In aviation safety, there is a concept of "Swiss cheese" model, where each successful layer of safety may not be 100% perfect, but has a different set of holes, so overlapping layers create a net gain in safety metrics.
One can treat current LLMs as a layer of "cheese" for any software development or deployment pipeline, so the goal of adding them should be an improvement for a measurable metric (code quality, uptime, development cost, successful transactions, etc).
Of course, one has to understand the chosen LLM behaviour for each specific scenario - are they like Swiss cheese (small numbers of large holes) or more like Havarti cheese (large number of small holes), and treat them accordingly.
LLMs are Kraft Singles. Stuff that only kind of looks like cheese. Once you know it's in there, someone has to inspect, and sign-off on, the entire wheel for any credible semblance of safety.
They probably already can for a lot of things, but "Safety" is really about accountability when things go wrong. As a society, I hope we don't end up at "AI isn't perfect, but it's better than people on average, sorry if it failed you, good luck with that."
It will only get better at generating random slop and other crap. Maybe helping morons who are unable to eat and breathe without consulting the "helpful assistant".
LLMs are very good at first pass PR checks for example. They catch the silly stuff actual humans just miss sometimes. Typos, copy-paste mistakes etc.
Before any human is pinged about a PR, have a properly tuned LLM look at it first so actual people don't have to waste their time pointing out typos in log messages.
Interesting concept, but as of now we don't apply this technologies as a new compounding layer.
We are not using them after the fact we constructed the initial solution. We are not ingesting the code to compare against specs. We are not using them to curate and analyze current hand written tests(prompt: is this test any good? assistant: it is hot garbage, you are inferring that expected result equals your mocked result).
We are not really at this phase yet. Not in general, not intelligently.
But when the "safe and effective" crowd leave technology we will find good use cases for it, I am certain (unlike uml, VB and Delphi)
> One can treat current LLMs as a layer of "cheese" for any software development or deployment pipeline
It's another interesting attempt at normalising the bullshit output by LLMs, but NO. Even with the entshittified Boeing, the aviation industry safety and reliability records, are far far far above deterministic software (know for a lot of un-reliability itself), and deterministic, B2C software to LLMs in turn is what Boeing and Airbus software and hardware reliablity are for the B2C software...So you cannot even begin to apply aviation industry paradigms to the shit machines, please.
I understand the frustration, but factually it is not true.
Engines are reliable to about 1 anomaly per million flight hours or so, current flight software is more reliable, on order of 1 fault per billion hours. In-flight engine shutdowns are fairly common, while major software anomalies are much rarer.
I used LLMs for coding and troubleshooting, and while they can definitely "hit" and "miss", they don't only "miss".
I was actually comparing aviation HW+SW vs. consumer software...and making the point that an old C++ invoices processing application, while being way less reliable than aviation HW or SW, is still orders of magnitude more reliable than LLMs. The LLMs don't always miss, true...but they miss far too often for the "hit" part to be relevant at all.
They miss but can self correct, this is the paradigm shift. You need a harness to unlock the potential and the harness is usually very buildable by LLMs, too.
Concrete examples are in your code just as they're in my employer's which I'm not at the liberty to share - but every little bit counts, starting from the simplest lints, typechecks, tests and going to more esoteric methods like model checkers. You're trying to get the probability of miss down with the initial context; then you want to minimize the probability of not catching a miss, then you want to maximize the probability of the model being able to fix a miss itself. Due to the multiplicative nature of the process the effect is that the pipeline rapidly jumps from 'doesn't work' to 'works well most of the time' and that is perceived as a step function by outsiders. Concrete examples are all over the place, they're just being laughed at (yesterday's post about 100% coverage was spot on even if it was an ad).
"Брюки цвета кофе" ("pants of coffee color") is natural, "коричневые брюки" ("brown pants") is natural, but "кофейные брюки" is not. In fact the latter would likely be interpreted as "coffee pants" or "pants made out of coffee."
"кофейные брюки" is totally ok. everybody will understand it.
it's just the way the russian language is. you can abuse it, you can come up with words that do not really exist in language and make no sense, yet, everybody will understand what you meant to say
> "кофейные брюки" is totally ok. everybody will understand it.
