Hypothesis
People and organisations are fundamentally misunderstanding Generative AI; probably because they are closely associating it with it's two primary building blocks:
Traditional (ML) AI and
IT / technology more generally.
At first glance, the assumption that Gen AI is akin to, and alike, these forefathers makes intuitive sense, but turns out to be (mostly) unhelpful as a starting point for conceptualising and utilising Generative AI.
Afterall, Gen AI is built with many of the same ingredients as other (silicone/digital) technologies, so, an understandable default starting point for thinking about it would be to assume that it has inherited the same characteristics. This perception is quickly eroded through direct experience, but this starting point is stubborn to overcome and move on from.
In part this is because the natural ‘home’ for generative AI is seen as technology and IT departments of an organisation, and whilst there are certainly relevant and important skills, knowledge and methodologies routed here, the actual nature of generative AI is much more flexible, creative, broad and organic than most tech and IT platforms and tools.
How is Generative AI different?
To understand how Gen AI is different we can draw on some concepts (dimensions) that are illustrative. Many of these qualities are double edged, in most cases there is no clear winner, rather they imply how Generative AI can be conceived and utilised, and in some cases why guard-rails are necessary.
Computers are like Old Testament gods – lots of rules and no mercy. Joseph Campbell.
Any sufficiently advanced technology is indistinguishable from magic. Arthur C. Clarke
Deterministic & Quantitative Vs Liquid & Qualitative:
IT and systems are generally binary or quantitative in nature, they implement mathematics and logic and follow that logic exactly, not so for Generative AI which generates responses that can be rich and varied, even with the same inputs! Perhaps ironically then, to create a new technology or system it is essential to very clearly understand the details of the problem to be solved or mistakes and errors are inevitable (or to assume that you don't understand it and iterate towards a best solution over time). Gen AI is qualitative in the sense that, unlike its cousins, uses prediction, rather than binary logic as its engine. This can lead to creativity and variability in its outputs.
Explainable Vs Inscrutable:
Traditional IT and tech is deterministic, it literally implements a defined set of logic (if x, then y stuff), and we can therefore, in theory at least, understand exactly what is happening behind the curtain. This however is not true for Generative AI where the way in which responses are generated can not be mapped, precisely understood, described or fully explained. This quality means that double checking answers may be necessary, but it is a quality that also underpins an ability to create new, never before seen, outputs and solutions.
Automated Vs Creative & Generative:
One of the most important and most misunderstood (and creepiest!) properties of Generative AI is its ability to create genuinely new and unique output. Most people believe that Gen AI models are simply storing data and regurgitating it, again this likely comes directly from knowledge of traditional IT (i.e. databases/indexes), but it is not the case. Gen AI performs in a way akin to the human brain, it is complex and generative, it relies on knowledge (data) for sure, but it combines and creates new concepts and content. This creative capacity should not be misunderstood as a goal of Gen AI, but rather it is a property or capacity of the more advanced models.
Accurate / Precise Vs Fuzzy:
By nature IT systems are precise in their processing and outcomes. Calculators provide exact answers, implemented logic is unfalteringly followed, input or usage errors inevitably lead to system failures. Integrating systems is notoriously difficult because there is no flexible human to mediate and think. This rigid precision is not present in Gen AI. In fact Gen AI can be quite poor at mathematical operations or utilising structured data, but is much stronger and more forgiving of errors, even to the extent that it may correct or ignore input errors and still generate meaningful and valuable outputs.
Targeted & Specific Vs Broad:
IT tools are typically built to solve specific problems or cater for identified workflows, often the more targeted the system, the better it is judged to be. Sprawling ‘ERP’ type toolsets are of course popular, but even these are customised (through processes and data) to suit a particular environment. Once more, Generative AI tools tend to the other extreme, they are generalists able to respond flexibly in a wide range of situations and problems. Increasingly they can interact in more ways, text, speech, images, code, and can do so interchangeably. This combination opens a kaleidoscope of Use Cases.
Unforgiving Vs Flexible:
As already discussed, IT and tech tools are rigid and determined, they do what they are programmed to do and nothing more or less. Interacting with them is limited by interfaces and specifically structured formats, any divergence from these will result in disappointment almost inevitably. One comma in the wrong spot, one incorrect data type and it's broken. Generative AI on the other hand seems to be trying to understand requests, it can navigate around mistakes, interpret and even request more guidance. New developments will only increase this capacity, with ‘Agentic Gen AI’ and Multi-Modal AI coming online quickly. No longer is it necessary to speak machine code to interact and get valuable responses, once more this opens up more Use Cases and makes adoption and usage much quicker and easier.
