cat_plus_plus

I am down to bundling forked inference engines in my huggingface models. They do work and do things impossible with base package, yet I don't have time to understand all the details of say sglang code and get it to a state where maintainers will probably accept the PR, for example make the patch work with every other model. Think of it from contributor point of view as well - I made something to work for myself, would love to share it with others but others don't pay my salary. I am not sure what the solution is, maybe have a big draft PR pile a peridodically AI-scan for top enhancements that seem to be a common need and are worth more knowledgable maintainer's time to inveatigate? At the same time, other users can also have their AI agents hunt there for personal fix ideas.

Interesting, but I wouldn't oversimplify this to one cause. Home cooking of meals and making of lemonade/fruit punch used to be a norm. Society chose to move towards taking a lazy way out with microwave dinners and takeouts. On the other hand, Asian countries have healthier fresh made bento boxes in convenience stores so it's not time saving alone. At the end of the day, companies deliver what consumers want and try to discern demand, what else are they supposed to do?

LLMs go off context. Dumping more things related to prompt into context gives them more to go off when giving final answer. If you don't like reasoning think "free association", writing a bunch of things related to what you are trying to do on sticky notes and then looking at these for inspiration.

Do you want your smart TV to become a persistent access point to your home network? If not, you should prefer to have an explicit unbypassable confirmation before another device is allowed to install arbitrary apps on it and pregrant them arbitrary permissions.

Yeah, so long as you can mooch off US healthcare innovation like COVID vaccines and weight loss drugs. There is a difference between going to doctor and getting effective treatment, as anyone who dealt with cancer in Soviet Union can testify.

Gemma 4 31B just because of faster MTP

I just hooked it up to my personal AI Japanese Teacher app, pretty good quality / natural sounding speech in mixed English / Japanese while running fast on CPU so I don't waste VRAM.

Drastically slower than Macs and NVIDIA unified memory boxes while not being any cheaper.

No senior architect around here is capable of understanding tinygrad

AI is like a therapist, it will only change your life if you want to change. I set up python scripts at work to automatically triage, assign and if straightforward fix bugs assigned to me so I have time for complex feature work. At home, I have a script that sends me daily invites to local events based on my interests. Also put together a Japanese voice tutor powered by several local models, can have long free form conversations to learn the language. Planning to go on a long trip around less visited corners of Japan once I feel I can enjoy conversations with locals. Speaking of trips, on my last vacation I had a coding agent find flights/hotels/activities for me using APIs and browser control and just clicked buy buttons. Agent then read my inbox for booking confirmations and added itenary to my Google calendar to stay organized. All in all pretty awesome. But you do have to explain what you want from AI in some detail, it can't decide how you want to live for you.

Hehe you think L9s do oncall? Senior not understanding is a problem all right. It's not necessarily junior's problem.

Write me simple AI inference code that has high performance in big batches.

The primary purpose of code review is to maintain existing hierarchy by preventing junior SWEs from getting promoted by committing code that is smarter than what the senior architect can understand.

Gemma4 31B with MTP enabled is faster and I feel a bit stronger at coding. Either one can run in 32GB VRAM or unified RAM with some tuning (3 bit weights, 8 bit kv cache)

It's not at all like e-mail spam. Vast majority of contributors made a change useful for themselves that they wish to share with others. It's better to think of this as an influx of new programmers or existing programmers picking up new domains. They can be taught to use coding agents better and are likely to stick with projects that facilitate this rather than shutting them out. Maybe it's best for everyone. Let Linux kernel be super locked down to l33t contributors only and let alternative OSes that nobody paid attention to before gain new developers.