If the context is clothes, people would likely be able to guess, sure. But consider another example "кофейная чашка" ("a coffee mug"). In this context, it would most certainly be interpreted as "a mug for coffee" and not as "a coffee-coloured mug." In other words, you must include the word "цвет" ("color") for it to be correct and unambiguous.
> it's just the way the russian language is. you can abuse it, you can come up with words that do not really exist in language and make no sense, yet, everybody will understand what you meant to say
I don't think this is unique to Russian. I'm sure you can do the same in English and Japanese at least.
It’s fine as an occasional stylistic choice, but using it repeatedly as a regular synonym for brown is a pragmatic and collocational error. The meaning is clear, but the wording is marked, and overuse makes the speech sound odd in everyday contexts.
I admit that. I also realize that tguvot is actually arguing in my favor, as he said that coffee color is distinct from brown, and therefore the inference is that they aren't synonymous. I would summarize that they are conceptually different, as "brown" is a real color, whereas "coffee color" is a marketing color.
Uh huh. Don't forget "aliceblue" and "rebeccapurple." But seriously, those are just arbitrary marketing aliases, aren't they. I remember e-shopping for sneakers, and every brand's "off-white" was a different color.
German inflection is pretty minimalistic. There are just four cases, and it's mostly the article that is being changed with only occasional and predictable changes to the noun itself. Meanwhile in Russian there are six cases and no article, so it's the word itself that has to change. Also there are three different declensions not counting exceptions.
Gender in Russian is much easier than in German though - most of the time you can tell it by the word itself
That's declension of nouns. Then there's the conjugation of the verbs, which is reasonably regular in German and similar to Latin (three persons, two numbers, three basic tenses each with a "perfect", two voices, across four moods).
What's difficult really depends on the languages you already know.
In addition to noun inflection, verb aspect, pronunciation stress, and punctuation trouble many native English speakers. That's in addition to all the simple irregularities, like irregular nouns and verbs.
Stress even troubles native speakers. When I lived there, I saw slideshow "where 's the stress?" quizzes used to fill time on screens in taxi buses, waiting rooms, and the like.
Stress is a bit of a rarer aspect, most words can be disambiguated with any stress placement, except for a few exceptions, i.e. зáмок (castle) /замóк (lock).
Punctuation is secondary, just put commas, colons and semicolons where you feel they should go, most Russians don't know any better themselves.
Noun and verb inflections you will master with enough practice, yeah.
Maybe overall a more difficult language than English or German, but not in the same league as Chinese or Arabic, in my humble opinion.
As an Arabic speaker I enjoyed learning Russian because we share verbless sentences, and you could just put the words together in any order and you get your idea across and you could be spot on too. So 'what is the time?'(Kotoryy chas) is 2 words as in Arabic for asking the time and other questions in conversation. And some Russian words have lovely music to my ears, as with ice cream and of-course, мороженое и, конечно.
> Stress is a bit of a rarer aspect, most words can be disambiguated with any stress placement
The difficulty is that the stress pattern is not fixed and needs to be memorized, and it often changes the inflection of the word. E.g. "домá" means "houses", while "дóма" means "at home". Another tripping point is that the stress placement is almost always different in Russian when compared to English.
I'm volunteering as an English teacher for Ukrainian refugees, and one of my rules of thumb is: "If an English word looks similar to a Russian word, then the stress is likely on a _different_ syllable". It works surprisingly well.
>If an English word looks similar to a Russian word, then the stress is likely on a _different_ syllable
Most of these are Latin and French loanwords where Russian (same as e.g. German) carried the accentuation over from the source language. English is the odd one out as it insists on putting the primary stress on either of the first two syllables, except in some recent loans (and those still get a secondary stress). With nouns the preference is for the first syllable. Russian surnames get similarly butchered, including notably Nabokov, which could have been adopted unchanged.
It seems like an extremely coarse classification. Category 3 contains languages with very different degrees of difficulty, while Bulgarian and Russian are both Slavic they are nothing alike in terms of difficulty since Bulgarian is the most analytic of Slavic languages (has the less inflection). That makes it extremely easy to learn compared to Russian.
