Sending good vibes.

Meaning that infrastructure like AWS is not something that Signal, or almost anyone else, could afford to just “spin up.” Which is why nearly everyone that manages a real-time service–from Signal, to X, to Palantir, to Mastodon–rely at least in part on services provisioned by these companies. 8/

What happened? :)

3/ NSO also emerges from the WhatsApp v NSO case with just an absolute TON of their business splashed all over the court records..

Ouch

E.g. check out the filings from Sept 15th 2025:

www.courtlistener.com/docket/16395...

2/ Altho massive punitive damages jury award against NSO Group ($167m) got reduced by the court, as is expected in cases where it is so large (to 9x compensatory damages)...

This is likely cold comfort to NSO since I think the injunction seriously dims value of NSO's spyware product.

No, but a lot of the "weights are not derived from training data" hinges on arguments that individual works disappear in the larger training corpus.

(My mind was blown by the "more women are willing to undergo a hormonal cycle and surgery than men are willing to donate sperm" - that was just unimaginable to me previously)

According to the podcast the bottleneck is exactly that: An insufficient number of volunteers in France at least.

So my information is from a French podcast called "Les couilles sur la table" (episode here: www.binge.audio/podcast/les-...) and it had the most insane data point: Up until last year, there were more women willing to donate eggs than men willing to donate sperm in France.

Random bizarre data point btw: In France, there's a 10k waitlist for receiving a sperm donation, average waiting time is ~2 years. So in theory, any man could participate in procreation, if not in fatherhood. Would one get a tax rebate then?

It's pretty nuts how much of an intrinsic bias toward geometric shapes the ReLU network seems to have. It extrapolated much more of the shape than I had anticipated (in Adam, the SGD variant looked more overfitty).

Ok, convergence with SGD was too painfully slow, so I did a few thousand iterations with Adam, then switched to SGD. The result looks like this:

Current state is this. It's early in the training process, though - I chose SGD instead of ADAM (which becomes unstable when trained to very low loss), and the weight initialization which led to many creases in the upper left corner is still visible.