The year is 2022, I am learning a new javascript framework. GitHub copilot is helping out with ghost code completion. I had been using it for only a few weeks and already I don’t like it. It was good at completing the code but I felt that I was not as active in the process as I was before. I would periodically switch it off to make sure I could still function without it. I found myself typing function names and pausing waiting instinctively for the ghost code. Later I would see various twitter users complaining of the same issues that the GitHub copilot was making them in their words “dumber”.
We are fortunate to live through one of the most interesting eras in tech. AI still being actively researched and developed, new novel blockchains released every other week. CPU performance still loosely tracking Moore’s law. It’s amazing how incredible it seems whenever you put it in the context that, half the technology we have now would blow people’s minds in 2014.
LLMs fascinate me the most of the all these innovations. I am not alone for sure, you cannot scroll X without seeing at least 2-3 posts related to ChatGPT or some other LLM every day. I have spent quite a some of time developing software with these AI tools, I’d estimate ~500 hours probably more. The most shocking bit for me was how good Anthropic models are. I mean other than developers using Cursor IDE there’s hardly any regular person talking about Anthropic. In hindsight there’s isn’t much else that the model is exceptionally better than ChatGPT other than domain specific tasks. For most people ChatGPT is just enough I guess.
I recently built a flutter app mostly with LLMs, it’s a sleep debt tracking app kind of like the Rise Science app. Took about 4 hours to build the app, I took a minor detour fixing a broken flutter health connect plugin. With the right architecture such as lovable.dev’s GPTEngineer, but for flutter you could go from app concept to prototype in hours. The flutter compiler is extremely good, prints multiple errors at a time. Unlike most compilers that stop after encountering the first error. This makes iteration extremely fast. The LLM can make changes to maybe 7 files at a time, forget to update imports or maybe mismatch a closing ). With just one failed compilation the LLM is able to see these errors and fix them. This is just a terminal based LLM too, if you integrate the LLM with an editor you get access to the LSP for error detection, access to all extensions written to help developers write better code. You hardly ever need to incur the the wait times of a failed compilation.
I used to be an AGI sceptic. I still am to an extend but not as strongly. The LLM architecture is just not capable of actually reasoning, or so I thought. The models are getting so good that although it’s not reasoning, it mimics it so well that I can’t tell the difference. It’s not capable of novel thought, but to think of it how many times do we come up with a truly novel thought? Not as often as we would like to admit honestly.
LLMs might end up being the most impactful invention of the 21st century. For most people that just means the chat interface on ChatGPT website. But it could be so much more. Excited for the future.