Hacker News: cat_plus_plus

New comment by cat_plus_plus in "AI assistance when contributing to the Linux kernel"

cat_plus_plus — Sat, 11 Apr 2026 17:17:43 +0000

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.

New comment by cat_plus_plus in "Reallocating $100/Month Claude Code Spend to Zed and OpenRouter"

cat_plus_plus — Fri, 10 Apr 2026 18:32:15 +0000

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.

New comment by cat_plus_plus in "The Claude Code Source Leak: fake tools, frustration regexes, undercover mode"

cat_plus_plus — Tue, 31 Mar 2026 23:15:02 +0000

Why are you assuming the actual implementation was authored by a human?

New comment by cat_plus_plus in "Montana passes Right to Compute act (2025)"

cat_plus_plus — Sat, 14 Mar 2026 16:57:44 +0000

So what does liberal even mean these days? California is passing bs like age verification in OS and Montana is protecting my right to leave the way I want in my own home, running whatever AI models suit me as long as I am not bothering anyone. That's just another "none of government business" personal freedom issue like pot or sexuality, why aren't blue states all over it. And yes, using tuned LLMs can be like an acid trip, but the distance between having a trip at home and tangible harm is much greater than in the case of access to guns, knives, power tools, cars and rodent poison yet at least some of these are widely available to law abiding citizens in every state. Government interventions can be staged at the points where there is evidence of actual imminent harm, like problematic public behavior. Why are Democrats the new "Reefer Madness" pearl clutchers and why should I still believe they have anything to do with living the way you want?

New comment by cat_plus_plus in "BitNet: 100B Param 1-Bit model for local CPUs"

cat_plus_plus — Wed, 11 Mar 2026 21:04:31 +0000

There are 1 bit average GGUFs of large models, not perfect quality but they will hold a conversation. These days, there is also quantized finetuning to heal the damage.

New comment by cat_plus_plus in "Open Weights isn't Open Training"

cat_plus_plus — Tue, 10 Mar 2026 23:18:55 +0000

Well, it's open training in the sense that the code is open source and you are free to fix it so it trains successfully. That's consistent with how open source works generally. In my experience unsloth is where new model training is usually fixed first.

New comment by cat_plus_plus in "LLMs work best when the user defines their acceptance criteria first"

cat_plus_plus — Sat, 07 Mar 2026 02:33:45 +0000

The actual task is usually to mix something that looks like a dozen of different open source repos combined but to take just the necessary parts for task at hand and add glue / custom code for the exact thing being built. While I could do it, LLM is much faster at it, and most importantly I would not enjoy the task.

New comment by cat_plus_plus in "LLMs writes plausible code, not correct code"

cat_plus_plus — Sat, 07 Mar 2026 02:27:56 +0000

Try writing code from description without looking at the picture or generated graphics. Visual LLM with a suggestion to find coordinates of different features and use lines/curves to match them might do better.

New comment by cat_plus_plus in "LLMs work best when the user defines their acceptance criteria first"

cat_plus_plus — Sat, 07 Mar 2026 02:16:10 +0000

That's very impressive. Your LLM actually wrote a correct code for a full relational database on the first try, like it takes 2.5 seconds to insert 100 rows but it stores them correctly and select is pretty fast. How many humans can do this without a week of debugging? I would suggest you install some profiling tools and ask it to find and address hotspots. SQL Lite had how long and how many people to get to where it is?

New comment by cat_plus_plus in "Where things stand with the Department of War"

cat_plus_plus — Sat, 07 Mar 2026 01:40:55 +0000

I have an NVIDIA Thor Dev Kit, a somewhat less known cousin of DGX Spark.

New comment by cat_plus_plus in "Where things stand with the Department of War"

cat_plus_plus — Fri, 06 Mar 2026 01:36:26 +0000

There are so many inference providers not working for Department of War. Even Alibaba and sure China has lots of issues but they are not bombing anyone now if that's your first priority. Or else, smaller US / European / Asian companies with pure civilian focus. SOTA open weights models they serve are perfectly suitable for coding and chat. I run a local Qwen3.5-122B-A10B-NVFP4 instance and it writes entire Android apps from scratch and that's a midsized model.

New comment by cat_plus_plus in "Did Alibaba just kneecap its powerful Qwen AI team?"

cat_plus_plus — Wed, 04 Mar 2026 17:09:51 +0000

Do you have evidence that they were sacked rather than resigned because they would rather work in a different direction from the one company is taking?

