morningsam

>the LLM constantly guesses arguments or dictionary formats wrong [...] it's better to avoid that whole class of runtime errors altogether.

Use Mypy in strict mode and run it in the post-turn hook of your LLM harness so the LLM has no choice but to obey it. And don't use overly general dictionary types when the keys are known at development time; use TypedDicts for annotations if you must use dicts at runtime.

LLM slop article. But as for the topic: To me, "Kanban with refinements and retrospectives" is the sweet spot. The concept of Sprint seems not only superfluous but to add unnecessary rigidity. But you do want everyone to understand the tickets, so refinements are useful, and you do want to try and improve "meta" issues where possible (and have the amount of effort going into that limited so you don't end up in the methodology equivalent of yak shave hell), so retrospectives are useful, too.

This sounds like it only changes the framing, but in reality it would lead to completely different behavior, so the "leave the gun alone" option would likely lead to far fewer deaths than the red button option, simply by virtue of organisms including humans being generally biased in favor of "do nothing" (= leave the gun alone).

You could do both experiments with dogs instead of humans and roughly 100% of dogs wouldn't manage to shoot themselves with the gun, whereas if you forced them to press one of the two buttons (e.g. keeping them in a room until they press one by chance), roughly 50% would press the red one. So the two experiments differ strongly w/r/t to how likely it is for a "non-thinking" organism to choose each option.

Made me stop reading a few paragraphs in. I don't have a "problem" in the ethical sense either, but as the sibling comment notes, the way LLMs write is rather grating. To make matters worse, a) people seem to use them to add pointless volume / "filler" to their texts, so now I have to wade through pages and pages of this stuff, and b) I have no easy way to distinguish between an article at least based on novel human insights vs entirely LLM-generated from a "write me something about X topic" prompt. I don't think it's a stretch to say that the latter just isn't worth reading given the state of the art.

What's the source of the claim that it was a Roblox cheat? Neither the report linked at the start of this article nor Context.ai's and Vercel's notices mention this.

>The technology is pushed forward by a simple psychological logic: every key global actor knows that if they don't build the technology, they will be outcompeted by other actors who do build the technology. No key actor thinks that they have the luxury of not building the technology even if they wanted to not build it.

I don't remember who, but someone made an interesting point about this around the time GPT-4 was released: If the major nuclear powers all understand this, doesn't that make nuclear war more likely the closer any of them get to AGI/ASI? After all, if the other side getting there first guarantees the complete and total defeat of one's own side, a leader may conclude that they don't have anything to lose anymore and launch a nuclear first strike. There are a few arguments for why this would be irrational (e.g. total defeat may, in expectation, be less bad than mutual genocide), but I think it's worth keeping in mind as a possibility.

Yudkowsky himself also posted a rebuttal today: https://x.com/ESYudkowsky/article/2043601524815716866

Exactly. This whole thing just seems like a repeat of Flappy Bird to me. What was the "lesson" of Flappy Bird for game developers? That you should make very small, very simple games? How has that worked out for the vast majority of copycats who tried? The truth is there isn't any lesson, other than "sometimes people play the lottery, get lucky and win". Most people who play won't, though.

And beyond the ethical points it makes (which I agree may or may not be relevant for LLMs - nobody can know for sure at this point), I find some of the details about how brain images are used in the story to have been very prescient of LLMs' uses and limitations.

E.g. it is mentioned that MMAcevedo performs better when told certain lies, predicting the "please help me write this, I have no fingers and can't do it myself" kinda system prompts people sometimes used in the GPT-4 days to squeeze a bit more performance out of the LLM.

The point about MMAcevedo's performance degrading the longer it has been booted up (due to exhaustion), mirroring LLMs getting "stupider" and making more mistakes the closer one gets to their context window limit.

And of course MMAcevedo's "base" model becoming less and less useful as the years go by and the world around it changes while it remains static, exactly analogous to LLMs being much worse at writing code that involves libraries which didn't yet exist when they were trained.

Was my first thought as well. I had no idea those little guys were based on something real.

Same in Germany, and not just for elementary schools but also secondary schools. At least that's how it was decades ago when I was a student, maybe it's different now.

>Typeless loosy goosy code that passes dictionaries all over the place is just not fun.

mypy --strict in CI & don't let dict[str, Any] pass review if the keys are constants, insist on a dataclass or at least a TypedDict.

Looking through the options listed under "Non-Durable Settings", [1] I guess synchronous_commit = off fits the bill?

[1]: https://www.postgresql.org/docs/current/non-durability.html

In German, we use "aufrufen", which means "to call up" if you translate it fragment-by-fragment, and in pre-computer times would (as far as I know) only be understood as "to call somebody up by their name or nummer" (like a teacher asking a student to speak or get up) when used with a direct object (as it is for functions).

It's also separate from the verb for making a phone call, which would be "anrufen".

