Feb 8, 2011

GC Lingua Franca(s)

Posted by in categories: futurism, open source

This is an email to the Linux kernel mailing list, but it relates to futurism topics so I post a copy here as well.
Science doesn’t always proceed at the speed of thought. It often proceeds at sociological or even demographic speed. — John Tooby

Open Letter to the LKML;

If we were already talking to our computers, etc. as we should be, I wouldn’t feel a need to write this to you. Given current rates of adoption, Linux still seems a generation away from being the priceless piece of free software useful to every child and PhD. This army your kernel enables has millions of people, but they often lose to smaller proprietary armies, because they are working inefficiently. My mail one year ago ( listed the biggest workitems, but I realize now I should have focused on one. In a sentence, I have discovered that we need GC lingua franca(s). (

Every Linux success builds momentum, but the desktop serves as a powerful daily reminder of the scientific tradition. Many software PhDs publish papers but not source, like Microsoft. I attended a human genomics conference and found that the biotech world is filled with proprietary software. IBM’s Jeopardy-playing Watson is proprietary, like Deep Blue was. This topic is not discussed in any of the news articles, as if the license does not matter. I find widespread fear of having ideas stolen in the software industry, and proprietary licenses encourage this. We need to get these paranoid programmers, hunched in the shadows, scribbled secrets clutched in their fists, working together, for any of them to succeed. Desktop world domination is not necessary, but it is sufficient to get robotic chaffeurs and butlers. Windows is not the biggest problem, it is the proprietary licensing model that has infected computing, and science.

There is, unsurprisingly, a consensus among kernel programmers that usermode is “a mess” today, which suggests there is a flaw in the Linux desktop programming paradigm. Consider the vast cosmic expanse of XML libraries in a Linux distribution. Like computer vision (, there are not yet clear places for knowledge to accumulate. It is a shame that the kernel is so far ahead of most of the rest of user mode.

The most popular free computer vision codebase is OpenCV, but it is time-consuming to integrate because it defines an entire world in C++ down to the matrix class. Because C/C++ didn’t define a matrix, nor provide code, countless groups have created their own. It is easier to build your own computer vision library using standard classes that do math, I/O, and graphics, than to integrate OpenCV. Getting productive in that codebase is months of work and people want to see results before then. Building it is a chore, and they have lost users because of that. Progress in the OpenCV core is very slow because the barriers to entry are high. OpenCV has some machine learning code, but they would be better delegating that out to others. They are now doing CUDA optimizations they could get from elsewhere. They also have 3 Python wrappers and several other wrappers as well; many groups spend more time working on wrappers than the underlying code. Using the wrappers is fine if you only want to call the software, but if you want to improve OpenCV then the programming environment instantly becomes radically different and more complicated.

There is a team working on Strong AI called OpenCog, a C++ codebase created in 2001. They are evolving slowly as they do not have a constant stream of demos. They don’t consider their codebase is a small amount of world-changing ideas buried in engineering baggage like STL. Their GC language for small pieces is Scheme, an unpopular GC language in the FOSS community. Some in their group recommend Erlang. The OpenCog team looks at their core of C++, and over to OpenCV’s core of C++, and concludes the situation is fine. One of the biggest features of the ROS (Robot OS), according to its documentation, is a re-implementation of RPC in C++, not what robotics was missing. I’ve emailed various groups and all know of GC, but they are afraid of any decrease in performance, and they do not think they will ever save time. The transition from brooms to vacuum cleaners was disruptive, but we managed.

C/C++ makes it harder to share code amongst disparate scientists than a GC language. It doesn’t matter if there are lots of XML parsers or RSS readers, but it does matter if we don’t have an official computer vision codebase. This is not against any codebase or language, only for free software lingua franca(s) in certain places to enable faster knowledge accumulation. Even language researchers can improve and create variants of a common language, and tools can output it from other domains like math. Agreeing on a standard still gives us an uncountably infinite number of things to disagree over.

Because the kernel is written in C, you’ve strongly influenced the rest of community. C is fully acceptable for a mature kernel like Linux, but many concepts aren’t so clear in user mode. What is the UI of OpenOffice where speech input is the primary means of control? Many scientists don’t understand the difference between the stack and the heap. Software isn’t buildable if those with the necessary expertise can’t use the tools they are given.

C is a flawed language for user mode because it is missing GC, invented a decade earlier, and C++ added as much as it took away as each feature came with an added cost of complexity. C++ compilers converting to C was a good idea, but being a superset was not. C/C++ never died in user mode because there are now so many GC replacements, it created a situation paralyzing many to inaction, as there seems no clear place to go. Microsoft doesn’t have this confusion as their language, as of 2001, is C#. Microsoft is steadily moving to C#, but it is 10x easier to port a codebase like MySQL than SQL Server, which has an operating system inside. C# is taking over at the edges first, where innovation happens anyway. There is a competitive aspect to this.

Lots of free software technologies have multiple C/C++ implementations, because it is often easier to re-write than share, and an implementation in each GC language. We all might not agree on the solution, so let’s start by agreeing on the problem. A good example for GC is how a Mac port can go from weeks to hours. GC also prevents code from being able to use memory after freeing, free twice, etc. and therefore that user code is less likely to corrupt your memory hardware. If everyone in user mode were still writing in assembly language, you would obviously be concerned. If Git had been built in 98% Python and 2% C, it would have become easier to use faster, found ways to speed up Python, and set a good example. It doesn’t matter now, but it was an opportunity in 2005.

