terhechte

There's an issue with this index where many models are listed in the model chooser, but missing from the index. E.g. Qwen 3.8 Flash Next is listed but missing, as are older Claude (4.6, 4.5, etc) or OpenAI models. This makes it a bit tricky to see the improvements over time.

I just double-checked my Samsung TV. I only bought it so it could connect to my Apple TV or Switch, so I don't need their software. Thankfully I was already cautious when I bought it and apparently never connected it to any of our networks.

It would make me so happy if this would allow running UTM.app with Hypervisor support on iPadOS. I don't want to travel with two devices, the would be enough, but sometimes I need a real desktop. Apple, of course, wants you to buy two devices. I do have UTM with Jit Linux running on my iPad, via altstore and entitlements and so on, but it is all a bit complex and still fairly slow.

Which part of "I use it every day" did you not understand? Its for me. I just decided to also make it available for others. The app is free.

I don't want to toot my own horn but I developed a mix of Photoshop / Pixelmator for AI assisted Photo editing and I use it every day instead of Pixelmator Pro [ 1 ]. I think it is in many respects the better app because I added many features I always wanted to have (replicate layers, any content as masks, crazy effects, etc). Granted it is buggy and still needs polish, but most software has bugs. So I've completely switched over to a complex AI generated app.

[ 1 ] https://pixellect.app/

Not really. “Flash” generally means fast or cost-efficient, not specifically MoE (Mixture of Experts).

- DeepSeek V4 Flash: 284B / 13B active MoE

- DeepSeek V4 Pro: 1.6T / 49B active MoE

- DeepSeek V4.1 Flash: 552B / 8B input, 16B output MoE

The weights have yet to be released. The previous MiniMax M3 had (I think) ~428B parameters. Since this one is called Flash and they advertise it as inference being 3 times faster, there's the possibility that this will be a ~150B size model which would (quantized) run on DGX Spark or 128GB Macs or the new AMD AI systems.

This was also my reasoning. The prompt even says so:

> You are the artificial intelligence in charge of a railway junction

Clearly more trains will pass through that junction. I'd wonder what happens if you remove that sentence or reword it.

Why are there no screenshots of the "IDE"? Would love to see what it looks like, what features it has, etc.

I released something similar today: https://pixellect.app/

Though it was done roughly 1-2 months ago, a very busy work / private life kept me from finally releasing it. Interesting to see this being released on the same day. Though Pixellect has a different mission. Its not a clean room photoshop clone but more between Pixelmator and Photoshop and also features (I think) a bit more AI Integration. However Photocraft looks much more complete in terms of typical Photoshop features.

It was only 4 months ago that AI development was in question because there were no vibe-coded Photoshops. Here they are, and I think more will come.

> Dots are rolling out in ChatGPT on web, mobile, and desktop starting today to Pro users in markets excluding the European Economic Area, Switzerland, and the UK

Living in the EU, I suspected as much. Still sad. I understand it is because we voted in a bunch of imbeciles, still sad though.

when Cappuccino was written, Apple was still using the Objective-C + AppKit stack. It had been honing this stack for years, every WWDC brought useful and deliberate improvements. Back then macOS development was much more niche (and tech in general). They could take their time. MacOS releases happened every couple of years, not every year. Not the kind of outcome we have nowadays where changes are rushed because of demand.

Cappuccino improved upon javascript by using a different language (objective-j) that added many features js didn't have, and it used the same (mostly) well-honed api as AppKit. In comparison to other js UI frameworks of its time, it was lightyears ahead.

It is the point. Also open source model enthusiast tell you otherwise, there is a coding quality gap between these models. If I use DeepSeek, I do so knowing that I have to limit to simpler tasks on smaller, well specified prompts. What the author did, letting the model do the planning, is not something DeepSeek will excel at. I'm using GPT (Terra, Sol, Luna), Claude (Opus 5, Fable), Qwen 3.8 and GLM 5.3 Flash daily and have to vary which model I use where because there's a huge intelligence step function difference here. That's why this article is so useless:

Imagine someone trying to make the case that riding bicycles is a terrible experience and their whole argument is that they took a random cheapo bike with flat tires and rode it for 3min and that wasn't fun. Sure, but if you buy a 25k carbon bike you will have a different experience. I'd not trust that person. If someone told me they have 10 bikes they ride daily and can explain the differences, in detail, between their bikes, and what they excel at. I'd trust that person's opinion.

