Synthesizer for Thought
The article delves into synthesizers evolving as tools for music creation through mathematical understanding of sound, enabling new genres. It explores interfaces for music interaction and proposes innovative language models for text analysis and concept representation, aiming to enhance creative processes.
Read original article
The article discusses the evolution of synthesizers as tools for creating music by understanding sound as a mathematical object. Synthesizers allow for the production of new sounds and genres by assembling sound from logical components, unlike traditional instruments. The ability to attach various interfaces to synthesizers expands the ways humans can interact with music. The text also explores the concept of instruments for thought, where interpretable language models enable detailed decomposition and precise steering of language concepts. It envisions potential interfaces like heatmaps for text analysis, spectrograms for visualizing writing structures, and semantic diffs for exploring alternative text forms. Additionally, the article suggests using icons and glyphs to represent abstract concepts discovered by neural networks and proposes a concept library for sharing linguistic features to enhance thinking and writing processes. The piece highlights the potential for new tools and interfaces to revolutionize how humans interact with and create content based on mathematical understanding of different mediums.
7 commentsIts one thing to have an academic approach to synthesis and sound design - its another thing entirely to have the funds to get ones own lab up and running and have some crazy AI thing built for you:
https://www.factmag.com/2017/07/14/watch-aphex-twin-midimuta...
Would be curious how these two projects collide/coallesce ..
A concept library for sharing and collaborating on units of meaning and style
For example, I'd love to make a todo app where the user doesn't need to worry about the exact wording when they add or search for something; where they can choose the amount of verbosity in the UI... but not be stuck with a todo database that can't easily be imported to a different app.
It would be fantastic for a new kind of webbrowser, too: smaller file sizes, and perfect support for all human languages.
This already exists in a few Ask your PDF tools, and I haven't found it useful.
The problem in text is not discoverability. Order just matters too much in text. Just like how plucking exact match search results is about as useful as semantic search for 99% of use cases, the context matters. You either have something you want or you don't. The issue with semantic matching and trying to display or use that information is that you end up with a lot of blur and would be better of feeding such information to a GPT model to summarize, or simply using a GPT model initially, to construct something more useful for you to work with. Or in other words - semantic comparisons and grouping of content creates too much bloat compared to the reasoning currently possible by GPT models, and there is no solution to this here.
I can't see the appeal in synthesizing from theme; music it doesn't matter if there's more drums or more trumpet, if it's brighter or more mournful, the details don't really matter - the mood and feeling matters. But text isn't like that for me, words have meanings, they come one after another, I don't think I want synthesized writing which has more "concept of being held" and less "discussions about parenthood"; what would it be - poetry? beat poetry? SEO marketing fluff which tries to convey a feeling without being nailed down to any specific claims or accuracy? That kind of writing is something to cut through to get to what's behind it and tiring to do it; Who would want to spend time reading it for pleasure like people listen to music for pleasure, and in what situations?
It starts with 'synthesizer for thought' but moves to synthesizer for words and writing. We already understand the concepts they mention, American Robin, being held, there being three of something. New synthesized thoughts as words, how do they become more meaningful than just mishmashing words normally - "a 13 dimensional concrete", "a society based on <untranslatable>", "French, but it's whalesong"?