Glyn Heath in his role as co-founder and director of Bayezian.
AI needs developers. It needs software application developers, data scientists, a few causal, predictive and generative AI theorists and of course a healthy contingent of business strategists to all come together and build composite AI teams that can deliver new intelligence functions into business.
As much as that part of the AI equation is now being accepted and understood, there is still debate surrounding exactly what AI will give us in the future. One man who thinks he knows how at least part of the next chapter in this story might play out is Glyn Heath, co-founder and director of Bayezian, an organization that describes itself as a specialist in accelerating and supporting individuals at the start of their career to incubate ideas and bridge the digital skills gap.
The Swinging AI Pendulum
Talking about what he says is a swinging pendulum that points to the direction of the “great AI debate” at any given moment in time, Heath highlights the fact that there are two major directions of thought. One week the consensus seems to be that AI will propel humanity into a world free from starvation, war and disease—it might even unlock the secret to eternal life if the latter can be conclusively solved. The next week, AI will cause the inevitable extinction of humanity and, for the ultra-pessimistic commentators, that event will be sooner rather than later.
While we can all agree that neither is a likely outcome to surface in the immediate future, it is perhaps concerning that with each swing of the pendulum, opinion becomes increasingly more polarized.
The Road To AGI
“Yes, there is a chance that a truly artificial general intelligence (AGI) could find a way to halt or reverse the aging process altogether and, while it’s doing that, systematically find cures for all known fatal diseases. However, it’s one thing to know how to do something theoretically; we have fundamental global societal issues to resolve before, or indeed if ever, that could become practical for all 8 billion of us. At the other extreme of the pendulum’s arc, it’s also possible that runaway AGI could trigger the irreversible demise of homo sapiens as a species,” offered Heath, even-handedly.
Between these two extremes, he views a range of much more likely outcomes which we still have time to determine. Of course, there is a degree of urgency and there is no room for complacency. But it’s a matter of degree; some things that are urgent require a response measured in milliseconds or less, a circuit breaker to prevent a person being electrocuted by a faulty appliance for example. Other urgent things require a response over several decades, like climate change, for example.
“That’s the macro view. What we need in the immediate term and on an ongoing basis is to deal with the incremental, but no less significant, developments that illuminate the track we’re currently on and which direction it’s headed,” advised Heath. “The GPT-4 developer tool enables bad actors to generate output that would otherwise be restricted and it has been demonstrated that existing guardrails can be bypassed relatively easily with modest technical expertise and at very low cost. It should be noted that full access to the GPT-4 developer tool is required to do this and that access is currently being restricted.”
Why does this debate even surface? Because the suggestion here is that ChatGPT can be provoked into going from trapping a very respectable 93% of bad responses to generating a very poor 95% bad responses to such prompts. Research suggests that the problem also exists in GPT-3.5 Turbo and Meta’s LLaMA.
Jailbreaking LLMs
The process for “jailbreaking” large language models—meaning the use of processes intended to make LLMs behind AI perform beyond their intended scope and take skewed actions—is simply to have one LLM issue a prompt and a different LLM tasked with responding to the other.
“This technique is called persona modulation and involves a human initially providing prompts to one LLM and using those responses to make another LLM adopt a particular persona which proffers otherwise restricted responses,” explained Heath. “This behavior is baked into the way LLMs ‘learn’ from huge quantities of conversational text and consequently could prove tricky to resolve without impacting the potential for good output.”
In fact, he says, a couple of potentially challenging issues are already starting to emerge as a result of these flaws being identified. Firstly, in attempting to bolster guardrails there is a possibility that we will inadvertently train a model to more clearly see what ‘good’ looks like and, by extension, more precisely what bad looks like and consequently become very good at being bad.
“This is still to be proven but clearly can’t be ignored until it's either discounted or fixed,” commented Heath, to provide real world balance here. “Secondly, in establishing what bad looks like and guarding against it, we potentially lose the chance of generating genuinely unexpectedly good output. So all this means that important work needs to continue apace on these important issues and many others in parallel with the astonishing progress that is being made in using AI for good.
Current Growing Projects
We have recently seen news of DeepMind’s GNoME project, the output of which is being made publicly available and is set to hugely accelerate breakthroughs in materials research. There is also talk of AuroraGPT, aka ScienceGPT, a trillion parameter generative AI model trained on scientific papers and results to aid in new, novel research. While the project is at an early stage with only 256 nodes of the 10,000 node supercomputer brought into use for early training, the future potential is enormous.
“Herein lies the dilemma,” argued Heath. “Are these specific bad aspects in their current form sufficiently compelling reasons to stifle development in AI generally? I’d argue definitely not. Or at least not yet. The types of bad output being talked about caused by jailbreaking LLMs—how to make a bomb or a lethal pathogen—can already be found on the public Internet or certainly on the dark web contrast that with the groundbreaking advances that we risk not making by imposing too many restrictions or indeed training out or at least slowing down the opportunity to do good.”
Understanding Differing Narratives
If we want to lessen the amplitude of these pendulum swings, it seems clear that we perhaps need to find a way to achieve consensus around some of the increasingly polarized narratives in the AI development space. While (as we said at the start) we will naturally look to software application developers and data scientists to work these problems through, there are suggestions that we may also need to involve lawmakers and even politicians in this whole debate and process.
“We can’t allow the extreme pessimists to create negative public opinion that results in draconian restrictions on the development of AI. If the Future of Life open letter that was started back in March 2023 is a reliable indicator, that should still be some way off. It has to date only garnered around 33,000 signatures, even though it is still open for signing. However, we’ve all seen how social and mainstream media can quickly amplify misinformation and exert disproportionate influence leading to some genuinely disturbing outcomes,” said Heath, who concluded by reminding us that the future of humankind, for better or worse, cannot be determined by whoever shouts the loudest or whoever has the deepest pockets.
The Way Forward With AI?
There is no universally (or even globally) agreed way forward with AI as a whole, but many commentators and analysts agree that we need to target bias and AI hallucinations from an open and collaborative stance. Equally, we need an intelligently nuanced and flexible approach to AI controls to ensure that only the harmful uses of AI are subjected to outright ban or extreme regulation. Additionally, we need large language models to become small language models for specific use cases, as organizations themselves also develop their own proprietary language model technologies—and we also need retrieval augmented generation for external data source ratification, validation and verification.
In short, we need a lot of technology to move in a lot of different directions for AI to grow the right way and for us to avoid not just the pendulum (as fans of Edgar Allan Poe will know) but to avoid the pit beneath it.
