As one of my podcasts approaches 1.5m downloads and the the other 50,000, I thought I'd share a few things that I've learnt over the past 3 years on building a successful podcast 👇

Santkumaras-The four adorable kids of the Universe.

The Kumaras r the four sages who roam the Universe as children from the Puranic texts of Hinduism named,Sanaka, Sanatana,Sanandana& Sanatkumara.They r described as the first mind-born creations &sons of the creator God Brahma.

Brahma,who had received the knowledge of Puranas from the Universal being(God Himself), imbibed this to his 4 Kumaras.
Sanat Kumaras then taught the Puranas to Narada.Narada transmitted it to Krishna Dwaipayana who in turn scripted them in monumental epic documents of 18 Puranas.

The Vishnu Purana is recorded in two parts,the Vishnu Purana & Naradiya Purana.The teachings of Sanaka of the Kumara brothers r contained in the Naradiya Purana which is also divided in two parts, the first part containing the teachings of Sanaka &second part,teachings of others.

Born from Brahma's mind,4 Kumars took lifelong vows of celibacy against their father's wishes.They r said to wander throughout materialistic &spiritualistic universe without any desire but with the aim to teach.All 4 brothers studied Vedas from their Childhood&travelled together.

The Bhagwata Purana lists the Kumaras among the 12 Mahajanas(great devotees or bhaktas) who although being eternally liberated souls from birth, still became attracted to the devotional service of Vishnu from their already enlightened state.

In 1937, a trucker got frustrated waiting for his cargo to be unloaded. He later sold his company for $1.1 billion.

Here’s the story of how Malcom McLean made a billion dollars and which startup could be following in his footsteps. (Inspired by our episode with @laurabehrenswu)

In the early 1950s, the shipping industry was a dying business.

New York’s docks were handling half as much domestic cargo in the early 1950s as they had been in the depressed 1930s.

In thirty years, no one had invested significant money in coastal shipping.

Why?

As McLean found out, loading ships was difficult and slow.

A truck or train would deliver items to the port. Each item was unloaded separately, recorded, and carried to storage.

When a ship was ready, each item was taken from storage, counted again, and hauled on-board.

The cost of this labor was significant. As one expert at the time explained:

“A four-thousand-mile voyage for a shipment might consume 50 percent of its costs in covering just the two ten-mile movements through two ports.”

Yet, despite the industry’s issues, it felt little pressure to change.

Foreigners were barred from operating domestically. Cartels ran international routes. And gov. subsidies eased the pain of labor costs.

Reshaping the industry needed an outsider.

Enter Malcolm McLean.

Lately there have been various rumours in the media about myself and Jumbo in relation to Williams. To clarify things, Jumbo is not and will not invest any money in my career. 1/2

Of course we’re collectively looking at career opportunities, however this is certainly not our philosophy and strategy. I hope to announce my future plans soon, be patient 😉 2/2

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.