5 tips for choosing the right AI model for your business

The expansion of generative AI (gen AI) has been pushed by high-profile giant language fashions (LLMs), resembling Open AI’s GPT-4o, Google’s Gemini, and Anthropic’s Claude. 

Nonetheless, whereas these bigger fashions hog the headlines, one other set of fashions has been gaining traction. Some specialists consider small language fashions (SLMs) may very well be the way forward for gen AI.

In line with analysis agency Gartner, whereas LLMs have historically dominated the event of language fashions, SLMs provide potential options to key challenges recognized by practical leaders, together with price range constraints, information safety, privateness issues, and danger mitigation related to AI. Enterprise leaders may need to decide on between bigger and smaller fashions as they discover gen AI. 

So which can win the battle? 5 enterprise leaders give us their opinions.

1. Contemplate domain-specific alternatives

Claire Thompson, group chief information and analytics officer at monetary companies large L&G, mentioned she expects small and enormous fashions to have a spot in enterprise actions. Nonetheless, she additionally thinks in the present day’s high-profile fashions may very well be tweaked for brand new use circumstances.

“I can see a state of affairs the place among the LLMs might begin to be skilled on particular matters to get extra element out of them, and I can see that starting to occur increasingly,” she mentioned.

Whereas there’s a hole for domain-specific fashions, Thompson instructed ZDNET she’s not sure if many corporations would dedicate human and monetary sources to in-house improvement.  

“I do not know whether or not you’d construct your individual,” she mentioned. “Once I discuss constructing fashions, it is extra about leveraging present fashions internally and utilizing your information in a safe atmosphere to realize outcomes.”

Nonetheless, whether or not giant or small, Thompson mentioned the longer term is about domain-specific fashions.

“I feel we are going to begin to get extra tailor-made fashions,” she mentioned. “You might see, for instance, the way you may tailor a mannequin round medical info, local weather matters and ESG, and asset markets. It is these particular use circumstances the place you might get extra bespoke fashions popping out.”

2. Choose the best horse for the course

Nick Woods, CIO at MAG Airports Group, is one other digital chief who mentioned the way forward for gen AI might be a mix of enormous and small fashions.

“I do not suppose it is one dimension matches all,” he mentioned. “And I feel the mannequin you choose depends upon the use case in your online business.”

Woods instructed ZDNET it is common to listen to professionals say the group ought to spin up an AI program. His response? “No, it is the very last thing we must always do.”

Woods mentioned executives ought to give attention to the enterprise transformation agenda and determine which instruments, together with gen AI, might help ship the best outcomes. “So, for instance, we might wish to run a small, particular mannequin on the sting to go and remedy a selected use case round one thing like recognizing when an air bridge has docked,” he mentioned.

“I would run one thing totally different when trying to create a mannequin for a query like, ‘What does world air visitors appear to be, and the way will it react to climate modifications?'”

In brief, mentioned Woods, selecting a mannequin is about selecting the correct horse for the course.

“I feel you will note many small fashions deployed on the edge at scale for specific use circumstances,” he mentioned. “That is nearly inevitable. Nonetheless, I nonetheless suppose you may see some huge fashions prevailing.”

3. Contemplate the context

Gabriela Vogel, senior director analyst within the Government Management of Digital Enterprise follow at Gartner, mentioned her conversations with CIOs counsel small, domain-specific fashions have an essential function to play — a minimum of within the shorter time period.

“The shoppers I converse with are looking for and create fashions utilized to a selected context,” she mentioned. “They don’t seem to be essentially huge, normal fashions, however ones particularly tied to small databases for a selected utility.”

Vogel instructed ZDNET that increasingly corporations are shifting from exploration to manufacturing gen AI companies utilizing SLMs.

“They’re making this shift as a result of they’ve examined loads,” she mentioned. “They’ve seen what works and does not with greater fashions, after which they’re attempting to go extra particular and apply that strategy. That is what I’ve personally seen with my shoppers.”

4. Go small to cut back hallucinations

Ollie Wildeman, who leads buyer satisfaction at Massive Bus Excursions, mentioned the selection of SLM or LLM depends upon the use case — and for a lot of corporations, the choice is prone to be smaller moderately than greater.

He instructed ZDNET how Massive Bus Excursions makes use of Freshworks Buyer Service Suite, an omnichannel help software program that features AI-powered chatbots and ticketing. The corporate additionally makes use of an AI-enabled digital assistant from Satisfi Labs that connects to its web site and offers with fundamental buyer queries.

“Satisfi’s AI know-how solely takes information from the particular corporations they work with,” he mentioned. “The corporate’s know-how isn’t related to large-scale AIs, like ChatGPT or different instruments — they’re doing it themselves.”

Wildeman mentioned this contained strategy creates enterprise advantages — executives may be positive how their information is used rigorously to supply outputs.

“In that approach, your information is safer as a result of you understand the place it is coming from and what processes they’re utilizing,” he mentioned. “Additionally, you get fewer hallucinations as a result of you understand the mannequin you are utilizing is designed for the kind of enterprise you are in.”

These outcomes lead Wildeman to conclude that smaller, domain-specific fashions shall be essential for enterprises.

“I feel for companies, the selection of mannequin goes to be extra particular, whereas most likely for the final consumer, these large free fashions that you just see in all places shall be common.”

5. Focus in your first-party information

Rahul Todkar, head of information and AI at Tripadvisor, mentioned the best mannequin for an organization won’t simply be a query of massive or small.

Professionals might attempt each fashions. Nonetheless, Todkar instructed ZDNET that purpose-built and customised fashions are the way forward for AI, whether or not they’re outlined as huge or small.

“Take the instance of Mistral 7B, which is a comparatively small mannequin within the context of different LLMs, but it surely does fantastically properly once you have a look at particular duties,” he mentioned. “So, to me, the longer term is about customizable fashions.”

Todkar suggests the important thing to AI success is guaranteeing the mannequin makes use of your information securely and successfully.

“It isn’t the coaching dimension or the options within the mannequin that matter, however moderately it is about taking that mannequin and making use of it in your context together with your first-party information,” he mentioned. “That is once you transfer past off-the-shelf fashions and might use the insights out of your information. So, the reply goes to be someplace within the center.”