Goerge Hotz Asks Mistral 7B to Make Q* Algorithm in Python and Runs It

by

Computer coding genius George Hotz asks the Mistral 7B model to make the Q* algorithm in python and run it.

This is near AGI level actions.

He added in a human in the loop safety feature.

It worthwhile to watch how George Hotz prompts the LLM.

9 thoughts on “Goerge Hotz Asks Mistral 7B to Make Q* Algorithm in Python and Runs It”

  1. Mixtral 8x7B is now downloadable over bittorrent, and appears to be ChatGPT-3.5+ class without any guardrails.

    It’s time be a bit worried, not of the AI apocalypse per se, but the mundane uses such as phishing.

    Mundane Singularity -> Mundane Dystopia?

  2. I think that there is a disconnect between how we interpret the various outputs of the various AI’s out there.

    If we tell DALL-E to make a picture of Elon Musk the Conquerer riding a rocket and landing successfully on Mars where Martians cheerfully greet him then we are making a request that is very open to creative interpretation on the part of the AI and on the part of the people who evaluate the picture the AI produces. In short there is room for error and interpretation.

    Code doesn’t work that way. It is either right or crashing. It isn’t like a poem that Grok writes. Code is a fundamentally different kind of text output and that is why ChatGPT is plateauing when it comes to coding challenges which are a low bar.

  3. Huh? The Q* paper had pseudocode for an A* variant. AGI consists of turning pseudo code in to Python (which itself is barely above pseudocode)? The pseudocode has checks for boundary conditions and other details. Pseudocode to Python is like Spanish to Italian.

    Apologies to the people who make a living with Python but it isn’t the be all end all of programming languages. In particular for AI demos it isn’t compiled and will defer crashes until specific parts of the program are executed. In other words just because the AI outputs text and it “runs” doesn’t mean that it is anything near correct. Maybe as a starting point do this in C or C++ or some other compiled language. Get the pointers correct and we can talk.

    By way of comparison I had an intern several years ago who I would give pseudocode to and she would implement it (in C++). The difference being that she understood what my code was actually doing and could find bugs in it before the first line of code was written.

    (For the uninformed, C++ code is generally 1000x times faster than Python and about 10x more complex).

      • I took a look at the video – the test case is extremely trivial and short.
        I’d speculate that Java and Python might have done better on a longer run.

    • Depends. I’ve done c++ since around 1994, python since around 2011. You can do some things in Python to get tremendous speedups (one that springs to mind is using some features/libraries that make function calls c-like in speed). The python interpreter is written in c (or c++?) and you can call c and c++ libraries from Python, so you can take full advantage of fast libraries from Python. Python modules like numpy, scipy, and pandas are optimized to be very fast and are feature rich (and have lots of pieces written in c/c++). 10x more complex is really bad news if it does the same job, and there really isn’t anything you can’t do with Python if more complexity is necessary. I can’t see writing the kernels of supercomputer simulation codes or GPU shaders in Python instead of c++ although it wouldn’t surprise me if someday some group decided to try to make that tractable. Personally my biggest gripe about Python is that there are times I wish it was strongly typed and you could do more static code debugging and analysis on it. Programmer productivity is usually a bigger consideration than raw code speed.

      • We don’t actually disagree. One of my comments is still pending. Python works fine in a heterogeneous computing world where it delegates the heavy lifting to libraries written in C/C++. That is how Python is meant to be used as a language the glues together work done by other languages. Nobody is going to run their AI code in Python (though they definitely will call AI code from Python).

        As you point out it is not strongly typed and that is why turning pseudocode to Python using AI is a bit of cheating. The Python code could be wrong in several areas and you wouldn’t ever know because parts of it work. Strongly typed and compilation of course aren’t guarantees but they are a check that Python lacks.

        “Programmer productivity is usually a bigger consideration” Not for the libraries that Python programmers call. Again it boils down to where you are in the call stack. If you are at or near the bottom of the call stack then performance and memory efficiency is incredibly important.

    • I love how people that are no programmers make up these comments because they read a few things in the internet. If you say python, JavaScript or any other high level programming language is pseudo code then you have absolutely no clue what you are talking about. It’s evident in your answer that you tried to skip these languages that were sub-par in your opinion to learn c++. Your little brain blanked out at pointers and now you are impressed. The underlying architecture of high level languages relies on c++ principles – they are just taken care of for you. You are definitely not an engineer.

Comments are closed.