The present generative artificial intelligence "gold rush" exists for a good reason: the Rule of Three. In most instances, a stable competitive market never has more than three significant competitors, the largest of which has no more than four times the market share of the smallest,” BCG founder Bruce Henderson said in 1976.
The caveat is that the rule does not work as well to industries that are unstable or heavily regulated, such as investment banking; consumer electronics; some parts of the IT software and services business; life insurance and parts of the telecommunications industry.
Sometimes known as “the rule of three,” he argued that stable and competitive industries will have no more than three significant competitors, with market share ratios around 4:2:1.
So one has to assume the same pattern will emerge for frontier GenAI models as well.
In 2023 alone, some 123 artificial intelligence foundation models, the building blocks of many modern AI applications, were released. In 2024, there may well be thousands of models in use, including both the smaller number of "foundation models" that lead the market, a larger number of general-purpose generative AI models that might be important in verticals, plus the many thousands of models that have been customized for use by specific enterprises.
Virtually nobody believes all the would-be foundation models will survive, long term. And that fuels the "gold rush" mentality: only a few foundation models are likely to emerge as eventual market leaders, as tends to be true in any market.
While application markets tend to exhibit more diversity over the long term, compared to operating systems or semiconductor chip ecosystems, a reasonable argument can be made that, over the long term, just a handful of leading foundation models will lead the market, as that is the pattern in computing in specific and almost all markets generally.
Among the dozens of large foundation models that seem to be most used are the GPT series (OpenAI); the Claude series (Anthropic); PaLM and Gemini series (Google) as well as the LLaMA series (Meta). But there also are many small language models developing that generally are designed for specific purposes.
In healthcare, SLMs might be used for medical document analysis, patient record summarization or perhaps research. In finance, SLMs might be used for fraud detection, sentiment analysis of financial news or risk assessment.
For customer service, SLMS might be used for chatbots. The point is that where LLMs were previously required, in the future SLMs might suffice.
SLMs will be favored because they cost less. Training and deployment are more affordable, for example, since the models do not have to be trained on much-larger datasets. In some cases, SLMs can be developed faster.
SLMs will be favored for industry-specific use cases, as well.
Because there is less processing, there also is less energy consumption. Some argue SLMs also enable more privacy.
No comments:
Post a Comment