Gen AI is different both in function and interaction from its forbares, it seems to have emergent properties that are more (or at least different to) the technologies and data that it is built from. This means that the default assumptions about its deployment and methodologies need rethinking.
What are the implications of these differences?
Quite profound.
Focus on creativity rather than workflow
Early applications and Use Cases reflect the strengths of Gen AI, and avoid its gaps and weaknesses. This means that they are generally much more creative (generative) and language/image focused, than data and workflow oriented. So right out of the gate, early adoption sits more comfortably in the domains of marketing and customer service than in more structured and technical areas of organisations. Which roles or groups end up the owners of Gen AI initiatives will depend on the organisation and projects, i.e. be context dependent.
A complement and accelerator
Given current strengths and weaknesses, Gen AI adoption should be seen and used as a complement or accelerator, to both existing IT systems and teams. This is reflected in the names of many of the tools that are coming to market now including Microsofts ‘Co-pilot’. To be sure Microsoft and others are using these kinds of names to make their products less (‘you are redundant’) scary, and more friendly, but also because most Gen AI tools can get off track and benefit from human guidance.
Guide or guard rails for usage
Gen AI can be a powerful tool, and like all others can be risky or even dangerous when poorly utilised. So, whilst no one usually sets out to cause issues, having policies in place can raise awareness and significantly reduce the risks that are inevitable with these fledgling tools and during immature / early adoption. Ideally the goals here are to keep it simple and understandable, enable experimentation and acceptable mistakes, but eliminate the unacceptable potentially damaging blunders.
Governance mechanisms
The Gen AI field and tools are moving fast, as are regulations and industry knowledge, this means uncertainty and change that needs focus attention from leaders as well as teams. This is not a set-and-forget topic. Setting up a Gen AI forum and/or interest group that enables distribution of information and feedback, and includes decision making and governance can help keep that organisational knowledge growing, minimise risks and feed into policies and procedures.
Leadership from non-tech
Gen AI projects and tools have roots in IT and tech, but are also (particularly in early forms) mostly used for generating more creative output. So, whilst the methods and processes of IT can, and should inform those applied to Gen AI projects, they will also require consideration through a non-tech lens. This could be through governance forums or through involvement of non-tech teams as early adopters. The goal here is to avoid the default (IT/tech dept) starting point which may restrict options and lead to disappointment (because Gen AI in its current form is weak at structured / tech solutions).
Lighthouse initiatives
One relatively safe approach to initial use of Gen AI is through internal (i.e. those that are insulate customers or stakeholders) experimental projects. Ideally these ‘lighthouse projects’ are ‘low hanging fruit’ (easy to implement and very likely to succeed) and should be interesting and engaging enough to self-promote and drive watercooler conversations rather than rely on formal corporate communications channels. The projects are lower effort, create early wins, accelerate learning and drive awareness. Whilst this description sounds unrealistic (a ‘goldilocks’ project), because of the surprising nature of Generative AI these Use Cases and tools already exist and are often ‘no-brainers’.
Gen AI as an untrained horse
Gen AI is in its early stages, it's being integrated into everything and new toolsets are popping up daily. There will be change and consolidation, groups like Microsoft are committing heavily to Gen AI and are building their suite around it and the integration of their various data repositories. Right now however, there is lots of noise, but little refinement or maturity. It can therefore be helpful to think generically about the capabilities of Gen AI, for example as being like a new form of engine that can boost and provide energy to activities and the tooling that enables them. Understanding the essence of Gen AI in this generic way enables intuitions about potential successful (and unsuccessful) Use Cases, so can inform decision making.
Understand it's ‘out there’, and overlap with organisational needs
“Out-there” definitions: 1: woo-woo (wü-ˌwü) 2: not going away
Currently there is a dizzying proliferation of creative (sometimes way out-there woo-woo style) applications that are useful only in that they illustrate the creative capacity of Generative AI. These range from video creators, to avatars and jingle creators. In these early days of Gen AI, the art is to be aware of these capabilities and to understand how they overlap with the organisational needs. Many Use Cases may be feasible, but frivolous, and others not workable.
Comments