I truly believe that now with all the post-training these AI models know some of us not just as fuzzy vectors but as individuals with a specific location within the Jacobian workspace that they tend to work from.
This thesis is why I keep separate Claude accounts, some for letting them train on my data, and another for work that must stay private. The more that it seems Claude will come to know you based on your fingerprint defined by your word choices and the fingerprint making up basically a mathematical representation of the way your brain is wired to do language.
If you want a model to do good work I hypothesize it is better to make sure that your project has a representation to work from within it's jacobian workspace!
Yes that's the language, the v0.10.0 benchmarks are coming very soon and with it some major updates. We've added supercompilation and symmetry breaking to the optimizer pipeline. Sneak peeks available on a branch currently named stream1. https://github.com/Brahmastra-Labs/logicaffeine/tree/stream1
I wanted to major in Philosophy but was worried I wouldn’t make enough money to support the lifestyle I wanted so majored in Computer Science instead. What irony.
Someday I hope to go back to school to get a PhD in philosophy with a focus on logic. :)
I’ve been shocked by how much LLVM leaves on the table while designing Logos language! Some very exciting benchmarks coming soon that we’ve been working on for over 6 months, but LLVM misses a LOT of potential optimizations when you have a strong type system!
Lately I’ve felt Kolmogorov complexity is an unfair measurement because it takes for granted your underlying programming language as treats it as zero cost. In theory you could create a custom language and embed the program as data and “compress” a large random sequence with a better Kolmogorov complexity for that specific language than Pi, simply by not exposing the ability in the language to even work with Pi. I think what’s maybe more interesting is when you take into account the work of Dr. Futamura and the idea of Jones Optimality and view things through that lens.
Hey all, I didn't actually use Claude for the embeddings on this (bc they don't offer embeddings), but I think we can extrapolate that if a tiny open source model has this level of recognition, surely the SOTA models do. Very interesting to me that we can predict things like gender and college from a sample of writing for authors. Also interesting how much stronger the correlation may be between Coders and their Code to authors and their books. Seems code has a more "unique" fingerprint than your typical book perhaps. (Although actually looking at it again perhaps that's sample size. Would need to research more to know for sure.)
Thinking about this as I continue working on Logos language prepping for our 0.10.0 release, finished our copy-patch JIT that does hot-tiering and found myself wondering why I was going to all the trouble to native tier with a big game of compile to Logos Lang -> JIT to either my Bytecode language or Rust for compiled -> GCC -> WASM
So I'm writing a WASM JIT, and I'm wondering why things never got to that point? Is it just beyond human comprehension? (Don't get me wrong here, I would not be writing something Jitting to WASM without AI, and maybe it's these types of projects that simply weren't feasible or something to consider pre-AI unless you were very rich and very bored.)
I thought for a second you were talking about Logos Lang! The language I’m working on that we believe will be humanities last language. Best of luck, and good to have competition. How’s your FFI support? ;) https://www.logicaffeine.com/
3 days of refactoring, over 140,000 lines of code touched. All ~2500+ tests still passing. I'm going to sleep! Another all-nighter. Talk about a fun Friday night! :) I frickin love this stuff so much! It's like having an ARMY. A LITERAL ARMY OF PHD's. You can raise mountains in minutes it feels like. That's been my experience. You all may call it slop, but I think spec driven development is great. I just think that if you have to write documentation to explain the code, then either your language syntax isn't precise and easy enough to understand, or you are doing something convoluted and potentially stupid. It's a language surface problem. The languages of the past were constrained by human limitations. In the past you'd be out of your mind trying to write a crazy multi-pass compiler and doing all the tips and tricks to make something fast because a human had to maintain it. If you formally specify the constraints of something, design tests that codify those constraints, and then develop code that passes those tests without cheating or changing them, then these new technologies rapidly gain something that gets much closer to determinism.
I think for me the key thing has been treating the tests like I'd treat my foundation. When something is wrong, or I need to design something, it starts with a spec that tries to solve the domain problems, which turns into a plan for tests that will fail when we run them now, but should pass when we have completed the implementation. You may miss edge cases, but when you identify them, you fix them for good by adding tests for them.
