Join our daily and weekly newsletters for the latest updates and exclusive content on leading AI industry coverage.
Imagine if, as a business leader, you described your symptoms to me and I used ChatGPT to generate a treatment plan and sent it to your local pharmacist without consulting a doctor. Or consider a trade where you get the world’s top data scientists, but all your business experts join your competitor, leaving you with only data and no context from experts.
In today’s AI era, discussions about opportunities, risks, and best practices for adopting generative AI, especially language models like GPT-4 or Bard, are everywhere. New open-source models, research breakthroughs, and product launches are announced daily.
However, language models are only useful when paired with knowledge and understanding. For example, someone could memorize all chemistry-related words from a dictionary but without understanding basic principles, that knowledge would be pointless.
Language models can mimic expert-written documents, frameworks, and recommendations. But when generating a new recipe, they rely on correlations from previous recipes, not on knowing what tastes good. They avoid unlikely ingredient combinations not because of taste knowledge, but due to the absence of those combinations in their dataset.
Good recipes generated by language models owe their quality to the experts who created the source data. Thus, expertise is crucial to making language models useful.
Correlation does not equal causation. Machines are great at finding correlations, but expertise is needed to interpret whether those correlations indicate true causation and should inform decisions.
As humans, we learn language first, then gain knowledge, and eventually understand cause and effect. By adulthood, we internalize expertise, combining language, knowledge (what), and understanding (why).
Without expertise, language models cannot make informed decisions. A language model deciding on its own is like giving car repair tools to someone who only knows car-related words but lacks practical knowledge.
To harness the potential of language models, start with expertise and work backward. Machine learning (ML) and machine teaching involve translating human expertise into machine language for informed decision-making or autonomous actions.
Expertise is more critical than data in AI and ML. Data identifies patterns, but expertise determines their usefulness. When an expert identifies a pattern beneficial for decision-making, it can be translated into machine language for autonomous decisions.
Building AI solutions should start by identifying the most crucial expertise within an organization and assessing the risks of losing that expertise or the benefits of offloading it to a machine. Next, determine which critical expertise can be translated into machine language using data and tools like language models.
Most organizations have already laid the groundwork for building expert systems. Language models can reference or be checked against programmed expertise.
In the next decade, market sectors will shift based on AI investments. Netflix’s streaming success and Blockbuster’s failure to adapt is a cautionary tale. Companies that wait to react to competitors’ AI advancements will likely find it too late to catch up.
Organizations best positioned for future success will invest in transferring operational expertise to machines and setting a vision for market leadership. They will explore and operationalize new discoveries to create tangible value.
Brian Evergreen is the founder of The Profitable Good Company. This article was co-written with Ron Norris and Michael Carroll from Georgia-Pacific.
DataDecisionMakers is a VentureBeat community for sharing data-related insights and innovations. Join us to read and contribute articles on the latest in data and data tech.