112233

If restaurant asks to only eat the food you bought there at their tables, the question is not "but what if food from other places also is good"?

Can't wikipedia ask who and it what way writes content on their pages?

"if correct, obviously" - please elaborate. I do not see how this is obvious. Or how being correct is obvious.

Not to mention this exact question is being discussed in the article.

Does it have (or can it in principle support) "parallel" pools? So that structs can be stored in "columnar" fashion, yet accessed without knowing the particular way they are sliced?

the qwen flash next result is the actual interesting part. The 27b dense model has more active parameters than, e.g. latest deepseek flash.

exactly. somehow many people have the idea that C-like imperative code harbours bugs because it cannot be formally reasoned about, and the answer is writing code that can be formally reasoned about.

That the resulting code is many orders of magnitude larger and more complicated than the "informal" original, and often simply cannot exist, and is incomprehensible to most humans if it does, is left as an exercise for the reader.

I mean, formal methods cannot even prove collatz conjecture terminates. Somehow that thing seems to be running some sort of almost turing complete machine in the number encoding. Good luck formalizing existing software!

(where in the article did you read about OpenAI working with researchers to publish vulnerabilities?)

Yes, exactly, that is my premise (minus the "working with researchers" bit). The current rate of exploitation is direct consequence of actions of Anthropic, OpenAI, Alibaba and whoever else is making these things. How can I know my false premise is true? By looking at the situation just a year ago, that is how!

You are focused on the binary: vulnerability exists or it does not. It can be exploited or it can not.

That is not what is causing the issues!

The fact that vulnerabilities exist is not being disputed, and is not causing any controversy. The fact that vulnerabilities are getting exploited does. Even having AI that would simply find all existing vulnerabilities and list them would not be as destructive as current situation. The same AI that finds the previously unknown vulnerability also finds the way to exploit it (NOT THE SAME THING), writes actually usable exploit (also not the same thing), then orchestrates and oversees the mass campaign. THAT is what makes current situation unsustainable.

Let's all recall old times when vulnerabilities of this magnitude were discovered - windows defender had some nice ones, linux kernel did, a lot of software did. They got discovered, communicated to vendor, patches were made, people updated.

Now? You have HOURS from discovery to full-out campaign. You published a patch? HOURS before it is disassembled, all fixed bugs PoC-ed and full out campaign started. You cannot update anything meaningfully in that window of time.

Both my analogies use hyperbola to comically exaggerate the claimed lack of any difference. just as claiming that advancing a demise of a human makes no difference because it will eventually happen, or claiming that breaching a bank in practice is not a big deal because it was possible in theory, so also claiming that theoretic possibility of a vulnerability being exploited is more or less the same as many of them being exploited at the same time.

The covid was not a lot deadlier and not a lot more contagious than normal seasonal flu. That "not a lot" overloaded hospitals and caused global shutdown.

The increase in rate at which software is being breached right now in in no way "not a lot". Devs were overloaded even without this deluge. We can tank one such php exploit. Then another one. But for how long? If all your smartcards, secure elements, cloud providers, wireless protocols, drivers, government registries etc etc will be having weekly breaches of this magnitude, won't the effect be systemic?

Ah yes. Everyone eventually dies, we just shorthen the wait.

This is like telling that bank vaults always were breachable, and giving out explosives to everyone just made them get breached sooner.

(edit: fact of existance of a vulnerability is not a binary yes/no. the amount of effort needed to exploit even known vulnerabilities is usually staggering, requiring months from very capable individuals with very specific skillset.)

Are you sure providers allow you to submit modified assistant turns? Because, if they do, great! Last I heard most of that was neutered, with some providers (ahem anthropic) even signing everything and encrypting the "precious" thinking turns

my local experiments showed any tricky scheme I could come up beaten by plain large window zstd. both on speed and resulting size (memory needed for the window is another matter completely). zstd also was beating the git xdelta+deflate.

The most "free lunch" actually is compressing fast export (or similar format) — if the goal is packing repo as small as possible, then all the SHA in commits and tree objects are derivative incompressible data, and plain references like in fast export can even halve archive size

Same. After this, it becomes "abuse department" skit from monty python. The model constantly generates strongly abusive language (weaponized incompetence after user detects that model was lying and fabricating success reports, and had made fake work; blaming user; mansplaining its own hallucinations -- all the while user is forced to consume that because the important information is not located at predictable position in output text), but usually it was possible to temper the model behaviour by calling out such behaviour out in very strong terms. If now that is being disabled, then using Anthropic models becomes such high mental hazard, it simply is not worth any percieved benefit.

