15 comments

  • ramon156 1 hour ago
    Cool stuff! I can see some GPT comments that can be removed

    // Increased for better learning

    this doesn't tell me anything

    // Use the constants from lib.rs

    const MAX_SEQ_LEN: usize = 80;

    const EMBEDDING_DIM: usize = 128;

    const HIDDEN_DIM: usize = 256;

    these are already defined in lib.rs, why not use them (as the comment suggests)

    • sloppytoppy 36 minutes ago
      Oh yea I'm totally running this on my hardware. Extra credit for "from scratch" in the title. The future sucks.
  • untrimmed 2 hours ago
    As someone who has spent days wrestling with Python dependency hell just to get a model running, a simple cargo run feels like a dream. But I'm wondering, what was the most painful part of NOT having a framework? I'm betting my coffee money it was debugging the backpropagation logic.
    • ricardobeat 1 hour ago
      Have you tried uv [1]? It has removed 90% of the pain of running python projects for me.

      [1] https://github.com/astral-sh/uv

      • DiabloD3 1 hour ago
        uv is great, but I think the real fix is just abandoning Python.

        The culture that language maintains is rather hostile to maintainable development, easier to just switch to Rust and just write better code by default.

        • trklausss 1 hour ago
          Every tool for the right job. If you are doing tons of scripting (for e.g. tests on platforms different than Rust), Python can be a solid valid alternative.

          Also, tons of CAE platforms have Python bindings, so you are "forced" to work on Python. Sometimes the solution is not just "abandoning a language".

          If it fits your purpose, knock yourself out, for others that may be reading: uv is great for Python dependency management on development, I still have to test it for deployment :)

          • aeve890 40 minutes ago
            >Every tool for the right job. If you are doing tons of scripting (for e.g. tests on platforms different than Rust), Python can be a solid valid alternative.

            I'd say Go is a better alternative if you want to replace python scripting. Less friction and much faster compilation times than Rust.

            • DiabloD3 22 minutes ago
              I am not a huge fan of Go, but if all the world's "serious" Python became Go, the average code quality would skyrocket, so I think I can agree to this proposal.
            • physicsguy 6 minutes ago
              Go performance is terrible for numeric stuff though, no SIMD support.
        • airza 1 hour ago
          There's not really another game in town if you want to do fast ML development :/
          • DiabloD3 54 minutes ago
            Dunno, almost all of the people I know anywhere in the ML space are on the C and Rust end of the spectrum.

            Lack of types, lack of static analysis, lack of ... well, lack of everything Python doesn't provide and fights users on costs too much developer time. It is a net negative to continue pouring time and money into anything Python-based.

            The sole exclusion I've seen to my social circle is those working at companies that don't directly do ML, but provide drivers/hardware/supporting software to ML people in academia, and have to try to fix their cursed shit for them.

            Also, fwiw, there is no reason why Triton is Python. I dislike Triton for a lot of reasons, but its just a matmul kernel DSL, there is nothing inherent in it that has to be, or benefits from, being Python.... it takes DSL in, outputs shader text out, then has the vendor's API run it (ie, CUDA, ROCm, etc). It, too, would benefit from becoming Rust.

        • Exuma 46 minutes ago
          i hate python, but the idea of replacing python with rust is absurd
      • TheAceOfHearts 34 minutes ago
        Switching to uv made my python experience drastically better.

        If something doesn't work or I'm still encountering any kind of error with uv, LLMs have gotten good enough that I can just copy / paste the error and I'm very likely to zero-in on a working solution after a few iterations.

        Sometimes it's a bit confusing figuring out how to run open source AI-related python projects, but the combination of uv and iterating on any errors with an LLM has so far been able to resolve all the issues I've experienced.

    • zoobab 22 minutes ago
      "a simple cargo run feels like a dream"

      A cargo build that warms up your CPU during winter while recompiling the whole internet is better?

    • Galanwe 1 hour ago
      > spent days wrestling with Python dependency hell

      I mean I would understand that comment in 2010, but in 2025 it's grossly ridiculous.

