Rodrigo Madanes, EY Global Innovation AI Leader.
Generative artificial intelligence (GenAI) models can create all kinds of outputs from massive amounts of data to help users with new and exciting applications. However, GenAI presents many challenges and risks that users must understand and address. The technology requires substantial computational resources, adequate ethical standards and sophisticated technical solutions to allow safe, fair and accountable use. GenAI also involves significant collaboration and innovation among different actors and entities.
Stepping back, here are some key themes in the space that are shaping GenAI’s future.
1. Large Language Models (LLMs)
There’s a continuous push to develop better LLMs. Historically, there was an urgency to build LLMs with ever more billions of parameters. Today, developers are paying more attention to the data quality and amount of data used to train the models. This is commonly referred to as scaling laws for LLM training. For example, Microsoft Research recently released a new language model called phi-1.5, which was trained on significantly fewer parameters (1.3B) than most recently released LLMs, and yet was able to achieve comparable results on a series of general-purpose benchmark tests.
In addition, companies are developing a diverse array of LLMs. They are building models with a larger context—or memory footprint to process data—and are investing in alignment, which lets the model better align with users and operate ethically.
2. Compute Scarcity
LLMs and other GenAI models require a lot of computing power to process the large amounts of data and parameters involved. However, large enterprise and AI lab demand for computing resources vastly exceeds supply, particularly for specialized hardware devices such as graphics processing units (GPUs) that power the technology efficiently and cost-effectively. Companies use GPUs for the process of running AI models to generate answers, called inference, as well as for training the models they create.
Compute scarcity poses a challenge for scaling up GenAI models and democratizing their access and use. Startups and academic institutions have struggled to gain access to high-performing GPUs, as legacy technology companies and newly minted unicorns have the capital needed to purchase the GPUs directly or pay expensive cloud computing costs to train their large parameter AI models.
Expect compute scarcity to grow as more complex and powerful iterations of AI emerge and proliferate, such as those that combine text, image and sound, which require even more processing power.
3. Orchestration Layer
AI developers are focusing more attention on improving the technology tier positioned above the LLM. This “orchestration layer” helps simplify and streamline GenAI application development and deployment, while enabling interoperability and modularity among different LLMs and data sources.
Orchestration layers can plug into multiple external databases and run service searches to turbocharge LLMs for business and general user use. They can also “chain” together multiple LLMs and write code that provides far more comprehensive results than simple LLM queries deliver.
This means that users can pose more conversational application queries at the orchestration layer that LLMs trained only on historical data cannot solve. They can, for example, connect to multiple external databases and find a flight to a specific location at a particular time and at a certain price.
LangChain, a framework that allows users to build GenAI applications using a graphical interface and a declarative language, is one example of a commonly used framework that is part of this new orchestration layer.
2. Open Source
More open licensing terms have made the open-source LLM space a popular option. Developing these LLMs involves companies sharing the underlying open-source software to build models.
Comparing open-source LLMs against their proprietary counterparts reveals the pros and cons for each. Proprietary products own the full tech stack supporting the LLM and can often “control” multiple aspects of the stack. Historically, open source has led at times to fragmentation, with spin-offs of the models or the open-source code. There are also discussions about whether open-source models can match the quality of proprietary products.
However, open source has many advantages for advancing GenAI research and development. It can foster more diversity, creativity, quality and accountability in GenAI models. Open source can also lower the barriers to entry and participation in GenAI for various stakeholders. The jury is out on how this competition will play out.
5. Guardrails
Guardrails are essential for ensuring the safety, fairness and accountability of GenAI models. However, they can pose challenges, such as how to define, implement and enforce them, balance them with innovation and freedom, and resolve conflicts or disagreements among different actors or entities.
As fears of AI’s capacity to inflict job losses, spread misinformation and cause harm to grow, national governments listening to calls for regulation have taken steps. In July, President Biden met with leading AI companies to secure voluntary commitments to help with AI risk management. In June, European lawmakers added rigorous amendments to a proposed EU Artificial Intelligence Act conceived to set some AI system development and use constraints.
6. Multi-Modal Models
There have been recent significant commercial advances in multi-modal AI, models that can learn and work with text, images and sounds. A growing proportion of research papers and developments in multi-modal has fueled progress in models that combine more of the senses to detect ever more powerful signals.
Multi-modal models have potential applications in education, entertainment, health care and many other sectors. They can also enable new forms of communication and expression that are richer and more natural than text alone. In one example, a medical model that was trained on scans and physicians’ notes can assess a patient’s condition based on imaging and clinical notes, as well as generate reports with text and images.
In the end, GenAI offers too much promise for firms to ignore or wish away. So, keep apprised of these trends, and others that emerge, to learn how to make this powerful technology work for your company.
The views reflected in this article are the views of the author and do not necessarily reflect the views of the global EY organization or its member firms.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
Read the full article here