this experiment cost me just $1.01 for 2h30 on a rental GPU. But the actual successful run was less than 10 minutes for both phases. The rest of the time I was spending fixing the code, tuning the params, train, and retrain. It took me about 6 hours to build and clean the two datasets, though.

For the next step, I'm thinking of improving the model accuracy, maybe with RL, but I would not go about shrinking the model size any lower. Prior to this, I've tried a lot of different model sizes on different kinds of tasks, from 135M to 4B. I'm not sure I like the performance of these small models for code generation :D