[Zero]

In your opinion, what does a language need to sustain intelligence? That language could be formal (or not) and be a tool for social interactions, inner discourse, and so on. What's the minimal set of features needed?

For example, distinguishing past/present/future seems useful. Some kind of "syntactic sugar" so to speak, to say "me", seems useful also. These are examples, and I'm looking for a common subset.

I know a language named Toki Pona. I love it, but I don't know if it's rich enough.

[infurl]

In your opinion, what does a language need to sustain intelligence? That language could be formal (or not) and be a tool for social interactions, inner discourse, and so on. What's the minimal set of features needed?

There is an important area of study called Universal Grammar that considers that question.

https://en.wikipedia.org/wiki/Universal_grammar

I think it is especially significant if you are working on symbolic AI because it would help to define what the software has to be able to handle. I believe that for artificial intelligence to be useful to us it has to be able to communicate with us.

Toki Pona looks very interesting by the way.

https://en.wikipedia.org/wiki/Toki_Pona

[MagnusWootton]

I dont know what mean by intelligence.  Do you mean all the way to AGI?

maybe prolog can do it,  but ppl havent worked it out.   thats it not the languages fault,  ppl havent put it in the variation required for it.

[Zero]

Infurl, very good pointer as usual, thanks a lot. You made me discover AUG which looks interesting too.

Intelligence, no I mean mere programs all the way up to human beings.
I've always been fascinated by the simple elegance of Logo's prefix syntax. A conlang that would be formal and easily usable by humans would definitely be a wonder, wouldn't it?

[Zero]

https://mw.lojban.org/papri/Lojban

I'll look into it, hoping it's not too fat. I like minimalism.

[Zero]

So... I'm still in the process of evaluating Lobjan to decide whether or not it fits my requirements. But I can already give a "first impression".

If you had a chance to study Ada, you know that feeling. Lobjan is really wonderful, clever, fat and boring.

In 2016, a 580 pages codex pops up, starting like, "[after decades of development,] the language described in this book will not be changing by deliberate act of its creators any more." ...then you learn that a committee made a xorlo reform, that's not in the codex. It wants to be frozen, but from the outside it looks like it's still having non backward compatible changes. I don't know, this is unclear.

One final thought. I just don't like it. I don't feel excited, you know... it's not sexy. It just feels awkward.
Here:

When two gismu are adjacent, the first one modifies the second, and the selbri takes its place structure from the rightmost word. Such combinations of gismu are called tanru.

Why on earth aren't they using yet another Lobjan word for the word "word"? that remains a mystery. So far, my opinion is that this language is probably a lot more appealing to computers. I'll keep studying it for a while, and will certainly learn a few interesting tricks.

But if this is Ada, I want C  I want Rust.   

[Zero]

Wow, look at this one: https://elefen.org/
I don't know how it feels to native English speakers, but to me (French), it feels very nice. I can understand immediately what I read!

If the grammar is very regular, it could be a good candidate as vocabulary supplier.

Edit: OK that's official - I'm in love!!! so exciting!

Doesn't answer my question though, as to what's the minimal set of features needed to sustain intelligence.


I think I'm gonna pick ideas from Lobjan, vocabulary from Lingua Franca Nova, and use Logo's prefix notation. 

[Zero]

Well no, LFN is too beautiful, I don't have the right to scorch this language. But it's pretty regular, and it should be "easy" to produce a PEG parser for it.

Facebookresearch's LASER supports Lingua Franca Nova. There must be other NLP tools as well. 

[infurl]

Since you're taking a look at artificial languages, have you reviewed Esperanto yet? That one has been around for as long as I can remember and I know people who have a very high opinion of it.

[Zero]

No, I instinctively discarded Esperanto without even thinking about it. It was created at the end of the 19th century, and while it is probably a very high quality conlang, I felt that it wasn't a good candidate precisely because it was created a long time ago. The background of the creators of the language is very important. For example, Toki Pona is made by someone who knows about Taoism, Lobjan is made by people who know about computers (maybe too much), and so on. The creator of LFN is C. George Boeree. He lived in our modern era, he knew about computers (well, as a lot of people do today), and was a psychologist and professor.

