What's on your mind lately?
Google Deepmind's Logan Kilpatrick recently said, "AGI is going to be achieved by a product, not necessarily a 'model'." I liked the quote, not necessarily because I have any clue about how AGI is going to be achieved, but because I think the whole industry needs to have better products. There have been so many model and feature releases over the past year and a half, but the single most useful product that shipped from any of the big players was NotebookLM, whose killer feature isn't the automated podcast, but the humble reference links to the supporting paragraph from the original document. It's the one feature that allows anyone using AI to do document research to mitigate the tendency of LLMs to get things wrong, even in RAG applications, it's a wonder it hasn't been universally adopted. Then again, we did put astronauts on the moon before we put wheels on luggage.
How can we better understand the current AI hype?
I really enjoy tech analyst Benedict Evans' take on things because he's always so even-keeled. He tracks the potential of generative AI and documents the way it has turned the whole tech industry upside down, even as he highlights how, paradoxically, we still haven't found great product-market fit for a technology that supposedly has the potential to change the world. A recent favorite from his blog last January that I've been thinking about since: "Are better models better?"
What's one fact about AI that everyone should know?
Most of the time, AI solutions are really workflow solutions. I was working with a regional newsroom in the Philippines for an AI accelerator program, and after talking them through how AI could help their editors and reporters turn around stories faster, they realized that for it to work, they needed to overhaul their internal systems. And so the whole project went from just figuring out the right prompts into changing their whole editorial and production process, from submission to publishing. It's also why newsrooms that have well-developed production systems and data pipelines have tended to do well in AI adoption.
What's a good hobby to pick up?
Cooking! You can be just as much of a nerd at it as other hobbies, especially when you get into kitchen tools: Dutch ovens, santoku knives, and sous vide cookers. The other day I caught myself thinking about how hard it was to implement AI applications in journalism without having good clean data, and how it was a lot like how good knife skills and mise en place really go a long way toward good cooking, before I realized how much of a nerd I was being.
