Journal Journal: AI Time Capsule
With the rate at which things are moving, I feel like making any statement on AI, opinion or fact based, is going to be outdated in a month. Given that this journal is old enough to buy a drink, I think it'll be fun to look back in another 20 years and see what was going on (assuming we don't get destroyed by T-1000s) while my unit tests are running.
Here's a few blanket statements:
- JESUS CHRIST AI IS FUCKING SLOW. At this point, I do not care if you release a newer more powerful model. Make it execute sub-second. In the beginning, I think we were all just blown away at how cool it was to get a human-like response back from a chat bot. Now though, we're trying to get productivity out of these, and when it takes an hour to churn out a new feature, I am no longer blown away. Don't get me wrong, this is amazing on all accounts. Such a feature would have taken me hours, days, even weeks or months to finish. Now I get them done in less than an hour. But it's time to improve. The AI sits there and thinks for fucking ever, and I have to babysit it because Cursor doesn't yet allow me to auto-approve (company controlled - though this is improving and I should be able to add my own auto-approve commands tomorrow). I also need to make sure it's on the right track because it can really go off the rails. So as I watch it's thought output, I'm scanning to make sure it's doing the right thing. If I see hints that it's going off the rails, I have to intervene. Sometimes I don't care and say fuck it, and let it take a stab at the problem, then review after. At least one time this resulted in having to blow away all the changes and completely rework it. But this comes down to having good documentation and Skills, and laying out good expectations in the prompt.
- We seem to be living in two parallel worlds. I went to the eMerge Americas conference in Miami. This was one of the most depressing things I have ever seen. I used to love conferences. In 2010, work sent me to Adobe MAX and I had the time of my life. Not only did I learn super cool tips and tricks and find out about upcoming features to ColdFusion, the product I supported at the time, but I got drunk for free, I flirted with some super hot conference attendees, I got a free Android and Google TV, and I'm pretty sure I even saw Weezer play. Fucking wild. In 2016 I went to Collision Conf, the conference that birthed Uber. It wasn't as good as Adobe MAX but still fun and I got drunk for free and saw a little bit of cool tech and even saw Mick Foley talk. But eMerge Americas... We've reached peak technocrat fascism. Everything surveillance, everything government contracting, everything AI. It was so transparently disgusting. From drones designed to suck from the government teat, to Israel dick rider and Federalist Society/MAGA judge Roy Altman on stage with former disgraced mayor Francis Suarez. And what tech issues were they tackling? Israel. Specifically, defending the genocidal apartheid nation and promoting Altman's book, Israel on Trial. Insane. Oh, and Erik Trump closed one of the nights. JFC. So this parallel world thinks AI is everything. They think it is the godsend savior and the solution to all of their problems - financial and otherwise. Then there's the other parallel world, where the rest of us are over it. Engineers know this shit can't do our job, but it is sure sucking the life out of us on the way to being laid off by management who do think it could do our jobs. Data centers are sucking the water out of lakes and rivers, and electricity out of the power grid, while we have to pay the difference. These two worlds cannot live in contention forever. Something's going to give.
- I don't think we can ever go back. As awful as AI is, I don't think it's going away. The genie is out of the bottle. I mean, hell, I can run a quantized LLM on my Macbook. It's slow and dumb, but it does work. So I don't think it's ever going to be outlawed or anything. As for the big companies, OpenAI, Google, Anthropic, Mistral, Meta... Maybe we can reign them in, but we'll need a completely new political system to do it.
- I have a love-hate relationship with it. On the one hand, I'm able to ship new features faster than ever. On the other, it hasn't really made me any less stressed or anything. I'm still at the keyboard for 8-10 hours a day, most days a week. Someone still has to guide these things and check the work and deploy the code and test and coordinate with other teams, and go to meetings.
- The models aren't getting that much better. From Opus 4.6 to 4.7 to 4.8, I'm seeing marginal improvements. They are very good and GPT 5.5 is good. But I feel like we're at that point in model power that CPUs were at in the mid to late 2000s, where for the most part, the more powerful the processor, the casual users barely noticed. Memory became more of the bottleneck. In LLMs, it seems price and speed is becoming the bottleneck.
That's about it. Love it or hate it, I think we're stuck with it. I just hope it gets faster, cheaper, and less compute intensive. And doesn't take our jobs.