A friend of mine asked me what I think the generative AI model landscape will look like in a few years. Will there be one model or many models?
Even with new model releases that have massive performance gains, there will be many different models as long as there are no massive breakthroughs that massively changes how AI is developed and works.
Why? It comes down to different business use cases having different requirements. The main requirements that we see today are the following.
This applies to all potential use cases. Over time companies will optimize their use case to use the model that gets the job done at the cheapest cost. Generally, the more parameters in the models the more expensive it will be to run. This means that there will always be a market for different-size models. Companies will be looking for the smallest model that performs their use case to a satisfactory level.
This means that there will always be a market for use case specific models.
Today we can see differences in models' performance on real-world tasks. If the current trajectory holds, we may end up having different models focused on different tasks. However, note that I believe this will change for LLMs and expect that a new model architecture will be released that is generally better at every real-world task and practically erases this dimension.
However, for the different image generation algorithms, the different approaches are suited for different use cases and ways of working. Because of this, it is unlikely that they will merge into only one way of creating images/videos.
This is one of the main concerns for companies in highly regulated industries. Regulations for industries such as healthcare and finance vary from country to country, and data must be processed locally in many places. This means that there will always be a separate market for those that provide local models vs. cloud-only models. The most advanced models are likely to only run in cloud environments due to security concerns of the models.
Similar to data privacy, there are companies that do not trust giving access to any of their internal data to other companies. This means that they will need to have control of the models to be able to use them. This is somewhat solved by cloud platforms making all different models available. For example, if you use Microsoft could services for other things, then there is no reason to not trust running AI models hosted on their cloud.
More and more people are waking up to the ethics of training AI models. Whether you are allowed to use copyrighted material or material that you don’t have consent for training varies from country to country, but is also a hot ethical debate topic.
I believe some companies and consumers will want to only use models trained ethically. This means that there will be a separate market for those models that get some form of stamp of ethical approval.
Another topic of hot debate in the AI community is the open vs closed models. There are different levels of transparency, some companies like Open AI are the opposite of open and hide everything from model weights to training material and model architecture. Others open one or more of these. One trend is to publish model cards that explain what data a model is trained on. Only a select few companies are fully open providing both source code, architecture, and training data. Link to LLM model directory here.
Even though it’s possible we will reach a situation where a single model theoretically could be used for all use cases, the differing cost and regulatory environments mean that there will always be many models available.