Unlike foundational models, traditional AI deep learning models that use algorithms such as recurrent neural networks and convolutional neural networks are narrow in scope and used for specific cases. For example, a manufacturing company might implement AI with deep learning computer vision systems for warehouse inventory management. The underlying models are trained on large and diverse datasets, so the use cases are broader. This does not necessarily benefit the quality of the output. However, if humans can fine-tune the model or be more specific in the instructions they give to the model, the underlying model can be more closely aligned with specific use cases.
It is only recently that companies have begun commercializing these models on a large scale. Our forecast is that the generative AI market will expand at a compound annual growth rate of 58% from 2023 to 2028, reaching $36.4 billion by the end of 2028. The underlying models segment, expected to generate $11.4 billion in revenue by 2028, is relatively unconsolidated due to the significant resources and budget required to pre-train models that can keep up with leading underlying models. The situation will continue. This increases the barriers to entry and means that large technology companies, which already have significant computational resources, will capture a large share of the underlying model market. Most startups build on these models by adding services or providing additional training to optimize performance.
The tech giant's advantage will be slightly offset by regionally based models trained on local language and context, such as Abu Dhabi's Falcon, an open source LLM. The advancement of regionally-based models is partly the result of political moves to build domestic technology industries in countries outside the United States, but biases arise when data collection and human review are conducted only in the United States. It also reflects concerns about the possibilities.
The power of agile engineering
Prompt engineering is about communicating orders and inputs in such a way that the underlying model works. Providing accurate and valuable prompts, usually in text format, is very important as it directly affects the quality of the model's output. For example, high-quality prompts typically include information about the role the AI model should adopt when providing an answer. The prompt requires him to focus on two things. A detailed, clear, and specific description of the task. Context including required format and length. According to the International Monetary Fund, the benefits of instant engineering are manifold. In particular, it increases the accuracy of the text produced, provides some control over the output, and, importantly, reduces inherent bias.
Key risks of generative AI and underlying models
Generative AI using foundational models has great potential, but risks and limitations require a human-in-the-loop approach. Some of the most common potential risks of foundational models include:
-
Hallucination: The ability of the underlying model to provide fictitious responses, which can result from a lack of context during the prompt, bias in the training data, low-quality training data, and other causes.
-
Inaccuracies in output (such as outdated or limited information): The current architecture of the underlying model means that it is susceptible to inaccuracies, which can be alleviated by adding deterministic controls such as vector databases that base their responses on real data. Masu.
-
Exploitation: Malicious uses of AI, such as generating deepfake images or videos, or using generative AI for cyber attacks.
-
Biased answer: The data on which the underlying model is trained contains human biases (such as gender, political, or racial), the training data is not updated regularly, or it is simply not diverse enough. , output that may be discriminatory or unfair. Other concerns about bias may be related to the entity behind the model. For example, companies and governments may decide to incorporate cultural sensitivities into their models, which can lead to bias.
-
Lack of transparency: Humans cannot explain how a model arrived at a particular solution or reproduce the model's responses.
-
Intellectual property concerns: The model may produce results similar to existing content, potentially leading to copyright infringement. U.S. case law indicates that machine-generated content cannot be protected by patents or copyrights because it is not created by humans.
-
Data privacy: Concerns that sensitive information may be leaked to externally available models.
-
Infrastructure requirements: Generative AI requires significant computational and network resource demands, increasing infrastructure spending and hampering sustainability efforts.
-
Copyright infringement: Several lawsuits in the United States focus on copyright infringement due to the alleged illegal use of media to train underlying models. The model creators claim fair use, but copyright holders such as Getty disagree.
Governments and businesses need effective and flexible governance frameworks to build trustworthy AI
As Professor Melvin Kranzberg said in 1986, “Technology is neither good nor bad, nor is it neutral.” This statement is especially true in the case of AI. The rapid development of AI does not yet have a commensurate level of oversight, whether at a supranational, national, or corporate level. However, things are changing, albeit slowly.
At the national and regional level, several AI regulations are under development, including the EU AI Act, a framework expected to be completed by the end of 2023. Similar regulations are being developed in the United States and Asia. Additionally, the development of generative AI has led to new guidance, such as the Organization for Economic Co-operation and Development's recently announced “His G7 Hiroshima Process on Generative Artificial Intelligence.”
A flexible approach is needed as policymakers and businesses need to build reliable AI frameworks to manage risks and other potential pitfalls that we are unaware of. Regulations and guidelines for AI governance can vary around the world, but they determine the rules of the game.
Outside of business, the rise of generative AI and its potential ubiquity offers great potential to help solve problems such as climate change, world hunger, disease, education and income inequality, and the energy transition. For example, technological advances could improve quantum technologies and enable “digital experimentation” of physical processes such as nuclear power generation. While the potential for good is virtually limitless, so is the potential for harmful outcomes, whether intended or not. As such, generative AI requires a robust, human-driven, regulated ecosystem to ensure that its highly disruptive nature leads to positive outcomes.
Related Research
-
451 Research's Generative AI Market Monitor, June 7, 2023, S&P Global Market Intelligence.
-
451 Research Machine Learning: Fundamentals, November 28, 2023, S&P Global.
-
“AI Governance Challenge,” November 28, 2023, S&P Global.
-
Generative AI use cases could power document and content management software, September 13, 2023, S&P Global Market Intelligence.
external investigation
-
FinBERT: A Large-Scale Language Model for Extracting Information from Financial Text, September 29, 2022, Contemporary Accounting Research.
-
“Technology and History: 'Kranzberg's Law'”, Technology and Culture 27 (3), July 1986, Melvin Kranzberg.
-
G7 Hiroshima Process on Generative Artificial Intelligence, September 7, 2023, OECD.
- Technology Innovation Institute Introduces World's Most Powerful Open LLM: Falcon 180B, September 6, 2023, Technology Innovation Institute.
- “Generative Artificial Intelligence in Finance,” August 22, 2023, International Monetary Fund.
