Why should we care so much about generative AI as opposed to the other forms of AI that have been around for some time? Because generative AI represents a massive phase shift, a new development in democratisation of a powerful technology, bringing capabilities to a vastly bigger audience than before.
It might feel like we’re in the middle of the hype-cycle for generative AI. Everyone by now has heard of it, most have had a play with it, and there has been a lot of initial enthusiasm. It also feels like some of the realities of the technology are starting to become clearer as it gets applied to more and more uses and the inevitable limitations become clearer.
However, in this piece I’d like to test some thinking, building on conversations with others, as to why generative AI isn’t like other AI.
How knowing something well can leave you surprised
One of the things I’ve observed is that the more people are expert in something, the more familiar they are with it, the more difficult it can be for them to identify a discontinuity – a phase shift or threshold point where things are no longer like they were before.
Take the infamous prediction in 1943 from the then CEO of IBM, “there is a world market for about five computers”. It can sometimes be easy to misread something you’re really close to, because you are more intimately familiar with all that’s happened, such that it often seems like part of an ongoing evolution rather than something materially different. If you’re knowledgeable about all the steps taken to get somewhere, you may not notice how big the journey has been.
While this may be a bit of a generalisation, I have noticed some scepticism or discounting of the significance of the technology, of seeing it as just another development to go with many others, from some who are more data and IT savvy.
For instance, for those who have been closely familiar with AI, they can point to all the other developments in AI (e.g. machine learning, reinforcement learning, etc.), and see generative AI in a broader context, rather than a sudden development seemingly out of nowhere. They may well be more attuned to the limitations – hallucinations, reliability, consistency, etc. – which mean they see it as one tool rather than (as some of us less expert might think) magic.
Of course this is very understandable, but I think it risks miscategorising this technology and its potential for transformative change. I’d like to walk through some arguments for why generative AI can be seen as a discontinuity and why that matters.
(Obviously we’re all vulnerable to over or under estimating the significance of events – predicting the future is a fraught activity! – so I’m keen to test the argument and would love to be told I’m wrong and why (then I can stop worrying about all of this)).
Commoditisation and democratisation of a new capability
Innovation – “implementing something novel to the context in order to achieve impact” – is crucially about adoption rather than just invention. It is the application and diffusion of an innovation that matters. An innovation can have a big impact in practice – e.g. computers have transformed how we work and interact. But the impact is as much about the spread of an innovation, and how it might be used in different ways by different groups, outside of the original envisioned uses – e.g. computers have been used in all sorts of ways beyond initially imagined.
One of the interesting things about the commoditisation and democratisation of innovations – rather than just the spread of the end benefits but the actual innovation itself – is that the more people who can use something, the more possibilities that can be considered, experimented with and tested.
I liked the way that FutureInitiative/LizTheDeveloper put it on TikTok: “So when something becomes more usable, and therefore we get more users, we get more imaginations involved. And the more imaginations involved we get, the more meme pools we pull in, the greater number of possibilities that now exist.” (The video is definitely worth a watch.)
When we consider current AI experts, it is natural that there are certain use cases and scenarios that dominate, as any grouping will reinforce some possibilities over others. Those most familiar with a technology are going to be most familiar with particular uses, particular norms and particular expectations about how that technology can and should be used.
However, when a technology or innovation stretches beyond that grouping, reaching new audiences, new perspectives, new imaginations, that ‘can and should’ might expand or pivot dramatically. Different uses become apparent, different needs come to the forefront, and different possibilities reveal themselves.
This is magnified with a tool like generative AI which doesn’t just ‘do’ things or provide outputs, but can be used in such varied ways, including being used to help you understand how to use it better. One of the most powerful things about some of the gen AI models out there is the ability to get it to help show you how to achieve your intent – removing some of the limiting factor of needing other people’s expertise to learn and make use of the tool (within limits of course, but far more permissive limits than with most things).
What does that really mean in practice though?
