Post

How to get URL link on X (Twitter) App

On the Twitter thread, click on or icon on the bottom
Click again on or Share Via icon
Click on Copy Link to Tweet
Paste it above and click "Unroll Thread"!
More info at Twitter Help

Carlos E. Perez

@IntuitMachine

Sep 18, 2020 • 7 tweets • 1 min read • Read on X

Scrolly

Meaning-making is all about discovering useful sign (see: Peirce) rewrite rules. #ai

The conventional artificial neural network (i.e. sum of product of weights) is a rewrite rule from a vector to a scalar. Each layer is a rewrite rule from a vector to another vector.

A transformer block is a rewrite rule from a set of discrete symbols into vectors and back again to discrete symbols.

Execution of programming code are just rewrite rules transforming high-level code to machine code for execution.

Re-write rules can do everything. The hard problem is discovering these re-write rules. The even harder problem is formulating a system that discovers these re-write rules.

Deep Learning networks learn re-write rules by adjusting weight matrices. No new rules are added, just the relative importance of rules are adjusted. Like biology, this involves a differentiation process and not an additive process.

DL networks only work if given sufficient diversity on initialization. Initializing all weights uniformly is a recipe for failure.

• • •

Missing some Tweet in this thread? You can try to force a refresh

This Thread may be Removed Anytime!

Twitter may remove this content at anytime! Save it as PDF for later use!

More from @IntuitMachine

Carlos E. Perez

@IntuitMachine

May 30

Humans tend to experience thought through one dominant frame at a time, which creates tunnel vision: the active viewpoint determines what we notice, ignore, and consider possible. Productive exploration with LLMs counteracts this by making frames explicit and allowing us to move deliberately across abstraction levels and perspectives. The value is not just more answers, but more ways of seeing.

Wordcells remain captured by the abstractions encoded in language, while shape rotators recognize abstractions as movable frames. They can zoom out, zoom in, and rotate perspective rather than merely elaborate the current verbal frame.

Most people use LLMs to complete their current frame. That feels productive because the idea becomes clearer, smoother, and more persuasive. But the highest-value use is to make the model attack the frame itself. Ask it to assume you are wrong, find the world where the opposite is true, identify what your framing made invisible, and force every abstraction back into a concrete test or decision. Agreement is cheap. Disconfirmation is where the value is.

Read 5 tweets