But I can ask 100 causally masked questions against common prefix, and get 100 answers, all in a single PP pass using any existing "classical" attention transformer model? Like, I had the impression that is what everyone was doing for classification already?
Is the difference "we did RL to tune logit distribution"? Because I really do not see anything new there. What is the difference?
ok, point camera at a screen. let's counter that by also taking 360° image with secondary camera. Ok build some props, let's add gps into it. fake gps. let's add inertial tracking and self-destruct on tamper.
Ok, but we really need a fake image, here is our badge, give us signing keys for special use case...
By narrow luck compiler writers so far have been the sane bunch, and have ignored C++ committee on many important points. Thus we still have explicitly non-conformant things like -fno-exceptions that lets one use C++ compiler on embedded.
But I wonder how long that can last, with the way C++ is going.
At one point, it will make practical sense to update codebase to some other language, rather than keep fighting this one
I have been saying that C++23, or maybe C++26 due to reflection, will eventually be the last standard that actually matters.
For a large number of C++ users, it boils down to what it offers beyond C, but not to the extent WG21 is driving it since C++20.
Also the major surviving three compilers have lost wind on their sails as the corporations sponsoring their development have switched focus to other compiled languages.
Other than the whole security debate, there are no features that would make C++ significantly better for LLVM, GCC, CLR, V8, CUDA,.. improvements.
In fact, some of those projects still require C++17.
If this sounds strange, how many care nowadays about ISO Fortran 2023, or ISO COBOL 2023, despite the amount of software written in them powering many busisesses, or Python libraries even, e.g. SciPy.
Or even with C, almost 20 years later many still reach out to C99, ignoring everything else.
Not to take away from your points; SciPy is now Fortran-free completely[0] (we are also requiring C++17 at most). NumPy never had it. BLAS is all C/Assembly in all optimized vendors. For LAPACK we are working on it [1].
Once there is enough pain, none of the talking points matter for any language. They don't and can't die but linger. I fear that time for C family might come in a decade which would be a shame given how magical Cpp compilers are, all that effort folks pouring in.
Not "this". By that logic most people do not need backups too. Rebuilding working service state after config change is non-trivial, requires using admin-only kernel interfaces, and said kernel does not care about these types of problems. Meaning average user ends up simply rebooting rather than trying to figure why enabling wifi makes printer stop working (and such). We cannot have network stuff controlled by multiple programs, because linux kernel is broken like that. This thing is the userspace part of the monolith.
> With Verus, however, developers can mathematically prove the safety of their unsafe Rust code
what about proving safety of not-unsafe code? The meme that rust is "safe" is becoming tiring. Does this thing allow proving absence of infinite loops? Bounded resource use? Correctness of comparison operations? Etc.
Also, why is there still no hardware tagging to simply prevent memory misuse at cpu level, if it actually is such an important issue?
Objection was against "memory safety" somehow having become "safety". Here is an epic tool that allows one to attach and formally prove assertions. Genuinely impressive and massively useful.
Yet the pitch is that it is needed for the "unsafe" unsafe code, to make it "safe". Not for all code, to make all code safe.
Sometimes you can use SMT for “theorem proving”. It is a rather broad term. I don’t think they added something much different than what they already had.
"but strong mode-decision pipelines can pick the right mode for the job and save lots of bits" - citation needed. just a bit above author claims lossless is not useful for web, disregards modular jxl that uses nearby pixel predictors (even though it can be lossy, too), then produces this claim against vardct. spoiler - dct frequencies ARE using "nearby pixel values"
great distinction. can we use it to freely produce and share mp3/h264 copies of any media? because it is provably impossible to reconstruct original based on that information.
But keep in mind: So far there is no way to train the models while completely avoiding memorization and only including generalization. That would be a great way to avoid any copyright issues, but all attempts I have seen so far were fairly limited.
Because I really want to see difference between approximating input signal using smooth orthogonal basis functions, and approximating those using quantization with rate/distortion feedback as "violates copyright, compression algorithm", and then approximating input signal using perceptron, with backprop loss based off original media as "not compression algorithm, skip jail"
reply