zionsati

btw how did you get 4.2k stars for your repo? I genuinely would like to learn.

omg thank you! Mac is turning into Windows where every post install needs a debloater!

Good idea! Any suggestions?

Thank you :)

Microsoft's support for WASI has been pretty slow. Wasm isn't just for the browser - it came from there but with WASI, Wasm has the opportunity to run just about anywhere, even more portable than .NET Core.

MSFT's interest in making a wasm profile has been somewhat lacking and I understand the reason - the size I managed to achieve would've been impossible if I went with the full BCL / CoreLib. In particular, I had to cut reflection and typenames, keeping only thin RTTI.

So yes to staying as close to the latest C# language features but also deliberately not supporting features that bloat the resulting wasm if there's better and more modern alternatives such as source generation in place of reflection.

Read the readme but still unclear - does this mean we can run DeepSeek v4.1 Flash on GPUs with 8GB VRAM?

Very nice! Is there a lower quant like Q4? There's research showing lower quants suffer from low confidence in its own reasoning and tend to overthink because of that. I wonder if this helps?

Yeah all good. They're all good questions.

My point around ts is people want strongly typed languages these days.

The bigger problem with HTML/CSS is, when you're trying to do more sophisticated things (see the immediate drawing page on the demo), you'll find you need to use a canvas/HTML5. That then becomes a mini UI framework of its own...

For simple webpages, HTML/CSS is probably still the best way to go. But I wouldn't say the same for webapps?

For a lot of front end devs, CSS/JS are frowned upon. We use it because we have to. In fact, we typically avoid JS by having TypeScript transpiled to JS. Not to mention CSS is still very inconsistent across the various browsers.

Porting HTML/CSS/JS to native apps is another big problem, which is why we ended up with Electron/WebView. The amount of RAM it gobbles up just to run simple apps is pretty horrible.

How does it handle dynamic complex sites E.g. webapps.

To be honest, I'd be happy to get 3k users for anything I've created!

Marketing is always the hardest part. I came across this medium article https://zach-saw.medium.com/the-internet-is-dead-96a3f3ed866... while researching why it is so hard. It talks about the difficulty of getting some eyes on your app / project. In particular, it said Linus needed to be a content creator first before creating Linux in this day and age which hit me pretty hard.

How about WPF/SwiftUI grade UI framework for the web?

Not everyone is building the same thing. What makes you think that?

I'm working on deprecating JS and HTML... :)

I did a bit of digging after this to see if my feeling was accurate, and lo-and-behold: https://deepswe.net/

Look at the gap between gpt5.4-mini vs deepseek v4 pro!

deepseek v4 pro - but honestly it is comparable to gpt-5.4-mini, far from GPT5.4 let alone GPT5.5! Its advantage is really just its pricing. I'm going to give Kimi K2.7 a try - with K2.6, even its cloud chat locks up all the time, so it really didn't give me much confidence at all for agentic coding.

Do LLM crawlers adide by this?

AI slop! lol

What I've experienced was if you let an agent "fix" the problems, and keep vibe coding away, at some point they'll stop working. Fixing one breaks another, and the tests you think you had as guardrails turned out to be useless because they're contradictory. The specifications (hence spec driven coding with AI) become very important. If you get that wrong, the agents will try their best to somehow fit the output to your requests. For simple UI projects, sure, you can quickly see it visually. For anything complex (most real world systems are), there's a lot of reading and re-reading AI produced code. This could be slower and more time consuming, and often cause mental fatigue because there's no way you can gauge the level of THIS particular AI's output, unlike when you're doing PR reviewing for a actual human. You tend to take shortcuts because you know them, how they code, and their strengths and weaknesses. With AI, it's random, and I've personally seen prod going down because of it.

So yes, I think the crowd is anti-AI where it matters. But claiming AI-assisted version could have been deployed 10x faster, I doubt you've really maintained any real world software systems.

Ooh that's good advice! Very useful for everyone in my team. Appreciate it!

Love it! Well done on the podcast!

I think that's all correct what they're saying. It's not contradictory - writing code has become a commodity to a certain extent. But someone's still responsible for the production systems when it fails. AWS had a couple of big outages last year because of AI. Even with humans reviewing it, review is review. If you're not writing the code, you'll not have spent enough time thinking about edge cases. That's where the most expensive part of software engineering, not coding. Coding being a commodity is indeed analogous to robots in the production line, but those robots have huge teams of mechanical engineers to keep them running. The calibration, the servicing etc.

What's most interesting is now that vibe coding is expensive, will the AI slop go away, gradually?

I'm facing the same problem. I'm spending a lot of time on side projects now precisely because my company is forcing our strongly encouraging everyone to use AI for coding. At some point I found myself complacent enough to accept AI code at face value which scared me. With my side projects, I still use AI, but only when I get stuck, to help me get the aha moment. But most of the time it's, hmm... That works but, I don't think we should do it this way. Very tempting to continue prompting and keep going from there though, I have to say.

It's a gradual and continuous "uh oh"... Mythos is a much bigger uh oh for me than any other models before it.

I'm more worried about the next generation. Kids growing up today outsourcing their thinking to AI completely...

I've always wondered how this world works. The person who created something used by companies making hundreds of millions in profits doesn't actually get a share of the wealth generated by his creation. And he has dedicated his life to this, and now having to deal with AI PRs / bug reports, like the rest of us mere mortals. Every cloud provider's biggest offering is Linux instances. But because of free open source software, big corps can pocket every single cent.

Ahhh a fellow Borland Turbo C++ Dev (we weren't called Devs back then!)

I agree. GUI these days are largely driven by how HTML works and we now have a whole generation of UI Devs (frontend devs) who've never known any other way.

I missed the good old ways of GUI.