What is also interesting is how written Russian was heavily influenced by old Bulgarian. In fact, written russian includes a lot of older written bulgarian vocabulary.
This results in a weird paradox: for literate Russians it is easy enough to read written bulgarian but almost impossible to understand the spoken language.
This happens in other languages too - danish and Norwegian are almost the same written, such that most products just combine the two on the packaging. But spoken it can be very difficult to comprehend
So... codified written languages are similar but real spoken ones have diverged? Is this only in the way things are pronouced or the differemce is deeper?
I speak Russian and some Bulgarian as third/forth languages, and while I agree that Russian is more difficult, I wouldn't say Bulgarian is "extremely easy" in comparison. It's maybe ~20% easier at best.
I think Bulgarian is considered the easiest Slavic language in terms of grammar because it has a simplified case system similar to how English dropped its cases over time.
On a superficial level that seems like a roughly correct ranking in my experience. On the other hand, I picked up one of the category 3 languages pretty easily. I think some of these are more "weird" to a native English speaker than "hard" per se.
The aspects that make languages difficult for a native English speaker vary quite a bit with the language. I would expect individual experiences with the languages to have high variance as a consequence.
As others hsve pointed out, it's a very coarse (and rather arbitrary) categorization.
E.g. both Turkish and Russian are in Category 3, but Turkish is trivial compared to Russian.
Turkish grammar is extremely regular, and follows easily defined rules that fit about two pages of easily digestible tables.
In comparison, Russian is a separate class tought in Russian schools for four years to native Russian speakers. And you still get people who can't properly inflect numerals, for example.
Turkish has a completely different vocabulary (loanwords aside) and a completely different grammar.
"I want to swim" in Russian is "ja hoću plavatj", "I" + "want" + "to swim". The only difficulties are the conjugation of "want" and the aspect of "to swim". In Turkish it's "yüzmek istiyorum", where "-mek" is "to" and "-um" is "I". Even if the system itself is straightforward, it's still alien to a native English speaker.
> Even if the system itself is straightforward, it's still alien to a native English speaker.
As a native Russian speaker who speaks English and Turkish:
The question isn't about alienness. It's about difficulty. Turkish is trivial compared to Russian. You can learn all the grammar rules you'd ever need in a week or so (though most study materials make it harder than necessary). The rest is just learning vocabulary. Which is just as alien to an English speaker as Russian.
As for the example...
Here's a valid three word sentence in Russian: Ya idu domoj (I'm going home).
Depending on context, mood, feel, etc. any permutation of those words is a valid sentence: ya domoj idu, idu ya domoj, idu domoj ya, domoj ya idu, domoj ya idu.
And that's before we get into inflections, conjugations, gender etc. that neither English nor Turkish have. Or somewhat arbitrary pronunciation rules (korova is pronounced kahrohva) whereas in Turkish every word is pronounced exactly as written (with very few quite regular contractions in regular speech like yapacağım -> yapıcam) etc.
Turkish is regular, has well specified rules you can learn in a week, is extremeley easy to read (pronounce as written, there's no floating/jumping/changing stress). Oh, and the alphabet is latin-based.
Russian: extremely complicated grammar using concepts entirely alien to English (declensions, inflections, conjugates, grammatical cases, genders, transgressives, and even plurals are weird), has free-form-not-really sentence structure, jumping stress. Oh, and a completely different alphabet to boot.
Not always well. One of the problems is that English speakers do not get taught the proper parts of speech. I learnt far more about English from learning other languages than I ever did in English class. We would get told off for bad spelling and grammar, but very little on the actual mechanics.
All through middle and high school, so for 7 years from around 10 to 16.
It did become one eventually in primary school, so probably the last 2 or 3 years there.
I have been generally successful at learning Russian as an adult, but tonal languages are something that I just struggle with on a fundamental level. I want to express meaning and connotation with tones, rather than denotation. On the other hand I've never been terribly motivated to learn a tonal language, so it probably could be overcome, but it's something that would take an immense amount of training to overwrite that tone=connotation/emotion/question instinct.
It is also quite frustrating when a native speaker is completely unable to understand something you say because of a tonal issue. To their ear it must sound entirely different, yet to a non-tonal ear it sounds like you're saying everything 'almost' exactly correct.
