alvaroir

I'm really impressed about this demo! Apart from the usual quality benchmarks I'm really impressed about the latency for audio/video: "It can respond to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds, which is similar to human response"... If true at scale, what could be the "tricks" they're using for achieving that?!

If the problem is that the "recommendations" are biased, wouldn't it be simpler to:

1) Remove the recommendation part. Go back to a simpler version of Twitter. 2) Still give the option for a recommendation system but, somehow, open the code, maybe some version of the data and publish documentation (like papers or something) detailing the training process of the current running version.

than build a recommender system marketplace? Also, in what sense would this be different from allowing third party Twitter clients + opening a ritcher API?

The ultimate idea of using ML is to "automatically" build the recommender system the user likes (measuring this with some particular metric like online time or retention) the most and also automatically adapt it as his/her preferences change. The problem to me is more the metrics chosen to be optimized.

However, I believe that in the end, and in order to be profitable, user retention and time on the platform will still be pursued. It doesn't seem like an easy fix to me.

Regarding the "free speech" part, I'm not an expert, but I'd say (and after having watched the TED interview) that countries' legislations will considerably constraint this.

I love the idea of a true free platform tho