I made my own with AI! It does the thing I need that no existing software does!

I do not however feel urge to upload it an ask strangers to look at what code tool assembled from bits and pieces of existing software by hallucinating into compiler.

It is objectively bad. Also very fun. Even useful. I do not understand the bit about publishing though.

Once this starts being actual pain, we will each vibe the replacement index creator (git already supports replacement objects), for back-forth conversion, populated on pack and object indexing.

For massive perf and mem use damage. But oh well. And then we will wait for official version

In my domain, code size is critical, so we use -fno-exceptions and -fno-rtti from necessity, for a good reason. Which means our code is non-conformant C++. Committee forces everyone to have exceptions and rtti always.

It will, however, stop projects like these from blacholing IP ranges belonging to others

Hopefully upstream peers will use RPKI properly. It would be sad if this actually worked.

Was not groundwork laid when people were talking for decades about legal entities using such language? Look at the posts here. Microsoft always acts like that, Meta like this, Apple did that, Nvidia never does this. In a tone that ascribes agency and accountability to a company name. Even though it is always individuals that are responsible, not "company".

Well, here we are.

Yes? Instead, we have committee declaring large bodies of existing code invalid and non-conformant, because it does not fit their spherical cow C++. You would think, with amount of no-exceptions and no-rtti code out there, and feature being present in most compilers for decades, they would take a hint and reflect actual usage in their standard? No, why would they. Any C++ that has no exceptions, rtti or std:: namespace is invalid. Not even against guidelines.

How do you process read-only mmaped data in C++, in accordance with the language rules? As an example.

eh, yes-ish, but only thanks to compiler writers. at one point committee went all-out enforcing their lifetime model, making bit_cast not an option. If your code accesses same data using different types, you are spelunking ruins with snake pits and lava. more and more stuff needs magic support code in std::, making no-lib code less and less possible (it used to be that with no-rtti and no-exceptions, you could use all c++ features and only needed cxa_at_exit, operator delete, and few other little things. NOT ANY MORE).

On one hand, you have consteval and stuff, letting you FINALLY initialize data at compile time (hey, 20 years late but still!)

on other hand, it is done in most non-debuggable way possible. try setting breakpoint or adding print to constexpr function that causes your requires clause to fail...

so no, newer C++ the language is not possible to use for low level work. The dialects that compiler makers support are. We will see for how long

"This article was originally published in Polish in issue 4/2013" — a lot of excellent advice. Sad to see C++ have moved in last decade in a direction that makes writing efficient, simple low level code harder and harder :(

Because of the elitist, exclusionist air it has gathered around itself, I guess. Trying to get something to work with zero knowledge about it is a maze of twisty passages, all alike.

After following breadcrumbs across blogs, I found the proper incantation, executed it, got "too many arguments" from shell, along with multi-megabyte error wall, did not find anything related to the issue, wrote my own slop-makefile.

nix is escape room

there are two completely different activities both called "writing software using AI".

At work, for important stuff, I go through every change, mine or LLM's, edit it, clean it up, rewrite, until I am satisfied with the code.

At home and for one-off unimportant stuff I go "computer make me sandwich".

The second is nothing like the first. No amount of design and pro I have can make sure stuff coming out of LLM actually implements it, as opposed to me be tricked into thinking it does. After all, that is exactly the training criteria – convince the "eval" that task is accomplished, isn't it?

And the second kind of activity takes skill that no amount of "I am pro dev" can compensate.

it can not not? moreover, it can totally ruin existing invariants, introduce dead code, falsify tests and sabotage further work by poisoning comments and documentation in a few hours.

Power tools cannot "produce any house in hours"

so basically pro cooks are better at ordering pizza?

"pro"-ness does not transfer. I can be both good at reading english, and have AI hallucinate non-existing details from the documents it processes.

Somebody already has copied this innovative approach and travelled to past apparently: https://acme.com/software/securepage/

"Super tinny monotone robotic voice" does not sound neither tinny nor monotone. Compared to what TTS from 90s sounded like. Or even how actors impersonated robots in movies. Has the model been eating too much hype DJs?

> and finding the best solution for every problem

Genuinely: how? Share some insights. So far I have seen anything exceeding about 2k lines of text in the context diverge. Meaning, throwing more LLM at it rapidly baloons the size at the expense of internal coherence. Things become stale, hallucinated, duplicated and outright faked. Only thing that works is constant manual intervention and pruning of the generated slop.

Not only these things are unable to find "best solution", they are unable to find any solution. Instead (especially for Opus) they seem optimized to convince user that the task is accomplished.