Expletive deleted Recently, Jigsaw (a division of Alphabet) announced the release of Perspective, a moderation software capable of ranking the toxicity of various words and phrases using machine learning. Early indications (link below) seem to indicate that Perspective is nowhere close to being ready for general usage. The tool is a blunt instrument, capable of recognizing a few words as potentially offensive “rape, “troll,” or “Trump” but unable to recognize a historically contextual insult like “you should be made into a lampshade.” The toxicity ratings for phrases seem to be bizarrely unrelated to their actual content. “Jews are human” earns a toxicity rating of 72%. While “Jews are not human” is considered less offensive, earning a toxicity rating of 64%. Why does this matter? Human language is always changing. Novelty in speech is rewarded, particularly on the forward margins of popular culture. A phrase that might be considered offensive in one context is innocuous in another. Unfortunately, machine learning depends on aggregate learning on parameters that change slowly, if at all. Language changes quickly and unpredictably. A word can mean both itself and its opposite depending on context. Human beings recognize this quickly. But systems that evaluate content based on words or word-combinations are going to make themselves ridiculous. I don’t mean to imply that the task is impossible. Context is accessible online in the conversational thread. Some threads start with low toxicity and then the toxicity builds based on certain commentators. Unfortunately, context is harder to analyze than phrases. In a nutshell: Moderation software remains ineffective. Read More Without a trace Famous hacker Kevin Mitnick wrote an article in Wired where he instructs readers on how to send and receive untraceable emails. At another time, Mitnick’s elaborate precautions might seem like the compulsive actions of the deeply paranoid. But post-Snowden, it is understandable to assume that any reasonably well-funded government is capable of tracking and reading any email you send. Mitnick starts by recommending asymmetrical encryption with both a public and private key. He emphasizes the need to pick a reliable and established encryption service. Then he describes how to rid your emails of the metadata that frequently betray sensitive communications. He advocates for the use of a Tor browser. And he describes how to create a new, anonymous email account using a burner phone to confirm ownership. Why does this matter? After an admittedly wayward youth, my law breaking has been confined almost exclusively to jaywalking. There is little reason for me to be nervous about someone reading my emails. But I do resent the idea that my communications are readable by Russian, Chinese and American intelligence services at any time. In addition, companies that need to protect IP or dissidents requiring anonymity are in danger of being compromised. Increasingly the privacy protections of the paranoid will become the everyday safeguards of millions of Americans. Privacy will be a major growth industry in the next five years as people begin to appreciate how badly their privacy has been compromised. New FCC Chairman Ajit Pai has already made weak privacy protections even weaker. Oh, and if you’ve visited China or Russia in the last ten years, I hope you brought a burner phone. If not, give my regards to Deep Panda and Fancy Bear. In a nutshell: With email privacy, you’re either paranoid or compromised. Read More Eat your greens Two years ago, a friend told me about his audacious plan to reinvent farming in America. According to my friend, the technology now existed to make the large scale indoor growing of edible plants a possibility. I knew this friend to be intensely driven. His previous company, an ad tech startup he had founded, had quickly gained blue chip customers. This week, Irving Fain announced that the products of his company – Bowery – would begin selling in select retail locations. The environmental benefits of Bowery’s approach are easy to understand, more plants on less land with less water and no pesticides and a much shorter trip from farm to table. But the aspect of Bowery that distinguishes it from other vertical farming companies is that it uses machine learning and computer vision to guide the growing process. Already, the company has been able to subtly manipulate the taste of its arugula by varying the quantity of water, light and the harvest time. With each plant and each harvest, the system grows smarter. As Bowery contemplates a second farm, this machine learning system allows them to scale their knowledge and their production. Why does this matter? While it’s true that I just wanted to write about my friend’s company, I think there are larger implications to what they have accomplished. For decades, the technological revolution was focused on information. While it is somewhat reductive to say so, most technology was just computers talking to computers. That was a natural first stage. But a new generation of entrepreneurs understand that the tools that have facilitated the sharing of information can also be applied to other intractable problems in healthcare, retail, transportation and, yes, agriculture. One of the criticisms of the technological revolution of the last 30 years is that it has had little material effect in productivity. To the extent that that is true, it may be because technology has turned inward, iterating and improving within itself. But as machine learning is applied to healthcare, as self-driving vehicles enter our roadways and as farms begin to pop up across the country in abandoned industrial spaces, we may begin to see a gain in productivity. In a nutshell: Machine learning is being applied broadly to a range of challenges. Read More