Generative AI, the latest advance in a decades-long journey towards machine intelligence, has exploded into the public discourse since the release of ChatGPT in November 2022. This milestone marked an important shift in the development of AI models, enabling non-technical users to engage with them in more natural ways – using plain language rather than code. As you no longer need to be an engineer or a researcher to interact with these models, the barrier to entry for development, use, and adoption of AI keeps falling.
This has caused a great deal of enthusiasm about increased productivity, novel products, analytical tools, and even the potential for the betterment of society. Yet, the question remains: is this progress as transformative as it seems?
AI is not new, of course. The original programs that we would call “AI” were developed more than 70 years ago for the sciences. The first such program, Allen Newell and Herbert Simon’s “Logic Theorist,” created in 1956 (!), was designed to prove mathematical theorems by mimicking human problem-solving and reasoning abilities. The program used a method known as „heuristic search“ to explore possible solutions, selecting paths that seemed most likely to lead to a correct proof, much like a human mathematician might. Its major success was a ground-breaking demonstration of how a computer could perform tasks traditionally thought to require human intelligence, leading to rapid development and many investments in “AI” over the following decades, both in the U.S. and around the world.
While most have learned of “neural networks” only recently, the original multilayer neural networks date back to the 1960s. Though – remarkably enough – they were trained using similar methods to today, these early efforts could not scale, for several reasons. Only following the advent of the Internet and the associated production of massive amounts of machine-readable data, together with improvements in the training algorithms and decrease in commodity computation costs, did these models start to become truly useful in the mid-1990s.
While the impact of AI has already been felt in many sectors, its adoption is decidedly uneven, as concerns about misinformation, job displacement, and the ethical implications of advanced artificial intelligence are vigorously debated across industry, academia, and policy venues. Anybody looking to understand the trajectory of development of this technology would be wise to pay attention to the rapidly evolving regulatory landscape, such as California’s proposition of SB 1047 and the specifics in the nascent implementation of the EU’s AI Act.
Advanced applications of AI are no longer a novelty in many industries. The first applications of machine learning and AI in the financial domain date back to the early-to-mid-2000s. At Bloomberg, the first products enabled by AI, such as topic modeling and sentiment analysis, date back to 2009.
In the following 15+ years, the financial industry itself and its appetite for advanced AI has changed dramatically. As new talent has entered the workforce, the demand for labor saving devices, tools for data analysis, and more sophisticated models has increased. The advances in open source technology, as well as the exponential growth of computational capacity and machine-readable data has led to a proliferation of sophisticated data-driven models and a shift towards tailored, bespoke advanced analytics as the amount and variety of data needed for rigorous decision-making has increased beyond the ability of any one human being to consume or process it. However, the adoption of generative AI models is still very much work in progress, as many institutions and industries continue to struggle to identify their killer app.
While these new generative AI models are powerful and increasingly capable, they also have important limitations. These models are enormous; they require increasingly large datasets and massive amounts of computation for training. By virtue of their size, as well as their mathematical structure, they lack interpretability required for high-stakes decision-making, where errors can have severe consequences. They are still fragile, often exhibiting unexpected failure modes like hallucinations, thus lacking robustness needed in financial workflows. They can encode inconvenient biases by memorizing training data (as they tend to be vastly overparameterized), and thereby can be unable to account for rare events or complex market dynamics involving regime shifts or other types of non-stationarity. They are not (yet) capable of complex reasoning, particularly involving mathematical rigor or temporal relations. As a result, while gen AI shows great promise and can be very useful as an assistant in a variety of workflows, its deployment in mission-critical domains, like finance, requires careful thought, rigorous testing, and ongoing human oversight.
Through the experience accumulated over 15 years of AI-driven product development, we at Bloomberg stand in a unique position to appreciate both the opportunity and the limitations of this new generation of AI. During this time, the industry has used AI in many financial workflows – to extract insights and value from unstructured data, to create novel bespoke analytics (everyone has a unique opinion on what “ESG sentiment” ought to mean, for example), and to improve productivity (e.g., via helping users better navigate complex documents like earnings call transcripts and analyst reports). Many of today’s applications of AI in finance use a retrieval-augmented generation (RAG) paradigm to sidestep the limitations of the current generation of gen AI models.
But this is not a panacea, and the development of these applications is a complex undertaking, far more involved than just typing prompts into ChatGPT or a similar chatbot. It requires deep domain knowledge on the part of the developers, product managers and data experts, carefully engineered datasets that are specific to the problem, and sophisticated UX design in the solution to provide attribution, transparency, and robustness needed to support high stakes decisions. With all of these ingredients, you can do remarkable things.
For example, by using these models as assistants for software development or data work like annotation or key facts extraction, analysts and developers can save up to 90% (!) of human effort on repetitive, tedious tasks. Even so, you do need to be a proficient user of these tools to be able to extract value from them. In that, these models are less revolutionary and more evolutionary: the right tools for the right job.
So what’s the verdict? What does the future hold for AI?
While it is becoming increasingly difficult to forecast the trajectory of AI development (the pace of improvement is frantic), I believe there are things we can confidently anticipate. The most recent versions of open source AI models (such as LLaMa 3.1) have essentially closed the gap in performance with the private models (such as GPT-4o). One of the striking things about the recent model releases was the realization that the smallest LLaMa 3.1 model (which still has 8 billion parameters) is practically on par in performance with the GPT-3.5 turbo. Size, clearly, is not everything. This means that, in the very near future, powerful models like these will be able to run on a laptop or smartphone, and will be personalized as an assistant to each individual. One has to wonder, naturally, about the implications of such assistants particularly in education and in the workplace.
We can also easily anticipate these models becoming increasingly multi-modal – that is, capable of dealing with disparate data types, not just limited to text. We already have models that can jointly understand images and text. What will the models that understand text, images, and structured data like time-series or tables be capable of?
Nonetheless, they will still lack the capability to set goals. This means human supervision and know-how will remain indispensable for the foreseeable future, regardless of how powerful AI gets. The key will be in how we integrate these powerful new tools into our industries and society in a responsible way.
While AI shows great potential to improve productivity and enable innovation, we must also be mindful of the ethical and social challenges it presents. Ultimately, the fate of AI will be shaped not just by technical progress, but also by our ability to thoughtfully guide its development to the benefit of everyone.