Their trackrecord is such that if I got a similar call my first question when possible would be how I was being reimbursed. They are welcome to fork anything of mine if they observe the license attached. I will take a look at any PR. I will NOT spend time explaining anything to their engineers unless reimbursed at my regular rates.
Backblaze can lose 3/20 hard drives and still recover the data from parity drives. They don't have to rely on recovering it from non-operative drives unless they become unbelievably unlucky.
I had an insight when learning about the Halting Problem. There is a class of programs for which the Halting Problem is solvable since you can use circuit analysis to see if well-formed loops halt. We should strive to write programs in that class since then we know that the execution, long as it may be, will yield a result.
I found after starting to apply mise en place that when I wasn't doing it I was still quicker than before. I fetched more things from each cupboard or fridge when I opened them, even if I wasn't measuring out the item into a container beforehand.
I wonder if you could get the intended result for any language with a simple ruleset. You don't need a language to reduce/eliminate miscommunications, you need intent and understanding.
Having a better language helps. Every language has some pain points that are responsible for more than their fair share of miscommunications. Just to use English as an example, although this is true of most natural languages, the overuse of the copula "is/are" is a consistent source of mistakes whether accidental or purposeful. So much so that computer programmers learning object oriented programming have to be taught the difference between "is-a" vs "has-a" relationships.
If you say "My coworker Taro is a samurai", what does that mean? Did he dress up as a samurai for the office costume party? Does he study older forms of Japanese martial arts? Was he descended from a noble family from the Japanese feudal era and have claim to an actual samurai title?
There are controlled dialects of English, like E-Prime, which forbid the use of the copula "to be" as ungrammatical, or at least heavily restrict it. You can say your coworker "dressed as a samurai", "trains in the arts of a samurai", or "traces ancestral lineage from a samurai clan", but you can't say he "is a samurai" in E-Prime.
Would this make communication more clear with fewer mistakes? I don't think this has been adequately studied enough to say for certain. But at least in certain domains like military speech and air traffic control we have examples of such enforced language simplification resulting in measurable decreases in miscommunications. I'm very curious to see if we can generalize those results to a full, general-purpose dialect or entirely new language.
> the overuse of the copula "is/are" is a consistent source of mistakes whether accidental or purposeful
I’m not sure I’d characterise these as ‘mistakes’, per se. They’re just a consequence of the fact that the copula is multifunctional — just like every other English word. In fact, you could similarly criticise almost any common word: ‘go’, ‘from’, ‘good’, ‘like’, ‘not’…
But no, "is/are" is not multifunctional in the same way that "from" is. If I say "she came from school" it is clear that we're talking about relative motion today or in the recent past from one nearby location to another. If I say "her family came from Japan" it is clear I'm talking about ancestry and/or a long ago emigration, but also spatially oriented. If I say "the party is from 3pm to 5pm" then I am using "from" to indicate a temporal rather than a spatial motion. But it still conveys the same basic meaning in all these cases, as the origin point of a motion or interval, and it is not usually the case that a single usage of the word could be confused for more than one meaning.
The IS-A vs HAS-A relationship is entirely different. It is the difference between existential quantification (some aspect of this thing resembles/has an X) and universal quantification (the entirety of this thing is fully captured by the meaning of X). Rigorously analyzed these are very different claims. In a strongly typed language they couldn't be substituted for each other.
To make this concrete, if I say "he is bad" then it is unclear whether I am saying that person is an intrinsically bad man, or if I am just commenting that the thing that he is doing right now is not morally justified.
This may seem like splitting hairs, but that's rather the point. It's at the edge of what we are comfortable thinking about in everyday life. But how much of that is because the language that we use--English--doesn't have these distinctions built into its very foundation? If we were native speakers of a E-Prime, maybe this distinction would be obvious and trivial. And maybe, just maybe, we wouldn't let politicians and con men off the hook so easily for equivocating language.
‘The copula can be used in many different situations.’
-----
Also, I think you underestimate the amount of variation in other words, e.g.:
> If I say "she came from school" it is clear that we're talking about relative motion today or in the recent past from one nearby location to another.
Not just recent past, but having any time near to the point of reference: ‘When she came from school after the bomb scare…’.
> If I say "her family came from Japan" it is clear I'm talking about ancestry and/or a long ago emigration, but also spatially oriented.
Or they might have arrived yesterday to visit her.
> If I say "the party is from 3pm to 5pm" then I am using "from" to indicate a temporal rather than a spatial motion.
But only because ‘3pm’ and ‘5pm’ themselves have clear temporal reference as opposed to spatial reference. For an ambiguous example, if I say ‘we drove from breakfast to lunch’, that might mean we drove starting at breakfast-time and finishing at lunch-time, or it might mean we started at a location where we had breakfast and drove to a location where we had lunch. (To more fully show that the latter is a valid interpretation, consider ‘we drove from breakfast to breakfast’: it makes no sense if ‘breakfast’ is a time, but it makes sense if treated with the sense of ‘place where we had breakfast’.)
-----
> The IS-A vs HAS-A relationship is entirely different. It is the difference between existential quantification (some aspect of this thing resembles/has an X) and universal quantification (the entirety of this thing is fully captured by the meaning of X).
