Coming off the 2023 “AI hype cycle”, how should business leaders approach 2024? Learn 5 practical AI strategies every business should consider.
Practical AI Strategies for 2024
2024 is here! So how should business and tech leaders approach 2024 and beyond?
After talking to hundreds of leaders and working on a variety of machine learning projects, including fine-tuning multiple open source LLMs, here are 5 sensible approaches leaders should implement now to maximize your ROI on data science, future-proof your organization, and deepen your moat in 2024 and beyond.
Strategy 1. Start Implementing Domain-specific LLMs
Here’s a fact: Open source models are rapidly converging on proprietary models in terms of performance and cost. At the same time, in the last 6 months, ~$1 billion in VC funding has flowed into open source platforms such as Hugging Face, Together AI, Replicate, and OctoML to support open source model builders. Mistral, an open source model provider, alone raised $415m. These are strong market signals.
What are the motivations to consider open source models? Organizations want to: (1) control their destiny/ avoid platform dependencies on proprietary models; (2) protect their privacy and data; and (3) customize models to their unique use cases and requirements.
Leaders and data teams must start implementing domain-specific LLMs to understand their benefits and costs. Practically speaking, this means examining LLM use cases, researching open source models, and gaining experience with the technique of transfer learning (see below). Teams that create structured, iterative, and repeatable processes building and deploying LLMs will generate significant value.
Strategy 2. Deepen & Augment Your Data Strategy
One common refrain we hear: “We don’t have enough data to get started”. This is unlikely to be true. Transfer learning is the practice of using a pre-trained model on a new machine learning task that has a lot less data. Data teams have the convenience of bypassing initial model training steps and recomputing one or more layers at the end of a neural network for a specific use case.
If you have industry experience and expertise, you most likely have enough first-party (proprietary) data to get started. Leaders should seek to augment with 3rd-party data sets and real-time sources from their industry or adjacent industries. Data quality, diversity, and provenance matter much more than big data and organizations that learn to create and apply meaningful corpora, data sets, and sources will accelerate their growth.
Strategy 3. Adopt a “Data Agile” Approach
Software engineering has benefited tremendously from agile methodologies: roadmapping, sprint planning, continuous integration/ deployment (CI/CD), etc. The same could be true of machine learning. For security and privacy reasons, there’s a default tendency to lock down data.
Yet, there are new tools and databases that enable secure data collaboration while supporting data augmentation strategies and boosting data quality. Data agility means teams can maximize the time value of data - the idea that data has its greatest value when it is first generated (and erodes gradually over time).
Data agile teams reduce time to insight for users while protecting privacy and maintaining security practices.
Strategy 4. Automate with Reinforcement Learning
Reinforcement learning (RL) is an incredibly valuable machine learning framework that is often underutilized in business. It’s particularly valuable to automate tasks, especially when supervised learning is not an option. It’s essentially training an algorithm to learn by trial-and-error.
RL offers a path to automating many labor-intensive tasks, specifically in marketing, cybersec, fraud detection, cloud spend optimization, and energy consumption. For example, RL algorithms can personalize and tailor content and recommendations to individual users in real-time, based on their interactions and behavior.
Strategy 5. Empower End-to-End Data Teams
It’s time to end silos that hamper so many data teams. Silos show up as database silos, specialization silos, and team silos.
Leaders should work to end messy “hand offs” where data teams hand off productionizing and deployment to separate teams like devops. The more dependencies and barriers to deployment, the less you’ll ship. Leaders have to empower data teams to own the end-to-end process from exploratory data analysis to deployment and scaling and monitoring. New tools and databases exist to accelerate the process, no matter your cloud provider. Give data teams the ability to make an impact!
Bottom Line: Ship More Models
“If it’s not in production, it doesn’t exist.” We heard this directly from a senior product manager responsible for deploying models in an insurance startup disrupting the industry with real-time machine learning and automation.
The organizations that are best suited for 2024 and beyond are those with a singular focus: ship more models. That means getting more value and insight from unstructured data and generating valuable feedback loops to improve models and deepen your moat. The longer you wait, the less experience - and data - you’ll have.
Now is not the time to sit on the sidelines, but to act. Your competitors already are.Photo by Ian Simmonds on Unsplash