Blog AF

Starting a company had been a pretty clear goal in my mind for years. Continue reading...

I’ll try to start posting links on a semi-regular basis, and make a habit of it again. Here’s a few bits I’ve read recently. Too many of them are about LLMs and Generative AI. Please help me change that — send good long-form my way.

• Software Systems in a World of LLMs
  Davis Treybig
  

  One of the most thoughtful pieces I’ve read about looking forward and the changes coming from AI. Treybig doesn’t focus on societal changes, but changes to the production of software systems themselves. Programming languages, development frameworks, and libraries are tools built to make it easier for humans to read and write code. Soon we’ll start seeing new ideas popping up here to make it easier for LLMs to read and write code, too.

• “Succession” and the end of the illusion of ethical capitalism
  Branko Milanovic
  

  On the show Succession being a much deeper critique of capitalism than you thought.

• AI and the American Smile
  Jenka Gurfinkel
  

  An interesting observation of cultural differences, looking at AI generated selfies to explain how nuance and slight variation get papered over by ML models. Training datasets make a huge difference in the output we see. A different flavor of the world globalizing and flattening, and the loss from that process.

• What literature do we study from the 1990s?
  Matt Daniels - The Pudding
  

  An analysis of university curricula and Goodreads reviews to understand the infleunce of books written in the 90s, and how we perceive them today.

• The Perils of Prudence
  Abraham Thomas
  

  Paradoxically, going faster can sometimes be the less risky move. This is true in startup-land, but also in Antarctic exploration.

• Multi-layered calendars
  Julian Lehr
  

  Exploring different UX ideas for how to improve calendars. I hate both Google and Apple calendar, and would love to try something like this out.

• AI and the automation of work
  Benedict Evans
  

  Automation, and thinking of LLMs as an army of interns who can perform most tasks reasonably with some supervision. Evans drives us towards Jevons’ Paradox, pointing that the falling cost of getting some things done with LLMs actually increases demand for those things, increasing, rather than reducing our use of them. From coal, to oil, Xerox machines, and typewriters… why is this time different?

• A Petabyte of Health Insurance Prices Per Month
  Adam Geitgey - Turquoise Health
  

  On data engineering, redundancy, and “looking for hundreds of needles in a haystack approximately the size of Mars.” Reminds me of my Apple days, and the challenges of working in insanely large scale systems. Turquoise is working on some cool stuff.

• Remote work blew $484M hole in SF tax haul, report finds
  Joe Eskenazi - Mission Local
  

  A few crazy stats about WFH and the evolution of post-pandemic SF. GDP keeps going up, but the concentration of large companies driving tax revenues puts the city at risk – the largest 100 companies (.7% of the total) account for nearly 60% of the business tax collected. Interesting to see SF politicians realize that Laffer curves are not idiotic.

• "Low-Resource" Text Classification: A Parameter-Free Classification Method with Compressors
  Zhiying Jiang, Matthew Yang, Mikhail Tsirlin, Raphael Tang, Yiqin Dai, and Jimmy Lin - Findings of the Association for Computational Linguistics
  

  A 14 line python script with 0 parameters that can outperform a bunch of the deep transformer-based LLM models everyone is so crazy about.

• Excuse me, is there a problem?
  Jason Cohen - A Smart Bear
  

  Most companies fail because they don’t solve actual problems. This is a good problem sizing framework, based on his experience building WP Engine.

• Hashing
  Sam Rose
  

  A visualization of hashing functions, what they are and what makes a good one. This would have been a great resource back when I was taking intro to programming classes.

• Building LLM applications for production
  Chip Huyen
  

  Most LLM applications I’ve seen so far are still just demoware, things that look cool and shiny but don’t really work at scale. This piece is a few months old, and good chunks of it already feel outdated.

• The Problem With LangChain
  Max Woolf
  

  A post pointing out problems with one of the most popular frameworks in AI-land these days, and the perils of picking the wrong abstraction.

• Changing my mind on remote, moving the team back to San Francisco
  Flo Crivello
  

  I’m bullish distributed-teams long term as companies grow, but early-on technical teams working on startups should be in one place, and that place should be San Francisco.

• This is the hometown of San Francisco’s drug dealers
  Megan Cassidy and Gabrielle Lurie - San Francisco Chronicle
  

  Part one of a very in depth report on the connection between Honduran migrants and San Francisco’s drug crisis. A strange, sad story about international trade, drug markets, and immigration.

It’s been a while since my last post. Here are photos from lots of travel, with a mix of film and digital from the first half of the year (plus a bonus last week of ‘22).

We start out in Chicago, go to San Francisco, then Yosemite, Austin, back to Chicago, then the North Bay, more SF, Phoenix, and more SF. I feel lucky I get to travel this much. More...

Two years ago, excited to begin a new journey, I called into an onboarding Zoom from the makeshift desk I set up in my kitchen at the start of the pandemic. Continue reading...

As Wolfram implies, building “one model to rule them all” is a stupid goal Continue reading...