Happy Sunday and Memorial Day, if you are in the U.S. Thank you to all of those who have served and continue to serve our nation.
The Five Best Things
Cloudflare Blog: Bringing AI to the edge with NVIDIA GPUs
Cloudflare is a very low level infrastructure company that provides content delivery networks (CDNs) and security (eg DDoS prevention) against bad actors, for small and medium businesses. It does so by operating a global footprint of datacenters with high performance hardware and software. This allows their customers to scale globally and be able to serve their own downstream customers with high availability and performance - at the flip of a switch. Cloudflare is attempting low-end disruption against the market leader, Massachusetts-based Akamai, with a product-led approach and catering to under-served market segments.
The announcement in the blog is an innocuous-looking partnership with Nvidia, to place GPUs in the 200+ datacenters that it operates, to perform machine learning at the “edge”. It’s a logical evolution to providing more intelligent services further up the stack for Cloudflare; in their own words
Previously machine learning models were deployed on expensive centralized servers or using cloud services that limited them to “regions” around the world. Cloudflare and NVIDIA are putting machine learning at the edge, within milliseconds of the global online population, enabling high performance, low latency AI to be deployed by anyone.
Because the models themselves remain in Cloudflare’s data centers, developers can deploy custom models without putting them on end user devices where they might risk being stolen.
Why do I think this is interesting? It presents a compelling alternative for edge ML vis-a-vis the large cloud providers of the world (disclaimer: I work for one), and allow companies to run ML models near - but not on - customer devices. This provides a stronger level of data protection, while maintaining high performance. With companies now having to operate in many different geographical regulatory regimes, if Cloudflare can abstract such complexities away, it will provide a very compelling value add. More about Cloudflare here.
Twitter engineer blog: Sharing learnings about our image cropping algorithm
Twitter’s image cropping ML algorithm came under a lot of criticism last year, when it was seemingly cropping large images and choosing to highlight white and male participants in cropped pictures, over others.
In order to produce these cropped images, Twitter used a “saliency” algorithm, trained on human eye tracking data, and predicts where a person might favor looking at a larger image, and opt to focus on that for the crop. After outcry, the algorithm was dropped and an investigation announced.
Last week, Twitter announced the results of the investigation
Here’s what we found:
In comparisons of men and women, there was an 8% difference from demographic parity in favor of women.
In comparisons of black and white individuals, there was a 4% difference from demographic parity in favor of white individuals.
In comparisons of black and white women, there was a 7% difference from demographic parity in favor of white women.
In comparisons of black and white men, there was a 2% difference from demographic parity in favor of white men.
Twitter will now allow full length pictures, i.e. without crops, and allow users to choose where to crop. Their follow through, transparency and decisive actions on their commitment must be applauded.
Brian Lui: Upside Decay
A great essay that shows how a series of small virtues can compound and result in outsize success for the participant; in contrast, the absence of such virtues can also compound in the other direction - a phenomena that the author labels “upside decay”
Upside decay is hard to spot. It’s invisible if we’re not specifically looking for it, because the absence of rare positive events is unexceptional. Even when people finally notice something is wrong, they’ll attribute it to malign actors, imaginary enemies or conspiracy theories.
You’d think that being on the lookout for missing positive tail events would be enough to spot upside decay. But it’s too late to be useful. We need to observe an organization over a long period of time or have access to a generous chunk of its history. By the time we identify the presence of upside decay, it will already be far advanced.
What’s worse, upside decay has corrosive effects on organizations. Trying to connect with people becomes sluggish and hard. Things which were easy become longer, slower, more tedious. Problems that are left unattended will metastasize. Motivated people can still push projects through and deal with problems, but it takes all their time and energy and burns them out.
The author uses China as an example to make his case. Under Xi Jinping’s leadership and policy decisions, the country may be headed towards a “middle-income trap”, wherein forcing public participation instead of private enterprise stagnates the pie for everyone. Without strong ties to rely on for partnership and support, it antagonizes and loses trust slowly over time. Costco and Berkshire Hathaway are presented as counter examples.
By now it’s well-known business lore that Richard Montañez dreamed up Flamin’ Hot Cheetos when working as a janitor at Frito-Lay, and decided to pitch the CEO on his idea - resulting in a promotion up the ranks.
The LA Times pieced together that this was all a sham; the person who spearheaded the development of the product was Lynne Greenfeld, a product manager who has since retired. She discovered consumer preferences for spicy flavors in convenience stores in Detroit, Houston and Chicago - which led to the insight to develop the product line.
Montanez, meanwhile commands up to $50,000 in speaker fees, wrote several books, and is developing a biopic based on his life with Eva Longoria directing.
A lovely piece about physicist and Nobel laureate Richard Feynman and his wife Arline, who was dying of tuberculosis as Feynman was recruited to and led the Manhattan project. A series of letters, including a gut wrenching on written after Arline’s death are available for public consumption.
“I know you will assure me that I am foolish and that you want me to have full happiness and don't want to be in my way. I'll bet that you are surprised that I don't even have a girlfriend (except you, sweetheart) after two years. But you can't help it darling, nor can I—I don't understand it, for I have met many girls and very nice ones and I don't want to remain alone—but in two or three meetings they all seem ashes. You only are left to me. You are real.
My darling wife, I do adore you. I love my wife. My wife is dead.
P.S. Please excuse my not mailing this—but I don't know your new address."
Helsinki Times: Researchers develop artificial intelligence that can detect sarcasm in social media Sure Jan
Meanwhile, per Slashgear: AI turns hit Spice Girls song into a gritty Nine Inch Nails cover Using OpenAI’s Jukebox model
New Yorker: The Challenges of Animal Translation A long piece about the use of AI to understand and translate the language of Dolphins and Octopuses
Arstechnica: Archaeologists train a neural network to sort pottery fragments for them Using CNNs (convolutional neural networks)
the program was about 96 percent accurate. That’s not an improvement over human accuracy, but it could offer a more efficient way to deal with the sheer number of potsherds some sites offer up.
“This will free up time and effort for archaeologists to concentrate on the meaning of the results,” wrote Pawlowicz and Downum.
WSJ: The Safe, High-Return Trade Hiding in Plain Sight YOLO some I bonds at 3.54%
FrameSynthesis: A fun Google maps based 3D driving simulator
NYTimes: A Culture of Fear at the Firm That Manages Bill Gates’s Fortune The terrible stories about Bill Gates since the divorce announcement continue; there seems to be a pattern of involvement with unsavory and criminal characters.
WSJ: What Honest Abe Learned From Geometry Lincoln studied Euclid’s geometry proofs to learn how to argue in court - and later in politics - with honesty and integrity.
What made Lincoln special was integrity, his belief that you should not say something unless you have demonstrated that it is right. Whitney writes: “It was morally impossible for Lincoln to argue dishonestly; he could no more do it than he could steal; it was the same thing to him in essence, to despoil a man of his property by larceny, or by illogical or flagitious reasoning.”
In Euclid, Lincoln found a language in which it’s very hard to dissimulate, cheat or dodge the question. Geometry is a form of honesty.
The ultimate reason for young people to learn how to write a proof is that the world is full of bad logic, and we need to know the difference. Geometry teaches us to be skeptical when someone says they’re “just being logical.” If they are talking about an economic policy, or a cultural figure whose behavior they deplore, or a concession they want you to make in a relationship, rather than a congruence of triangles, they aren’t just being logical. They want you to mistake an assertively expressed chain of opinions for proof of a theorem.
Disclaimer: The views and opinions expressed in this post are my own and do not represent my employer.