trane_project

I've never had an Intel laptop work well in the efficiency and thermal department, Apple or not. I used to blame Apple too, but seeing the difference, it's hard to argue who the main culprit was. Can't design around a bad foundation.

The thermals were all Intel’s fault. My 2019 MacBook from work is an oven. I can’t tell whether my smaller m4 max is turned on by touching it.

I have been mostly working on a set of courses based around backing tracks I got from a course called “improvise for real”. So I only had to make the structure of those backing tracks explicit, which only required knowing the circle of fifths.

I am the creator of Trane (https://github.com/trane-project/trane/) which can replace Anki and similar systems for domains that have a well-defined hierarchy between subskills (turns out that's most of them, including learning vocabulary).

There are several unresolved issues with Anki, SuperMemo, at al.

1. Emphasis on memorization. I wanted something that could be applied to domains (music) whose scoring is not based on memorization, but mastery.

2. Lack of any hierarchical information limits how well it works to learn much larger skills. It's trivial to replicate the functionality of Anki with Trane: just have one lesson where you put all your exercises. Anki cannot replicate what Trane does, which is to use the dependencies of the many sub-skills to track and limit progress until mastery is achieved.

3. Emphasis on creating your own exercises. You are expected to come up with your own decks, which takes a lot of time. For complicated skills, you need to be a domain expert to really know how to do this.

Trane is pretty much done, and I have used it to teach myself music. It does not have a UI at the moment, so I am pretty sure I am the only user. Not that I mind, because I rather not work for free.

Currently working on a literacy tutor that will be built on top of it. When it's complete, it will be able to guide all students from learning their A-B-Cs to reading and writing at college level, backed by the most up-to-date research on reading and writing acquisition. Hopefully I will be able to release an MVP in the middle of the year.

Mexico City is great for this because it points you to the central square. You can see the avenues spiraling out of the square, some of which follow the same routes as the avenues that lead to the city-island of prehispanic times (Calzada de Tlalpan, for example).

The company got nearly 2B of dollars in investment, was acquired for half the price, and the founders somehow managed to make a profit. No wonder he's a fan. Which other system would allocate resources so effectively?

Serious question, but how exactly does Atlassian hope to recoup that billion dollars?

Loom is useful, but given that Atlassian's investors are expecting a large return, do they really hope a screen recorder to eventually add 2B+ dollars to their value? The AI features will eventually be commoditized, so that can't act as a moat.

Thanks for pointing me out to this. This is a proposed definition of intelligence. Is it the same as the real thing, though? Even assuming that it was:

> Like Solomonoff induction, AIXI is incomputable.

That would mean that computers can, at best, produce an approximation. We know the real thing exists in nature though, so why not take advantage of those competencies?

My point is that if you use biological cells to drive the system, which already exhibit intelligent behaviors, you don't have to worry about any of these questions. The basic unit you are using is already intelligent, so it's a given that the full system will be intelligent. And not an approximation but the real thing.

I think most AI research up to this day is a dead end. Assuming that intelligence is a problem solvable by computers implies that intelligence is a computable function. Nobody up to this day has been able to give a formal mathematical definition of intelligence, let alone a proof that it can be reduced to a computable function.

So why assume that computer science is the key to solving a problem that cannot even be defined in terms of math? We had formal definitions of computers decades before they became a reality, but somehow cannot make progress in formally defining intelligence.

I do think artificial intelligence can be achieved by making artificial intelligence a multidiscipline endeavor with biological engineering at its core, not computer science. See the work of Michael Levin to see real intelligence in action: https://www.youtube.com/watch?v=Ed3ioGO7g10

They wanted me to roll out a fix first. Apparently rolling back first was not moving fast enough for the interviewer’s liking.

I interviewed there once and they asked me what I would do if a service broke after a deployment. I said the first step was to revert to the last known good version and then investigate. Color me surprised when that was not the answer they expected.

The mongols sacked Baghdad, killed everyone, and destroyed all institutions of higher learning and the people working in them. Then the age of exploration permanently shifted the center of world power away from the Middle East and the Silk Road. Then European colonial powers subjugated those territories. Afterward, Saudi Arabia started exporting its very "particular" (to put it lightly) branch of Islam. Sprinkle some US intervention and you get to the present day.

