4 ms·

> Recommendation algorithms hold far too much power and influence in our society to remain as secretive corporate IP. I don't think that simply making the algo

> Recommendation algorithms hold far too much power and influence in our society to remain as secretive corporate IP. I don't think that simply making the algorithm public would do much. Okay, now you know it's an ML algorithm applied to X data, what does that change? It's just as toxic as before. Maybe it should be required to show a _non_-recommendation feed as an option? Recommendation feeds create self-reinforcing filter bubbles by design.

urthor 5y ago
Oh I don't mean to publicize just the algorithm. You would need to publicize the weights and input data of the algorithm. You can't just say "see this whitepaper of the textbook for algorithm." Obviously there's a lot that goes into this. But the flip-side is that the recommendation algorithms are so crucial we have to tackle this.
- phailhaus 5y ago
  > You would need to publicize the weights and input data of the algorithm Again, I don't think that really tells you anything, and I'm telling you now there is _zero_ chance that these companies open source their training data (that's their secret sauce). But even if you could, what can you possibly do with this information that would be useful? It's completely inscrutable for laypeople, which means it will only be analyzed by ML-competent engineers and mathematicians. And if it's a deep-learning algorithm, it's completely unexplainable and will yield practically no useful insights. I don't think there is an "ideal" recommendation algorithm. There is no way to say "ah if we just tweak these weights, everything will be fixed!" Every single one has tradeoffs and will direct you towards filter bubbles. The problem is that many sites give you practically no other way to see content, which means you're constantly being manipulated by an algorithm that is itself learning from how you interact with what it proposed to you. That feedback loop is toxic, and steals user agency.

urthor 5y ago

Oh I don't mean to publicize just the algorithm. You would need to publicize the weights and input data of the algorithm. You can't just say "see this whitepaper of the textbook for algorithm." Obviously there's a lot that goes into this. But the flip-side is that the recommendation algorithms are so crucial we have to tackle this.

phailhaus 5y ago
> You would need to publicize the weights and input data of the algorithm Again, I don't think that really tells you anything, and I'm telling you now there is _zero_ chance that these companies open source their training data (that's their secret sauce). But even if you could, what can you possibly do with this information that would be useful? It's completely inscrutable for laypeople, which means it will only be analyzed by ML-competent engineers and mathematicians. And if it's a deep-learning algorithm, it's completely unexplainable and will yield practically no useful insights. I don't think there is an "ideal" recommendation algorithm. There is no way to say "ah if we just tweak these weights, everything will be fixed!" Every single one has tradeoffs and will direct you towards filter bubbles. The problem is that many sites give you practically no other way to see content, which means you're constantly being manipulated by an algorithm that is itself learning from how you interact with what it proposed to you. That feedback loop is toxic, and steals user agency.