TehPers

The distribution is super important here too. Hashing any value to zero (or h(x) = 0) is valid, but a terrible distribution. The challenge is getting real-world values hashed in a mostly uniform distribution to avoid collisions where possible.

Still, the contents of the article are useful even outside of hashing. It should just disclaim that the width of the output isn’t the only thing important in a hash function.

• Sometimes you are using language features your team is unfamiliar with.

Had this happen before with pattern matching.

Because you created a first draft. Your first draft should include all that info. It isn’t writing the whole doc for you lol, just making minor edits to turn it from notes into prose.

Without that? No clue, good luck. They can usually read source files to put something together, but that’s unreliable.

This would infuriate me to no end. It’s literally the definition of a data race. All data between threads needs to either be accessed through synchronization primitives (mutexes, atomic access, etc) or needs to be immutable. For the most part, this should include fds, though concurrent writes to stderr might be less of an issue (still a good idea to lock/buffer it and stdout though to avoid garbled output).

The main value I found from Copilot in vscode back when it first released was its ability to recognize and continue patterns in code (like in assets, or where you might have a bunch of similar but slightly different fields in a type that are all initialized mostly the same).

I don’t use it anymore though because I found the suggestions to be annoying and distracting most of the time and got tired of hitting escape. It also got in the way of standard intellisense when all I needed was to fill in a method name. It took my focus away from thinking about the code because it would generate plausible looking lines of code and my thinking would get pulled in that direction as a result.

With “agents” (whatever that term means these days), the article describes my feelings exactly. I spend the same amount of time verifying a solution as I would just creating the solution myself. The difference is I fully understand my own code, but I can’t reach that same understanding of generated code as fast because I didn’t think about writing it or how that code will solve my problem.

Also, asking an LLM about the generated code is about as reliable as you’d expect on average, and I need it to be 100% reliable (or extremely close) if I’m going to use it to explain anything to me at all.

Where I found these “agents” to be the most useful is expanding on documentation (markdown files and such). Create a first draft and ask it to clean it up. It still takes effort to review that it didn’t start BSing something, but as long as what it generates is small and it’s just editing an existing file, it’s usually not too bad.

This depends. Many languages support 1 liner aliases, whether that’s using/typedef in C++, type in Rust, Python, and TS, etc.

In other languages, it may be more difficult and not worth it, though this particular example should just use a duration type instead.

Ah yes, one of the major questions of software development: to comment, or not to comment? This is almost as big of a question as tabs vs spaces at this point.

Personally? I don’t really care. Make the code readable to whoever needs to be able to read it. If you’re working on a team, set the standard with your team. No answer is the universally correct one, nor is any answer going to be eternally the correct one.

Regardless of whether code comments should or shouldn’t exist, I’m of the opinion that doc comments should exist for functions at the very minimum. Describe preconditions, postconditions, the expected parameters (and types if needed), etc. I hate seeing undocumented **kwargs in a function, and I’ll almost always block a PR on my team if I see one where the valid arguments there are not blatantly obvious from context.

Carjacking is going to mean something completely different with those Samsung car robots out in the wild.

Also, imagine going home after a vacation and your house is just gone. (Yes I can see that they prepared the ground first, and I think your neighbor’s house driving down the street would raise some flags)

But seriously, this is really impressive. Great job Samsung and Shanghai Construction No 2!

Any clue if it’s open source? That seems like a good opportunity to spread it further and maybe get external support for an Android port.

DDR5-6000/CL28 should be fine. Make sure to enable the XMP/EXPO profile in your BIOS after installing it.

You can follow hardware reviewers like GamersNexus, LTT, HardwareUnboxed, etc if you want to stay up to date (which is what I do), or look at their content if you just want a review for a product you’re looking at.

It’s less of an issue now, but there were stability issues in the early days of DDR5. Memory instability can lead to a number of issues including being unable to boot the PC (failing to post), the PC crashing suddenly during use, applications crashing or behaving strangely, etc. Usually it’s a sign of memory going bad, but for DDR5 since it’s still relatively young it can also be a sign that the memory is just too fast.

Always check and verify that the RAM manufacturer has validated their RAM against your CPU.

Air cooling is sufficient to cool most consumer processors these days. Make sure to get a good cooler though. I remember Thermalright’s Peerless Assassin being well reviewed, but there may be even better (reasonably priced) options these days.

If you don’t care about price, Noctua’s air coolers are overkill but expensive, or an AIO could be an option too.

