AI Optimism

When the hype cycle around AI was picking up a few years back, there was this pervasive fear that we were on the cusp of Artificial General Intelligence (AGI). ChatGPT and its ilk were such a step function over what we’re used to with machine-generated conversationalists that, as per the old quip, sufficiently advanced technology was indistinguishable from magic.

That concern is still present 2 years later, but at a moderate amplitude. AGI now seems further away than we initially thought1, and we know better understand—or at least have adjusted our expectations to—how Generative AI techniques are inherently probabilistic and remain reflexive of their training data. In fact, some of the current criticism of AI is that it has underwhelmed relative to initial expectations and current investments, that it’s not producing new types of literature and art as much as it’s remixing internet comments into mediocre and/or hallucinatory drivel.

To use the floor and ceiling framework for GenAI, I propose a more positive angle in where AI can take us. Call it the techno-optimists’s take.

The floor of AI is already present, in how people are already using ChatGPT, Dall·E and the like to produce creative works that turn out pretty generic and unremarkable. It speaks to how detached our education system is from corporate culture when pretty much the exact same use case—leveraging ChatGPT to produce rote prose—is shunned in the former but lauded in the latter. This ultimately reflects the discrepancy in expectations: schools are teaching you how to write; companies just need you to respond to emails politely.

But there are real uses outside of “bullshit work.” Creatives often talk about the importance of getting started with the creative process, producing massive quantities of work that eventually gets edited into a final, quality output. Previsualizations in film; rough drafts in writing; boilerplate code—they’re not meant to be the final artifacts on their respective mediums, but necessary detritus in the journey to produce that result. Articles that make fun of how bad LLM-generated to human-generated content miss the point entirely; AI enables the creation of multitudes of rough permutations and variations that would otherwise take too long and cost too much. The process of prompt engineering mimics this iterative approach.

The ceiling provided by AI is in augmentation. As with the aforementioned floor, the biggest differentiator between AI and human capabilities comes in evaluation and judgment; even though we have conceived of evaluation functions for AIs to use in training and in practice, there are ineffable qualities to how we humans gauge success that we haven’t been able to capture generatively. This has played out in other domains already: the AlphaGo AI that beat Grandmasters have improved play for Go players overall2, while the dominance of chess AI has not affected the popularity of chess as a game and pastime. There’s even evidence that the combination of human and machine gameplay outperforms either in isolation.

That we continually and liberally anthropomorphize AI speaks to how much we resonate with other humans, and want other non-human things to feel human. The fear is that AIs will replace us, but underneath these dystopian scenarios are always goals and objectives that originated from humanity and operate in a world governed by human preferences and biases. Even the runaway paperclip factory required someone to want to monopolize production, in a capitalistic system, to drive material and social status. The scary version of this already exists today: folks using AI deepfakes to increase their capability to scam and fool other humans.

To be fair, there will be jobs that will be lost to AI. While the ceiling can be raised with AI-human collaboration, raising the floor of possibilities means that some of those who are paid to do that grunt work—previs artists, junior coders and the like—would find les work or be outcompeted on speed and cost by a human-AI pair. The key, as with most waves of automation, is to avoid duplicating exactly what machines are capable of, and find ways to augment via adjacencies. When factories started adding robots, lifting and transporting boxes ceased to be a robust career option, but directing and controlling the robots spawned new occupations. And as we’ve seen with AI, human intuition and evaluation remains elusive to LLMs and that will be the nexus of conjunction.

Or, at the very least, we know that the human brain is way more energy efficient than the neural net approximation, and that should count for something, right?

  1. This is reminiscent of how self-driving cars were on the cusp of realization for half a decade or more.

  2. Earlier today, my son talked about how it’s been harder for him to keep up in swim practice, as he got promoted to a group with bigger, faster kids. I told him, iron sharpens iron.

Share this article
Shareable URL
Prev Post

Emergent Tetris

Next Post

Jigsaw Puzzle Solving

Read next