PickleBrine — LessWrong

LESSWRONG
LW

Replying toHow AI Is Learning to Think in Secret

It seems to me people are still anthropomorphizing (or maybe "phrenomorphizing" might be more apt) the chain-of-thought "reasoning". With respect to the issue of AI alignment I don't think it matters that much that they do this, LLMs don't have egos or the potential to set goals, have motives or intents, etc. They just learn a kind of mess of alien abstractions that optimizes their text-producing behavior for some training data. The real issue is that these abstractions probably often do not correlate with the actual ideas that are represented by word tokens, and so you get hiccups like hallucinations or these chain-of-thought snippets that make it look like it's actually thinking... (read more)

PickleBrine's Shortform

PickleBrine

20d

This is a special post for quick takes (aka "shortform"). Only the owner can create top-level comments.

PickleBrine20dQuick Take

Hello - I’m just here to drop a somewhat vague/incipient idea for an AI model and see if there are any existing frameworks that could be used.

The general idea is to view agent action and perception as part of the same discrete data stream, and model intelligence as compression of sub-segments of this stream into independent “mechanisms” (patterns of action-perception) which can be used for prediction/action and potentially recombined into more general frameworks as the agent learns.

More precisely, I’m looking for:

The method of pattern representation
An algorithm for inferring initially orthogonal/unrelated patterns from the same data stream
Some manner of meta-learning for recombining mechanisms

One promising suggestion I received elsewhere was to use reservoir computing/liquid state machines for the time series pattern recognition.

(For a conceptually similar model look at Friston’s “Active Inference”.)

Replying toIs intelligent induction even possible?

PickleBrine1mo

Is intelligent induction even possible?

Yeah that was one of my first thoughts, but does it actually mean that informative inductive hypotheses can't be conjectured more efficiently than guessing? I can imagine such an agent's average efficacy across all problem types being fixed but still better than dart-throwing.

Is intelligent induction even possible?

PickleBrine

1mo

In the course of doing some research into A(G)I models I've found myself stuck on one conundrum. One of the core feautures of general intelligence seems to be something like domain-independent pattern finding - a truly *general* intelligence would be able to "suss things out" so to speak in a variety of different domains with roughly similiar efficacy.

The problem that strikes me here is something like the *pragmatic* (as opposed to epistemological) problem of induction. This being the issue of building a model that can intelligently (that is, better than brute force) extract patterns from raw data with little to no bias for its mode of presentation. After combing through some existing... (read more)

Intro & Proposal for AGI Model

PickleBrine

9mo

Hello all, this is my first time posting in this forum, so I look forward to your feedback.

First, some brief background info about myself: I have an undergraduate background in Computer Science & Philosophy and am considering pursuing a Master's in AI. Apart from my formal education I have also engaged in a significant degree of self-study. My motivation in participating in this forum is, as you might assume, to share my ideas and discuss topics in artificial intelligence. To start off, I'll share with you an outline for an AGI model.

My proposal here is inspired by compression-theoretical models of information and detection of patterns over string streams. A lot of work... (read 790 more words →)