Not much of a turning away, but a systematic and violent dismantling of institutions of learning (and a supporting culture) that took centuries to build. No surprise that those conditions didn't happen again in the same place. In most places, such a golden age never happens.

Yes, but not the kind of engineering mentioned in this article. Recently I went in a bit of a Youtube fueled permaculture/afforestation binge.

Reforesting without planting trees by encouraging shrubs to grow by pruning them (which are really trees with too many branches to grow tall): https://www.youtube.com/watch?v=RBP2uRQk5pQ

Increasing water retention and underground storage by building simple rock dams: https://www.youtube.com/watch?v=c2tYI7jUdU0

Creating forests quickly with the Miyawaki method: https://www.youtube.com/watch?v=y9c_Zlmqcgw

The solutions are there and rely on accelerating positive natural processes. But they don't funnel public money into private pockets, so let's fund harebrained schemes instead.

I looked this up and found this report: https://nces.ed.gov/pubs2019/2019179.pdf

Only 12.5% of the population scores in the levels 4 or 5 (they had to group them together because there were so few is my guess). This is a disgrace. There's no reason why every adult should not be able to read proficiently. We are talking about reading, not some obscure skill.

I wonder what these figures would be in Cuba. From what I remember reading, they were much higher because of widespread literacy campaigns.

Edit: Found the source of this figures (https://www.oecd-ilibrary.org/education/oecd-skills-outlook-...) which has data from more contries in the OECD. Japan scores the highest, followed by Finland.

My implementation of mastery learning (along with other related ideas): https://github.com/trane-project/trane/

It bothered me enough that there was not something like this already available, when the idea seemed so simple, that I ended up doing it myself. The user experience is not very polished yet, but I managed to turn the courses in here https://improviseforreal.com/learning-materials into courses for this software that can teach you ear training and to improvise on any instrument in all keys, all modes, and most common chord progression. Currently, I am doing it for piano, but I've only reached about 15% of the total jam tracks so far. Obviously, I am not distributing the tracks myself.

I am pretty sure the main issue is that there's no one funding implementing these ideas. We've known about mastery learning, spaced repetition, interleaving, etc. for decades, but it's not all been put together into a coherent system. Something like https://mathacademy.com/ is similar, but it's not open sourced and cannot be used to create your own materials. No need for LLMs or anything fancier to be involved when there's so much low-hanging fruit that's not been implemented yet. The core of my software is just a depth-first search over a graph, lol.

I think they meant incest. If we already did our share of it in the distant past, then fun with your cousin leads to a baby with a pig's tail. Or at least it did for the Buendía family: https://en.wikipedia.org/wiki/One_Hundred_Years_of_Solitude

The whole thing about how GPL propagates to unrelated works that happen to use a GPL licensed software is a misunderstanding. One that the FSF foundation is happy to propagate, but not one that would hold in court.

The concept of derived work in copyright law has nothing to do with how the binaries are linked together nor is an entire work derived from a GPL library just because they happen to call it at one point. Lawyers look at this very differently.

See https://www.linuxjournal.com/article/6366

See my example. I am trying to extract the value of an optional field in the enum. That's where the unwrap comes in.

Not sure if I disagree but your example is terrible. Let patterns in if statements greatly simplify things. The alternatives are a lot more convoluted.

See example from code I wrote this week: https://github.com/trane-project/trane/blob/master/src/data/...

Without the let, I'd have to do a match statement followed by an unwrap just to check if a field in an enum is set.

My bad. It's the music lesson.

The spirit of music is supposed to be the sequel but I haven't read it yet.

I don't disagree that this type of transcription leads to good results, and in fact, it's what you should eventually do, but I've found a different approach to transcription that to me is easier, more fun, and more powerful.

Take the song, load it up in transcribe (https://www.seventhstring.com/) or similar software, pick a part. Up to here, this method and yours is the same. But rather than worrying about the exact notes being played, study the context by just singing and playing over the song and seeing what works and what does not.

There's a lot more information than just the notes that were played. So this type of transcription gets you to navigate the same context the musicians were playing in. And as you do it, you'll start hitting the actual notes, and eventually you can close in on them. What's great is that it's fun, playful, musical, and you can do it even with songs that are way out of your level when transcribing note by note or learning from a score.