> I want to express meaning and connotation with tones, rather than denotation. On the other hand I've never been terribly motivated to learn a tonal language, so it probably could be overcome, but it's something that would take an immense amount of training to overwrite that tone=connotation/emotion/question instinct.
Why would you want to? Pitch also provides connotations / emotions in Mandarin.
> It is also quite frustrating when a native speaker is completely unable to understand something you say because of a tonal issue.
That will never happen. Your bad pronunciation can aggravate other problems, but if your sentence is otherwise good, ignoring the tones will still leave it fully intelligible.
(I once asked a student in a Chinese school whether a particular class wasn't occurring, and he responded "poss". After some confusion, he was frustrated that the pronunciation difference between "poss" and "pause" should make such a difference in communicating with an English speaker.
But of course, it doesn't. If "pause" were a valid way to respond to "is chemistry class happening today", I would have had no difficulty understanding "poss". His problem was in bad knowledge of the language, not bad pronunciation.
You appear to be making the same mistake here. If you try to communicate, and fail, that is not evidence that you are qualified to diagnose what the problem was.)
Only somewhat related: I was surprised by how simple and sound vietnamese grammar is when read through the latin alphabet. Tones are only a problem when speaking but it's increadibly easy to start understanding signs and labels in the country. Slavic and baltic languages i can read are MUCH harder to start with.
So i kind of suspect it might also be the case for chinese: tones and the alphabet are obscuring a clean grammar.
Conveying what I've heard from a few Vietnamese that also speak Chinese, so not any kind of firsthand experience since I speak neither: Vietnamese is more difficult to speak but is a simpler (less expressive) language.
I agree that written Vietnamese is relatively straightforward. It isn't that difficult to read to the eyes of someone used to latin script.
Fiendish logographic writing system (Chinese) vs fiendish grammar (Russian). I'm not a fan of Pinyin transliteration aesthetically.
Russian has a lot of words I can recognise in it. Not just loanwords either but words such as brat, dva, kot (brother, two (twa), cat). The other problem is the tonal system although Mandarin balances that out with simple grammar. Mandarin strikes me as mostly vowels and Russian as strings of consonants.
> It is not that different from German in this matter.
I've met several Germans who spoke Russian fluently, none of them has really mastered the instrumental case, not even a friend of mine who worked at the German embassy in Moscow. Although you might say it's a minor grammar difference, this particular grammar case seems hard to grasp for people who are not accustomed to it through their native language.
Also, from my personal experience, quite a few Germans who learnt Russian had a real struggle understanding the concept of perfective/imperfective aspect.
These kinds of grammatical difficulties are typical for people who are learning only their second language after their native language.
After learning 3 or more languages that are not closely related, one is usually exposed to most grammatical features that can be encountered in the majority of the languages, so usually grammar no longer poses any challenges, but only memorizing the unfamiliar words and pronouncing sounds that do not exist in the native language.
>It is not that different from German in this matter.
Russian inflection changes the stress. In German it's fixed. Inflectional forms are much more varied in Russian. Colloquial German is much more analytical (past tense is almost always "ich habe" + participle). German has devolved to basically 3 cases at this point (with genitive dying out), compared to Russian's 6. But conceptually, they're very similar indeed.
If you just want to be understood, Russian is not very hard. I think it's true for any language. To master it, however...
> Russian grammar is inflectional, yes, but that's about the only difficult part of the language.
That's saying that getting to the lunar orbit is the only difficult part in landing on the Moon. The whole complexity of inflectional languages is in the inflections. It's also why Slavic (or Turkic) languages form such a large continuum of mutually almost-intelligible languages.
Compared to inflections, everything else in Russian is simple. The word formation using prefixes and suffixes is weird, but it's not like English is a stranger to this (e.g. "make out", what does it mean?). The writing system is phonetic with just a handful of rules for reading (writing is a different matter).
Add baltic languages to the mix as well! Lithuanian is like a slavic language with all the inflection drama but with additional word types that are currently mostly gone from slavic languages.
Well, Lithuanian is also a Proto-Indo-European language. But the one that somehow got sucked into a time warp from the past. And it even has a tonal pitch accent in addition to the stress pattern, just to make it more interesting.
