Grade Awarding 2021 - a thread.
1. For pupils and parents; it sounds daft but there isn’t much you can control right now. Teacher will play the main role in awarding grades, and there may well be some time or formal exam for subjects. But we don’t know exactly what yet.

Hey, for people interested in this stuff, here's a thread about a change we made to running our facebook account:

So, about 5 years ago, we got on facebook. Why? Well, facebook has a big audience and they like to consume on facebook. So, we started posting comics there, even though we got paid nothing and facebook made money off the free content.

Why? Because back then, once in a while, you would say "also hey, I sell books" or "hey, I'm going to be signing at this event" and your facebook followers would actually see it.

Effectively, the implicit old arrangement was "facebook lets me reach readers efficiently, and I supply facebook with free content they run ads on."

Over time, like a bad business partner, facebook basically made it impossible to reach your audience without *paying them*. Thus, you supply them free content on which they run ads, and you get nothing in exchange.

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.

(THREAD) Pre-election Russia collusion may take down Trump. It's equally possible the Trumps' pre- and post-election collusion with Saudi Arabia, the UAE, and Israel will do it—a course of collusion also connected to the Kremlin. I explain here. I hope you'll read on and retweet.

1/ I must stress how unbelievably complex the "Grand Bargain" theory of the Trump-Russia case is—a different thing from saying it's not substantiated. It's substantiated in *almost every single particular*—it just *also* happens to be very confusing. Not byzantine—just confusing.

2/ The basics: Saudi Arabia, Israel, and the UAE all view Iran as their chief regional enemy. Iran is propped up by Russia. Therefore the Saudis, Israelis, and Emiratis all need a US government willing to find a way to get the Kremlin to *stop* supporting Iran in the Middle East.

3/ The best way to get Russia to stop supporting Iran—or reduce support—was/is to drop all sanctions on Russia over its 2014 annexation of Crimea, as that'd be worth *trillions* to the Kremlin over the next decade. Everyone knew that Clinton wouldn't do this—and that Trump would.

4/ Per the NYT, on August 3, 2016, Donald Trump Jr. met secretly at Trump Tower with a Saudi and Emirati emissary, George Nader, as well as an Israeli intelligence expert, Joel Zamel, with *significant* ties to both Israeli intelligence *and* Russian oligarchs allied with Putin.