Nick Fisher

You're just not cracked enough to understand, bro.

If you look closely, you can spot my reflection too.

Results from Hunyuan World Mirror (original video captured with an iPhone 17 Pro).

It's still falling over with >60 images and I haven't had much success trying to stitching rooms together manually.

I keep coming back to LongSplat because it's exactly what I need but no luck so far.

Hunyuan World Mirror is actually very fast for a small number of images, but I have a 2 minute video I'd like to try and convert to a splat, I'm not sure there's any hardware that could actually run that model with a video that size.

3d-models.hunyuan.tencent.com/world/

pytorch3d is notoriously finicky about exact dependencies but I don't know what's going on with Lightning, it comes with a single conda environment that seems to hate literally everything I throw at it.

I'm giving up and going back to Colab for this one.

Unfortunately my ancient GTX1080ti is too slow to train LongSplat (well, not literally, but it will take ages due to lack of tensor cores for fused SSIM).

I've been futzing around for a couple of hours with Lightning's cloud notebook, but keep running into build incompatibilities.

Breaking: we release a fully synthetic generalist dataset for pretraining, SYNTH and two new SOTA reasoning models exclusively trained on it. Despite having seen only 200 billion tokens, Baguettotron is currently best-in-class in its size range. pleias.fr/blog/blogsyn...

The general concept probably isn’t unique to Reddit, Twitter hashtags, YouTube Live, Twitch, even radio/TV all had some form of live Q&A. Works best on Reddit bc good questions float/stick to the top, and they (probably) had a marketing team actively scouting for celebs/politicians.

From the LocalLLaMA community on Reddit Explore this post and more from the LocalLLaMA community

Just saw there’s a Reddit AMA with the Kimi/Moonshot team:

www.reddit.com/r/LocalLLaMA...

I’m too tired to figure out time zones but I guess it’s starting soon.

Discord is actually a phenomenal technical resource (I’ve sourced a lot of contract work there too). It is a shame that it’s all locked away and I know the enshittification is going to start any day now, but it is such a good product at its core.

Not something I expected an instruction for.

Also some great little tidbits about history, like Descartes sticking an ox's eye in a wall in 1637, scraping off the tissue at the back and looking through it in a darkened room to figure out the shape of the lens.

I just discovered this "First Principles of Computer Vision" series which looks like an incredibly well-designed course on the basics of lenses, image features, convolutions and more advanced image processing.

Definitely worth bookmarking if you're new to the field.

fpcv.cs.columbia.edu

I've never tried to implement a multi-step asset transformation pipeline before, but nowadays I'd probably just ask Claude/Kimi etc do it for me. The biggest problem with build systems (CMake included) is arcane syntax, which isn't something I want to spend my own brain power on.

Also, a 90-day expiry for prepaid credits is really crappy from Google, in fact it feels like something that is probably illegal in certain jurisdictions.

I've mostly been working with Modal/Colab Pro notebooks, but after yet ANOTHER batch of Colab compute units expired, I've reverted to running something on my local Linux machine (GTX1080ti). It's very dated but this thing is a workhorse, I think it's been with me for almost 10 years now.

It might feel that way, but remember what it was like doing the same thing circa-2010! Things are definitely better than they used to be.

I’m seeing various mentions of Nano Banana 2 - does anyone have access? One of my ongoing projects could benefit from improved image style transfer.

I’m looking at structure-from-motion models and a thought popped into my head - how similar is the binocular field from human to human? Is it possibly like colour blindness where some people perceive depth/perspective differently from others?

I know a couple of people doing this, it's quite lucrative.

YouTube video by droidcon & fluttercon Using Dart FFI for Compute-Heavy Tasks - Robert Odrowaz-Sypniewski | Fluttercon EU 2025

There's some useful benchmarks here to compare Dart vs Dart-with-FFI performance.

www.youtube.com/watch?v=eJR5...

What was the use case?

I always say that CMake is like democracy - it's the worst form of build system, except for all the other ones.

I haven't enabled it for exactly this reason, I felt like it was going to be a time-sink until the kinks are ironed out.

I prefer my coffee app-less, but I'm interested to hear about your dev experience anyway.

Say what you will about CMake, I don't think there's ever been a case where I wasn't able to hammer it to do exactly what I wanted.

It might not be pretty, but it works.

I've had countless conversations with technical people where I try to convince them that "AI girlfriends", "AI therapists" etc are a really bad idea.

I really don't think the answer to emotional problems is "more apps".

The tax office is definitely an exception.

On the other hand, my passport cost more than $500, and corporate tax returns still have to be filed in paper with no editable PDF. So it's not all roses.

e.g. right now I'm browsing breakdowns of waiting times for public elective surgery by procedure, hospital, improvements over time, etc.