Using a human language analogy, your suggestion would be to mimic conversations and mine would be to babble in other people's conversations until you become fluent. It's clear in human languages that one is much faster than the other. Sure babies are geared for it, but I think the babbling would work for adults too if they became cute as a baby or puppy, got over themselves, and could shamelessly babble onto other's conversations.

The best book on the matter I've read of how to learn by ear is "The Gift of Music" by Victor Wooten. It's kind of weird in that it's written as a fiction book, but there are definitely music lessons. And if some of them feel too out there, just listen to him play. He's a master, not a stoner pretender.

There's also "Thinking in Jazz", but I am only a fourth or third of the way and so far it's mostly about the historical and cultural background of improvisers. It's not gotten too into the weeds musically speaking.

That's pretty much what I've been trying to do with https://github.com/trane-project/trane/

I wanted something like you describe, but as far as I know nothing existed. So I've been hacking at this and the basic idea does work. It's now just a matter of designing the courses and polishing the user experience.

I am just coming up with the structure for how to define what music would depend on each other. Trying to do it based on music theory would be ideal, but probably beyond my capacity. So I think the historical development of the genre you are trying to learn is a good proxy. For jazz, for example, this would be something like learning African music first, then spirituals, then blues songs, then new orleans jazz, then basic standards and so on. Trane works based on a graph, so the progression does not have to be linear.

It's pretty early stages at the moment. Only one course for now since I've been trying to work out the process first: https://github.com/trane-project/trane-music/blob/master/cou...

These "transcription" courses first ask you to loosely sing the song, then loosely improvise over it with your instruments (you can customize your own), then sing in different keys and do it more thoroughly, then improvise more closely to the actual song. The last step is what is normally called transcribing, but the course is meant to progressively lead you to that. The whole process is meant to recreate the apprenticeship process that all the early Jazz masters went through.

Ideally there's a graphic interface that downloads the music and lets you loop, slow down, and change the pitch. But for now, there's only a command-line interface and the user has to do that themselves. Not ideal, but it works.

GNU Typist was good enough when I had to learn to type on a similar keyboard.

Ironically, I am now very slow to work with normal keyboards.

I'd love to work full time on my open source project (see profile) but I'll have to keep working for the man for now.

It works and is useful for me already, but without the time needed to polish it and add more content, won't be that useful to other people.

One can dream.

And the pay is about the same as my current job. I was thinking of applying again now that I have more rust experience just to see but then I remembered how off putting the process was.

Rob Pike once told me I was a moron (not those exact words) but that intent when I asked a question in some performance Golang slack about how big a Golang map is in memory (as in the total bytes used by it internally).

He went to a previous commit from the one I had linked, saw my stupid attempt at estimating it, and proceeded to go on a rant in which he simultaneously called me an idiot and pointed to the answers in the runtime code. Pretty entertaining.

Of course, the code was stupid because I wrote it knowingly so, just so I could fill the function with compilable code and go debug something else. I guess not writing a comment is on me for not knowing Rob Pike was going to go out of his way to find it the next day.

Fun day at the office, that one. So yeah, I am not the greatest fan of Go or him either. But that was from way before when I had to write a JSON-LD parsing library and cast stuff to `interface{}` everywhere.

I write personal stuff in Rust now, but would happily do it in any other language with a proper toolchain and type system. Go has only one of those, so I don't dislike it. Just don't really enjoy it.

I paid for it for a month or so but thinking of going back now that Google is getting increasingly shittier.

I'll put it this way: It's not a lot better than 2013 Google, but it's looking a lot better than 2023 google. And this removal (just an A/B test for now, since it works fine in incognito mode) would make Google pretty much useless for me. I guess I didn't realize how bad the results have become until this option was taken from me.

Kagi is expensive for something that used to be free, but I wouldn't have to worry about core functionality being taken away to optimize their search results for ads, since there aren't any.

Another data point for the operator being ignored. Probably an unfortunate participant of an A/B experiment. But my search quality is close to the point of searching for alternatives.

This talk is some of the most mind-blowing stuff I've seen. Thanks for sharing.