Aditya

Blog, Notes, and Other Things

15 Jan 2023

OpenAI, Microsoft, Apple VR & Other Things

Hey there!

Welcome to the second edition of the newsletter, which contains news that caught my eye (mostly those that broke in the last week, but no hard constraints) along with an assortment of links and my brief notes to anything that I have found interesting.
Let’s get to it!

News πŸ”—

Movies and Documentaries πŸ”—

  • China’s Unmarried ‘Leftover’ Women
    • A VICE documentary covering one of the surprising side effects of China’s one-child policy. China has a relatively high human-sex ratio i.e, more males than females. Somewhat surprisingly this has also increased pressure on women to get married due to a lot of reasons including societal pressure and media. This New York Times Op-Doc focuses on a single woman and is a much more intimate look into the effects on the individuals.

Podcasts πŸ”—

  • Creators, Creativity, and Technology with Bob Iger
    • Bob Iger’s conversation with Chris Dixon and Sonal Chokshi. Mostly about his first term as Disney’s CEO.
  • Cyberbunker | Part 1, Part 2
    • Covering the story of Cyberbunker a bulletproof hosting service provider that used to host The Pirate Bay and one of WikiLeaks mirrors among other sites from its inception to its end.

Videos πŸ”—

  • HACKENBUSH: a window to a new world of math
    • A video that starts out with a fun game, brings in some basic math ideas to simplify and understand the game, introduces notation to keep up, and suddenly explodes to a lot of notations until you get lost and it finally closes off with a book recommendation. It’s a beautiful video accompanied by pleasant music where at times I felt like maybe the abstractions being built are a bit too stretched and janky, maybe because I don’t fully understand the scope of its usage. Nonetheless, the video tries to show how accommodating the notations and abstractions built are to various other games.
  • How to Build a Startup Without Funding by Pieter Levels
    • Pieter Levels who has built multiple companies (notably nomadlist, remoteok) talks about his process of building and launching stuff.
  • History of the entire Church, I guess…
    • Before watching this I didn’t even know Christianity had so many denominations. With my very limited understanding of religions, it looks like, unlike Hinduism where the groups get created primarily out of the abundance of gods and thus different worshippers of different gods (I’m simplifying obviously, there are other different reasons groups get created), in Christianity where there is primarily one sole entity the denominations seem to get created on the different interpretations of the saying of god or who/what god is. Of course, it’s not as simple as that, as the video gets to there are denominations that have formed during different times in history to align the concept and make it more amiable to the time.
  • Scientists whose lives were total dumpster fires
    • Among others, it covers the life story of Guillaume Le Gentil, an astronomer who wanted to observe the transit of Venus to measure the distance from the Earth to the Sun but to put it lightly the circumstances didn’t allow for it.
  • A very interesting article on the reinsurance industry. It provides an introduction to what it does and more interestingly taking a specific example of Berkshire Hathaway how it does. One of the highlights for me was how not over-specifying risk is an important thing when dealing with unknowns. Although Ajit Jain (Berkshire’s key underwriter) talks about how they start out with as much science as possible, post a certain point over-specifying an unknown is an act in self-delusion. The general tendency when over-specifying is in some sense similar to removing the slack out of a system, the more granular you go the more you think you’ve covered the space and the more confidence you get in removing broad-level estimates.

  • Research and work on defensive tactics against AI have come back into the spotlight, in particular people have been worried about the problem of how to identify AI-generated content. There are efforts on various fronts on this problem such as Have I Been Trained which allows you to search through a public text-to-image dataset to see if your content has been used to train these large models. But one of the more interesting approaches is “radioactive data”. The idea being you embed some pattern into your content ideally non-noticeable to humans but which will have a recognizable effect on the trained models and this minor pattern can then be used as a sort of marker to identify if the output of a model has been trained on the tainted content. Scott Aaronson has similarly been working in OpenAI to do something similar for text generated via GPT, although I don’t know much detail about this, he suggests it’s a much more “wrapper” kind of functionality outside of the core model. To me this is distinct in that the approach is stemming more from model producer angle, rather than the “radioactive data” approach which is more stemming (at least theoretically ) from the content creator angle

    • Similar to the news article this week on ChatGPT being used to create malware, these tools are starting to set up new grounds for adversarial long-running cat-and-mouse games.
    • On a slightly unrelated note, an interesting change that seems to be happening in AI research compared to earlier is that in some of the research in the periphery of AI such as fitness evaluation of AI for specific tasks, political inclinations of models, seems to be looking at the core model as essentially a black box, and then using that model as a system under test to identify various characteristics of the model. It’s probably a reaction to both the complexity of the model as well as lack of direct access to these large and expensive models or maybe looking at the model as a black box is probably the right abstraction for these problems. If it is the latter then this might have been the abstraction level for a lot of older research as well, but I personally haven’t been too deep in the space to know about these research.
  • An article from MIT Technology Review on Bitcoin mining in Kazakhstan. I would have never guessed that Kazakhstan was the second-largest country in terms of hash rate in 2021. But as the article mentions it’s not really surprising

    Howson compares bitcoin mining to the disclosing tablets that dentists used to give to schoolchildren in the UK, which dye areas of tooth decay in bright colors. β€œI think that’s what Bitcoin does,” he says. β€œIt swishes around the world and it highlights areas where there are geopolitical tensions going on, and there is poverty and corruption."

  • If you like me have largely ignored (either intentionally or unintentionally) the whole ongoings of “Twitter files” then this New Yorker commentary briefly sums up the entire fiasco. And just as you might have guessed, it’s a whole bunch of obvious stuff a private company would have done including debates on what to under-expose in its feed and being a media company having ties with the government to do preferential propaganda.

  • This week, I also learned about the concept of Search, Experience, and Credence Goods. The explosion of experience goods i.e, goods and services which can only be evaluated post experiencing it, also explains the existence of reviewers as a job to be done and why they as a group became more important with the larger experience economy expanding.

  • NYTimes covers the trend of engineering organizations investing in what is essentially engineer-focused content marketing. The practice of blogging about tech challenges that you faced while building has existed for quite some time now, but the fervor with which it is approached now is at least to me relatively new. Startups and large companies alike explicitly seem to focus on this as a metric and some companies go even as far as hiring ex-journalists to create some of this content.

  • Henley & Partners released 2023 editions of their Henley Passport Index

That’s all for now, feel free to share this with anyone who might find it interesting. You can reach out to me on Twitter if you want to discuss anything and you can always subscribe via email to this newsletter or follow using either site-wide RSS feed or Newsletter only RSS feed.