BUILDING THE LLM - PART III
Bill Inmon
Founder, Chairman, CEO, Best-Selling Author, University of Denver & Scalefree Advisory Board Member
BUILDING THE LLM – Part III
By W H Inmon
When LLM’s first appeared on the earth, the value of the LLM in addressing the challenges of text became apparent. The hue and cry was that we needed LLM’s. An LLM was a language model that could be used to parse and disambiguate any text.
But reality quickly entered the arena. Language is very large and complex. Even if you could create a real LLM, it would be so large that processing the LLM through the computer would be very expensive and slow (if it could be done at all). In addition, language is very complicated and trying to construct an LLM is an extremely arduous task. Furthermore, language is constantly changing, which implies that the LLM needs to be constantly maintained.
From a theoretical standpoint, an LLM was a good idea. From a practical standpoint, an LLM was terribly impractical, both to build and then to use.
Very quickly the world discovered that it needed SLM’s – small language models – rather than the impractical LLM. ?An SLM cut down some of the complications of the LLM.
But it turned out – from a practical standpoint - that the SLM was only slightly better than an LLM. SLM’s were still very large and consumed resources like crazy. And SLM’s were subject to the same challenges as LLM’s except only to a smaller extent.
It took a little longer for the world to discover that if LLM’s weren’t the answer than SLM’s weren’t the answer either.
The next step in the evolution of language models was to discover that the SLM needed to be aligned across different industries. Instead of building one SLM for the world, the SLM needed to be broken down into applicability to an industry. Typical industries such as banking, insurance, retailing, real estate, restaurant management, airlines, and many more industries needed their own industry language models. By reducing the SLM to an industry language model, the size of the model and the complexity of the model was greatly reduced and the usability of the model was elevated. This reduction in the size and scope of the language model made the industry language model workable and manageable.
But there was a fly in the ointment. The industry language model for an industry was repeatedly being built over and over by each company in an industry. The industry language model for bank ABC was almost exactly the same as the industry language model for bank XYZ. The industry language model for airline QWE was almost exactly the same as the industry language model for airline RTY, and so forth. Building individual industry language models for each company in the same industry was a colossal waste of time.
Into this arena came the generic industry language model. The generic industry language model saved huge amounts of time and resources. With the generic industry language model it was not necessary for every company in an industry to build their own version of the industry language model. But the generic industry language model still needed to be customized.
The good news was that customizing a generic industry language model was an easy thing to do.
The progression/evolution of language models is shown by –
领英推荐
There are many implications to the evolution of language models.
LLM’s started in an untenable position. LLM’s were too big to ever be used even if they could be built. LLM’s were impossible to construct completely and accurately. And LLM’s were constantly changing, so LLM’s were a moving target. SLM’s had the same problems as LLM’s but on a smaller scale.
Industry language models were practical and useful. But there was no need for everyone to build their own unique industry language model.
And in the final analysis, industry language models still needed customization.
With customized generic industry language models, the world reached a point of stability.
?
Bill Inmon lives in Denver with his wife and his two Scotty dogs – Jeb and Lena. Lena sleeps in her cage each night. She feels protected there and won’t sleep anywhere else. And each morning Bill goes down and lets Lena out of the cage. Lena always greets Bill each morning.
?
?
?
Head - Data Engineering, Quality, Operations, and Knowledge
2 周in short - truth eventually catch up
Such a great history lesson and tutorial in 14 paragraphs and a graphic.
Transformational Leadership | Data and Analytics Strategy | Artificial Intelligence
4 周Love this, Bill. I went on a similar journey where the excitement of possibility completely clouded our ability to see pragmatic usefulness for the business. Eventually, the models got smaller and their applications got more finite. Instead of replacing complete value chains and departments of people, we are replacing activities in a process and busy work. Progress for sure, but not the transformation we were imagining. Reminds me of all those January gym memberships people buy every year. The vision in our head never quite matches the results we get.