#Titan

ABC Bullish in 0.5%

November Future

So it turns out that Google Chrome was making everything on my computer slow *even when it wasn’t running*, because it installs something called Keystone which is basically malware.

I made a website because this shouldn’t

Wired first reported on how bad Keystone was 11 years ago when they put it into Google Earth (they seem to put it in all their popular

The fact that Keystone hides itself in Activity Monitor is bizarre. (The only sign of it was excessive CPU usage of WindowServer which is a system process).

I don’t know if Google was doing something nefarious with Keystone, or a third party figured out how to (which Wired warned about). But either way, I’m not inclined to give Google-the-organization the benefit of the doubt (despite the many good people who work on Chrome)...

...since it's been a decade+ and this still hasn't been "fixed".

There is no reason for auto-update software to need to do what Chrome/Keystone was doing. It also has a long history of crashing Macs.

Ok, I'll hop on this Elizabeth Warren/Native American testing bandwagon (gently and non-politically).

I've been a @23andMe customer for a while, and have followed their ancestry updates closely.

All is more or less as expected....except for this bit about Native American.

The family is almost completely composed of Spanish peasants (from various regions) who emigrated, along with a massive wave in the late 19th-cent./early 20th-cent., to Cuba back when it was a booming economy (richer than Spain's) and worth emigrating to (Communism killed that).

Also, the native population of Cuba was annihilated early on---was the first place the Spanish colonized after all. Having a native background in Cuba would be like having the same in, say, Massachusetts, particularly if you're (say) mostly Irish. Just really, really unlikely.

(Note: the North African/Arab background is less mysterious. The Iberian peninsula was part of the Muslim world for centuries. It would be odd *not* to have some Arab/Middle Eastern background coming from Spain. Given the family is mostly from Northern Spain, it's small though.)

I have a Spanish passport, have been back to the ancestral villages in Spain, seen the church where my grandmother was baptized, my grandfather told me stories about growing up as the child of Spanish colonists in rural Cuba. The native bit just clashes with all the family lore.

#Thread The legend of Ekadash Mukhi Hanuman ji...

Mahadev himself narrated the story of Ekadash Mukhi Hanuman ji to Maa Parvati. Once a great monster named Kalkarmukh performed the severe penance of Brahma ji.

Pleased with the austerity of the demon, Brahma ji gave him a boon that the person who will have eleven Heads like you and who was born on the day of your birth, only he/she will be able to kill you. No one else will kill you.

After getting a boon from Brahma ji, Kalkarmukh created a ruckus in all three lok. He invaded Swarg and snatched all the rights of the Devtas and became himself the king of Swarg.

All Devtagan came to Shri Rama and asked him for help. Hearing the agony of the Devtagan, Shri Ram called his supreme devotee Hanuman ji and said, "Pawan Putra, Sankatmochan".

Important paper from Google on large batch optimization. They do impressively careful experiments measuring # iterations needed to achieve target validation error at various batch sizes. The main "surprise" is the lack of surprises. [thread]

https://t.co/7QIx5CFdfJ

The paper is a good example of lots of elements of good experimental design. They validate their metric by showing lots of variants give consistent results. They tune hyperparamters separately for each condition, check that optimum isn't at the endpoints, and measure sensitivity.

They have separate experiments where the hold fixed # iterations and # epochs, which (as they explain) measure very different things. They avoid confounds, such as batch norm's artificial dependence between batch size and regularization strength.

When the experiments are done carefully enough, the results are remarkably consistent between different datasets and architectures. Qualitatively, MNIST behaves just like ImageNet.

Importantly, they don't find any evidence for a "sharp/flat optima" effect whereby better optimization leads to worse final results. They have a good discussion of experimental artifacts/confounds in past papers where such effects were reported.