A Mac Studio m4 max in low power mode consumes ~50-100 watt on 100% gpu consumption. On GPU idle it is ~10 watt. If we imagine improved models (e.g. qwen 3.8 is a 5b moe which is way faster to process) and improved chip performance on 2nm and half a day of usage (the AI will sleep when we sleep then):

((75 watt / 2 [faster models]) / 2 [faster chips]) / 2 [half a day of usage]) => ~ 10 watt.

AI for everyone does not imply pC mAsTeRraCe for everyone.

Man I used to develop SAAS software using Cappuccino. Since I knew Objc (and JS/CSS) really well it was such a beautiful system for development. I'm sad it never went anywhere

In 2004 I was working in digital signage for museums and for one of them we needed to drive 16 displays from one computer. The article mentions this card:

> On the 17th of December in 2004, the company announced the QIP LP PCIe where the QIP meant Quad Information Display. This was the industry’s first PCIe x16 graphics card, and it was a derivative of the Parhelia. [...] The card could drive four displays with a maximum resolution of 1600 by 1200 per display,

I think we had 4 of these in one system and it was the only card on the market that allowed for this. We ran this on Linux and getting the bios set up correctly to support this was a nightmare. My memory is fuzzy though

I've just used it for a fairly complex refactoring of the UI in a SwiftUI / AppKit app. It managed the refactoring in blazing colors, and the resulting UI looked really good. It was also quite fast. I'm impressed.

I tried to install this the other day with Altstore and gave up after 30min. I learned there are two altstores, and the one I could install didn't support the "source" listed on that repository. Then I looked into the other options and all required some terminal messing and rooting and what not. I thought in the EU we should be able to just install software now.

I created something similar, fully open source, using the ideas behind scrolling window managers. There's currently no website as I'm actively using it to work on a different project. It has features such as a shortcut (command+.) to open an overlay menu that allows you to switch to any split just by typing a letter and a number. Throughout the app. It also has optional vim bindings, support for different models and agents (e.g Claude, Codex, etc via ACP), and much more. There're bugs, but if you run into any, feel free to let AI create a PR. Goes without saying that it's vibe coded.

https://github.com/terhechte/Cormac

Fuck ack. I'm working on a new benchmark that combines strong visual requirements with tool and coding requirements. I haven't even tested Sol yet, but between Sonnet, Terra & Luna I already see much better results from OpenAI's models. I'm not releasing anything yet as I still have issues in my harness that need to be fixed.

Which setup did you use? MLX/GGUF, Quant, Engine (e.g. llama.cpp or MTPLX, etc)? There’s so much variety these days.

Its working fine; to the detriment of the whole club of stochastic parrot folk

https://news.ycombinator.com/item?id=49069787

This is incredible. Even though Apple seems hell-bent on repeating their iPad mistake on Vision Pro ("A baby computer because our users are morons and we have to keep them safe from any theoretical harm. Also, any kind of software development outside of macOS is strictly forbidden"), there is a small tinker community that works on genuinely making iPadOS useful. Another example (also shared here recently) is this:

https://www.lispm.net/apps/uhf-x11/

For any horrible web administration UI:

- Apple Appstore Connect (gazillions of forms of metadata to release an app) - AWS - DigitalOcean - Google Play Store

Whenever I dread logging in because I know the simple sounding task requires me to click through countless menus I use an agent browser. With confirmations of course. However, while the agent clicks through these (oftentimes dog slow) UIs I can do other things. Once it requires permission, I read, decide and act.

You can run DeepSeek Flash just fine on 128GB using lower quants. Antirez' DwarfStar supports it really well. The bigger the model, it seems, the less it is affected by high quantization. I know some people are working on quantizing GLM 5.2 so it can fit into 128GB (using additional tricks to select only some experts, etc). There's a lot going on.

I'm using this in one of my side projects, Emyn ( https://github.com/terhechte/Emyn ) a macOS virtual camera app for composing camera video, app windows, backgrounds, effects, notes, and captions into a polished live presentation feed.

It works very well, the integration is much easier than before, users have model choice. So happy that this exists!

I'd rather have a ai-written article then no article at all. I think this is a real struggle. Writing an article takes time that could be spent on improving the product. Would I prefer a human article, yes. But if the alternative is "no article", then that's how it is.

It is really slow on openrouter and I ran into many http errors.

There's a couple of great images in this article:

https://nmb.bm/history/look-down-look-down/