    • codetiger 1 hour ago
      I guess, resource utilization like GPU, etc
    • taminka 2 hours ago
      lowkey ppl who praise cargo seem to have no idea of the tradeoffs involved in dependency management

      the difficulty of including a dependency should be proportional to the risk you're taking on, meaning it shouldn't be as difficult as it in, say, C where every other library is continually reinventing the same 5 utilities, but also not as easy as it is with npm or cargo, because you get insane dependency clutter, and all the related issues like security, build times, etc

      how good a build system isn't equivalent of how easy it is include a dependency, while modern languages should have a consistent build system, but having a centralised package repository that anyone freely pull to/from, and having those dependencies freely take on any number of other dependencies is a bad way to handle dependencies

      • dev_l1x_be 1 hour ago
        > lowkey ppl who praise cargo seem to have no idea

        Way to go on insulting people on HN. Cargo is literally the reason why people coming to Rust from languages like C++ where the lack of standardized tooling is giant glaring bomb crater that poses burden on people every single time they need to do some basic things (like for example version upgrades).

        Example:

        https://github.com/facebook/folly/blob/main/build.sh

        • taminka 13 minutes ago
          i'm saying that ease of dependency inclusion should not be a main criterion for evaluating how good a build system is, not that it isn't the main criterion for many people...

          like the entire point of my comment is that people have misguided criteria for evaluating build systems, and your comment seems to just affirm this?

        • huflungdung 1 hour ago
          [dead]
      • quantumspandex 1 hour ago
        Security is another problem, and should be tackled systematically. Artificially making dependency inclusion hard is not it and is detrimental to the more casual use cases.
      • itsibitzi 1 hour ago
        What tool or ecosystem does this well, in your opinion?
        • taminka 8 minutes ago
          any language that has a standardised build system (virtually every language nowadays?), but doesn't have a centralised package repository, such that including a dependency is seamless, but takes a bit of time and intent

          i like how zig does this, and the creator of odin has a whole talk where he basically uses the same arguments as my original comment to reason why odin doesn't have a package manager

      • IshKebab 1 hour ago
        This is the weirdest excuse for Python's terrible tooling that I've ever heard.

        "It's deliberately shit so that people won't use it unless they really have to."

        • taminka 11 minutes ago
          i just realised that my comment sounds like it's praising python's package management since it's often so inconvenient to use, i want to mention that that wasn't my intended point, python's package management contains the worst aspects from both words: being centralised AND horrible to use lol

          my mistake :)

      • jokethrowaway 1 hour ago
        Is your argument that python's package management & ecosystem is bad by design - to increase security?

        In my experience it's just bugs and poor decision making on the maintainers (eg. pytorch dropping support for intel mac, leftpad in node) or on the language and package manager developers side (py2->3, commonjs, esm, go not having a package manager, etc).

        Cargo has less friction than pypi and npm. npm has less friction than pypi.

        And yet, you just need to compromise one lone, unpaid maintainer to wreck the security of the ecosystem.

        • taminka 4 minutes ago
          nah python's package management is just straight up terrible by every metric, i just used it as a tangent to talk about how imo ppl incorrectly evaluate build systems
  • techsystems 2 hours ago
    > ndarray = "0.16.1" rand = "0.9.0" rand_distr = "0.5.0"

    Looking good!

