cat_plus_plus

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.

The preponderant bottleneck is inventing new architectures to make AI actually good at human and superhuman tasks. For example, AI agent harnesses add tool calls and long term state management, allowing AI to autonomously complete complex tasks. Once these are in place, finetuning models with examples of good tool calls helps, but somebody first needed to invent the fundamental capability. Now try to implement humanlike long term memory for AI to be your coworker or life assistant working on tasks that last month. Even if necessary low level technologies are already there, structuring them to be practically useful is non trivial.

Models consume a lot of memory and power to slowly generate autocompletions of existing content one token at a time. Letting this text control important things in real world is a human decision and control of nuclear weapons or penetration testing agents should be well regulated. But these should already be well regulated without AI. So how about we focus on drawing down nuclear arsenals, which is a present danger one idiot (or faulty AI) away from world shattering consequences rather than fearmongering about unknown future. Before effective AI regulation can be drafted, we need to anyway know more about AI specific dangers rather than humans committing already very common crimes like hacking for ransom with new tools?

In 2026 there is a considerably cheaper/quicker solution, but that in no way invalidates OSS maintainers' right to enjoy a summer vacation without interruption.

SGTM, if I am worried about a curl exploit, I will type details into Zoo Code prompt and it will disappear in about 30 seconds and then I can upload a PR for others concerned. Enjoy your vacation and I will enjoy security for a lot cheaper than an enterprise contract!

There are priorities here which are not mutually exclusive. If you are building the next big thing in your garage, polishing it before establishing it's in fact the next big thing is a waste of time. If it really is, early adopters / investors will seize on it like they did on AI. Once that is established, yes there is space for adults in the room to adopt proper procedures for large scale production. I do believe AI can help with establishing the next big thing, and if other's don't, it's a irreconcilable difference of opinion.

There are amputees that do snowboarding with specially designed prosthesis and boards, so there is certainly a way to take load off weak knees with appropriate gear. OP is just, quite reasonably, not prioritizing this minor dream enough to invest so much time and money in it at the expense of other priorities.

Does "athlete" in your original post have to be only martial art/contact sports? For example, are you medically unable to do absolutely any strength exercises, maybe upper body only, with appropriate support/isolation? A lot of times it's a matter or re framing rather than giving up goals, a lot can be done at home with a pair of adjustable dumbbells.

Lean into medicine, tech and science? Zepbound + strength training is a life transforming miracle and they are coming up with even more effective drugs and modern AI deep research (think coding agent pouring over mountains of data) is amazing at coming up with an investment plan to eventually own a home, for example putting up a downpayment and running a rental in a really cheap area and then eventually using income to finance something more convenient if you are not willing to move.

My point is not that all problems can be solved, but I think what OP is saying is that life is a matter of focus. I am putting plans of a soaring corporate career on ice not because I am 100% sure I couldn't swing it but because trying to would take focus away from things I want more like lifting heavy weights and tinkering with tech at home. But for absolute top priorities, I don't think it's ever worth giving up on the concept. Hard limits can be handled by reframing what success looks like and path to get there, not giving up on the essence.

Just don't blame me if your coding agent curses CEO and bypasses presubmits with dirty hacks a year later, I never volunteered to be a role model.

That blog post is human prompted, anyone who has experience with AI knows the difference between AI originated content (tables and bullet points) and AI spicing up a human prompt with detailed roasting instructions. Been there, done that (harmlessly like mocking concepts not targetting individuals).

I have a 128gb box to grow my own tokens at home. Gemma 4 31B uncensored works great.

Yes, I spend my days writing lots of code using AI (I do rigorously review it, it's still much faster than hand typing) and I get paid enough for it to pay mortgage and send kids to college.

You just need to work on your agent design and prompting skills, modern LLMs are crazy good at all the things you listed with the right context and tools.

China's AI vassal state. But so far in a good way because it's open source and everyone benefits.

What should people who can't afford to hire a white collar professional do?

It's the latest form of elitism - like in less sunny countries it's fashionable to be tanned because it means you are rich enough to have time to hang around on the beach while in sunny countries it's fashionable to be light skinned because you are rich enough that you don't have to work in the sun. Disdain for AI is a luxury belief of those who are either talented enough to draw / write / code without or are wealthy enough to not have to.

My news reading AI policy - I will use whatever source for news, AI or human, that gives me the best news. A lot of times it's human, like I appreciate diverse human comments on ycombinator. CNN used to have comments threads on articles and then gave that up because some comments were spicy, so I stopped reading. I don't remember the last time I went to arstechnica, I guess because they didn't standout compared to just asking Grok what's new in tech or browsing Reddit. If they could have used more AI to make their site more interesting, they should have.

I don't eat grass.

That's a dim view, people also contribute to make projects work for their own needs with hopes to share fixes with others. Like if I make a fix to vLLM to make a model load on particular hardware, I can verify functionality (LLM no longer strays off topic) and local plausibility (global scales are being applied to attention layers), but I can't pretend to understand full math of the overall process and will never have enough time to do so. So, I can be upfront about AI assist and then maintainer can choose to double check, or else if they don't have time, I guess I can just post a PR link on model's huggingface page and tell others with same hardware they can try to cherrypick it.

What's missed is that neither contributors nor maintainers are usually paid for their effort and nobody has standing to demand that they do anything they are not doing already. Don't like a messy vibe coded PR but need functionality? Then clean it up yourself and send improved version for review. Or let it be unmerged. But don't assign work to others you don't employ.

On the other hand, companies like NVIDIA should be publicly taken to task for changing their mind about instruction set for every new GPU and then not supporting them properly in popular inference engines, they certainly have enough money to hire people who will learn vLLM inside out and ensure high quality patches.

I just got MiniMax $200/year token plan. Usually it works fine for daily coding, if it gets stuck I pay for some Claude API calls through Roo gateway. Unlike other plans, this one officially supports running OpenClaw or other API workflows and doesn't suspend you long term if you use too many tokens, just set rate per few hours.