Low and no-code options are arriving fast
Over the past few months there have been several signs that suggest that, accelerated by the arrival of generative AI, the near future will be a lot more ‘no-code’ or ‘low-code’ than now: a situation where previously complex IT tasks that were once limited to those with significant domain expertise and experience are going to be possible for laypeople (such as myself!).
For instance, Monster API offers a no-code large language model finetuning platform for customising AI models. OpenAI launched ‘GPT Builder’ allowing customers to configure customised GPT models and share those with the world through the ‘GPT Store’. Now you can access tailored GPTs on fitness coaching, cooking, first aid, sex education, creative writing, therapy – and nearly anything else you can think of.
OpenAI also offers the ability to create AI Assistants without needing a huge amount of technical know-how (though that is still very necessary to make the most of them). The Rabbit R1 product is an example of how we might interact with AI assistants/agents in new ways, requiring no particular technical expertise.
All of a sudden, it has become possible for someone with no formal IT experience or training to leverage the not inconsiderable power of these new AI models. And where you get into trouble, you can often use these same tools to help critique what you’ve done and offer advice on how to improve. The multi-modal functions available in ChatGPT, for example, mean that you can just take a snapshot of what you’ve done, feed it to the model, and then ask for tips or constructive advice on how to achieve your intent better.
There are many elements that will still require knowledge of course, but all of a sudden a much, much larger audience understands what might be possible, and what they can ask for (or about).
And the direction of travel for these technologies suggests that this is the only beginning of this transition. As the models become more powerful and as more and more people customise the models to provide more tailored and better advice about how to do this effectively, more and more options are likely to open up to people previously seen as unskilled on IT.
AI is no longer just about data and IT, it is about capabilities and possibilities
One aspect that I think might be currently ‘over-weighted’ in the dialogue about LLMs and foundation models is seeing it primarily through the lens of data or IT.
Data has been critical to the development of these tools and good data is important for many other reasons.
However … there are some early signs that suggest that while important, data is not the be-all and end-all for many applications of the technology. A recent study, for instance, highlighted that when comparing GPT-4 with a more specifically trained model, “These models have been shown to outperform models fine-tuned with domain-specific data on some tasks, but still fall short on others, particularly when deeper semantics and structural analysis are needed.”
In short – the capabilities achieved with the existing data may be sufficient for all sorts of uses, regardless of what comes next.
It is also a challenge to how IT is viewed, because the ubiquity, the uses and the capabilities of these tools open up new way of using them – ones that may often circumvent current processes and rules around IT. I think this is especially the case for the ideation and problem-discovery end of the policy process.
The options for experimentation, prototyping and simulation just jumped in a big way
As outlined by Mark Dodgson, David Gann and Ammon Salter in their book Think, Play, Do: Technology, Innovation and Organisation two decades ago, new technologies allow for new ways of ‘play’ including simulation and prototyping.
The capabilities of generative AI tools are already a huge step up from what was possible in terms of low-fidelity prototyping. Being able to use the multi-modal capabilities of these models, it is now possible to prototype much faster and much easier. Whether it’s a mock-website, a new service or a nascent policy idea, the ability to engage in a discursive process with a tool that can simulate and run through different lines of thinking at pace for little cost is a massive breakthrough. Of course, care still needs to be taken and no public servant should be entering any sensitive information into the public tools, but even the ability to get feedback on a slide deck (for a public event), to create scenarios about a publicly known policy problem, to have something that can help you brainstorm is incredibly powerful. One of the most valuable aspects I have found, is using these tools to be a critical thought partner, to be able to quickly and easily create counterfactuals, and to poke holes in an argument.
Often the limiting factor in the policy side of government is the time it takes to get across issues and understand the dynamics at play. If we imagine a situation where we have widespread access to these tools and they can be easily tailored to specific issues (and ideally can be used securely), then we can foresee a situation where we can develop much richer scenarios and proposals at a much faster rate.
With regards to services, these tools offer the promise of being able to develop much more sophisticated prototypes at same or lower cost and time. This is likely to result in a higher quality of insights about the needs and wants of our stakeholders.