[Zero]

I'm still in the process of "parsing" (what an appropriate word) the grammar of Lingua Franca Nova.

Then when I'm done, I will define a strict subset of LFN, and make several tools for it, mainly to read/write a directed graph representation of a situation that's described in LFN.

Then, use it as the global language of Ma-ali, the island of my roguelike (which is intended to become a testbed platform for Gofai).

Then, retire.

[Zero]

I'm about to define a new controlled language, designed to be used for symbolic AI inner discourse, social interaction, and so on. I call it µlfn (Micro-LFN).

So, here is my strategy.

There are 2 stages: definition of the vocabulary, and definition of the grammar.

For the vocabulary, I will include
- The LFN translation of every word in the Toki Pona dictionary. This is because Ai need something to talk about: the Ma-ali island will have plants, animals, fire, water, ...etc. A very simple world, but it still needs words to talk about these things.
- From LFN, what I call "tool words": prepositions, conjunctions, ...etc.
- From LFN, words related to the mind, like thinking, hesitating, wondering, learning, being surprised, ...etc. But I don't know yet how I will extract them from the LFN dictionary.

If I can, I'll try to stay under 250 words.

For the grammar of µlfn, fortunately I can use the extremely well written official LFN grammar book. For now, I will include only the rules I need to make µlfn as capable as Attempto Controlled English, trying to keep things simple (one right way to say something should be enough). I should be able to hand-craft a complete PEGjs parser out of it. Then if I see a need for more rules, I could add carefully selected ones.

The advantage of staying strictly in a subset of Franca Lingua Nova is, you can always augment Micro-LFN later gradually if you need to: you don't have to invent anything, you're already into a well-designed human language.

[Zero]

Natural languages are very strange landscapes...

I'm working on the PEG parser for Micro-LFN, half-way there and it's already a huge beast. I have to go through so many special cases and exceptions, for a language that still (even though it's formal) looks a lot like a natural one, I can't help thinking: there's definitely something we don't understand in human thoughts structures.

[HS]

Since languages map onto the world with low resolution, I think it generally takes a lot of interpolation to convert symbols into their meaning. What might be happening is that people are trying to do most of the work on the language side, by subdividing symbols into ever finer increments like calculus. But a picture is worth a thousand words, it's easier to describe things simply with the right method. What if you tried to animate language with an approximated human thought process, instead of using linguistic rules to try and recreate a human thought structure?

The closest measurable manifestation of a fundamental human thought structure/process might be our movement. What stands out to me about movement is delegation of control, which lets us focus mainly on high-level concepts, and incremental recalibration, where an initial estimate becomes finetuned as real-time data is gathered during an action. Maybe those kinds of algorithms could help out where it's difficult for existing linguistic rules to describe things simply.

[MagnusWootton]

I think something along these lines->

What we say to each other is simple, I think.
Its not the amazing thing, the amazing thing is how quickly we can change our minds, and spontaneously come up with new models of communication instaneously.     (a telephone robot only has 1 unchanging model.)

Thats what I'm going for,   something that is similar to a robotic telephone program, except it gets generated by a model generator, and continually updates it with a new strategy for combatting its environment with what it says,  based apon a simple detected goal.

This predefined goal is tricky to come up with,  and will simplify what it ends up saying,  but I think it could still work.    What if the goal is to get the character to do something with an obvious object,  in maybe a simple game world.      it could end up generating something interesting.

But how do you spontaneously generate?   brute force search,  is one way, possibly maybe the only way IMO.

Only problem is your only allowed up to 28 or so single bit parametres, before it gets too computational.

Cool thing is its different than GPT-3,   its a different approach,  could be worth something?  Im going to find out.