OK. I’m tired of the emotional reasoning, so I went to the CDC’s website and sorted through the data myself. Here are the cold, hard facts as of 9/4/21.

TOTAL CASES (US): 40,048,255
TOTAL DEATHS (US): 643,858

1) The Temple of the Divine Claudius was a vast temple complex on Rome's Caelian Hill, overlooking the valley that would later be occupied by the Colosseum. Though almost nothing now remains of the structure, its scale and grandeur once rivalled anything in the city.. #LostRome

2) The 'Claudium' was intended as a sacred complex to revere Claudius and the imperial cult. Construction began soon after the death and deification of the emperor in 54 AD, initiated by his widow Agrippina who was likely responsible for his supposed death by poisoning. #LostRome

3) The sprawling temple complex sat on a gigantic podium that spread across the Caelian Hill, measuring 180 x 200 metres; with towering retaining walls of travertine that raised the structure 20 metres above the valley below. #LostRome

4) The elevated sacred area was likely entered by steps facing the temple on the west side. The temple itself had a projecting hexastyle porch and was surrounded by gardens and a grand portico, probably the 'porticus Claudia' mentioned in Martial's 'On the Spectacles'. #LostRome

5) Statues of both Agrippina and Claudius’ ill-fated young son Britannicus, sculpted in sumptuous dark Egyptian greywacke, were discovered among the ruins on the Caelian Hill in 1651 – and had almost certainly been displayed in the temple precinct. #LostRome

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.

1/ Throughout my startup career, I’ve seen many founders have an aversion to sales. This is a mistake that can kill your startup.

2/ Sales is often viewed as either a saving grace or proof that the product isn’t good enough (because it should sell itself). Neither are ever true. Some common mistakes that result in...

3/ Mistake 1: Hire a sales rep before reaching product/market fit to get your initial batch of customers. This is a mistake because founders need to work through their MVP with early adopters to truly understand what it is they’re selling.

4/ Mistake 2: Reach product/market fit, need to scale, and rely entirely on self-serve. For enterprise products that require big commitments and internal shifts, almost no product is self-explanatory enough to sell itself.

5/ Mistake 3: Make a first sales hire who isn’t scrappy enough to help mold the sales process from scratch. Some salespeople are amazing at their jobs, but not cut out to establish the processes that others end up following. This skillset is what @rdedatta calls a “sales ninja”.