New comment by cat_plus_plus in "Did Alibaba just kneecap its powerful Qwen AI team?"

cat_plus_plus — Wed, 04 Mar 2026 16:48:40 +0000

Somewhat of a devil's advocate here, because I am very familar with corporate idiocy. But how do you define a non-sociopathic corporate scenario where a company makes a lot of money from a good product they develop? Even if done in maximally practically and emotionally intelligent way, this still requires changes from research phase no?

New comment by cat_plus_plus in "Did Alibaba just kneecap its powerful Qwen AI team?"

cat_plus_plus — Wed, 04 Mar 2026 16:41:41 +0000

Well, Alibaba hiring a Gemini guy to run Qwen suggests they want to make Qwen into a big consumer / enterprise business like Gemini. I am not sure that I blame them even if it clashes with how their top researchers were hoping Qwen would be run. Most obviously it's natural for a company to want to make money on things they paid for developing. But also, the world needs more competition in AI businesses just like it needs competition in AI research. I wouldn't mind Qwen code to grow into a commercial grade competitor to Claude code that is better, faster and cheaper. I am sure the talented researchers can find a new home in Moonshot AI or even US college or startup.

New comment by cat_plus_plus in "Addressing Antigravity Bans and Reinstating Access"

cat_plus_plus — Sat, 28 Feb 2026 17:30:41 +0000

The problem is that Google treats its customers as college kids who can be banned from a college maker lab for using too much 3D filament rather than entrepreneurs who are trusting their livelyhood to a service provider that promises to be reliable. If War Department uses too many Gemini tokens, do they cut them off, make them go through recertification process and permaban the next time around?

Which means that anyone serious about AI and not going local route should be using a provider with better reputation. I don't know if Alibaba, Z.ai or moonshots AI are also known for hair trigger responses, could be decent options for coding AI otherwise? If not, time to look for smaller providers with good reputation?

New comment by cat_plus_plus in "Stoat removes all LLM-generated code following user criticism"

cat_plus_plus — Sat, 14 Feb 2026 18:28:54 +0000

"it's worth considering that there are many people with incredibly strong anti-LLM views, and those people tend to be minorities or other vulnerable groups."

I have pretty low expectations for human code in that repository.

New comment by cat_plus_plus in "We mourn our craft"

cat_plus_plus — Sun, 08 Feb 2026 16:17:01 +0000

What are you inputs and outputs? If inputs are zip files and outputs is uncompressed text, don't use an LLM. If inputs are English strings and outputs are localized strings, LLMs are way more accurate than any procedural code you might attempt for the purpose. Plus changing the style of outputs by modifying inputs/weights is also easier, you just need to provide a few thousand samples rather than think of every case. Super relevant for human coding, how many hobbyists or small businesses have teams of linguists on staff?

New comment by cat_plus_plus in "I am happier writing code by hand"

cat_plus_plus — Sun, 08 Feb 2026 16:10:41 +0000

Good for artists I guess, I wouldn't know because I am not one. The best I can manage is drawing a stick figure of a cat. Years back I was working on a Mac app and I needed an icon. So I talked to an artist and she asked for $5K to make one for me. I couldn't justify spending so much on a hobby that I didn't know if it would go anywhere so I wrote a little app that procedurally generated me some basic sucky icon. I am sure Gordon Ramsay is also not impressed with cooking skills of my microwave, I just don't know how his objections practically relate to getting me fed daily.

New comment by cat_plus_plus in "I am happier writing code by hand"

cat_plus_plus — Sun, 08 Feb 2026 16:01:10 +0000

I am TL of an Android app with dozens of screens that expose hundreds of different distinct functions. My task is to expose all of these functions as appfunctions that can be called by an LLM in response to free form user requests. My current plan is to build a little LangGraph pipeline where first step is AI documenting all functions in each app's fragment, second step is extracting them into app functions, then refactoring fragment to call app functions etc. And by build I mean Gemini will build it for me and I will ask for some refinement and edit prompts.

I also like writing code by hand, I just don't want to maintain other people's code. LMK if you need a job referral to hand refactor 20K lines of code in 2 months. Do you also enjoy working on test coverage?

New comment by cat_plus_plus in "We mourn our craft"

cat_plus_plus — Sun, 08 Feb 2026 15:42:24 +0000

I have no idea what everyone is talking about. LLMs are based on relatively simple math, inference is much easier to learn and customize than say Android APIs. Once you do you can apply familiar programming style logic to messy concepts like language and images. Give you model a JSON schema like "warp_factor": Integer if you don't want chatter, that's way better than Star Trek computer could do. Or have it write you a simple domain specific library on top of Android API that you can then program from memory like old style BASIC rather than having to run to stack overflow for evwery new task.