nmfisher

I've had similarly good experiences with GLM-4.5 for smaller projects/requests. Unfortunately that did degrade with larger contexts, so I'm still treating it as a good fallback for Sonnet 4, rather than a full-blown replacement.

I know I’m the odd one out, but I really don’t find Bluey that wholesome (with the exception of two episodes - the rain one and the Bingo sleeping/space one, which I do think are fantastic). The others are very frenetic - it feels like a pure hit of sugar in television form. It also often shows a lot of bad behaviour that kids can interpret as funny (the cousin running away with the phone after being told, the old lady buying the scooter).

Peppa might be “empty” but I don’t worry that it’s inadvertently steering him in the wrong direction. The Peppa books are also far, far better than the Bluey books.

She didn't touch on how visually captivating the colour scheme is for kids. It's unsettling just how much it draws their attention, it's like a drug. It's also (partly) why I prefer Peppa Pig for my 3 year old, it's much easier to him to naturally disengage after 15-20 minutes.

I found Gemini CLI to be totally useless too. Last week I tried Claude Code with GLM4.5 (via z.ai API), though, and it was genuinely on par with Sonnet.

I still remember how amazing an unsalted boiled egg tasted the first time I did a 72 hour fast.

I think the problem was that you were forced to learn formal grammar, not that you were forced to read Huck Finn.

Millions of people are using (and paying for) LLMs for their daily work. The number of people using crypto as an actual currency is a rounding error by comparison.

There's definitely similarities when it comes to the wave of hype and greed behind them both, but the fundamentals really are completely different.

I don't think it's literally "winner takes all" - I regularly cycle between Gemini, DeepSeek and Claude for coding tasks. I'm sure any GPT model would be fine too, and I could even fall back to Qwen in a pinch (exactly what I did when I was in China recently with no ability to access foreign servers).

Claude does have a slight edge in quality (which is why it's my default) but infrastructure/cost/speed are all relevant too. Different providers may focus on one at the expense of the others.

One interesting scenario where we could end up is using large hosted models for planning/logic, and handing off to local models for execution.

Link?

I'm a heavy user of exhaustive pattern matching with sealed classes, definitely a great feature. You can achieve the same outcome as union types, but it's a lot more leg work, e.g.:

```

sealed class MyParam {}

class Bar extends MyParam { String val; }

class Baz extends MyParam { int val; }

void foo(MyParam param) {

  switch(myParam) { 

      case Baz(val: final val):

      case Bar(val: final val):

  }

} ```

compared to:

```

void foo(Bar | Baz myVar) {

    switch(myVar) { 

        case Bar:

        case Baz:

    }

} ```

I hadn't seen the extension_type_unions package, though, I'll check it out.

Off the top of my head:

- Properly cross platform (including first-class JS/WASM interop)

- AOT & cross compilation

- Hot reload

- Static typing with many nice features (generics, extension types, enhanced enums, etc)

- Expressive without being overly verbose

- Well-supported by the language & SDK teams (i.e. it's not dead)

- Easy build/package system

The only things I really feel are missing are (1) union types and (2) access modifiers like protected/friend.

Probably because the last time the USA tried something like this, they lost 8 soldiers and about the same number of helicopters.

https://en.m.wikipedia.org/wiki/Operation_Eagle_Claw

If I see any story about OpenAI, I assume it's an intentional 'leak'. I know all companies spend money on paid PR but OpenAI (and Sam Altman) take it to a completely different level.

I suspect that generating textures directly in UV space is a bit of a dead end because UV maps are object-specific, so flux/SD/etc are going to have a hard time understanding the inputs.

I've been experimenting with image-to-image (and video-to-video) for basic texture projection, which I think shows promise:

https://bsky.app/profile/nickfisherau.bsky.social/post/3lqrl...

I just started diving into this today too:

https://github.com/YixunLiang/UniTEX

which works with in "volumetric space" (for lack of a better word), which I think makes a lot more sense.

I'm a happy Supabase customer. It's not just postgres, it's postgres + auth + functions + storage all in one place, with a very good local Docker environment and a variety of nice client SDKs.

It's a quick/convenient way to get a "full" backend up and running. The other option (Firebase) can't be self-hosted and has some absurd pricing footguns.

How long did it take to sift through those?

Do you have any idea whether Substack drives traffic to, or sucks traffic from, your blog?

My brother is an orthopaedic surgeon, so I’m curious to know more. Do you have a website?

A new reasoning model release doesn't grab my attention as much as the fact that it's a Mamba-based model that's competitive with conventional architectures.

(Strictly speaking I'm a bit late, because it seems the non-reasoning version was released earlier in March, but I didn't see it until today).

I just gave it a unit test I was in the middle writing and asked it to fix it - no context/documentation/interfaces, literally just the test. It performed amazingly well - even where it didn't get it right, it made a very reasonable guess (e.g. calling dispose() when the actual method was destroy() - but since I hadn't given the interface, it had no way of knowing).

I'm going to road-test this a bit further via Cline (where I alternate between Gemini Flash and DeepSeek).

I had exactly the same issue, seems like it's practically useless for humanoid illustrations/characters.

This is exactly how Flutter (now) works on web - the browser loads a JS/WASM blob, and your app renders everything to a canvas - and you're right, almost everything you mention is problematic.

There is a time and a place for that paradigm, but it's niche. The lack of SEO alone is usually enough to stop it in its tracks.

Bridging 3D editing with AI image/video-to-video. I have a lot of the disparate elements (timeline, video export, texture projection, etc) but I'm still playing around trying to find the killer use case.

https://mixreel.ai

Website is just a placeholder, but I'm documenting my progress on Bluesky at https://nickfisherau.bsky.social/.

Exactly this. Every time I revisit a Jetbrains product, I uninstall it within 5 minutes. It doesn't matter how great the features are, it's just sluggish.

People can rag on Electron apps all they like, but VSCode on modern hardware is very snappy. Jetbrains is a noticeable downgrade.

Why are NFL teams banned from having an official account on Bluesky?

Interestingly, Werner Heisenberg was decidedly non-Nazi (and was regularly attacked as a “white jew”), and even though he had ample opportunity to leave, he chose to stay to work on the German nuclear fission program.

I don’t think it tarnished his scientific legacy, but it definitely created some friction in the post-war years.

Where are the details in Chinese?

GPU export controls are basically the USA “cheating” to try and stay ahead.

Looks like it didn’t work though.

Normally I’d give the company the benefit of the doubt for the reasons you point out, but not in this case. If you’re asking increasingly convoluted/unrealistic/impractical questions, you can’t complain that the response wasn’t practical.

It gives off strong “your answer might be correct but it’s not on my answer key so I’m marking you wrong” teacher vibes. Avoid at all costs.

Completely tangential, but from personal experience, the performance of Google Search had been degraded for at least 2 years prior, so when it was banned it was already mostly useless anyway.

GMail continued for a few years after that - I'm not even sure if it was ever actually banned, or just suffered the same fate (death by a thousands timeouts).

Gemini Flash is my go-to for many client projects, so there's at least a few people paying for it. If it fits your use case, it's easily the best value for money, and often makes something commercially viable that wouldn't otherwise be.

Sample size of 1, obviously, so YMMV.

I also pay for Claude Pro for personal coding use, so I'm aware that Flash isn't exactly a drop-in replacement for a more powerful model. From my limited testing, though, Pro 2.0 is almost indistinguishable from 3.5 Sonnet.