It's always good to periodically reevaluate your priors. Here's my current evaluation for AI. My guiding technical compass for the tech has always been that it is very granular, very expensive search. It's expensive because it's granular, and also because it's statistical. The granular part is obvious. We went from indexing website metadata relationships to the relationships of virtually all data quanta (tokens) accessible by internet. It started with words, and is bleeding over to things like 3D models, games, pictures, videos, audio, etc. If its digital, it has been chopped up into little particles and indexed. The search tools don't simply know a good website to find information anymore, it cut out the middle men and just has the information. I don't need to find the crystallographic data files attached to a paper from the 90s for a specific material. The model already found it and indexed it. Ironically, the one thing these models are still terrible at is actually telling you where it found all the information it gives you! Coding needs special mention, because it seems to be the biggest application getting economic attention now. In my own experience, and listening to the AI psychosis of others on social media (yes I get it too), the models are now good enough to autocomplete a lot of mundane coding tasks. Probably old news for some, but I needed to experience it and see it for myself. It feels like a graphical website editor like Wordpress or Dreamweaver or any other modern variants on steroids. Even a model like Gemini thinking or pro, that no one really talks about, is good for dealing with a few thousand lines of code at a time. That being said, it's far from a one-shot process. It still takes a lot of time and energy to dial things in. What comes out is much like the default template in a graphical design tool: looks like it could be useful but mostly just pablum slop. Sure, I can throw in some new functionality very quickly, but efficiency, usability, polish, etc is still a challenge that takes time. I'm building google app script workflows for folks at work, and I still ask them for a week or more for any nontrivial changes. I want to address the obvious counterargument to this last point. The AI booster will say "okay but you are just six months behind. Soon you will just ask the agents to make it and it will just happen." I do not think that is true today for production code (or at least production code that will survive long term), and do not think it will ever be true. The primary reason it's not true is liability. If it doesn't work, everyone will blame me, not the AI. I incur all of the risk, so I MUST own verification. Unit tests are good practice, but not a solution, as they can be gamed (like all of the benchmarks that the models game). Nevertheless, I'm building things relatively quickly that I otherwise would not be able to that I hope will streamline some processes to improve outcomes and hopefully save me some time in the future. I am also using it to help me develop hypotheses and build and evaluate initial sanity checks to see if they merit future effort. Like in the past, most ideas I have turn out to be either not new, or not valid. This is true of most research, and a lot of time is spent simply searching for a good hypothesis. In pursuit of finding good hypotheses, the very granular, very good search has enabled me to learn about the existence of and learn how to use a lot of physics simulation packages floating around on the internet. I can very quickly build up basic competency in using these tools to do initial idea testing. Currently, it has not made me more productive. I have incurred a time and energy debt putting some of these tools together, and if they do not work out, will result in lower productivity for me. Again, more risk I am incurring today for the hope of payoff later. I think this echoes the feeling folks get of AI burnout, where the barrier is now lower to jump into new areas, but the barrier for eventual payoff is just the same. To me it feels like we are doing to knowledge work what capitalism did to software development: as hardware became cheaper and faster, instead of pushing those benefits into products, we used it as an opportunity to create shittier products. We don't have to make sure a rocket can land on the moon powered by the equivalent of a Gameboy if 1TB of ram costs $1000 and fits in the palm of your hand. But you could. Then there is the whole issue of cost. Nobody knows what these models will cost to operate at steady state, once the massive subsidies dry up. Remember $8 Uber rides from the airport? Those are $100 now. And the competition is all dead. Meta is supposedly giving the equivalent of $100 a month in compute to each Muse agent people spin up. That's a free iPhone for every user per year. Insane amount of subsidy. You better believe they are silently adding that to the customer tab with the intent to go after it later. I still think the most likely place this all ends up is Google winning the game of chicken and maintaining their monopoly on search. Everyone else goes to zero. And what is left is a Google search that now costs Google 10-100x the resources to operate. Will their customers and their consumers pony up 10-100x the resources for what is admittedly much better search? And then long term, this still is not a sustainable technology. By absorbing all of the information accessible on the internet, they have cut out all content creators. Remember when I said it was ironic that the models are bad at citing sources? I think it's by design, to avoid lawsuits for theft. But even if they avoid lawsuits, they will kill the golden goose. People will simply choose to no longer publish things online. The internet as public space for information will die. This can be avoided by compensating content creators directly, but this is a one two punch for Google, who is already incurring massive costs just serving the data it stole. So Google will likely come out of this with drastically reduced profit margins (they are already seeing positive free cash outflows!), but an expanded monopoly that encompasses not just search, but also publishing. This will create new political and legal issues in addition to the stealing, since most of these massive tech giants argue they are not liable for what they deliver because they are not publishers. At the end of the day, who is liable for what these massive databases eat up and burp out? So at the end of the day, although a lot has happened in the last year or so, my priors still haven't really changed all that much. We are all currently enjoying our $8 Uber airport rides in the form of massively subsidized compute, knowing it's just a matter of time before we all get fucked.