Three issues with generative AI still need to be solved
Disclosure: Qualcomm and Microsoft are clients of the author.
Generative AI is spreading like a virus across the tech landscape. It’s gone from being virtually unheard a year ago to being one of, if not the, top trending technology today. As with any technology, there are issues that tend to surface with rapid growth, and generative AI is no exception.
I expect three main problems to emerge before the end of the year that few people are talking about today.
Generative AI uses massive language models, it’s processor-intensive, and it’s rapidly becoming as ubiquitous as browsers. This is a problem because existing, centralized datacenters aren’t structured to handle this kind of load. They are I/O-constrained, processor-constrained, database-constrained, cost-constrained, and size-constrained, making a massive increase in centralized capacity unlikely in the near term, even though the need for this capacity is going vertical.
These capacity problems will increase latency, reduce reliability, and over time could throttle performance and reduce customer satisfaction with the result. The need is for more of a more hybrid approach where the AI components necessary for speed are retained locally (on devices) while the majority of the data resides centrally to reduce datacenter loads and decrease latency.
Without a hybrid solution — where smartphones and laptops can do much of the work — use of the technology is likely to stall as satisfaction falls, particularly in areas such as gaming, translation, and conversations where latency will be most annoying. With translation, this will be especially problematic because the way translations are done will always introduce some latency. If the AI system adds more, it is could make the related tool unusable.
Qualcomm has released a performance report that looks particularly troubling in this regard, but it is likely only the tip of a nasty performance problem to come.
The language models generative AI uses include information that has not been fully vetted. Just this week, Elon Musk threatened to sue Microsoft for its use of Twitter to train its generative AI model. Given that Twitter’s data comes from its customers (and wasn’t created by Twitter), the lawsuit is likely to fail on its merits. That is, if Musk ever files the suit to begin with; his history is longer on threats than actions. It still highlights a growing concern about who owns the results AI tools generate.
You can derive a lot by successfully analyzing readily available information. I was a competitive analyst for a time, and it was amazing how much we could find out about a competitor’s activities using publicly available information. But that research was all done by hand using comparatively little data. These new AI models can pull petabytes of data now and will quickly grow to use exabytes and even zetabytes in the future.
The ability of these tools to uncover secrets from businesses, governments, and individuals will be unprecedented and the security technology needed to mitigate the problem not only doesn’t exist today, it may also be impossible to create.
Worse for some is that these tools can analyze data after the fact, often decades after the fact, surfacing things that were thought to be safely buried.
Generative AI tools can interact with others on our behalf and present a very different personality than our own. The tools can recreate our image, voices, and even our unique mannerisms while acting as proxies.
I was once an actor. One of the problems actors have is that others tend to conflate an actor with a role that has little to do with who they really are. In relationships, your significant other might well be in love with a character you play, not the person you are.
With generative AI, this will be a problem at scale as we increasingly allow these tools to interact with co-workers and maybe even interact on our behalf on dating sites. The disconnects between who people think you are (based on an AI proxy) and who you really are could damage trust and make lasting relationships, both personal and career, problematic.
We often go through a lot of trouble when building a relationship to hide our faults; AI tools could make that even easier. As a result, we may find it nearly impossible to trust those around us.
Despite these issues, generative AI has the potential to massively improve productivity, act as proxies, provide near instant translation, and deliver answers to questions that have been unanswered. But the issues with latency, security, and trust are real. And they come in addition to fears about job losses that have been hanging over the technology since it arrived.
I’m not arguing against generative AI; I doubt we could stop its advance even if we wanted to. But we have to begin thinking about how to deal with these before the damage is becomes unmanageable.