Merchants of dichotomy

When a former colleague, who was almost always against what seemed like knee-jerk tooling suggestions and who constantly advocated for thoughtfulness and long-term planning, left to focus on an AI-centric developer tooling project, his reports were taken aback, saying that they thought he was anti-AI all along. That did not surprise me. When all the discourse you are exposed to about a novel technology you barely understand is people sorting themselves into binary camps, you will inevitably start seeing things through the same lens.

If you are anything like me, you must be thoroughly worn to a frazzle from the hopelessly religious discourse on both sides of the artificial intelligence aisle. The idea to write something on this subject came to mind back in April but I kept putting it off amidst other life concerns I had to deal with. More recently, Anthropic announced that Claude will be watermarking its output and some of the reactions I have been reading on the pro side of things have been hilarious but have also served as a reminder of how fast things are moving in this space and how my original thoughts risk being overtaken by the zeitgeist if I continue to sit on my hands so here we are.

I have been fascinated by artificial intelligence and machine learning in general for a little over a decade. Seeing various subsets like computer vision, speech recognition, recommendation systems, and anomaly detection become useful in real-world applications has had me excited for the future and the prospect of automating away everyday drudgery. Then came the launch of ChatGPT and language models that have been reinforced in post training to not just complete sentences, but to do so in a conversational manner.

Very quickly, I started pondering the potential downsides of what I have long held as an absurd proliferation of LLMs in pretty much everything today. For starters, it did not seem like an efficient use of resources. Running inference on LLMs uses up a lot of energy, and even though there have been efficiency gains in the past few years, they get outdone by the fact that the technology has increasingly used up more tokens to approach the same problem presented to it in order to achieve higher quality results. What this means in practice, like I articulated to someone in a threads post recently, is that we are building new infrastructure at a pace detrimental to decades of climate progress.

Environmental concerns aside, there is also the individual. One of my earliest fears was an ever-increasing gap between the haves and have-nots. Let me explain further. People short on means, curiosity, or the willingness to sit with a problem have an incentive to go all in on LLM usage for everything in a misguided bid to project the skillset or talent they think they should have. Those under less pressure of circumstance or temperament will, however, figure out how to augment themselves in a way that is not a total surrender, but an understanding of where LLMs are useful and where they are not, and all things being equal, their proficiency becomes a superpower of sorts. Unchecked, this hardens into Plato's cave, with the puppeteers running the show and the prisoners (i.e. the over-users) mistaking the shadows for their own competence.

Then there is what I have called The Great Norming, which is the convergence of expression where rampant LLM usage inevitably collapses the user's voice into a slurry of a billion other people, effectively rendering them invisible. A fuller treatise is underway and, if you are online enough, you have no doubt experienced an uncanny feeling from recognizing a certain register that has become pervasive in writing across multiple sources in diverse contexts.

I do believe it can be helped. LLMs are here to stay, or at least until the next breakthrough relegates them to the background. Technological progress eventually tends towards efficiency, but the incentives at play right now with LLMs push us to throw more resources at the problem in a way that negates short-term efficiency gains. This ultimately calls for better architectures that scale well and are in a different efficiency class than what we have today.

That said, LLMs today have proven to be useful in multiple consequential contexts, especially in language wrangling, where novel use cases emerge, and in self-verifiable endeavors like software development and mathematics. On language wrangling, I have found them useful as articulation oracles, where you ask the model to reflect a compressed thought you have articulated to confirm if you did a good job or if there are meanings and implications you may not have accounted for.

I would like to believe I have a fairly decent method for organizing my thoughts in my journal over the years, and I have longed for a reliable way to map them. Enter LLMs and vector databases. Using the aforementioned software development capabilities, I built a proof of concept with an ongoing index of my publicly available utterances across Twitter, Threads, and my blog. I can now run a semantic query on any subject and surface whatever takes I have had regarding it and how they have evolved over time. While writing this piece, I scanned my corpus for AI/LLM references and reveled in my prescience. Okay, navel-gazing time up!

There is a need for more involved (and informed) regulation. The folks at the helm in the United States government today do not seem to care much besides not wanting to cede a perceived lead in a race against China. Given how things have gone with the unmitigated RIFs across the industry and the mixed messaging from executives of the leading AI labs, I am curious as to how this all works out for everyone as I remain skeptical regarding inequality of access and the possible long-term impact detailed above.

For all I have covered here, there are still a lot of threads that I have not pulled, and intentionally so. It speaks to how the conversation regarding artificial intelligence is more nuanced than the everyman purports it to be.

The extreme folks on either side of the discourse are under-informed (or in some cases, ill-informed) to varying degrees. There are pros, yes, but also a slew of cons we are yet to fully grapple with and hedge against. There is a lot of capital interested in pushing full steam ahead to something that, while gratifying for some, will irrevocably reshape the systems we depend on.