This doesn’t sound quite right to me. In my view, both ‘has’ and ‘is’ are basically relationships, rather than quantifiers. Consider a sentence like ‘I have the keys’: this merely states a relationship between two objects, rather than quantifying over any set. A sentence like ‘The morning star is the evening star’ is similar in this regard. It is true that a sentence like ‘I have a key’ has existential meaning — but I suspect the quantification is linked to the indefinite article ‘a’ more than the verb.
> To make this concrete, if I say "he is bad" then it is unclear whether I am saying that person is an intrinsically bad man, or if I am just commenting that the thing that he is doing right now is not morally justified.
This ambiguity isn’t limited to the copula, though. ‘I like him’ has exactly the same kind of ambiguity.
> If I say "she came from school" it is clear that we're talking about relative motion today or in the recent past from one nearby location to another.
Unless we mean that she came from one school of philosophy or art to another. Or she came from the school long ago to move somewhere else.
> If I say "her family came from Japan" it is clear I'm talking about ancestry and/or a long ago emigration
Unless her family came from Japan as part of a cruise before going to Thailand. Or her family moved out of Japan because they were American military stationed in Okinawa.
Getting rid of copula gets rid of only a small portion of the ambiguity in natural language.
Yup. One man's "overuse" and "source of mistakes" is another man's way to draw attention towards a region in the latent^H^H^H^H^H^H idea space.
This is useful and often good enough, because a) the rest of what's been said and situational context will help home in on more specific associations (and if not, one can always ask for clarification), and b) we often don't need to resolve specific ideas - "my friend is a samurai" may just be an invitation for the other person to reply "oh, fun, speaking of more ancient cultures, have you read about ...", and then continue jumping around big clusters of associations, without ever properly resolving any.
This is kinda what we call "legalese". It's a sort of formalized subset of English (or whatever language) that leans on standardized turns of phrase, it tends to set up definitions for terms that are then used throughout a document, etc. All in order to reduce misunderstanding and (hopefully) be easy to interpret in the event of a dispute.
However, we have whole judicial systems that spend a non-trivial fraction of their time interpreting legal verbiage. So clearly it falls short at least some of the time, otherwise courts could be, in part, automated away. Maybe that's because it's too hard or not possible with natural languages? Or the legalese ruleset just isn't refined enough?
I think that when law is introduced the consequences are not clear so the small print is used to introduce modifications to the rules, so the problem is about adapting a rule to the everyday use of it.
I suspect for clarity of communication we'd need to start with clarity of thought - the reason legalise doesn't eliminate ambiguity is because people aren't effectively omniscient, regardless of the level of their training. Add to that a shifting environment (new considerations etc) which can make previously valid documents ambiguous and you have a recipe for difficulties.
tl;dr cognitive ambiguities and changing circumstances are what make this hard, not language as a medium
I agree and disagree. I do think that the purpose of legal documents is to take a given set of inputs and dictate a predictable output. But (1) sometimes ambiguity is deliberate (for instance, to kick the can on a business point and hope that it never actually manifests itself after the deal is signed) and (2) as you note, sometimes totally unexpected circumstances arise.
Wouldn’t any constructed/logical language (is that the right term?) also be susceptible to unpredictable future developments?
Yes, you could, because any natural language can be used to teach mathematics. Lojban is, like any notation, a convenience for people who understand the concepts, but the concepts can be expressed in a natural language just as precisely, albeit less concisely.
In fact, as Florian Cajori stated in his book "A History of Mathematical Notations", the rise of mathematical notation was opposed by people who preferred the older, natural language style of doing mathematics, what Cajori termed the "struggle between symbolists and rhetoricians."
So, yes, logic is mathematics, and humans did mathematics in natural language for a very long time before we invented the conlang of mathematical notation. Moving more of the ideas into what is, ultimately, a more expressive notation is not a fundamental shift.
I don't think that the current models are at "knowledge capacity". So far all evidence points to training on more data on the same size model gives better results.
Both increasing the amount of parameters and the amount of training tokens improves results (more precisely: lowers training loss), and costs computing power. For optimally improving loss per training computing power, model size and training tokens should be increased equally. That's the Chinchilla scaling law. (Though low loss is not always the same as good results, the data quality also matters.)
An interesting corollary of this is that if you want to reduce the model size you can compensate by training for longer to achieve the same accuracy. Depending on your training:inference ratio this may be more optimal globally to reduce your total compute costs or even just reduce your frontend latency.
Yeah, though I have not seen a formula which takes the number of expected inference runs into account for calculating the optimal data/parameter balance.
You have the right comparison. Developers need to remember that nobody had a job punching holes in cards after the revolution of personal computing. Going to be the same here. The jobs will revolve around operating the new "loom" which we have been given.
It's not clear to me that the new "loom" won't just operate itself completely.
That's the problem with comparing this to previous tech advances. Past advances made people much more efficient at some task, but it still required the people to work the new thing. But now we have a new thing that might be able to work itself.
The "loom" in this case is kind of a loom that helps people create other tools. Creating tools is generally not an automated task and with AI it will be easier.
Difference is that compressive strength generally decreases over time. Here we have a product where it increases over time. Changes the equation on how you build and maintain it.