You can “leak” memory in GC, but that just means that you are still holding a reference. GC requires the system to have a fuller understanding of the code, which enables features like reflection. It is helpful to consider that GC is a step-up for programming like C was to assembly language. In Lisp the binary was the source code — Lisp is free by default. The Baby Boomer generation didn’t bring the tradition of science to computers, and the biggest legacy of this generation is if we remember it. Boomers gave us proprietary software, C, C++, Java, and the bankrupt welfare state. Lisp and GC were created / discovered by John McCarthy, a mathematician of the WW II greatest generation. He wrote that computers of 1974 were fast enough to do Strong AI. There were plenty of people working on it back then, but not in a group big enough to achieve critical mass. If they had, we’d know their names. If our scientists had been working together in free software and Lisp in 1959, the technology we would have developed by today would seem magical to us. The good news is that we have more scientists than we need.

There are a number of good languages, and it doesn’t matter too much what one is chosen, but it seems the Python family (Cython / PyPy) require the least amount of work to get what we need as it has the most extensive libraries: I don’t argue the Python language and implementation is perfect, only good enough, like how the shape of the letters of the English language are good enough. Choosing and agreeing on a lingua franca will increase the results for the same amount of effort. No one has to understand the big picture, they just have to do their work in a place where knowledge can easily accumulate. A GC lingua franca isn’t a silver bullet, but it is the bottom piece of a solid science foundation and a powerful form of social engineering.

The most important thing is to get lingua franca(s) in key fields like computer vision and Strong AI. However, we should also consider a lingua franca for the Linux desktop. This will help, but not solve, the situation of the mass of Linux apps feeling dis-integrated. The Linux desktop is a lot harder because code here is 100x bigger than computer vision, and there is a lot of C/C++ in FOSS user mode today. In fact it seems hopeless to me, and I’m an optimist. It doesn’t matter; every team can move at a different pace. Many groups might not be able to finish a port for 5 years, but agreeing on a goal is more than half of the battle. The little groups can adopt it most quickly.

There are a lot of lurkers around codebases who want to contribute but don’t want to spend months getting up to speed on countless tedious things like learning a new error handling scheme. They would be happy to jump into a port as a way to get into a codebase. Unfortunately, many groups don’t encourage these efforts as they feel so busy. Many think today’s hardware is too slow, and that running any slower would doom the effort; they are impervious to the doublings and forget that algorithm performance matters most. A GC system may add a one-time cost of 5–20%, but it has the potential to be faster, and it gives people more time to work on performance. There are also real-time, incremental, and NUMA-aware collectors. The ultimate in performance is taking advantage of parallelism in specialized hardware like GPUs, and a GC language can handle that because it supports arbitrary bitfields.

Science moves at demographic speed when knowledge is not being reused among the existing scientists. A lingua franca makes more sense as more adopt it. That is why I send this message to the main address of the free software mothership. The kernel provides code and leadership, you have influence and the responsibility to lead the rest, who are like wandering ants. If I were Linus, I would threaten to quit Linux and get people going on AI wink There are many things you could do. I mostly want to bring this to your attention. Thank you for reading this.

I am posting a copy of this open letter on my blog as well ( Reading the LKML for more than one week could be classified as torture under the Geneva conventions.


Comments — comments are now closed.

  1. Joel Pitt says:

    Unfortunately the author guy seemed to fail to comprehend much about the engineering challenges involved with OpenCog. We spent a fair bit of time exchanging emails on our OpenCog mailing list and he was already decided and mostly wanted to proselytise SciPy. He seems to think that, simply by virtue of more people using something, it will become faster. GC is a hard problem and different projects have different demands for memory usage. The wrong choice can cause unsatisfactory performance.

    Bazaar (a Git competitor) is written in Python and it’s slowness is actually a negative factor for many users. I love Python to prototype in, but like any engineering challenge you use the right tool for the job.

    Python isn’t the right tool for the core of OpenCog.

    (On the other hand, something like Erlang might be if we had the resources to reimplement the core… but alas, we don’t ;-( )

  2. Keith Curtis says:


    “Failed to comprehend OpenCog engineering challenges” is a blanket statement coming from someone working on AI. Is knowledge really so binary?

    There are plenty of ways to compile Python today. Bazaar / Bzr has gotten faster, but you should compare Mercurial. Mercurial is much faster than CVS / SVN, and is fast enough compared to GIT. It is 1% C and 99% Python. You need to look at the success stories, and learn from the failure stories. If you looked into Bzr’s perf problems, you would find they have nothing to do with Python, and everything to do with the algorithms. It is tricky to build an always fast source control system.

    So when you next say that Python isn’t right for the core of OpenCog, you are drawing that conclusion from your inaccurate data.

    You do have the resources to re-implement the core. How many months work is a Python port? Have you figured out how much faster it will make your current dev workitems? I didn’t really believe you’ve thought this through much at all. The code has been C++ since 2001, and you are acting on inertia.

    You don’t see the benefits and you are afraid of bad performance, so you feel you don’t have time.