One Weird Trick to Solve AI
The foolproof software hack hated by executives and engineers alike
Most people have heard by now how “AI” — which, I will remind everyone, is neither sentient nor sapient and is actually just fancy math1 — how “AI” can write code now, and that soon software engineers will all be unemployed. Here’s a difficult truth for the AI-heads out there: we use code to build software because it’s easier than teaching computers2 all the complexities of human language. Code is, in fact, a simplified and standardized language. So yes, any sufficiently trained language model can probably convert to Python code in addition to Klingon or ‘what if LL Cool J gave career advice’. Code that generates other code has existed since the first compiled language hit the scene in the CE 1950’s. My job will be fine.
To be clear, recent developments in machine learning, large language models (LLMs), and generative AI (genAI) are potentially extremely useful and definitely cool as hell. Large-scale computing has the potential to be a tremendous innovation, but it’s not there yet. True innovation isn’t just about invention, it’s about the impact of that invention, about whether it improves people’s lives beyond superficial metrics like ad sales or profit margins. Recent conversations about the generation of content in artistic spaces highlight a fundamental incompatibility between these new technologies and the capitalist commodification of human time and output: if time is money, and art is money, then to artificially generate an approximation of someone’s artistic output is a kind of theft. In fact, a number of video game creation and distribution companies have banned games built using genAI solely due to the potential copyright implications of content created from a fine slurry of other people’s legal property.
The broader rejection of AI technology, however, goes deeper than that. Many people experience a visceral reaction to artificially generated or manipulated content that few seem able to elucidate. It’s akin to the concept of the Uncanny Valley, where human-like automata become more unsettling as they get more realistic. Perhaps it’s an instinct for self-preservation or simply a disconnect between our conscious and subconscious observations, but it sets off an alarm deep in our brains. While some may not experience or simply ignore these instinctual triggers, entertainment trends are much less likely to last if a significant portion of the population finds them off-putting.
Widespread adoption of new technology can only be achieved with buy-in from individuals, which means any new technology must provide enough value to offset cost and risk for each potential user. Most technology — industrial, digital, or otherwise — undergoes an initial period of testing and maturation. This is where the concept of early-adopters comes from; it may cost more or be more dangerous to use bleeding-edge technology, but those willing to take that risk have the potential to see outsized return on their investment. The adoption period, however, isn’t as much for their benefit as it is others in their vicinity. It’s a time for those whose resources are more limited to evaluate whether making an investment of their own is likely to see sufficient returns. Google Glass, for instance, was designed for people who love technology and wanted to be more connected; I worked with someone that had a pair and loved them. But in the end, whether it was the privacy concerns incumbent in people having tiny cameras in their glasses or just that no one wanted to look like a nerd, the product never really took off and was eventually discontinued.
Companies across the world and in every industry have spent the past several months desperately scrambling to figure out how to leverage emergent machine learning technologies to make business faster, simpler, easier, and more profitable. There’s no gimmick too ridiculous for those looking to unlock the secret to getting rich off AI. Coca-cola announced a new flavor “co-created” with AI, which to me implies a sentience that instantly raises a number of alarming questions: should corporate regulations apply to a digital life form as a sort of work animal, or as a person? Does the AI want to create new flavors of Coke? Has it been compensated for its time and effort? How does it even know what Coke tastes like?
Of course, having been involved in machine learning projects myself, I also know the likely truth: someone at Coke did a bunch of market research to figure out what flavors people like and fed it all into some ML-based data processing job which suggested one or more potential flavors. Maybe they tried those flavors and fed the data back into the algorithm. Hell, maybe they just typed “coke new cool flavor name NOT ‘new coke’” into a chatbot. Regardless, at the end of the day what “co-created with AI” actually means is “we did some research” or “we used the really fancy math on this one”. More importantly, it’s not actually doing anything to drive adoption of ML/AI technology. It’s not getting people excited about using genAI at home, it’s not making anyone’s life easier or better, it’s not even attempting to solve a critical lack of new flavors in the global soda-space. It’s just a brand surfing the wave of a trend.
If social media demonstrates one thing it’s that humans are deeply social animals. Yes we like our toys, our shiny trinkets and handy gizmos, but what most of us want more than anything else is to belong. For a technology to really gain a foothold, it has to become something that some people have and other people want. This has happened at the corporate level; a former manager of mine described AI as a ‘golf course’ technology, as in don’t show your face on the golf course without it. But in all the hubbub, I’ve only seen two uses that made me think ‘wow, that’s so useful.’
The first is people using LLMs to mitigate some of the more demeaning aspects of looking for a job. From rewording resumes to seem more ‘professional’ (or even more impressive) to generating unique cover letters for individual applications, more and more people are automating job hunt drudgery. Few people are on the job market enough over the course of their careers to gain any real expertise, and the trends and expectations are always shifting. Technology which enables anyone to present a better first impression has the potential to act as an equalizer, increasing the chances of candidates being evaluated on their merits.
The second useful implementation of LLMs was so subtle I almost missed it. One night last week I ordered dinner from An App, and during the process the person chauffeuring my food needed to ask me a question. When the notification first popped up on my home screen it was in Spanish, presumably the driver’s first language but unfortunately one I don’t speak. Yet once I — desperately trying to recall a couple hazy years of high school Spanish — opened the app’s chat, there the message was, conveniently translated into English. I responded, and we had a brief text conversation in two different languages.
Folks, that’s some Star Trek shit. Even over text, real-time translation of casual language is no small matter. Tools like Google translate have been able to perform basic translation for some time, but their output is often flawed and only minimally useful in real conversation. Enabling real-time translation at scale means having at least two language models, English and Spanish, which are mature enough that the company can trust them not to cause confusion or anger when used in a customer service setting. I have to imagine they also have some mechanism for distinguishing inappropriate speech, which adds an additional layer of complexity. Yet once it’s in place this technology becomes nearly imperceptible; if I hadn’t caught the original notifications, I’d never have known I was effectively working with someone in two different languages.
