Posts about software and AI.

2/ It all revolves around an algorithm that measures whether a note is “helpful” or not. First, users rate notes as “helpful,” “somewhat helpful,” or “not helpful.” Then the algorithm uses these votes to train a model that evaluates both how helpful the notes are and the “quality” (or trustworthiness) of the voters. 🧠

3/ But votes alone aren’t enough: you need cross-partisan consensus. That is, support from both left- and right-leaning users. If there’s no agreement between both sides, the note isn’t published. 🤝 This consensus acts as a quality indicator: if it convinces everyone, it’s probably reliable information.

4/ Also, your influence depends on your track record. If your votes tend to align with the consensus, your weight in the system goes up. If you’re always disagreeing with everyone, your impact goes down. This prevents biased users from distorting the results. 🎯 (Obviously, this isn’t risk-free)

5/ Abuse is penalized: if a note is marked as offensive by cross-partisan consensus, those who rated it as “helpful” lose credibility. 🛑 That way, the system discourages toxic behavior and rewards constructive participation.

6/ Algorithm summary: • Only notes with consensus between left and right get published. • Reliable users have more weight. • Abuse is punished. It’s simple in concept, but it raises some questions: • Is this balance sustainable? • Isn’t reducing everything to “left” and “right” a bit simplistic? 🤔

7/ In any case, the system is a fascinating experiment. It’s not perfect, but it seems like a good foundation for managing misinformation on large platforms. And realistically, it works pretty well in practice. What do you think? 💬

8/ If you want to dig deeper, this post from Less Wrong explains it better than I can: lesswrong.com/posts/sx9wTyCp5kgy8xGac/co…

I used to use SMLs (small LLMs) to unify reports into a universal format and then give them the style I wanted. Although this approach delivers the best results, it has several problems:

1) Long wait times for users ⏳ This is probably the biggest problem for me. I want a smooth, uninterrupted experience. Having users wait for AI to generate reports has been the main reason to look for a change.

2) Inconsistencies in AI-generated reports, making year-over-year comparisons difficult Novus lets you compare annual reports and highlight differences, but the old model created inconsistencies and made it hard to make a coherent comparison

3) Sky-high costs 💸 AI drives costs way up. Even though they’ve come down a lot, generating hundreds of thousands of reports is still a burden I can’t ignore.

What’s my current solution? 🔍 A new processing method: I “normalize” and integrate the reports into their original style. That way, we reduce formatting issues, save costs and, best of all, it’s MUCH FASTER! 🚀

Abandoning the other approach? Yes and no. The limitations of the other approach will be solved over the coming months, so I’ll probably pick it back up when the technology meets my quality requirements for a better user experience. 😉🔜