The gap between LLM fluency and AI expertise
How come that modern AI solutions have become omnipresent, the number of AI experts sharing their insights is growing and companies still struggle to create a positive return on their AI investments?
Where modern AI thrives
Creating solutions with AI (or Machine Learning) is and always has been a challenge about a) leveraging data in the right way, b) understanding the concrete business goal and c) integrating the solution seamlessly in the processes in the bigger picture.
The emerging technologies around LLMs have great value in translating human natural language into machine language. This creates opportunities for novel ways how humans can interact with software through novel user interfaces.
Further, having machines "talk" in our natural human way to us allows us to exchange on more complex topics that require a logically stringent thinking. However, LLMs have learnt two levels of logic: a) building grammatically correct sentences and b) knowledge on a vast domain of topics.
Modern LLMs excel at grammatical correctness - which is a remarkable achievement. But it also hides that they often lack the depth needed for reproducing correct knowledge in specialized topics. They've learned the language of expertise without necessarily acquiring the expertise itself.
Overcoming limitations
Overcoming knowledge gaps through retrieval-augmented generation (RAGs) is popular, and works well for injecting specific information. These and other approaches are constantly improving to cover a growing number of use cases. However, they still face limitations, especially when we ask for interpreting and applying domain-specific expertise. To overcome these limitations requires more than just advanced tools, but advanced technical expertise.
The emergence of LLMs has left large parts of business and society with the impression, that these can be easily be applied to whatever given challenge there is. This is driven by their abilities of building sound sentences in correct grammar, and the simplicity with which they can be used in quick proof-of-concepts (PoCs).
Yet, creating business value reaches beyond these aspects. The abilities to optimize AI on a concrete goal to master it in a production-ready way have not changed - you still need to leverage data, understand the business goal and integrate it wisely into a bigger solution. The quick speed to realize PoCs is a strength - but does not free from the hard work to integrate them in a productive setup. This creates efforts that can easily be overlooked in the overall business case.
Organizations are increasingly deploying LLM solutions without traditional data science expertise. Driven by the wish to not miss out on this important trend, and pulled by the promise of quick business value, this urge to act fast is understandable. And it is right to act fast instead of overthinking problems. However, this approach needs to be framed by the right skills.
Beyond traditional data science skills, novel skills arise that need to be established. These include AI-Ops to monitor performance of AI, guardrails to mitigate critical risks, and ethical AI practices to only name a few. These are investments that are costly in the short term, but pay off in the long run.
Towards identifying the right challenges
To really pay off, we must also become better in framing the problems. We need to understand what LLMs can and cannot do. As every emerging technology, they have their strengths. Good AI strategies can only be built when we are clear about these strengths. This challenge becomes clear when we examine four criteria that determine whether LLMs can deliver quick value:
- How sophisticated does the solution need to be to deliver value - how do you define "good enough"?
Providing answers in customer service is less sophisticated than analyzing and diagnosing medical treatment for patients. - Does the value-add primarily stem from text processing capabilities?
The summary of a document works on other data than the prediction of equipment failures. - How deep, or specialized, does the embedded knowledge need to be?
Providing feedback on general economic topics requires different knowledge depth than interpreting tax law in a specific situation. - How tolerant are you about failures in your solution?
An imperfect formulation that is tolerated in marketing can have immense consequences in financial reports.
Growing the right skills
LLMs' lack of reproducing knowledge at high quality for all topics is hidden in plain sight. Approaches to circumvent these are generally known - but not apparent to many people working on these topics. LLMs truly create great value for some challenges in novel ways. But this does not translate into a super-intelligence that provides quick solutions for all challenges humans face.
We need to understand what these technologies' strengths are to apply them wisely. If we want to create business value with modern AI, a strong foundation of established data science skills along new ones to operate these technologies will create an environment where AI's promises find a way into the real world.
The true challenge for creating business value is not chasing a mythical super-intelligence, but to strategically augment our own. By building a foundation of essential data skills - both traditional and new - we can finally transform AI's promise into real-world business value.