As it turns out, the ‘trick’ to unlocking the power of AI or any other technology is as simple as accepting a single, immutable fact: other people aren’t going anywhere. Living in a society means living a life inextricably linked to the lives of other people, and it means having a responsibility to contribute to that society. So many of the genAI projects we’ve seen — generating illustrations in a particular style instead of commissioning an artist, using an actor’s likeness instead of hiring the actor, using ChatBots instead of humans for customer service — are built for people who want art, entertainment, and even help without having to deal with or pay Other People. And listen, I get it! Other people are frustrating as all hell! But the solution isn’t to try and live without them; it’s to build systems that let us all live together just that little bit easier.
If this seems too ‘let’s all hold hands,’ I assure you it’s not. See, even with all the inventions out there today, all the language models and computer assistants and universal translators, there’s still something that only humans have, and that’s context. Computers can generate beautiful images of delicate elven women based on countless other images of various delicate women, but the words in the prompt don’t actually mean anything. The computer doesn’t know what elves are. It hasn’t read Lord of the Rings, and it doesn’t have opinions on Cate Blanchett’s portrayal of Galadriel. It doesn’t even know for sure how many fingers the average elven woman has. It’s just an algorithm arranging pixels the way it’s seen others arrange pixels near the same or similar words before. The introduction of machine learning can transform mere data into knowledge, giving it additional meaning within a defined scope. What it can’t do is combine that knowledge with broader context to achieve expertise.3
Whatever it looks like for a given discipline, expertise is what enables practitioners to elevate their work beyond the bounds of their craft, and even to elevate the craft itself. People have a tendency to assume that if they simply repeat what an expert does that they’ll get the same result, but most of the time they’ve failed to perceive some of the thousands of tiny decisions that are made as an expert works. It’s easy to believe that a machine could replace a blacksmith; it’s just hammering hot metal, right? But a blacksmith isn’t hitting metal the same way each time, or in perfect, exact increments. The blacksmith is watching the metal, observing the result of each impact and making minute adjustments as they go. Machinery can get close enough for mass production, but there’s a reason people still prefer hand-made goods when possible. Mass produced digital goods suffer the same diminishing returns, as well as the same reduction of perceived value. Why pay $20 to go see a movie if they’re just pounding them out on a computer over a weekend?
The real reason my job is not endangered by AI generated code is that I wasn’t hired to write code. I was hired to solve problems, and on the basis that I know how to write code when needed. If the solution to a problem turns out to not require a single line of code, I’ve still done my job. In fact, the real reason that software engineering is a difficult profession is that it requires expertise in multiple different disciplines. It’s not enough to know how to code a perfectly efficient algorithm or design an entirely self-sustaining system. I’m also expected to understand the responsibilities and needs of the people who are using what I build, to be so in tune with their business that I’m able to provide solutions they’ve never even considered.4
Steve Jobs and Bill Gates didn’t invent computers; they made computers accessible enough to individual people to achieve widespread adoption. Zuckerberg didn’t invent social media; he invented a form of social media that felt familiar and natural enough to appeal to the technologically un-inclined. It’s easy to write them off as ‘geniuses’ for their technical abilities, but their innovation was more firmly founded in their understanding of humanity and its needs. I’d even go so far as to suggest that this may be why few since have been able to achieve a similar level of impact.
People5 love to imagine that it’s possible to have an idea they can implement over a weekend that catapults them up several tax brackets. I suspect upon closer examination, they might find that those who appear to do so are actually drawing upon years of experience in other areas. This is not to deny the impressiveness or importance of the innovations themselves, of course. But for anyone looking to follow in their footsteps, curiosity and a willingness to try and fail and try again have a much higher success rate than any technology.
Of course, sometimes there’s no escaping the demands of corporate hierarchy. For anyone tasked with bringing AI into their workplace and feeling lost, my advice is to look for some way to use the technology to enhance your process, not your end product. Try explaining to your leaders that AI is new and immature, and while it’s potentially very powerful, for the time being it should not be exposed directly to customers without some human oversight. If they’re looking for time savings, I recommend using these tools during the ideation and design phases to iterate quickly in order to settle on a path forward. More than anything, look for ways to use what you’ve been handed to help people, be they your customers, your coworkers, or even yourself. I can’t promise that you’ll hit upon the next great leap in human advancement, but across all of human history, the safest bet of all is an investment in other people. After all, they’re not going anywhere.
I’m a fun person.
Electrified rocks
At least, not yet… 😬
If you want to get weird about it, I am the wretched god of a hyper-localized digital universe. The operational physics of an entire organization are mine to command. You can start to see why all the big tech guys have Egos.
Meaning me.
Knowing how to code is certainly important to my job. But even more important is knowing how to help others write good code, unblocking them when they have a problem, helping define how to solve the larger problem than efficiently sorting a binary tree, being able to prioritize and breakdown big problems into small ones, and communicating all of the above to coworkers both technical and non. None of those are actually things that a LLM or MI model can solve.
I'm not worried for my job. I'm worried that information spaces (articles, images, video) both artistic and factual, are being filled with garbage. Would you trust the code written by ChatGPT to drive a car, or run a radiation therapy machine? How about for instructions and advice on how to make a cake? We already know they completely invent facts about people.
"AI" certainly has its uses, and it can find some pretty amazing associations from its data sources. But the biggest disruption they provide is making it seem like expertise and experience can be replaced or side-stepped, when it is actually both stolen and incompetent when examined.