It's not a "proto-indo-european"! It did change, and change massively. But it seems that because of how the number of native speakers was always low, and thr populayion was concentrated in a relatively isolated part of Europe, this evolution was more self-contained and reflective.
The language seems to have more archaic features and forms than, say, closely related slavic languages, and its vocabulary has more similarities to old sanscrit than one's average european language.
For curios language learners this means that the grammar is harder than even (already hard enough) slavic grammars.
Wow, I had no idea. This sounds extremely interesting. I need to read more about Lithuanian language (at least grammar, sadly I don't have time to learn yet another language)
English phrasal verbs make completely zero sense since there is no logic involved
If English was logical "make out" would mean somethibgg like "throw out". But "to make out" means something else obviously. And you dont throw out your trash. You throw them away...
The only difficult part of Russian is writing it. Most native Russian speakers, myself included, can't write properly even after completing 11 years of Russian language in school. Hundreds of rules nobody remembers.
I think as a native speaker it's different to you.
Native English speakers make spelling mistakes quite often. But as a language learner I struggled with everything, except spelling - I always knew how to spell a word, even if I don't know how to pronounce it. It's the opposite of native speaker experience.
English spelling is one of the hardest parts of the language to learn because the spelling represents ~16th century pronunciation. However what we gained is a common orthography for all the different dialects and accents of English. I can barely understand some people from Appalachia or Western England when they speak, but if they write it down it’s no problem.
> English spelling is one of the hardest parts of the language to learn because the spelling represents ~16th century pronunciation.
English spelling doesn't represent any pronunciation. English spelling represented pronunciation before the Normans, and afterwards was turned into something that would allow Norman speakers to do nearly-intelligible imitations of unpronounceable English words. Even worse, 1) French spelling also had drifted far from pronunciation (although not as far as now), and 2) English picked up a ton of that French and further mispronounced it.
Such as how place names that now end in "-shire" pre-conquest ended in "-scr," which is how they're still pronounced.
> However what we gained is a common orthography for all the different dialects and accents of English.
True, but those dialects came after the spelling changes. Vowels in English multiplied out of control and became more of a system of how vowels could relate to each other rather than specific sounds, like in (very regular) Old English when a long or doubled vowel was simply the same vowel sounded longer. Germanic vowels are crazy and got crazier.
To understand somebody's English, you listen to them for a while and figure out what they're doing with their vowels - we know from experience that some vowel sounds move together with each other, so when we hear X we can guess Y, and we then look for exceptions and mergers. Once we've figured out the vowels, the words become clear. A fun example is when you compare the Canadian accent to the US accent, and you see some words rhyme in both British and Canadian English that don't rhyme in US English.
IIRC, English is often described as having between 16 and 22 vowels, depending on who is speaking it. Writing that would be hellish, and as you say, you'd have to change spellings when you crossed rivers. English orthography is more like Chinese orthography than one would think.
I should have said it represents the pronunciation of English in and around London in the 16th century.
English spelling wasn’t normalized until long after the Normans. Norman scribes did their part but it was the printing industry in London that crystallized it.
Some of it is. Some of it is arbitrary. The "h" in "ghost" is said to be partly accidental. The "b" in "debt" due to folk etymology. There will be numerous other examples.
In his book The Chinese Language: Fact and Fantasy, John DeFrancis calls the English orthography the worst among the alphabetic ones, and Japanese the worst among the logographic ones.
Re English, maybe among major languages. Faroese orthography is bad in phonetic terms, but Faroese is not a well known language. I'm sure other smaller languages have even worse systems than English.
Among the major languages, French is also pretty awful. Its orthography is much less practical than Spanish or Italian.
Tibetan orthography is notoriously bad, but is neither alphabetic nor logographic. This is a result of Tibetan changing a great deal since it was first transcribed.
Sure, there are some recent exceptions. But nearly all words that contain vowel pairs like “ea” and “ai”, for example, represent a much older pronunciation regardless of the current one. Words like “hear” and “wear” would have rhymed at one point. Most words ending in “ed” would have had that syllable fully pronounced instead of reduced to a “t” sound.
