From Hype Machine to Workhorse: Practical AI Is the New Black

The initial frenzy around generative AI, fueled by the impressive (and somewhat unsettling) capabilities of chatbots like ChatGPT, is starting to subside. Businesses, the ones who actually pay the bills, are getting honest about AI. They want solutions, not sci-fi parlor tricks.

ABBYY, a company with over 35 years of experience in AI, is uniquely positioned to benefit from this shift. “We were using language models before they were cool," quips Maxime Vermeir, ABBYY's senior director of AI strategy.

The insights and lessons have equipped them with proven AI solutions and a customer-centric track record.

Taming the LLMs: Scaling, predictability, and ethics

Generative AI’s supersized language models are impressive but messy. Scaling them up is a technical headache.

Vermeir points out that while robust, scaling LLMs introduces technical and ethical complexities. One primary challenge is ensuring predictable results, and don't even get started on the ethics of it all.

"90% of POCs that companies have actually executed with large language models have failed. And one of the reasons is that it was very hard to get predictable results.," says Vermeir.

Companies need AI they can trust, not a digital wild card. This requires not just high data quality but the right process.

"So, whenever people try to leverage AI, it is to solve a particular problem. There's no point in trying to solve a problem or part of the problem of the business process without really knowing what kind of actual business processes you have," says Vermeir.

As a result, he sees the lines between hyped-up generative AI and the workhorse stuff of intelligent automation blurring, especially in the use of process mining, to understand how processes and AI models can impact each other.

Compliance in a regulated world

Highly regulated industries like banks, healthcare, etc., are no strangers to maximizing AI's potential. But many of these former models had discrete outcomes.

Generative AI's stochastic nature sees many highly-regulated companies sit on the sidelines. The problem is compliance, especially with many regulators ramping up their efforts. For example, the EU's AI Act is like a ton of silicon bricks dropped on companies, and everyone's scrambling. Many companies can't afford to play Russian Roulette with trust and reputation.

That has made transparency and demonstrable compliance in AI models non-negotiable. Yet, Vermeir sees this as an inherent problem in many large language models. So, ABBYY is working closely with its customers to improve transparency.

"Even right now, we're working on making sure that we can actually offer our compliance sheet almost like a checkbox," says Vermeir, who is looking to share the lessons and best practices with their other customers so that they have confidence that their AI models are compliant.

Smarter, leaner, greener AI

Vermeir highlights techniques for making LLMs more practical, including pruning, knowledge distillation, and quantization. Each aims to improve efficiency, reduce cost, and minimize the environmental impact of AI computations.

ABBYY is also pushing hard on small language models (SLMs). It utilizes 85 SLMs for efficient document information extraction—proving that sometimes smaller is better. These SLMs are becoming even more powerful with the incorporation of transformer-based technologies.

Carlsberg, for example, is using ABBYY IDP solution to streamline their purchase order process, saving 140 hours per month “in their entire purchase order process.” Who doesn't want a faster route to an ice-cold beer?

The future: RAG, more process mining, and fine-tuning

Retrieval-augmented generation (RAG) is the next frontier. Vermeir sees it as a game-changer. And rightly so, as ABBYY's document digitization expertise positions them well for RAG systems, where structured data feeds into more effective language models.

He also sees process mining as crucial for AI optimization. Understanding how AI will impact a business process before implementation saves time and resources, allowing for simulated testing before deployment. It's like a crystal ball for efficiency freaks.

The future, Vermeir says, lies in fine-tuning pre-trained models to solve industry-specific problems. "Companies are definitely looking for vendors that can offer that ensemble of capabilities to solve their specific needs."

Partnership, not DIY

The days of every company trying to DIY their AI are fading fast. Vermeir reckons businesses want turn-key solutions tailored to their specific woes. Fine-tuning pre-trained models is where it's at.

That's ABBYY's sweet spot—their track record and tech stack make them a compelling partner for anyone ready to ditch the hype and make AI deliver on its promises. Part of their value proposition is their expanding partner network.

"So, you're not left to your own devices. We have consultancy capabilities and an amazing partner network that I think is a perfect combo of USPs," says Vermeir.

Image credit: iStockphoto/Thinkhubstudio