According to this research study, these are the two main ingredients: 1) Homophily in the interaction networks 2) Bias in the information diffusion toward like-minded peers https://www.pnas.org/doi/10.1073/pnas.2023301118
Substack Notes is their way to venture into the social platform realm, which differs from the personal email inbox model, which isn't a social platform. With that in mind, for ingredient #1: people can only see notes from people they follow. People will overwhelmingly only follow people that align with their ideas/political leanings/beliefs. This creates the homophily in interaction networks. For #2, users can easily share biased information to their like-minded peers without pushback/opposing views since their social network is comprised of people with the same viewpoints.
To be clear, algorithms can definitely create echo chambers as well but ideally an algorithm based feed will promote discourse and dialogue about ideas.
An anecdote: I follow quite a few people on TikTok that post about housing policy in NYC. TikTok's algorithm exposes these videos to people with varying viewpoints, which creates a ton of dialogue on proposed solutions and pushback on ideas that could potentially harm certain demographics/historical areas, etc. This pushback is very important but is only possible when there is a chance for those with opposing views to discover it, which means there needs to be an algorithmic recommendation feature.
If TikTok had no FYP and instead people could only watch videos from people they follow, this would create a closed loop system and those opposing views for housing policy would not be nearly as prevalent.