From db2fdf3d737de1fd093bb9a95037bb38859dbd4a Mon Sep 17 00:00:00 2001 From: James Calligeros Date: Sat, 19 Jul 2025 22:59:06 +1000 Subject: [PATCH] policies: add slop policy Signed-off-by: James Calligeros --- docs/project/policies/slop.md | 83 +++++++++++++++++++++++++++++++++++ mkdocs.yml | 2 + 2 files changed, 85 insertions(+) create mode 100644 docs/project/policies/slop.md diff --git a/docs/project/policies/slop.md b/docs/project/policies/slop.md new file mode 100644 index 00000000..01537fed --- /dev/null +++ b/docs/project/policies/slop.md @@ -0,0 +1,83 @@ +--- +title: Generative AI Policy +--- + +It is the opinion of the Board that Large Language Models (LLMs), herein referred +to as Slop Generators, are unsuitable for use as software engineering tools, +particularly in the Free and Open Source Software movement. + +The use of Slop Generators in _any_ contribution to the Asahi Linux project is +expressly forbidden. Their use in any material capacity where code, documentation, +engineering decisions, etc. are largely created with the "help" of a Slop Generators +will be met with a single warning. Subsequent disregard for this policy will be +met with an immediate and permanent ban from the Asahi Linux project and all +associated spaces. + +## Illegal output +All of the popular Slop Generators are trained on an incomprehensibly large corpus +of text. There is ample evidence across the Web of this training material including +copyrighted material, brazenly stolen by the Slop Generator proprietors with +impunity. Due to the nature of Slop Generators, they are prone to regurgitating +their training corpus almost verbatim. This presents a challenge for FOSS projects +in that the use of generated slop is highly likely to violate intellectual property +law by way of regurgitating the aforementioned stolen training material. This +likelihood is proportional to the specificity of the problem area. + +Asahi Linux is a _highly_ specific project, working in esoteric problem spaces +on publicly undocumented hardware. Given the techniques used by Slop Generator +manufacturers, it is not impossible for them to have confidential or leaked +material owned by Apple or its vendor partners in their training corpi. It is +therefore likely that Slop Generators will regurgitate this when queried in just +the right way. We already forbid the use of illegally acquired or leaked +documentation and tooling (e.g. Apple's internal repair diagnostic tools). This +also applies to regurgitated slop. + +FOSS projects like Asahi Linux cannot afford costly intellectual property lawsuits +in US courts. The current political situation in that nation also makes it +incredibly unlikely that any FOSS project would win such a suit regardless of +the quality of its defence. + +## Waste of resources +Slop Generators consume an unfathomable amount of resources we can scarcely afford +to waste. Training, and to a lesser extent inference, require enormous amounts of +energy, water, land, and hardware. Manufacturing the hardware itself requires enormous +amounts of energy, water and minerals. All parts of the Slop Generator supply chain +are environmentally intensive. These resources are better used on quite literally +anything else. + +## LMGTFY +An emerging trend we have observed is people copying user questions or posts into +a Slop Generator, then replying to the post with the generated slop. This is +occurring with increasing frequency, particularly on Reddit. For some people it +is tempting to "help" others and answer questions by feeding them to a LLM and +then posting the answer as-is, or lightly edited at best. If this is you, please +realise that others also have access to the same models as you do, and if they +wanted an answer from one, they could have asked it themselves. Doing this is +exactly as helpful as posting a LMGTFY link, and everyone else _will_ view your +actions as if you did exactly that. + +## It's just matmul +It is very easy to get caught up in the hype that bad actors have built around +Slop Generators. The anthropomorphic presentation of Slop Generators as "agents" +or "assistants" is a very deliberate attempt to manufacture consent for their +integration into workforces at the expense of human interaction. The implication +of some higher degree intelligence or sentience is very much deliberate, and it +is very much false. + +Make no mistake, they cannot think. They cannot reason. They cannot take into +account context. They don't "know" things or have a sense of humour or any of the +other human-centric qualities bad actors would have you believe of them. Slop +Generators are a [chain of matrices in a stochastic system](https://en.wikipedia.org/wiki/Markov_chain). +The output of a Slop Generator is nothing more than a statistical calculation, +where the next word to be generated is decided by an opaque probabilistic +function dependent on previously generated words. This is fundamentally the +same mathematics that is used to predict the weather. + +A Slop Generator cannot assess the veracity of its claims, nor can it ever +tell you that it simply does not know something. Slop Generators +are often _confidently incorrect_ as a result, and require brow-beating +to admit a mistake. They are therefore highly inappropriate tools in contexts where +truth and correctness are of utmost importance, and when the user is not already +highly knowledgeable and confident in the problem area. This presents a bit of +an issue for Slop Generators; if the user is already highly knowledgeable and +confident in the problem area, then why ask the Slop Generator in the first place? diff --git a/mkdocs.yml b/mkdocs.yml index 3e74e3e5..802c6187 100644 --- a/mkdocs.yml +++ b/mkdocs.yml @@ -18,6 +18,8 @@ nav: - When will Asahi Linux be done?: project/when-will-asahi-be-done.md - References: project/references.md - Asahi Linux Board: project/board/asahi-board.md + - Policies: + - Generative AI: project/policies/slop.md - Platform documentation: - Apple Silicon Subsystems: platform/subsystems.md - Apple Platform Security Crash Course: platform/security.md