Thoughts on AI

I was watching the Blu-ray extras included with the release of Oppenheimer, one of which was a Meet the Press panel. It included the film’s director, one of the authors of the original biography, and a few physicist consultants on the film included a present-day director of Los Alamos. At one moment, the moderator was discussing the relevance of Robert J. Oppenheimer’s story today, and brought up the emerging technology of LLM chatbots, demonstrated with ChatGPT, or more broadly (and a perfect buzzword), AI. This was back in July 2023, when ChatGPT was based on the now-primal GPT-3, couldn’t process images, video, or audio, couldn’t generate images or video, or crawl the web to know anything from after its 2021 knowledge cutoff. Even though the rest of the panel wasn’t all that enlightening, this moment really made it sink in for me that we all lived through one of those rare-ish moments in history where you couldn’t see where a new technology would go and how it would affect society. I think the moderator knew, at least more than I did, the upheaval the world was about to face (and ~2.5 years later, is still facing).

I graduated with my Bachelor’s in computer engineering in 2025. With ChatGPT’s release in late 2022, everyone in my grade started college by doing all work the “normal” way for the first year and a half. As ChatGPT.com initially spread, it was still just a tool that could be to generate a mostly-accurate function or paragraph as a starting off point. I still remember the first time I saw someone on their laptop in public, using ChatGPT to complete an assignment, plain as day. I remember thinking to myself that if they just submitted the output of the LLM directly, it would be acceptable but uninspired and inaccurate at times. Good enough if it’s some gen-ed class that they didn’t really care about. But as developers started releasing more advanced models, its usage evolved along with it.

Soon, it was a lot more than “good enough” compared to the pre-LLM days. With GPT 4/4o, it could process images and perform web searches. Competitors started releasing their own models, some of which were highly specialized for certain tasks such as writing, coding or even deep-dive research into academic papers. In terms of software development, a whole ecosystem of AI-coding assistants such as Github Copilot began to spring up. Some of could integrate with “thinking” models and work on a large problem or project continuously until it believed it had been completed (“agents”). The term “vibe coding” was coined in February 2025. And with every new model release, its general knowledge and competence in completing specific tasks increased as well.

It’s undeniable that AI is here to stay; too many people are reliant on it for so many reasons that nothing short of a solar flare that fries all electronics on Earth will keep people from getting it one way or another. Some people, such as certain tech subcommunities on Twitter, are AI fanatics who believe that LLMs usefulness will grow exponentially, that we will develop AGI, and then ASI, which will solve all physical and metaphysical mysteries of the universe and transfigure all of us meat creatures into higher states of being. Others are more pessimistic and think that while models will continue to improve objectively at benchmarks and in subjective experience, we’ve gotten to a significant portion of what the intelligence cap is with the current transformer and attention mechanism architecture. Others are vocally opposed to AI’s infiltration into life, whether because they are creatives whose livelihoods are threatened (understandable; I’ll address this later), or for less coherent reasons, like the strangely persistent myth that data centers “use up” a significant amount of water, never to be seen again. It’s one of those things that has embedded itself in the minds of a critical mass and that level-headed people simply have to deal with, like the belief that nuclear power is significantly worse for the environment than coal.

I’m someone who is entering into the engineering world and is still adopting to the existence of these tools, some of my initial thoughts from the GPT-3 era remain the same. I still think that AI tools are the best as another item in the toolbox for knowledge-gathering, similar to web search engines. Neither search engines or LLMs are perfect for this, the former for being able to point to an incorrect source, and the latter having a habit of confidently lying (hallucinating)¹. People also aren’t perfect either, and depending on the subject matter results can very wildly; LLMs and people have this in common. If the subject matter is favorable however, both LLMs and people can almost completely dependable. I can almost always trust an LLM to know how to implement a binary search algorithm in a programming language of my choice, because there are so many examples of that on the dataset it was trained on. Likewise, I can almost always trust a person who makes pizzas for a living exactly how hydrated doughs should be for certain desired textural effects in the crust, because of their experience in that field. On the other side, the inherent randomness used in LLMs can make results and helpfulness unpredictable run to run, particularly when asking for conciseness. With people, there’s a different issue; the same person might be consistent with their responses, but multiple people can have wildly varying responses based on their life experiences; an example that comes to mind is piano pedagogy.

In my view, it’s at its best as a highly-specifiable search tool for existing knowledge or explainer tool for an existing document. As a source of general knowledge that can accumulate ways of doing something from every corner of the internet and other written texts, it’s unbeatable.

Then, there’s the jobs question. For software developers, there’s talk of coding models replacing the types that created them in the first place. However, real-world analysis is less conclusive; not many lost software jobs are directly attributable to LLMs, and the recent shrinking of the job market is more due to familiar factors such as the end of the ZIRP era, lowered demand for junior staff, and a general hiring freeze exemplified by the drop-off in work-from-home jobs. It seems, for now, that the rush for enterprise users to “accelerate” productivity by inserting “AI” wherever possible, has come out a wash for now. New SOTA models perform better than before but are still susceptible to the same hallucinations as GPT-3 was. In my opinion at least, it’ll take at least a similar generational revolutions like the transformer, probably multiple, for models to be able to understand and adapt to the world as we can. Only then will they become viable human replacements in the workplace.

As for the arts, we’ve already seen AI-generated movie posters in actual, non-publicity-stunt use. There’s been a few small controversies on AI-generated voices already; two that come to mind are recreations of the late Anthony Bourdain’s voice for a documentary about him, and for enhancing the accuracy of Adrian Brody’s Hungarian speaking in The Brutalist (2024). The live performing arts are certainly the most resistant to AI takeover. Part of the enjoyment of live performance is the thrill that comes from seeing performers do metaphorical or literal tightrope walks; that excitement simply doesn’t exist when its a system repeating an action it did before for a pre-determined outcome. I, personally then, don’t feel that threatened by AI taking the jobs of musicians, at least for now. As for music writing, I’ve also heard the outputs of music-generating models trained to generate finished audio or note information for “classical” music (most notably NotaGen for the latter); the tendency of AI to make an “average” of all its input data, therefore generating something that doesn’t stand out at all, is readily apparent to me and probably anyone who doesn’t have a financial stake in the technology.

Maybe the most apt metaphor for AI as it currently stands is that it is a master of savants. A savant has a mind-boggling intuition for something, but only at a surface-level and algorithmic way. One savant could instantly tell you what day of the week every Christmas was for the past 1000 years, another one could solve every Wordle in 20 seconds, and another could guess the location of anywhere on Earth based on how the sky looks; a highly developed transformer model would have a solid shot of doing all of those. If an LLM had consciousness, it would rightfully see its capacity for encyclopedic knowledge and even general pattern recognition ability as far above any human. But could it write a Pulitzer-prize winning novel? Probably not.

1. You could argue LLMs are worse for this, as they often don’t back up their claims with sources unless explicitly asked to. Even then, they could be making the wrong conclusions from those sources or completely hallucinating them as often happened with early models. ↩