Why is an industry poised to transform begging to be saved from itself? And why is there a litany of commentators simultaneously shaking the chains and predicting doom?
There’s a particular dichotomy in wishing there was a better general understanding of technology outside the walled garden of the industry and then – the moment those walls are temporarily breached – instantly regretting that wish. But that’s where we are right now with Artificial Intelligence and its current appearances in mainstream media reporting, the governmental regulatory lens, and conversations you overhear on the bus.
There’s a lot of careers possible in AI; predicting doom is just one of them
It can sometimes be a challenge to distill down big technological change into something that is broadly understood by an audience that doesn’t necessarily have the breadth of context that you or I might have had the privilege to have gained along the way. Luckily for media organizations, there are always those ready to step into those dangerous waters and try to make sense of it for the masses. Unluckily, within that set, there’s a handful for whom this is an opportunity to extrapolate current events into something much more significant and frightening for the audience, simultaneously creating chunky headlines. For example, this from Channel 4 News recently, where the presenter attempts to balance two views from guests, one measured and thoughtful, the other predicting the end of civilization.
Even where the underlying discussion is more balanced – like this coverage on the same day from BBC News – the headline is still very “end of the world”. Even a level down into the broad tech press and the prevalence, if not toward the end of the world, but something close to that state, was the lead.
You’re shouting at autocomplete
Over-reactions to technology stories that gain mainstream traction are not unusual (nor is the desire to grift for personal benefit from the event, sadly), but what sets this cycle apart is the relative accessibility of the technology coupled with the extreme nature of the reaction.
The current cycle of stories is largely sparked by generative AI (Large Language Models) LLMs and the best-known articulation of those techniques through a chat interface, ChatGPT. OpenAI – home of the GPT foundation LLMs and the maintainers of the ChatGPT chatbot through which the output of those models can be requested – has been not only a contributor to the art of the currently possible, but also a signatory to the statement from the “Center for AI Safety” which prompted the most recent glut of apocalyptical reportage.
That statement – “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” – promises profundity but only realizes the shallowest returns. GPT at present offers as close a risk to nuclear war in 2023 as the film “War Games” proposed online gaming in 1983 (although you could certainly imagine Dr Stephen Falken as a signatory to that contemporary statement).
Yet, the public at large using ChatGPT, Google Bard, or one of the other nascent public trial implementations of Generative AI could be forgiven for feeling disconnected from the news cycle that they’ve helped to create (much as users of TikTok might ponder how exactly they are contributing to global, geopolitical uncertainty).
No, LLMs are not going to enslave or enfeeble humanity
The reality is – as often reality turns out to be – more mundane. What foundation LLMs are far more likely to lead to in the short term is the realization that general models are good for general problems, and to get specific, you need to…. well, be more specific. When I describe the current wave of foundation models as “autocorrect,” there’s a degree of eye-narrowing among the audience, but that is the sort of general problem that models of this type are good at—predicting how I might want to finish a sentence by being able to parse through similar sentence roots they have seen before, and increasingly in popularity, do the same with common coding languages.
In the short term, we’re likely to see a shopping about of rapidly evolving foundation models, custom prompting and prompt tuning, and a probable aggregation of all of these (metaprompts? Is that a thing?) in the same way as the metasearch engines provided a result mash-up for a brief period in the days pre-Google (remember Dogpile?). If this seems like a mess, then you’d be right. Big platforms are hedging their bets somewhere between bolt-on partnerships, “bring your own LLMs,” and offering model marts as an adjunct to app stores. As of right now, it’s no more clear how the economics of the situation will resolve themselves than how easy or otherwise it will be for median organizations to start building their own custom models that will fit the industry-specific use cases that will emerge.
For OpenAI themselves, having raised $11.3bn to date, solving the economics of the situation can perhaps wait for a little while. However, with ongoing costs of an estimated $1m per day to continue to stand up their current GPT for OpenAI – aside from any other operating and development overheads – that these LLMs can provide many enterprise functions such as reading invoices, to document summarization, it they may well turn out to be far more costly at scale than specialized alternatives.
When overpromising gets out of hand, pull the emergency cord?
That the calls for restraint in the use of AI come not only from academia, government, and non-governmental oversight bodies but also from the very software vendors at the forefront of its development might seem to be curious. Surely, suppose you have a significant competitive advantage in a fully realized, manageable, and deployable technology that can match the grandiose promises of a new technical-industrial age that you’ve made in everyday use. Why would you campaign for it to be controlled? The answer lies in part within that question.
It’s certainly worth considering the fact that these LLMs were only really tested at scale very recently. The sheer volume of errors (sometimes referred rather euphemistically as “hallucinations”, which does nothing to calm the fears of those who bestow human characteristics upon these models) and omissions may well be higher than the developers expected; these ‘mistakes’ are certainly fuelling some of the current worry, distrust, and criticism that fuels this fire. The apparent openness to regulation is in part to mitigate the weight of questions coming down the road, especially where existing rules of data residency and processing (e.g. GDPR) were not designed with this set of use cases in mind (i.e. first party data that might be sent to models in the form of prompts).
Guardrails for all?
Let’s be clear, although we’re not buying the imminent threat to humanity posed by the specific wave of technology that has promoted this glut of mainstream coverage, that’s not to say that we’re against regulation (and we’re all here having this conversation as a result of the “great unbundling” forced upon IBM through antitrust suits kicked off in the late 1960s). The way in which AI is purported to be employed in public policing, for the targeting of automated conventional defense systems, for example, falls outside of how both are both currently regulated and – just as significantly – generally consented to by citizens.
The challenge that regulators will face will be impossible to overcome if they focus on the technology itself rather than how it should or should not be used. For a number of reasons, first, tech moves at speed; regulators do not. Any tech-specific regulation may quickly find itself dated and redundant because AI is simply a blanket label for a vast range of technology techniques, models, approaches, and structures. LLMs garner attention because of their general accessibility to demonstrate easily digestible and commonly discussed use cases in a way that the aforementioned policing and military examples are not.
Sleep well and don’t fear the models
As we mentioned in a recent analyst report, everyone who currently supplies you with software and services (as well as everyone who doesn’t but would like to) will want to talk to you about their Generative AI story. Our advice in that report is partly to be cool; very little of this is close to being in GA (general availability, in that it’s a thing you can buy and use) let alone demonstrably beneficial just yet. That doesn’t mean that the stories are not worth listening to or are in any way based on falsehoods. It’s just that you’ve got to find what is useful, manageable, and interesting.
What they are demonstrative of is that in a period of economic contraction, there’s an ever-greater need to find something extra special to spark interest in doing something bigger than merely keeping the lights on and getting everybody paid (the latter of which seems to be much more complex than it really should be, given the massive increases in revenue generated by many in the pandemic years).
Most importantly, separate the prophecies that foretell the end of humanity from the current wave of big, interesting models in your head. There’s enough to deal with working out how the latter makes sense without worrying about the former.