HackerTrans
TopNewTrendsCommentsPastAskShowJobs

cmdr2

no profile record

Submissions

New token system and scaling is 10x trash

github.com
3 points·by cmdr2·เดือนที่แล้ว·1 comments

How the PolyBlocks AI Compiler Works

docs.polymagelabs.com
1 points·by cmdr2·8 เดือนที่ผ่านมา·0 comments

GPU Compilation with MLIR

stephendiehl.com
2 points·by cmdr2·9 เดือนที่ผ่านมา·0 comments

comments

cmdr2
·5 เดือนที่ผ่านมา·discuss
> If you as a person outsourced your thinking than it's you who will suffer.

Using a code generator != outsourcing your thinking. I know that's the popular opinion, and yes, you can use it that way. But if you do that, I agree you'll suffer. It'll make sub-optimal design decisions, and produce bloated code.

But you can use code generators and still be the one doing the thinking and making the decisions in the end. And maintain dictatorial control over the final code. It just depends on how you use it.

In many ways, it's like being a tech lead. If you outsource your thinking, you won't last very long.

It's a tool, you're the one wielding it, and it takes time, skill and experience to use it effectively.

I don't really have much more to say. I just spoke up because someone who built something cool was getting beat up unnecessarily, and I've seen this happen on HN way too many times recently. I wasn't pointing fingers at you at any point, I'm glad to have had this discussion :)
cmdr2
·5 เดือนที่ผ่านมา·discuss
> this is presumably a 16k loc project that was commited in a single commit two days ago. So the author never commited any draft/dev version in the time

It's quite common to work locally and publish a "finished" version (even if you use source control). The reasons can vary, but I highly doubt that Google wrote Tilt Brush in 3 commits - https://github.com/googlevr/tilt-brush

All I'm saying is assuming everyone one-shots code (and insulting them like people do on HN), is unnecessary. I'm not referring to you, but it's quite a common pattern now, counter to HN's commenting guidelines.

> found AI-like comments in the code

Sure, but respectfully, so what? Like I posted in a [separate comment](https://news.ycombinator.com/item?id=47057690), code generators are like power tools. You don't call a carpenter sloppy because they use power tools to drill or cut things. A sloppy carpenter will be sloppy regardless, and a good carpenter will obsess over every detail even if they use power tools. A good carpenter doesn't need to prove their worth by screwing in every screw by hand, even if they can. :)

In some cases, code generators are like sticks of dynamite - they help blow open large blocks of the mountain in one shot, which can then be worked on and refined over time.

The basic assumption that annoys me is to assume that anyone who uses AI to generate code is incompetent and that their work is of poor quality. Because that assumes that people just one-shot the entire codebase and release it. An experienced developer will mercilessly edit code (whether written by an AI or by a human intern), and edit it until it fits the overall quality and sensibility. And large projects have tones of modules in them, it's sub-optimal to one-shot them all at once.

For e.g. with tests, I've written enough tests in my life that I don't need to type every character from scratch each time. I list the test scenarios, hit generate, and then mercilessly edit the output. The final output is exactly what I would've written anyway, but I'm done with it faster. Power tool. The final output is still my responsibility, and I obsessively review every character that's shipped in the finished product - that is my responsibility.

Sure plenty of people one-shot stuff, just like plenty of Unity games are asset flips, and plenty of YouTube videos are just low-effort slop.

But assuming everything that used AI is crap is just really tiring. Like [another commenter said](https://news.ycombinator.com/item?id=47054951), it's about skilled hands.

> something that the original author did not even bother to write

Again, this is an assumption. If I give someone bullet points (the actual meat of the content), and someone else puts them into sentences. Did the sentences not reflect my actual content? And is the assumption that the author didn't read what was finally written, and edit it until it reflected the exact intent?

In this case, the author says they used AI to generate the ASCII art in question. How does that automatically mean that the author AI-generated the entire readme, let alone the entire project? I agree, the knee-jerk reactions are way out of proportion.

Where do you draw the line? Will you not use grammar tools now? Will you not use translation tools (to translate to another language) in order to communicate with a foreign person? Will that person argue back that "you" didn't write the text, so they won't bother to read it?

Should we stop using Doxygen for generating documentation from code (because we didn't bother with building a nice website ourselves)?

Put simply, I don't understand the sudden obsession with hammering every nail and pressing every comma by hand, whereas we're clearly okay with other tools that do that.

Should we start writing assembly code by hand now? :)
cmdr2
·5 เดือนที่ผ่านมา·discuss
For what it's worth - for people who're in this space - this project is awesome, and I hope you keep going with it! The compiler space for GPUs really needs truly open-source efforts like this.

Code/doc generators are just another tool. A carpenter uses power tools to cut or drill things quickly, instead of screwing everything manually. That doesn't mean they're doing a sloppy job, because they're still going to obsessively pore over every detail of the finished product. A sloppy carpenter will be sloppy even without power tools.

So yeah, I don't think it's worth spending extra effort to please random HN commenters, because the people who face the problem that you're trying to solve will find it valuable regardless. An errant bold or pipe symbol doesn't matter to people who actually need what you're building.
cmdr2
·5 เดือนที่ผ่านมา·discuss
> This is obvious AI slop from the readme

I keep hoping that low-effort comments like these will eventually get downvoted (because it's official HN policy). I get that it's fashionable to call things AI slop, but please put some effort into reading the code and making an informed judgment.

It's really demeaning to call someone's hard work "AI slop".

What you're implying is that the quality of the work is poor. Did you actually read the code? Do you think the author didn't obsessively spend time over the code? Do you have specific examples to justify calling this sloppy? Besides a misaligned "|" symbol?

And I doubt you even read anything because the author never talked about LLMs in the first place.

My beef isn't with you personally, it's with this almost auto-generated trend of comments on HN calling everyone's work "AI slop". One might say, low-effort comments like these are arguably "AI slop", because you could've generated them using GPT-2 (or even simple if-conditionals).
cmdr2
·10 เดือนที่ผ่านมา·discuss
> Google is acting here as the owner and maintainer of a services ecosystem... > they increasingly experience difficulty to contain issues within that ecosystem and prevent them from spreading (piracy, malware, hacking,...)

I wonder if the smartphone app industry is big enough now that allowing just two corporates to govern them is no longer fair or democratic.

It has outgrown the "ecosystem" word a long time ago. It's a genuine industry now.

Apps are such a fundamental part of most people's lives now (whether they like it or not), and these two companies have a disproportionate amount of power over an entire industry.
cmdr2
·10 เดือนที่ผ่านมา·discuss
https://pypi.org/p/torchruntime might help here, it's designed precisely for this purpose.

`pip install torchruntime`

`torchruntime install torch`

It figures out the correct torch to install on the user's PC, factoring in the OS (Win, Linux, Mac), the GPU vendor (NVIDIA, AMD, Intel) and the GPU model (especially for ROCm, whose configuration varies per generation and ROCm version).

And it tries to support quite a number of older GPUs as well, which are pinned to older versions of torch.

It's used by a few cross-platform torch-based consumer apps, running on quite a number of consumer installations.