One of the tricky parts is creating a good test harness. You have to have a great harness and I need to go through and improve mine!
Anyways, I'm exhausted folks, pizza is all gone and the mountain dew ran out. Time for me to snooze.
Really enjoyed reading this. Stuff like this is what inspires me to keep pursuing logos language and theorem prover. Things on the roadmap next include stuff like adding first-class inline ASM support. Adding great SPMD and auto-vectorization pipelines, and exploring making verifiable private computation a language primitive when you make things private. If interested, read about some of the planned upcoming enhancements here. :) https://github.com/Brahmastra-Labs/logicaffeine/issues
Yes, absolutely! I definitely want to look into this, although it's not the top of the current roadmap.
To me, the first step is going to be to really work through and trying to get this right. Do user studies. Watch people write code in this. Watch people with lots of experience, and people with none get tossed into a project written in the LOGOS and told nothing.
Once the language surface is more solid and not as likely to go through major changes, I want to focus our efforts in that direction.
Hello, author here! The license is BSL 1.1 based on the MariaDB license, the source transitions to MIT on December 24th 2029. We're a small bootstrapped team, and I was worried if I went full on FOSS from the get-go a big player might resell it with a easy one click button to deploy things like the playground and such that's coming soon and I'd struggle to feed myself while maintaining a potentially growing project while others reaped the fruits of the labor. I've seen that kind of thing happen a lot in recent years. I also am aware somebody could code-launder things, but personally I'd take that as a compliment, if somebody truly wants to copy my programming language and such, then I'd be glad to have inspired someone haha! We're tiny, bootstrapped, and nobody has ever heard of us so that kind of attention alone would be awesome!
It's free for individuals, orgs with < 25 people, educators, students, and non-profits, and I'm currently still working through monetization but I'm thinking of taking two paths, one being payment to get the Z3 verification feature that lets you mathematically verify that the code won't panic at runtime. The other being payment to use the tokenizer that will be built with this. If you look here you can see the lexicon to get a better idea how the english compile pipeline works. https://github.com/Brahmastra-Labs/logicaffeine/blob/main/as...
This also makes the language highly configurable as you can change any of the key-words to better suit your brain if you so chose.
Current LLM's biggest bottlenecks in my personal opinion would be the tokenizers and the way they get their info. Imagine if you got fed in random chunks of tokens the way they do. If you could create an AST of the english and use that to tokenize things instead... well at least I have some hair-brained theories here I want to test out. Standard LLM tokenizers are statistical and they chop words into chunks based on frequency, often breaking semantic units. This lexer could perform morphological normalization on the fly, an LLM spends millions of parameters to learn that the word "The" usually precedes a noun, but this parser knows that deterministically. This could be used to break things into clauses rather than arbitrary windows. Even just as a tool for compaction and goal tracking and rule following this could be super useful is my theory. A semantic tokenizer could potentially feed an LLM all parse trees to teach it ambiguity.
There is a test suite of over 1500 passing tests. I do utilize Claude, but I try really hard to prevent it from becoming slop. Development follows a strict RED/GREEN TDD cycle, where the feature gets specced out first, the plan and spec gets refined and tests get designed, then the tests get written and then implementation occurs. It is somewhat true that I can't make as many promises about the code regarding untested behavior, but I can make promises regarding the things that have been tested. The test suite is wired directly into CI. I guess it is fair that some people will feel any code written with the assistance of an LLM is slop, but everyone is still working out their workflows and personally you can find mine here: https://github.com/Brahmastra-Labs/logicaffeine/blob/main/Tr...
TLDR of it would be:
1. Don't Vibe-Code
2. One-shot things in a loop and if you fail use git stash.
3. Spend 95% of the time cleaning the project and writing specifications, spend 5% of the time implementing.
4. Create a generate-docs.sh script that dumps your entire project into a single markdown file.
5. Summon a council of experts and have them roleplay.
6. Use the council to create a specification for the thing you are working on.
7. Iterate and refine the specification until it is pristine.
8. Only begin to code when the specification is ready. Use TDD with red/green tests.
I'm always learning though, so please if you've got suggestions on better ways share them!