In short, a situation where public servants have easy access to generative AI assistants could mean a lot more simulation and experimentation, and therefore a lot more learning, faster. In turn, this should hopefully result in better outcomes. And while data was critical to arriving at this point, it probably won’t be the limiting factor as to what happens next for many use cases – the capability threshold is already significant enough to make the difference.
As with all disintermediation and democratisation, this could be uncomfortable for traditional processes (and their owners)
Another couple of observations I have made over the years is that:
- Disintermediation – making things possible without those who were previously integral to the process – is never fun for those who were intermediaries. It is very easy to treat that process as a threat that has to be stopped (however I’m not sure there are many good case studies of where that approach has worked).
- Commoditisation and democratisation – things being made accessible to a larger cohort – is also often uncomfortable for those who previously had the exclusive skills or resources. Again, it is very easy for those seeing the tool opened up to others as either a threat or a worrying sign, anxious that those newly accessing the capability won’t have the right respect, care or sensibility to use it appropriately (and there is also the potential loss of status).
Generative AI is going to challenge existing process and system owners within bureaucracies and large organisations, such as the APS, because it’s going to challenge the expertise and gatekeeping role of IT and data owners/stewards. (Of course, in the medium term I suspect it’s going to challenge all of us, because these same generative AI capabilities are going to render moot a lot of the traditional gatekeeping of policy development through specialist argot/jargon.)
While there will be a critical need to ensure compliance, it is undeniable that for some uses – such as early stage ideation and ‘playing’ in the policy space – many of these risks will be manageable outside of the agency environment using highly capable tools that don’t need particular data other than stories (most of which are already floating around the public domain). And, over time, some areas will build up the competence and confidence to make the case for why alternative approaches and uses might be appropriate, and manageable at relatively low cost and with minimal technical debt.
This has been a bit of a ramble, so to distil:
- Generative AI brings AI to the masses, which is going to reveal a lot of use cases, applications and possibilities that will be new to how AI domain experts have thought about it previously
- Generative AI is going to help introduce and accelerate low-code and no-code options for IT
- That’s going to open up new options for a range of actors
- The capabilities already possible mean that in some ways generative AI has moved beyond data – there are many uses that won’t require further or specialised data to add value, because the tools have already achieved a level of capability that is sufficient
- An example of this is what they can allow in terms of simulation, prototyping and critical thought partnership, things that can offer a lot of value in the early stages of the policy process (where security, data confidentiality and other concerns are often less pertinent)
- Disintermediation and democratisation are often uncomfortable transitions for those who previously had critical roles or expertise that was critical.
- None of this is advocating a particular path, but rather reflecting on previous innovation journeys and thinking what they might mean in practice.
- Generative AI is likely a significant discontinuity, so none of us should be too comfortable or too glib in assuming what happens next.
- I welcome people ripping holes in my argument, I’ll just pretend you’re an LLM helping me improve my thoughts.
New risks will emerge and new treatments will evolve
None of the preceding things are to say that IT and data skills and expertise won’t still be critical. Just as the sharing of medical information through the internet has given us all the ability to search our symptoms until we diagnose ourselves with some form of cancer, doctors are still vital (well, for now…).
While not necessarily fully predictable, new risks will emerge with this further commoditisation of IT. Already there have been issues where the tendency for generative AI to hallucinate could be exploited to create new security vulnerabilities. While some of these are likely to self-resolve as people get more familiar with the core issues, and as the tools themselves are reinforced against such things, there are likely to be others that might require a reframing or rethinking of how some core IT security elements are done.
Some things that spring to mind are that given the high compute needs of these new tools, the deeper level of interdependency they might bring about as they are integrated, and the oversight of a rapidly evolving technology that will continue to throw up new questions, new challenges and new expectations.
I don’t think these are the only possibilities of course, and things could take many an unexpected turn as these tools reach a much larger audience. But I do think it’s an interesting future to start speculating about, as it promises to be one quite different from what we’ve had to date.