titzer

30 ms·

601.
by titzer 2y ago
> You also don’t need to do dynamic linking. This is a big one. Linking against libc on many platforms also means making your binaries relocatable. It's a lot of unnecessary, incidental complexity .
602.
by titzer 2y ago
There are several advantages to using kernel syscalls directly: 1. No overhead from libc; minimizes syscall cost 2. No dependency on libc and C language ABI/toolchains 3. Reduced attack surface. libc can and does have bugs and potentia
603.
by titzer 2y ago
> all those targeted ad algorithms are effectively just really dumb. They are. I've had a gmail account for almost 20 years. Hell, I worked for Google for 10 years. They know who I am. They've seen me search the internet and
604.
by titzer 2y ago
> Basically removing the dealers and therefore the OEM's stranglehold on software would take the teeth out of emissions enforcement. I don't buy that. This strangehold is the only way that VW managed to cheat emissions for year
605.
by titzer 2y ago
Model year and engine config (e.g. cylinder count) has usually been enough for auto parts for decades.
606.
by titzer 2y ago
From the writeup, it appears that microcode doesn't completely flush indirect branch predictions when instructed to do so, which leads to known cross-process attack techniques working again. Frankly I'm not surprised. Beyond the i
607.
by titzer 2y ago
Non-browser environments are a little behind on the Wasm standard, but not by much. E.g. wasmtime has now landed support for Wasm GC. AFAIK they implement all phase 4 proposals. Wizard implements all the Phase 4 proposals as well. The Wasm
608.
by titzer 2y ago
> crash the WASM instance I guess it depends on how you get to said crash, but no, data races on Wasm shared memory cannot "crash" anything. At worst racy reads/writes can produce garbage (primitive) values and put garbage
609.
by titzer 2y ago
It's also worth noting that the NaCl and PNaCl teams were integrated into a large Wasm team at Google and brought their expertise to the project. While we didn't all 100% agree on every decision made in Wasm design, we were intima
610.
by titzer 2y ago
Wasm bytecode verification is more strict than JVM bytecode verification. For example, JVM locals don't have declared types, they are inferred by the abstract interpretation algorithm (one of the reasons for the afore-mentioned O(n^3)
611.
by titzer 2y ago
I worked on JVM bytecode for a significant number of years before working on Wasm. JVM bytecode verification is non-trivial, not only to specify, but to implement efficiently. In Java 6 the class file format introduced stack maps to tame a
612.
by titzer 2y ago
> you essentially can't target it with a single-pass compiler, That might be true if your source language has goto, but for other languages that start with structured control flow, it's possible to just carry the structure thro
613.
by titzer 2y ago
> has to first be written in a language that already has a compiler You only need an interpreter for that language.
614.
by titzer 2y ago
A minimal quine just prints itself out as "source" code. You can choose the source language to be whatever you like, such as a minimal Turing complete combinator. So all you need is an interpreter for the base level, which could b
615.
by titzer 2y ago
Well, Quine is a thing.
616.
by titzer 2y ago
> FP hardware usually special cases them to some slow but correct silicon This is basically only true on Intel chips. Subnormals are not that slow on modern ARM or AMD hardware.
617.
by titzer 2y ago
The future is looking like an incredibly huge pile of crappy cloud software that no one asked for, no one understands, perpetually stops working, and no one can fix except by replacing it by an even crappier, faster breaking piece of junk.
618.
by titzer 2y ago
> with enough time If a textbook is a megabyte (2^20) and might take a week to read and grok , then then it would take 2^20 weeks to read one terabyte (2^40), which is 20 thousand years. ChatGPT-3 was trained on 570GB of text data, ac
619.
by titzer 2y ago
Yes, the fuel costs less than half a cent per kWh. All the other costs dominate.
620.
by titzer 2y ago
Please do not post facts you got from asking an LLM.
621.
by titzer 2y ago
342 Watts per square meter on average over the entire Earth, averaged over the entire year. https://www.nasa.gov/wp-content/uploads/2015/03/135642main_b... Put some panels on your roof and it's not
622.
by titzer 2y ago
It depends on how high-level the abstraction you're using is. Things like Halide do pretty significant reorganization of your code and data to achieve parallel speedup. And of course if you are using Tensor libraries you're doing
623.
by titzer 2y ago
I generally concur. Compilers that do advanced optimizations represent both control flow and data flow of the program as a graph where the nodes can represent basic blocks or operations in the program. One of the most graph-like IRs is call
624.
by titzer 2y ago
This is much nicer in functional languages that have tuples. Then an array of tuples (if normalized or flattened) is both stored efficiently and ready for vectorization. You really don't want to model mathematical objects like vectors
625.
by titzer 2y ago
The whole idea that we're going to build software systems using natural language prompts to AI models which then promptly (heh) fall on their face because they mash together text strings to feed to a huge inscrutable AI is lazy and stu
626.
by titzer 2y ago
So yes I sympathize that people didn't read the instructions, but I feel that that is partly out of habit that said instructions don't exist. We could work towards a better status quo where the configuration steps are simple, easy
627.
by titzer 2y ago
> no one told her she needs to message us Did you have that written down somewhere? In instructions , which was my point. Sorry for the snark, but I think "no one knows how to do anything" coupled with "Oh, the idiot did
628.
by titzer 2y ago
Threaded dispatch is worth 15-30% in Wizard's fast interpreter.
629.
by titzer 2y ago
People gonna be people. It's up to engineers and product designers to make things user friendly but also safe-by-default. If something needs to be configured, then provide instructions on how to configure it. Instead of pretending that
630.
by titzer 2y ago
It's so wild that the bar for AI performance is both absurdly high and absurdly low at the same time. To specify an output format (language or grammar) for solving a computational problem is one of the oldest exercises around. On the o

More ›