In the early days of this newsletter, we briefly touched on the choice before any corporate wanting to work with generative AI technologies: should they build their own generative AI models, opt for commercial solutions, or leverage open-source alternatives? This choice can significantly impact a company’s AI strategy, resource allocation, and competitive advantage. In this issue, we will explore the pros and cons of each approach, delve into the process of building your own AI model, and examine real-world case studies to provide practical takeaways.
Focus On: Commercial, Open Source, or In-House
One of the first things to assess at the start of any generative AI journey, from pilots to transformative initiatives, is whether to build the model in house or if looking at third parties, whether to go commercial or open source. Let’s quickly summarise pros and cons for each:
Commercial Models
Commercial generative AI models, such as OpenAI’s ChatGPT or Anthropic’s Claude, provide ready-to-use solutions with robust support and regular updates. These models are pre-trained on vast datasets and can be quickly integrated into existing workflows, although only a subset of these models can be fine-tuned.
Pros:
- Near-instant deployment
- Ongoing support and updates
- Reliability and performance
Cons:
- Limited customization options
- Potential data privacy concerns (these are all “black box AIs”)
- Recurring licensing costs
Open Source Models
Open-source generative AI models, such as those available on platforms like Hugging Face, offer a middle ground between commercial and in-house solutions. These models are freely available, community-supported, and can be customized to some extent. In many case state backed entities (like in the case of UAE’s Falcon) or large corporates (like the in the case of Meta’s LLAMA) are behind these models, so don’t necessarily think open source means produced by indie developers as reality is far from that.
Pros:
- Cost-effective: you only pay for the power to run the model, not to licence it
- Community-driven improvements
- Flexibility for modification and fine tuning
Cons:
- Limited specialised support
- Potential security vulnerabilities and most are still “black box AIs”
- May require significant in-house expertise to implement effectively
In-House Development
Building your own generative AI model allows for maximum customization and control over the entire process. This approach can lead to highly specialised solutions tailored to specific business needs.
Pros:
- Full control over model architecture and training data
- Ability to incorporate proprietary data and domain knowledge, embedding IP into the model itself
- Potential for unique competitive advantage and long term moat against competition
Cons:
- Hugely resource-intensive (time, expertise, and computing power)
- Ongoing maintenance and updates required
- Higher upfront costs
Building Your Own AI Model: A Closer Look
Those brave enough to venture into the build your own AI model space, have a few options at their disposal. Some enable relatively easy implementations, others offer the greatest flexibility at the expense of time and cost. Let’s see these options in more detail:
No-Code/Low-Code Platforms
These platforms, such as Pecan, offer a user-friendly interface for building AI models without extensive coding knowledge. They’re ideal for business professionals and data analysts who want to quickly prototype and deploy AI solutions.
Key Features:
- Drag-and-drop interfaces
- Pre-built templates and workflows
- Integration with common data sources
Example: A marketing team could use a no-code platform to build a customer churn prediction model, leveraging existing CRM data without the need for advanced programming skills.
Automated Machine Learning
Automated Machine Learning (AutoML) platforms like Google AutoML and Azure AutoML offer a balance between ease of use and customisation. These tools automate many aspects of the model development process while still allowing for some level of control.
Key Features:
- Automated feature engineering and selection
- Hyperparameter tuning
- Model evaluation and comparison
Example: A financial services company could use AutoML to develop a fraud detection model, streamlining the process of algorithm selection and optimization while still incorporating domain-specific knowledge.
Traditional Programming and Machine Learning Libraries
For organizations with advanced technical expertise, building AI models from scratch using libraries like TensorFlow, PyTorch, or scikit-learn offers the highest level of customisation and control.
Key Features:
- Complete flexibility in model architecture
- Fine-grained control over training process
- Ability to implement cutting-edge techniques
Example: A pharmaceutical company might build a custom deep learning model for drug discovery, incorporating proprietary molecular data and specialized algorithms not available in off-the-shelf solutions.
In the News
Several companies have successfully built and implemented custom generative AI models to address specific industry challenges, demonstrating you can effecitvely mix commercial, open-source and in-house models depending on the use case.
Lifespan, a healthcare system, utilized GPT-4 to simplify consent forms, enhancing patient understanding while reducing document length. Minijob Zentrale (MJZ) in Germany partnered with IBM to create an AI editorial assistant that significantly cut down the time required for content creation, while Booking.com developed an AI model to provide personalized travel itineraries, leveraging their extensive travel data to enhance user experience. Final example, Moderna employed custom AI solutions for various tasks such as reviewing clinical data and summarizing contracts, leading to improved efficiency in their pharmaceutical operations.
Takeaways
Based on a few conversations I recently had with subject matter experts in this space, here are key considerations for business leaders when deciding the best AI model approach:
Assess Your Data Assets
If your company possesses unique, high-quality data that would provide a competitive advantage, building your own model might be worthwhile. Proprietary data can lead to more accurate and specialized AI solutions, effectively building a moat by embedding IP into the model itself.
Evaluate In-House Expertise
Building AI models requires specialized skills. Assess your team’s capabilities in data science, machine learning, and software engineering. If you lack the necessary expertise, consider upskilling your team or partnering with IT/data companies with a proven track record.
Consider Time-to-Market
If you need to deploy an AI solution quickly, commercial or open-source options may be more suitable. Custom development can take months or even years, depending on the complexity of the model.
Prioritize Scalability and Flexibility
Consider how your AI needs might evolve. Building your own model can provide greater flexibility to adapt to changing business requirements, but it also requires ongoing maintenance and updates.
Address Regulatory and Ethical Concerns
In industries with strict regulations or when dealing with sensitive data, building your own model may provide better control over data privacy and compliance. Ensure that any AI solution, whether built in-house or sourced externally, aligns with ethical AI principles and relevant regulations. When going commercial or open-source, keep in mind many consider black box AIs as having no future. These models are ok for pilots and learning experiences, but for mission critical production workflows I recommend partnering with providers who open up their models to independent verification.
Start Small and Iterate
If you decide to build your own model, consider starting with a smaller, focused project using no-code or AutoML platforms. This approach allows you to gain experience and demonstrate value before committing to more complex, resource-intensive development.
Follow me
That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.
Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.