I despise "AI" for quite a few reasons: It's built on theft, it empowers the fascists and oligarchs, its masters seek to dis-empower or replace human workers and creatives, its name is a deception as well as its primary use case, etc. This community doesn't need a rehash. I personally despise AI because I love the programming craft and I worry about a future where code is only generated, or worse: generated autonomously. Don't get me started on "AI first" companies. Fuck that.

"AI" is an anti-human technology.

Now, separate "AI" and all its awfulness from LLM as an algorithm/data structure. Can LLMs be ethical? I honestly don't know whether the good can be isolated from the bad. I started to brainstorm this out below, but the more I write, the less convinced I am that there's a middle way. I'm afraid that much of the perceived benefit of LLMs is derived from the universal theft of training data.

Dear reader, please consider the following a brainstorm only from a non-expert Anti who's trying in good faith to find a path.

--

Here are some possible ethical use cases:

1. Natural Language Interface - Like a Terminal Interface (TUI) or Graphical User Interface (GUI) or Command Line Interface (CLI), but instead discerns user intent from human language
2. Pattern Recognition - Some of LLMs' legitimate accomplishments have been their ability to pore over decades of human work and detect patterns that otherwise would have been missed. Examples: Recent Erdős and Knuth news. LLMs are reasonable at code review and bug/security flaw detection
3. Summarization/Search - LLMs and their precursors have been rehashing summaries of well-tread topics in training data for years. Crafting summaries for human consumption seems a 'ok' use case, with the understanding than hallucinations are unavoidable. Examples: API documentation, code examples, encyclopedia-like snippets

IMO, an ethical LLM solution might have attributes like these. Disclaimer: I'm not an expert so some of this may be nonsense ("brainstorm"):

1. Public audit trail of training data
2. Author consent, voluntary or paid, for participation in training data
3. Harnesses should have a query-able manifest of valid operations. All user input should map to one of them
4. Harnesses should strictly require human acknowledgement before executing an operation, and especially when interacting with external systems
5. Human-first output - should encourage human learning and thought, not seek to replace it
6. Signed output - this one is tricky. I don't know how to accomplish it. It would be great if LLM output could be signed in a way that excluded it from future training. The signature would also serve as notice to humans that the content is explicitly from an LLM. Web browsers could then have configurations to filter LLM content out so that users can consent to consume it. This solution may not be part of LLMs themselves
7. Limited topic/training data - imagine an LLM that's only for recipes or only for a specific programming API or a specific new site. A smaller model should use fewer resources

I have high doubts that these qualities can be achieved due to complexity and cost. Such is the price of legitimacy.

--

OK, that's all. I'm going back now to stewing in my disdain for "AI".

• minor tweaks