Milan Weibel https://weibac.github.io/
Milan W
Karma: 395
No-self as an alignment target
keep in mind that one persons modus tollens is anothers modus ponens, and i provided no indication as to what update i prefer people make from reading my observation
“agentic” and “power seeker” (when applied to a person) form a pair of russell conjugates
I am interested in the space. Lots of competent people in the general public are also interested. I had not heard of this hackathon. I think you probably should have done a lot more promotion/outreach.
Here is a customizable LLM-powered feed filter for X/Twitter: https://github.com/jam3scampbell/Promptable-Twitter-Feed
This reads like marketing content. However, when read at a meta level, it is a good demonstration of LLMs being already deployed in the wild.
Maybe for a while.
Consider, though, that correct reasoning tends towards finding truth.
In talking with the authors, don’t be surprised if they bounce off when encountering terminology you use but don’t explain. I pointed you to those texts precisely so you can familiarize yourself with pre-existing terminology and ideas. It is hard but also very useful to translate between (and maybe unify) frames of thinking. Thank you for your willingness to participate in this collective effort.
Let me summarize so I can see whether I got it: So you see “place AI” as body of knowledge that can be used to make a good-enough simulation of arbitrary sections of spacetime, where are events are precomputed. That precomputed (thus, deterministic) aspect you call “staticness”.
How can a place be useful if it is static? For reference I’m imagining a garden where blades of grass are 100% rigid in place and water does not flow. I think you are imagining something different.
I think you may be conflating between capabilities and freedom. Interesting hypothesis about rules and anger though, has it been experimentally tested?
Hmm i think i get you a bit better now. You want to build human-friendly and even fun and useful-by-themselves interfaces for looking at the knowledge encoded in LLMs without making them generate text. Intriguing.
I’m not sure I follow. I think you are proposing a gamification of interpretability, but I don’t know how the game works. I can gather something about player choice making the LLM run and maybe some analogies to physical movement, but I can’t really grasp it. Could you rephrase it from it’s basic principles up instead of from an example?
Build software tools to help @Zvi do his AI substack. Ask him first, though. Still if he doesn’t express interest then maybe someone else can use them. I recommend thorough dogfooding. Co-develop an AI newsletter and software tools to make the process of writing it easier.
What do I mean by software tools? (this section very babble little prune)
- Interfaces for quick fuzzy search over large yet curated text corpora such as the openai email archives + a selection of blogs + maybe a selection of books
- Interfaces for quick source attribution (rhymes with the above point)
- In general, widespread archiving and mirroring of important AI safety discourse (ideally in Markdown format)
- Promoting existing standards for the sharing of structured data (ie those of the semantic web)
- Research into the Markdown to RDF+OWL conversion process (ie turning human text into machine-computable claims expressed in a given ontology).
Study how LLMs act in a simulation of the iterated prisoner’s dilemma.
A qualitative analysis of LLM personas and the Waluigi effect using Internal Family Systems tools
Reversibility should be the fundamental training goal. Agentic AIs should love being changed and/or reversed to a previous state.
That idea has been gaining traction lately. See the Corrigibility As a Singular Target (CAST) sequence here on lesswrong. I believe there is a very fertile space to explore at the intersection between CAST and the idea that Instrumental Goals Are A Different And Friendlier Kind Of Thing Than Terminal Goals. Also probably add in Self-Other Overlap: A Neglected Approach to AI Alignment to the mix. A comparative analysis of the models and proposals presented in these three pieces I just linked could turn out to be extremely useful.
What if we (somehow) mapped an LLM’s latent semantic space into phonemes?
What if we then composed tokenization (ie word2vec) with phonemization (ie vec2phoneme) such that we had a function that could translate English to Latentese?
Would learning Latentese allow a human person to better interface with the target LLM the Latentese was constructed from?
Anthropic is calling it an “hybrid reasoning model”. I don’t know what they mean by that.
i think my preference is “both at once” or something like that