That vast majority of words among all English dialects are spelled similarly and go back to the 16th century or thereabouts.
Your experience as a native speaker is completely different from learning the language from scratch as an adult, to the point that it's almost irrelevant. Writing Russuan is not that difficult, it's just the only part that you had to actually do any work to learn
Never thought the difference mastering writing can be so significant. Just like to add what I understand regarding this. It's rather about not making any mistake writing by hand ca. 1-2 DIN A4 pages while someone reads a text (slow enough). I can't remember exactly but making only one (or two) mistake(s) and it is not anymore excellent (just good). Making 4-7 mistakes and it is not good (just sufficient). Making few more and it is bad which means failed. It's a long text with a very short path to fail.
Ukrainian is less difficult to write. There are claims that standardization/reform of Russian made it more artificial (far from natural people language) with overtaking too many words from Latin languages. When I read / listen to Belorussian I think they have even more luck with matching pronunciation/writing than Ukrainian. Which suggests this language is even closer to the common roots old language. (I'm not a linguist.)
Poles will hate me saying this, but I've always really struggled with their orthography, even though I am used to the Roman alphabet. I can see what is going on in Belarusian, Russian and Ukrainian, maybe even Czech to some extent. Polish is bizarre. Szcz is one letter in Cyrillic. I'm still baffled by l with the line through it.
But if cost of the space GPUs is higher then the land GPUs, why would demand matter? Is there limited land? Are space GPUs better for some reason, like perhaps they can't be regulated as easily or because our AGI overlords will be less vulnerable to mobs with pitchforks?
What’s software that would benefit from running in space? The only thing I can imagine is processing of data generated in space so you need less downlink or can reduce latency, everything else can be calculated wherever you want, no?
I think the point the original guy is hand wavingly getting at is the point of something like this is to avoid the possibility of say a FBI raid or Nuremburgish trials for a vast AI surveillance processing facility hub for other down looking satellites if they were to lose their newly acquired power, or similar technocratic ramblings / ideas like it would survive the end of society.
Its like that scene at the end of Real Genius, "Maybe somebody already has a use for it, one for which it's perfectly designed." Lets look at the facts: Impossible to raid, not under any direct legal jurisdiction, high bandwidth line of sight communications options to satellite feed points that would be difficult to tap outside of other orbital actors, Power feed that is untethered to any planetary grid or at risk of terrestrial actors, etc.
That’s not how it works. Your state is responsible for your activities in space, so if you annoy other countries enough, your own country will regulate you. If they don’t, you could have just built the same thing on the ground in this country.
It's definitely much easier and much much cheaper to send a single rocket there blowing the assembled rather large target into still sizeable chucks of orbital debris than it is to deploy and assemble the thing there in the first place. And there are a few terrestrial actors rather capable of this. More than there are who could make it happen under whatever optimistic assumptions anyway.
In itself, a structure of this size in orbit is an efficient catcher of micrometeorites and orbital debris. Over "non-eternal" timeframes you don't even need a bad actor with good rockets.
Nevermind that in such a case, the eventual fate of these sizeable chunks of orbital debris is to become rods of god ... just without particular steerability.
At this point I'm going to assume that anyone pushing datacenters in space wants to host child pornography. That's the only realistic workload that ticks all the boxes for orbital datacenters.
I don't think it would "solve" little any of the legal issues with Child Pornography (not if the owner lived on earth, at least), but it would make a great and politically convenient target for space to earth weaponry.
Oh, fully agreed. Orbital datacenters don't solve many to any engineering problems either, so I figure its adherents are as much into legal problem solving as they are engineering problem solving.
I wouldn't hold my breath for it. It was suppose to be released in 2020. It's end of 2025 and current release date is 2027 (and who knows if it'll be pushed back again).
One can treat current LLMs as a layer of "cheese" for any software development or deployment pipeline, so the goal of adding them should be an improvement for a measurable metric (code quality, uptime, development cost, successful transactions, etc).
Of course, one has to understand the chosen LLM behaviour for each specific scenario - are they like Swiss cheese (small numbers of large holes) or more like Havarti cheese (large number of small holes), and treat them accordingly.
reply