Semantic Compression: AI is the "Mp3" of Human Thought

Typical mp3 soundwave overlaid on a FLAC soundwave. The highs are gone, the lows are reduced by the algorithm.

The public has spent decades getting used to the "flatness" of an MP3. To save space, the algorithm performs perceptual coding—it literally throws away the frequencies it thinks our ears won't miss. It shaves off the sonic highs and lows for the sake of a smaller file. Audiophiles stuck to records when CDs took over the industry in the mid to late 80s and then stuck to CDs when MP3s became the more popular form of music media in the 2010s. I began avoiding listening to music, and my Spotify Wrapped will attest that I listen mostly to podcasts. Then I was introduced to Tidal.

The first time I listened to a song on Tidal, I selected a song I know well - it was Prince’s Let’s Go Crazy. As a trained violinist, I heard the difference immediately. On the high fidelity FLAC file, I could hear the "air" around the voices in the harmony so that I could discern which voices were Prince over-dubbing and which were his backup singers; I could hear sharp, staccato percussion that literally made me look out the window - all the sounds that a compressed file simply deletes.

Just like MP3s shaving off the highs and lows, I’ve been thinking about how AI summaries do the exact same thing to our ideas.

AI is the MP3 of thought. Most models are built to find the "consensus," the most probable middle ground — and in doing so, they normalize the narrative. They discard the outliers, the specific cultural nuances, and the "dissonance" that actually makes an idea resonate.

If we rely solely on AI to summarize our research or generate our insights, we are effectively listening to our business strategy in 128kbps. It sounds "fine," but it’s flat.

Real innovation always lives in the outliers. It lives in the "highs and lows" that the algorithm thinks are expendable. Of course, it’s more risky – a “normalized” version will likely resonate with a greater number of people.

I will not be compressed. Let’s stay high-fidelity.