Using First Principles to More Readily Move GenAI from Prototype to ProductionUsing First Principles to More Readily Move GenAI from Prototype to Production

The traditional rules and plays of enterprise software development can be thrown away by the disruptive influence that is Generative AI.

January 15, 2025

4 Min Read
GenAI
Image courtesy of IBM

At a Glance

  • Not all models are born equal but the right path is more about the right set of practices/tools than the specific LLM choice
  • Watch the video interviews with IBM experts and Omdia’s Chief Analyst for AI below to learn more

These solutions simply won’t sit still long enough to accommodate a traditional 6-12 month development cycle before becoming wholly obsolete.

The speed, lure and promise of these new large language models (LLMs) may also lead many to believe strange things – that frontier-scale models, like OpenAI GPT, could be plugged into any number of use cases using nothing more than some clever prompt engineering – or that that LLMs can be readily chained together to automate complex workflows autonomously, a notion commonly referred to as "agentic" computing.

The reality is of course, not nearly that simple – and no enterprise wants to see today's seemingly solid expectations turn into tomorrow's technical debt.

What, then, is an enterprise to do? There is no standing still, and there is no going back, something highlighted by a recent Omdia study of more than 1,500 enterprise IT practitioners, which found that 52% of companies currently allocate between 11% and 30% of their AI budget to GenAI, signalling increased confidence in its transformative value. For many companies, this still nascent technology already serves as a critical area of investment.

To help demystify the picture and bring practical advice to the community, Bradley Shimmin (Omdia Chief Analyst for AI) sat down with three IBM experts to get their take on the different key aspects of finding the right approach.

Model Selection: Dr David Cox (VP, AI Models; IBM Director, MIT-IBM Watson AI Lab, IBM Research)

How are models evolving, what makes a model ‘fit for purpose’, and what are the implications for models amidst the focus on AI agents?

Openness: Anthony Annunziata (Director, AI Open Strategy and the AI Alliance, IBM Research)

Why is open/openness such an important idea, how can it help developers, and how can partner organizations help support the enterprise build applications?

Responsible AI and Governance: Dr. Maryam Ashoori (Senior Director of Product Management and Head of Product, watsonx.ai, IBM Software)

Why should enterprises pay attention to responsible AI, what capabilities are key for ensuring governance, and how do we ensure adherence to company values?

Conclusion

Picking the right GenAI model for the right job, helping developers build outcomes more effectively, and focusing on accountability, transparency, etc. all sound reasonably straight forward. But what are the first principles underlying each? What are the core truths that will help companies maximize their GenAI investments when every day GenAI seems to reinvent itself, technologically speaking.

Across all three IBM conversations above, there are three principles that emerge:

  • First, every LLM is unique. They are by their very nature fit for purpose depending upon how they're constructed, trained, and aligned – each with price and performance considerations. Companies must carefully select each model, ensuring that it can adequately complete a given task, be that extracting named entities from raw text, generating code, for playing a particular role within a complex, agentic system. . Choosing models that have been pre-trained on a transparent and open data sets, therefore, is essential for avoiding technical debt in the future, as well as maintaining consistency over time, and ensuring alignment with corporate policies and external regulations.

  • Second, companies must invest in developers’ skills and tools.  For example, the adoption of an open AI platform that's capable of speeding development without locking users into a specific set of tools, and typically one with a thriving open-source community that beats at the very heart of AI as both an engine of innovation and as a means of engineering trust and ensuring security and robustness.

  • Lastly, companies must broaden their view of responsible AI. Many already emphasize the governance of model outputs using both humans and LLMs – a major lever to de-risk GenAI systems and ensure some degree of compliance with policies and regulations.  However, the true task of de-risking GenAI extends back to the planning stages of development – from understanding the data engineering pipeline in training, to to tracking iterations in a well-documented, explainable, auditable manner. Such transparency will prove crucial for building trust in GenAI systems and ensuring that they are used ethically and responsibly over time.

Sign Up for the Newsletter
The most up-to-date AI news and insights delivered right to your inbox!

You May Also Like