timestretch

Their models have been great, but I wish they'd include the number of parameters in the model name, like every other model.

I wonder, how long it will be supported? These APIs seem to have a very short shelf-life.

I'm surprised to see an example from Apple Music. That app is an perfect example of design neglect too. The ability to filter your song list for a specific track is now hidden in Apple Music.

To reveal the filter field, you must select the "View" menu, and "Show Filter Field". Worse each time you relaunch Music, this field is again hidden, and you have to select the menu item again.

Sounds interesting. What model did you get? Does it have a stylus?

The hardware in the iPhone is probably powerful enough to run all software he'd run on his Mac if it were similarly unhampered by iOS.

They want to sell you two devices, and a monthly subscription to iCloud to interoperate between them.

Other ridiculous things: The "Apple Magic Keyboard Folio" for the 10th gen iPad has a better keyboard than the iPad Pro.

MacOS allows me to run any software I want. It has a terminal, and allows me to run scripts. It has a shared file system and offers true multi-tasking.

You can already run some iOS apps in MacOS. I'd like a single computer where I can switch between mobile / pencil usage and desktop keyboard / mouse usage.

I'd buy an iPad immediately if I could run MacOS on it.

I've thought of switching to Linux, but until BBEdit is ported, I'll stay on the Mac.

I love BBEdit and use it daily as my main text editor, and have since classic MacOS days! I'm hoping they add a couple of features:

- Ability to auto-format using Prettier.

- AI Large Language Model auto-complete for code. The field is moving fast, but I'd love to be able to use open models here. I don't want to send all my code to GitHub co-pilot.

BBedit has an excellent built in text-filter functionality and I've written scripts to send selected text as a prompt to various LLMs, but built-in would be ideal!

There is a lot of ongoing research into making language models that can run well on a wider variety of hardware. It seems VRAM is the main limitation at this point.

You can already run smaller language models on your own hardware if you have a GPU with sufficient VRAM. For example, with quantization, you can run gpt-neox-20b (512 token context window) or gpt-pythia-13b (full context window) on an RTX 3090 with 24GB VRAM. Quantization allows you to run the model with less memory, where each parameter utilizes 8 bits or 4 bits instead of 16 or 32 bits.

Another possibility is to use reinforcement learning with human feedback to tune smaller models to give results comparable to larger models.

I've also been using RWKV with good results. It is a language model that uses an RNN and only needs matrix-vector multiplication instead of matrix-matrix, so inference runs much faster. The 7B model uses about 14GB VRAM without quantization. A 14B model is currently in training, but progress checkpoints are available. You can also do inference on a CPU, although it is much slower than GPU.

A hybrid approach where you could "sketch" with SolveSpace, but then generate OpenSCAD style code that you could later edit would be amazing.

The main advantage of OpenSCAD is that it doesn't "blow up" when you change a variable to be out of some range where things work. In OpenSCAD, you can put it back if it does cause a problem.

Changing a dimension in SolveSpace sometimes throws off the whole design. SolveSpace's precision, constraints, and great keyboard user interface make it a fun program to design parts in. It is among the best open source tools out there.