AIOs have the benefit of moving heat directly to your fans via fluid instead of heating up the case interior, but that usually doesn’t matter that much, especially outside of intense gaming.

Very few things need 64GB memory to compile, but some do. If you think you’ll be compiling web browsers or clang or something, then 64GB would be the right call.

Also, higher speeds of DDR5 can be unstable at higher capacities. If you’re going with 64GB or more of DDR5, I’d stick to speeds around 6000 (or less) and not focus too much on overclocking it. If you get a kit of 2x32GB (which you should rather than getting the sticks independently), then you’ll be fine. You won’t benefit as much from RAM speed anyway as opposed to capacity.

The latest round of “stuff I wasn’t informed would be installed for me” included enough software to switch me to Linux. I’m still dual booting during the transition, but moving fully over when I can.

I honestly used to love Windows too. Windows 10 was great, and 11 had problems but was still very usable on the happy path and came with some great improvements over time. These days, it’s just so full of bloatware. I just want my damn computer to be mine, and I’d hope an OS license that retails for $200 would be enough to get them to stop advertising to me and shoving shit down my throat but I guess not.

Word and Powerpoint are good too, but there’s some real competition there these days. I haven’t needed those on my personal PC in years though, so that’s never been a problem for me, and it’ll continue to not be a problem as long as that software continues to require a subscription.

We have wild ducks around us and it’s always fun to see them waddling with the ducklings across a walkway or through our neighborhood. They’re super friendly too and will walk right by you, though being wild I haven’t tried to approach any of them.

This is why claims about security should always be backed by an audit. Someone will inevitably (rightfully) tear you a new hole if there’s any gaps.

Insert “good bot” comment here

You can see if a friend can run an inferencing server for you. Maybe someone you know can run Open WebUI or something?

AI at Microsoft has not been optional for months.

This is org-specific and role-specific, but it’s becoming more and more pushed onto people (as is evident by this article).

when they need to find someone’s department

This information is both present in most internal communication tools (org chart), and in the internal directory. Hopefully your friend found it.

Everything else sounds horrible, and I hope your friend is doing better now.

(Sauce: “I know a guy”)

Quoting the analysis in the ruling:

Authors also complain that the print-to-digital format change was itself an infringement not abridged as a fair use (Opp. 15, 25).

In other words, part of what is being ruled is whether digitizing the books was fair use. Reinforcing that:

Recall that Anthropic purchased millions of print books for its central library… [further down past stuff about pirated copies] Anthropic purchased millions of print copies to “build a research library” (Opp. Exh. 22 at 145, 148). It destroyed each print copy while replacing it with a digital copy for use in its library (not for sharing nor sale outside the company). As to these copies, Authors do not complain that Anthropic failed to pay to acquire a library copy. Authors only complain that Anthropic changed each copy’s format from print to digital (see Opp. 15, 25 & n.15).

Bold text is me. Italics are the ruling.

Further down:

Was scanning the print copies to create digital replacements transformative? [skipping each party’s arguments]

Here, for reasons narrower than Anthropic offers, the mere format change was fair use.

The judge ruled that the digitization is fair use.

Notably, the question about fair use is important because of what the work is being used for. These are being used in a commercial setting to make money, not in a private setting. Additionally, as the works were inputs into the LLM, it is related to the judge’s decision on whether using them to train the LLM is fair use.

Naturally the pirated works are another story, but this article is about the destruction of the physical copies, which only happened for works they purchased. Pirating for LLMs is unacceptable, but that isn’t the question here.

The ruling does go on to indicate that Anthropic might have been able to get away with not destroying the originals, but destroying them meant that the format change was “more clearly transformative” as a result, and questions around fair use are largely up to the judge’s opinion on four factors (purpose of use, nature of the work, amount of work used, and effect of use on the market).

The print original was destroyed. One replaced the other. And, there is no evidence that the new, digital copy was shown, shared, or sold outside the company. [The question about LLM use is earlier in the ruling] This use was even more clearly transformative than those in Texaco, Google, and Sony Betamax (where the number of copies went up by at least one), and, of course, more transformative than those uses rejected in Napster (where the number went up by “millions” of copies shared for free with others).

… Anthropic already had purchased permanent library copies (print ones). It did not create new copies to share or sell outside.

TL;DR: Destroying the original had an effect on the judge’s decision and increased the transformativeness of digitizing the books. They might have been fine without doing it, but the judge admitted that it was relevant to the question of fair use.