Anthony Loss, Director of Solution Strategy, Clearscale, offers a view from the ‘genAI trenches’
Going into 2024, generative AI technology transitioned from a shiny new type of solution that most folks are still trying to wrap their heads around, to a product that is becoming an everyday part of life. The question most people want to answer today isn’t “What is generative AI?” The typical digital citizen has already experimented with genAI tools, and a fair number of us are using them on a daily basis.
Now, what folks really want to know is what’s coming next in the realm of genAI – how will it evolve beyond its current strengths and limitations, changing the way we work and think?
To what extent do regulators need to impose guardrails around genAI technology?
I don’t have any magic ability to answer those questions. I’d like to think, however, that I have more perspective on the future of genAI than most folks, because I’ve spent months in the trenches helping my company build a tool that integrates generative AI into cloud applications. My experience with genAI runs much deeper than carrying out discussions with ChatGPT or asking GitHub Copilot to write some code. I’ve actually seen how the genAI sausage is made.
Based on that experience, I’ve developed some strong perspectives on where generative AI stands today and what’s coming next. Here are my thoughts.
- Stop obsessing about Artificial General Intelligence
There has been a lot of chatter lately about Artificial General Intelligence (AGI), meaning an AI model or tool capable of simulating all facets of human intelligence. Some folks have even speculated that the Drama in late 2023 at OpenAI stemmed from the company’s having achieved AGI, although there’s no hard evidence that this actually happened.
To me, asking how close we are to AGI is not the right question. The answers will always vary because there are different takes on what, exactly, counts as AGI. I tend to think we’re far from true AGI because Large Language Models (LLMs) are not capable of reasoning, and I consider reasoning to be a critical component of AGI. But one could argue that a model incapable of basic reasoning could still count as AGI.
Debates like those, though, are mostly beside the point of what matters in practice. What we should really do is analyze AI solutions based on how useful they actually are to the people they are intended to serve, not how closely they resemble AGI (however we choose to define it). It doesn’t really matter whether a given tool is capable of AGI or not if it’s serving its intended purpose well.
We should fixate less on achieving AGI and more on improving the quality of the AI solutions we already have.
- AI hallucinations aren’t always bad
Ask most folks about the shortcomings of Gen AI, and one of the first things they’re likely to mention are so-called hallucinations. Hallucinations happen when genAI models produce information that is false.
AI hallucinations are indeed a problem if people take the resulting data as fact, but the interesting thing about hallucinations is that they’re not always a bad thing. On the contrary, hallucinations are an important part of the ability of AI models to generate novel stories or ideas. Sometimes, you do want your model to make stuff up, if your goal is to get it to say something no one has said before.
What this means is that rather than seeking to prevent hallucinations, AI developers should focus on controlling them. AI models that are incapable of hallucinating would be a bad thing because they’d never say anything novel. As long as users can reliably control when a model says something new and when it presents only true information, they can leverage it to suit varying needs.
- To regulate AI, focus on models, not concepts
There has been much discussion about how to regulate genAI models, but to date, very few regulatory frameworks have actually appeared.
To my mind, the best path forward on the regulatory front is to ensure that regulations prevent harmful use of AI technologies, while simultaneously keeping the door open for new inventions and innovations. To do this, regulations should focus on specific models, rather than
concepts or practices. Policies that categorically forbid a certain type of AI development or prevent AI from being used in certain contexts run the risk of strangling innovation. But if a model already exists and we know its capabilities and limitations, it makes sense to regulate what the model is and is not allowed to do.
- Embrace AI as a complement to jobs, not a threat
Worries that humans will lose jobs to AI are understandable given the powerful capabilities of genAI technology. AI can already do many things faster and more effectively than humans, and it’s only going to get better with time.
But that doesn’t mean we should resign ourselves to a future where human workers are irrelevant. Instead, we should focus on upskilling humans so that they can work more effectively with help from AI. AI will make some types of jobs irrelevant, but it also creates opportunities for many workers to become much more efficient.
What this means is that smart employees should focus their energy and skills on learning how to use AI tools to become better at doing things that AI can’t do on its own. As long as enough people take that approach – as opposed to assuming that AI is ushering in some kind of dystopian future where humans serve no useful purpose – AI will become a net benefit for workers, not a threat.