    • kachapopopow 2 hours ago
      I was slightly curious: cargo tree llm v0.1.0 (RustGPT) ├── ndarray v0.16.1 │ ├── matrixmultiply v0.3.9 │ │ └── rawpointer v0.2.1 │ │ [build-dependencies] │ │ └── autocfg v1.4.0 │ ├── num-complex v0.4.6 │ │ └── num-traits v0.2.19 │ │ └── libm v0.2.15 │ │ [build-dependencies] │ │ └── autocfg v1.4.0 │ ├── num-integer v0.1.46 │ │ └── num-traits v0.2.19 () │ ├── num-traits v0.2.19 () │ └── rawpointer v0.2.1 ├── rand v0.9.0 │ ├── rand_chacha v0.9.0 │ │ ├── ppv-lite86 v0.2.20 │ │ │ └── zerocopy v0.7.35 │ │ │ ├── byteorder v1.5.0 │ │ │ └── zerocopy-derive v0.7.35 (proc-macro) │ │ │ ├── proc-macro2 v1.0.94 │ │ │ │ └── unicode-ident v1.0.18 │ │ │ ├── quote v1.0.39 │ │ │ │ └── proc-macro2 v1.0.94 () │ │ │ └── syn v2.0.99 │ │ │ ├── proc-macro2 v1.0.94 () │ │ │ ├── quote v1.0.39 () │ │ │ └── unicode-ident v1.0.18 │ │ └── rand_core v0.9.3 │ │ └── getrandom v0.3.1 │ │ ├── cfg-if v1.0.0 │ │ └── libc v0.2.170 │ ├── rand_core v0.9.3 () │ └── zerocopy v0.8.23 └── rand_distr v0.5.1 ├── num-traits v0.2.19 () └── rand v0.9.0 ()

      yep, still looks relatively good.

      • imtringued 44 minutes ago

            cargo tree llm v0.1.0 (RustGPT)
            ├── ndarray v0.16.1
            │   ├── matrixmultiply v0.3.9
            │   │   └── rawpointer v0.2.1
            │   │       [build-dependencies]
            │   │       └── autocfg v1.4.
            │   ├── num-complex v0.4.6
            │   │   └── num-traits v0.2.19
            │   │       └── libm v0.2.15
            │   │           [build-dependencies]
            │   │           └── autocfg v1.4.0
            │   ├── num-integer v0.1.46
            │   │   └── num-traits v0.2.19 ()
            │   ├── num-traits v0.2.19 ()
            │   └── rawpointer v0.2.1
            ├── rand v0.9.0
            │   ├── rand_chacha v0.9.0
            │   │   ├── ppv-lite86 v0.2.20
            │   │   │   └── zerocopy v0.7.35
            │   │   │       ├── byteorder v1.5.0
            │   │   │       └── zerocopy-derive v0.7.35 (proc-macro)
            │   │   │           ├── proc-macro2 v1.0.94
            │   │   │           │   └── unicode-ident v1.0.18
            │   │   │           ├── quote v1.0.39
            │   │   │           │   └── proc-macro2 v1.0.94 ()
            │   │   │           └── syn v2.0.99
            │   │   │               ├── proc-macro2 v1.0.94 ()
            │   │   │               ├── quote v1.0.39 ()
            │   │   │               └── unicode-ident v1.0.18
            │   │   └── rand_core v0.9.3
            │   │       └── getrandom v0.3.1
            │   │           ├── cfg-if v1.0.0
            │   │           └── libc v0.2.170
            │   ├── rand_core v0.9.3 ()
            │   └── zerocopy v0.8.23
            └── rand_distr v0.5.1
                ├── num-traits v0.2.19 ()
                └── rand v0.9.0 ()
      • cmrdporcupine 1 hour ago
        linking both rand-core 0.9.0 and rand-core 0.9.3 which the project could maybe avoid by just specifying 0.9 for its own dep on it
    • tonyhart7 2 hours ago
      is this satire or does I must know context behind this comment???
      • stevedonovan 2 hours ago
        These are a few well-chosen dependencies for a serious project.

        Rust projects can really go bananas on dependencies, partly because it's so easy to include them

      • obsoleszenz 2 hours ago
        The project only has 3 dependencies which i interpret as a sign of quality
  • Snuggly73 1 hour ago
    Congrats - there is a very small problem with the LLM - its reusing transformer blocks and you want to use different instances of them.

    Its a very cool excercise, I did the same with Zig and MLX a while back, so I can get a nice foundation, but since then as I got hooked and kept adding stuff to it, switched to Pytorch/Transformers.

