6 min Applications

Persona CTO: LLMs are ‘just’ sophisticated APIs

Persona CTO: LLMs are ‘just’ sophisticated APIs

Everybody wants to talk about AI. But while the consumer/user world is busy working out whether its job is going to be made redundant (spoiler alert: mostly it won’t, you’ll just get AI tools to accelerate & automate existing workflows and tasks), the technology industry itself is busy working out what ingredients we should put into the AI mix. That question is largely a factor of what Large Language Model (LLM) approach and vector database any given team uses to build its AI model, to what degree it embraces open source tools, how far it has worked to put guardrails in place to secure Personally Identifiable Information (PII) and how much it is willing to drive towards narrower industry- and task-specific language models in the immediate future. 

But is it all that complicated, or are AI LLMs really just another way of thinking about Application Programming Interfaces (APIs) as the essential ‘glue’ that today bonds software application components, services and entire operating systems together?

Fundamental misunderstandings

Charles Yeh, CTO and co-founder of online ID service verification company Persona thinks that this isn’t an outlandish suggestion and that there are some big misconceptions out there in this space right now. 

“Although Large Language Models (LLM) have proved their value in processing text and automating basic functions within applications, there’s a fundamental misunderstanding about what LLMs can and can’t do – as evidenced by the countless startups putting wrappers around LLMs, only to find out they have built software around the model in order to make the software functional,” said Yeh.

A trained software engineer himself, Yeh suggests that LLMs are not replacements for standalone software systems and that they are more like smart APIs. What he means is that, just as APIs serve as bonding bridges between various software components, LLMs are tools that augment human-machine interactions, offering a blend of capabilities most suited for tasks for which probabilistic certainty is acceptable.

A spectrum of precision 

“Everything we can enable from a technical perspective can be mapped along a spectrum of precision on one end… and a probabilistic one on the other,” explained Yeh. “LLMs generally find greater success and applicability for use cases on the probabilistic side. So with all this in mind, there are two major gaps that clearly distinguish LLMs from traditional software-coded systems. LLMs excel at understanding and generating text, but they don’t inherently comprehend the unique business rules, business processes, or strategies specific to an organization or application. In fact, they lack the capacity to incorporate complex business logic that underpins many software systems.”

As a working example here, let’s say that in e-commerce, a standalone software system may need to manage inventory, calculate shipping costs and apply discounts based on intricate rules – tasks that go beyond the capabilities of a language model. While it can be tempting to think that the model can be trained to understand complex business logic, this likely requires writing excessively long scripts. 

At Persona, Yeh and his software developers are experimenting with using LLMs to assist with Know Your Customer Anti-Money Laundering (KYC/AML) compliance, one aspect of which involves checking individuals and their known addresses against names and addresses on sanction and Politically Exposed Persons (PEP) watchlists. 

But, says the team, something as simple as matching an individual’s past address contains a lot of context. Some use cases allow for small typos in the street name or postal code (which happens all the time) while others strictly disallow typos. This gets even more complex when accounting for text extracted from a driver’s license image which, depending on the country, tends to have common extraction issues.

“You could type those rules into the prompt, but different models interpret them differently – and this presumes that all of your addresses are formatted the same. When you think about incorporating international addresses and factoring in their uniqueness and intricacies, your prompt gets even more unwieldy and unnecessarily complex,” explained Yeh, in real terms.

Prioritising nuanced business needs

Effective software systems obviously prioritise user needs and preferences. They utilise user data and feedback to tailor experiences and services accordingly. The ‘problem’ with LLMs (from Yeh’s perspective if not others) is that they lack the inherent ability to gather and interpret user data to make informed decisions about prioritising features or content. User-centric prioritisation requires a separate layer of software that can manage user profiles, preferences, and feedback.

“To continue on with the matching experiment we’ve run at Persona, we’ve also tried to use LLMs to match users and businesses for a variety of use cases. This has been challenging because different users have different requirements or want different things,” clarified Yeh. “For instance, a customer that provides loans has very different requirements because of the inherent risk involved and the KYC/AML obligations they must meet. But an e-learning platform that just needs to ensure that a student completing a course is who they say they are has a much lower risk exposure for getting it wrong. The LLMs lack the logic to parse business needs in the right risk context.”

But LLMs can…

This is of course not meant to be a total downer on LLMs, we know how important they are for the new breed of generative AI as it joins the reactive and predictive Artificial Intelligence functions that we already have in place. Wrapping up his points, Yeh insists that ultimately, writing a prompt inherently relies on the LLM’s interpretation for processing… but in many cases, you want there to be no room for interpretation. 

“The only way to precisely funnel a customer’s business needs all the way through is through precise code. LLMs can be used for probabilistic use cases or situations where customers do not know exactly what their business needs are. LLMs excel when the requirements are imprecise,” concluded Yeh.

If there’s one word that comes up a lot right now in the AI LLM space it is ‘nuanced’ i.e. getting the new breed of AI to really start to appreciate intricacies over and above and outside of the expansive world or essentially large LLMs (after all, the clue is in the name). As we move from imprecise to precise, we must do so with software engineering that is akin to precision engineering… and that’s surely a goal worth pursuing.