sanxiyn

26 ms·

421.
by sanxiyn 3y ago
Which spirit? The spirit of the GPL is four essential freedoms. https://www.gnu.org/philosophy/philosophy.html You should be able to cite the number between 1 and 4.
422.
by sanxiyn 3y ago
Terminating the subscription and preventing you from receiving the binary is Red Hat's prerogative. GPL does not say why Red Hat should not do so, and philosophically, it also does not contradict any software freedom.
423.
by sanxiyn 3y ago
No, 100% not. When you change the source, you are not obligated to give the change to anyone else.
424.
by sanxiyn 3y ago
No, I don't think it is against GPL, certainly not the letter, and not even the spirit. GPL gives you source. GPL does not give you stream of updates.
425.
by sanxiyn 3y ago
Agreed. Red Hat is simply doing what grsecurity did in 2017. It is legal and does not violate any license. https://lwn.net/Articles/721848/
426.
by sanxiyn 3y ago
Eh well, it is very close to just working. From "Training LLMs with AMD MI250 GPUs and MosaicML": > It all just works. No code changes were needed. https://www.mosaicml.com/blog/amd-mi250
427.
by sanxiyn 3y ago
No. Bootstrap is involved only if you are making a breaking change. For normal development, this doesn't change anything.
428.
by sanxiyn 3y ago
Zig already bootstraps from its C backend, so the second issue is not a problem.
429.
by sanxiyn 3y ago
MLIR is "LLVM done better", in fact by the same person. It fixes many of unforced LLVM problems, for example LLVM's inability to parallelize code generation.
430.
by sanxiyn 3y ago
I am not sure when you last checked, but for a long time SoCs like Qualcomm Snapdragon included Wi-Fi and Bluetooth on chip.
431.
by sanxiyn 3y ago
I agree the post could include more concrete metrics than self-assessment, but more than 1000 data points, even self-assessment data points, are usually not called anecdotes.
432.
by sanxiyn 3y ago
I am somewhat pessimistic because LLVM is significantly behind GCC in debug information quality (for example, see https://robert.ocallahan.org/2018/11/comparing-quality-of-de... ) and there is limit to what Rust ca
433.
by sanxiyn 3y ago
Given that Linux kernel is using async Rust which is implemented on top of kernel workqueue, and tokio won't be usable in kernel, it is a good thing tokio is not blessed. That's what enables both userland and kernel to use async R
434.
by sanxiyn 3y ago
Which preprocessor do you recommend for writing C?
435.
by sanxiyn 3y ago
Cyclone is garbage collected. Yes, Cyclone is probably the single biggest influence on Rust. No, but Rust had to introduce theoretical novelties on top of Cyclone. (Rust was very reluctant to do "research", considering its job to
436.
by sanxiyn 3y ago
While I think n>1000 data points on Rust learning curve is informative, I think build time complaint is less so. This being Google, it probably means something like "this C++ build takes 24 hours locally, but thanks to magical distr
437.
by sanxiyn 3y ago
Program termination is a solved problem with gas metering. Ethereum popularized the idea, but the idea itself is as old as hills.
438.
by sanxiyn 3y ago
This is such an obvious solution that I wonder why eBPF exists at all. WebAssembly is better for the purpose in like every way? Be against Not-Invented-Here, don't reinvent the wheel.
439.
by sanxiyn 3y ago
As a concrete example, if you look at an electric grid map of South Korea (here is one: https://www.e2news.com/news/articleView.html?idxno=236289 ), you will find that POSCO, Hyundai Steel, and Samsung Electronics have
440.
by sanxiyn 3y ago
Not this, but using the token selection restriction approach, you can let LLM produce output that conforms to arbitrary formal grammar completely reliably. JavaScript, Python, whatever.
441.
by sanxiyn 3y ago
For various reasons, token selection may be implemented as upweighting/downweighting instead of outright ban of invalid tokens. (Maybe it helps training?) Then the model could generate malformed JSON. I think it is premature to infer f
442.
by sanxiyn 3y ago
It's not temperature, but sampling. Output of LLM is probabilistic distribution over tokens. To get concrete tokens, you sample from that distribution. Unfortunately, OpenAI API does not expose the distribution. You only get the sample
443.
by sanxiyn 3y ago
Note that this (token selection restriction) is even available on OpenAI API as logit_bias.
444.
by sanxiyn 3y ago
This is wrong, byte-level models work fine, even if not as well as word-level models. From comparison of byte-level models and word-level models, we know tokenization part is responsible for minuscule part of performance.
445.
by sanxiyn 3y ago
Yes, "ByT5: Towards a token-free future with pre-trained byte-to-byte models" for example. https://arxiv.org/abs/2105.13626
446.
by sanxiyn 3y ago
In a trait-like system (including Java interface), there can be only one implementation of a trait for a type. That's why it is global, unlike ML module-like systems. Global is another word for anti-modular. For example, in the real wo
447.
by sanxiyn 3y ago
The usual alternative is a committee, see for example Apache Software Foundation's Project Management Committee Guide: https://www.apache.org/dev/pmc.html
448.
by sanxiyn 3y ago
Well yes, but you can't use the algorithm in the dragon book to generate three address code to generate SSA, because SSA has additional constraints. The dragon book really dropped the ball here, there is no excuse why it doesn't i
449.
by sanxiyn 3y ago
Yes.
450.
by sanxiyn 3y ago
Besides devoting too many pages to parsing, I think compiler practitioners have low opinion of the book because it isn't useful for them. For example, the second edition claims to be updated to modern optimization techniques, and in a

More ›