    • icemanx 1 hour ago
      correction: It's a cool exercise if you write it yourself and not use GPT
      • Snuggly73 1 hour ago
        well, hopefully the author did learn something or at least enjoyed the process :)

        (the code looks like a very junior or a non-dev wrote it tbh).

  • Charon77 2 hours ago
    Absolutely love how readable the entire project is
    • emporas 2 hours ago
      It is very procedural/object oriented. This is not considered good Rust practice. Iterators make it more functional, which is better, more succinct that is, and enums more algebraic. But it's totally fine for a thought experiment.
    • koakuma-chan 2 hours ago
      It's AI generated
      • Revisional_Sin 1 hour ago
        How do you know? The over-commenting?
        • koakuma-chan 1 hour ago
          I know because this is how an AI generated project looks. Clearly AI generated README, "clean" code, the way files are named, etc.
          • magackame 1 hour ago
            Not sure myself. Commit messages look pretty human. But the emojis in readme and comments like "// Re-export key structs for easier access", "# Add any test-specific dependencies here if needed" are sus indeed.
          • cmrdporcupine 1 hour ago
            To me it looks like LLM generated README, but not necessarily the source (or at least not all of it).

            Or there's been a cleaning pass done over it.

            • koakuma-chan 1 hour ago
              I think pretty clearly the source is also at least partially generated. None the less, just a README like that already sends a strong signal to stop looking and not trust anything written there.
        • GardenLetter27 1 hour ago
          The repeated Impls are strange.
          • magackame 1 hour ago
            Where? Don't see any on latest main (685467e).
            • yahoozoo 1 hour ago
              `llm.rs` has many `impl LLM` blocks
    • yieldcrv 2 hours ago
      Never knew Rust could be that readable. Makes me think other Rust engineers are stuck in a masochistic ego driven contest, which would explain everything else I've encountered about the Rust community and recruiting on that side.
      • jmaker 2 hours ago
        Not sure what you’re alluding to but that’s just ordinary Rust without performance or async IO concerns.
      • GardenLetter27 1 hour ago
        Most Rust code looks like this - only generic library code goes crazy with all the generics and lifetimes, due to the need to avoid unnecessary mallocs and also provide a flexible API to users.

        But most people aren't writing libraries.

  • Goto80 2 hours ago
    Nice. Mind to put a license on that?
  • kachapopopow 2 hours ago
    This looks rather similar to when I asked an AI to implement a basic xor problem solver I guess fundementally there's really only a very limited amount of ways to implement this.
  • capestart 36 minutes ago
    Very cool project, always nice to see deep learning built from scratch in Rust without heavy frameworks.
  • jlmcgraw 1 hour ago
  • abricq 2 hours ago
    This is great ! Congratulations. I really like your project, especially I like how easily it is to peak at.

    Do you plan on moving forward with this project ? I seem to understand that all the training is done on the CPU, and that you have next steps regarding optimizing that. Do you consider GPU accelerations ?

    Also, do you have any benchmarks on known hardware ? Eg, how long would it take to train on a macbook latest gen or your own computer ?

  • ndai 2 hours ago
    I’m curious where you got your training data? I will look myself, but saw this and thought I’d ask. I have a CPU-first, no-backprop architecture that works very well on classification datasets. It can do single‑example incremental updates which might be useful for continuous learning. I made a toy demo to train on tiny.txt and it can predict next characters, but I’ve never tried to make an LLM before. I think my architecture might work well as an on-device assistant or for on-premises needs, but I want to work with it more before I embarrass myself. Any open-source LLM training datasets you would recommend?
  • enricozb 2 hours ago
    I did this [0] (gpt in rust) with picogpt, following the great blog by jaykmody [1].

    [0]: https://github.com/enricozb/picogpt-rust [1]: https://jaykmody.com/blog/gpt-from-scratch/

  • bigmuzzy 2 hours ago
    nice
  • Emma_Schmidt 29 minutes ago
    [dead]
  • trackflak 2 hours ago
    [dead]