Now as for why, I imagine that it's just money. Anthropic presumably just got done training Mythos and Opus 4.7. that must have cost a lot of cash. They have a lot of subscribers and users, but not enough hardware.

What's a little further tweaking of the model when you've already had to dumb it down due to constraints.

That being said, GitHub does not even offer a time sorted search. Meaning that most of the results are going to be quite old and useless.

Second, API keys being shared on GitHub is quite an old problem. People setup automated scans for this sort of stuff. Me removing my comment isn't going to help anyone who already posted their API key online.

Edit: self censor based on a request

https://github.com/ggml-org/llama.cpp/blob/master/tools/serv...

You can't take someone's money and then not only not give it back, but also give it away.

The amount of times that the initial HTML, CSS and JS came through, but then choked on fetching the page content was insane. Staring at a spinner is more insulting than the page just not loading.

That being said, I'm not a huge fan of RSCs either. Dumping the entire VDOM state into a script tag then loading the full React runtime seems like a waste of bandwidth.

I notice a lot of examples just vaguely mention "oh, you can have others review your previous changes while you continue working", but this one doesnt make sense to me. Often times, the first set of commits doesn't even make it to the end result. I'm working on a feature using lexical, and at this point I had to rewrite the damn thing 3 times. The time of other devs is quite valuable and I can't imagine wasting it by having them review something that doesn't even make it in.

Now, I have been in situations where I have some ready changes and I need to build something on top. But it's not something just making another branch on top + rebase once the original is merged wouldn't solve.

Is this really worth so much hype?

I've never seen a service so opposed to me using it. There's no option for dev accounts either.

For GO, switching to Linux (with an AMD card) was a free performance boost. I gained like 30fps.

For early CS2, the performance on Linux was terrible.

Now, the peak fps is slightly worse, but the frame pacing is much more stable. Eg: you get less fps, but also less fps drops.

Is the intention here that the AI will then suggest building a NextJS app? I can't quite describe why, but this feels very wrong to me.

- Nice idea, though after playing Turing complete, I would like to skip the beginning and move to stuff that makes GPUs different to CPUs. But it's understandable.

- I'm not smart enough to intuit NAND from transistors. I'm also not sure I will be alone in that. It's such a weird difficulty wall.

- Speaking of, the difficulty is all over the place. Though easy mode is appreciated.

- Even with a n key rollover keyboard, I couldn't complete the capacitor refresh level. It seems like it speeds up and certain capacitors already start empty.

- The routing for wires is no good atrocious. Any level with more than 8 components will end up impossible to read.

- It doesn't help that you can't color code or even path wires manually.

- Might be Firefox only, but I had a hard time selecting the connection points.

- Dragging the mouse along the edge should pan. Otherwise you have to drop the connection and zoom.

- I appreciate the added "show solution". But it's not really giving you a solution. It's just a better hint.

- An option to show all tests or at least more tests would get great.

llama.cpp is about 10% faster than LM studio with the same options.

LM studio is 3x faster than ollama with the same options (~13t/s vs ~38t/s), but messes up tool calls.

Ollama ended up slowest on the 9B, Queen3.5 35B and some random other 8B model.

Note that this isn't some rigorous study or performance benchmarking. I just found ollama unnaceptably slow and wanted to try out the other options.

Much less than just not considering morals/ethics, it's seen as a weakness here.