That's assuming the AI owners would tolerate the subsistence farmers on their lands (it's obvious that in this scenario, all the land would be bought up by the AI owners eventually).

Seems like this was requested in 2021 and is currently in beta testing for select ecosystems only: https://github.com/dependabot/dependabot-core/issues/3651

>Though I have wondered about the idea of programming something like Dependabot, but telling it, hey, tell me about known CVEs and security releases, but otherwise, let things cook for 6 months before automatically building a PR for me to update.

Renovate can do both of these things already:

https://docs.renovatebot.com/configuration-options/#vulnerab...

https://docs.renovatebot.com/configuration-options/#minimumr...

There's a huge difference between believing someone will fail and hoping that they will.

>Spawning a PYTHON interpreter process might take 30 ms to 300 ms

Which is why, at least on Linux, Python's multiprocessing doesn't do that but fork()s the interpreter, which takes low-single-digit ms as well.

>I grumpily ordered a replacement key for 15 euros.

A single key for 15€?! I remember ordering one from an online shop specialized in replacement laptop keys at some point in the 2010s and it was like 2€ total. Browsing through similar shops now, it seems like the minimum is 5€ per key nowadays, but still a far cry from 15€.

>After spending more than 100 euros on plastic keys, which would soon break again, I calculated that my keyboard had 90 keys and that replacing them all just once would cost me 1,350 euros.

Someone who breaks keys this often could just buy the whole keyboard assembly FRU for ~30-50€ and take spare keys out of that, assuming it's not always the same ones that break.

I was wondering what the catch was :-)

GPLv3 reimplementation linked in that thread: https://github.com/joeedh/pigment-painter

I think this will become more common, not just in developing countries but also in the West, where phones are practically mandatory to have and laptops aren't, so lots of kids will get their first programming experience on phones.

>In the context of getting one by surprise in simple arithmetic they are approximately as bad as zero.

I don't think so, because getting 0 in a larger expression might yield a result that looks plausible, leading to hidden bugs. Inf and NaN both are good because they necessarily propagate all the way up to the end result, making it obvious that something went wrong.

For it to be effective, all it needs is its complement: An easily recognizable green label saying "Doesn't connect to the internet", which is only allowed on the boxes of devices for which this is the case.

Maybe some more levels in the middle like "only connects to the internet for firmware updates" (yellow) and "doesn't require internet access for core functionality" (orange). Basically Nutri-Score [1] for hardware.

[1]: https://en.wikipedia.org/wiki/Nutri-Score

>Thats why RedHat is a business. They're not selling Linux, they're selling the reliability, longevity, services, support etc.

>In truth the license doesn't matter.

It's funny to bring that up in the context of Red Hat who have started to circumvent the GPL by terminating their relationship with anyone who tries to actually make use of the rights granted by it. "The license doesn't matter" because they've found a loophole in it, but it clearly does matter in that they had to do so in the first place and weren't able to adhere to its spirit due to business concerns.

[1]: https://sfconservancy.org/blog/2023/jun/23/rhel-gpl-analysis...

[2]: https://opencoreventures.com/blog/2023-08-redhat-gets-around...

>Starting things and not finishing them is just... The default.

Exactly what I was going to write. I don't need a manifesto telling me to do this, this is what I do naturally unless I force myself to stick with something.

If anything, the "finish rarely [NOT never!]" bit is what needs emphasizing, yet the entire page is about the "start often" bit instead.

Not just considered in PEP 722 - the uv feature is just an implementation of PEP 723 [1] (PEP 722's successor/competitor), which was accepted. Other tools like pipx support it as well.

[1]: https://peps.python.org/pep-0723/

CPython doesn't, but there is Mypyc [1] which compiles statically typed Python to faster C extensions leveraging the type information. As usual, this comes with tons of limitations [2]

[1]: https://mypyc.readthedocs.io/en/stable/

[2]: https://mypyc.readthedocs.io/en/stable/differences_from_pyth...

I don't think it would've been very difficult to include volunteer networks in the database and allow users the option to fall back on them if an EU-funded network is not available (including a warning about potential eavesdropping on unencrypted communications).

NIH as far as the eye can see:

Germany already has a grassroots, volunteer-run network of free WiFi hotspots called Freifunk [1], which has pretty decent coverage in a lot of the larger cities. I'm sure similar initiatives exist in other EU countries.

Does Wifi4EU leverage this in any way? Nope, there is no way for volunteer-run networks to get included in the app [2]. Instead, it looks like municipalities have to apply for Wifi4EU funding (which they currently can't because "The next call has not been announced."), set up brand new hotspots themselves, and only then are eligible for inclusion in the app's database. [3]

[1]: https://freifunk.net/

[2]: https://forum.freifunk.net/t/wifi4eu-vs-freifunk/21686

[3]: https://wifi4eu.ec.europa.eu/