ChatGPT has demonstrated significant advancements over its inaugural year of iterations, exhibiting improved responsiveness to various prompts and directions while incorporating new features to mitigate hallucinations.
Looking back on the evolution since the renowned LLM debut to the public on November 30, 2022, we engaged in a discussion with Lucy Tancredi, FactSet’s Head of Strategic Technology Initiatives, and Ruggero Scorcioni, the Director of Machine Learning. They shed light on the pivotal developments and the future trajectory of artificial intelligence at FactSet. (Please note that this article excludes any updates regarding OpenAI’s leadership and governance alterations.)
Throughout its initial year of expansion, what notable capabilities has ChatGPT acquired?
One significant enhancement addresses a critical critique from the original ChatGPT release, where its unawareness of recent events stemmed from a training data cutoff in September 2021. The latest model, GPT-4 Turbo, now incorporates training data up to April 2023.
Moreover, the model facilitates the processing of lengthier text prompts exceeding 300 pages at once, offering quicker response times and a more cost-effective approach for input and output tokens. This expanded prompt capacity revolutionizes tasks such as summarization, analysis, or translation of extensive texts like research papers, books, code repositories, or meeting transcripts. Users can now complete these tasks in a single step, eliminating the need for multiple batch submissions, which were time-consuming and potentially less effective.
Individuals utilizing ChatGPT Plus for real-time data no longer require a web plugin for activation. Unlike conventional search tools that yield multiple result pages for users to navigate, the LLM formats internet-derived responses according to the user’s preference—be it in a bulleted list or concise paragraphs.
While plugins do not entirely eradicate hallucinations, they notably reduce their occurrence. Additionally, plugins grant users access to third-party data sources such as grocery orders, email platforms, and travel reservations.
Tailored versions of ChatGPT empower users to create specialized GPTs without the need for coding. These GPTs can serve as virtual assistants in various aspects of daily life, such as technical support, email editing, or prompt engineering coaching. Businesses can also develop or provide employees with the capability to build GPTs for internal functions while safeguarding proprietary data. (Note that data will still transit to OpenAI or Azure servers.)
In August, ChatGPT Enterprise was introduced to provide businesses with enterprise-grade security and privacy features. It now includes financial and legal safeguards like Copyright Shield, addressing concerns about potential copyright violations—a significant reassurance for businesses cautious about adopting AI technologies due to legal uncertainties.
The introduction of GPT-4V enables users to interact verbally with ChatGPT and share images through an intuitive interface. This feature allows visually impaired individuals to request assistance in reading maps, summarizing menus, or locating products in grocery stores. Travelers can utilize this functionality to translate foreign language content and signs. An impressive demonstration showcased the model’s ability to generate functional code from visual mockups, even interpreting spontaneous design decisions made during a whiteboard sketch session.
The integration of the DALL-E 3 image generation model in ChatGPT and ChatGPT Plus has streamlined the process of creating images and designs. The latter autonomously scours the web, comprehends images, and generates text-image combinations without necessitating specific details selection.
Software developers can leverage ChatGPT to expedite tasks like code explanation, bug identification and resolution, and software quality assurance testing. It also alleviates the burden of documentation writing during code development.
Users have the flexibility to configure personalized settings, enabling ChatGPT to offer tailored assistance. For instance, users can customize preferences to reflect expertise in machine learning (while excluding marketing) and opt for concise, bulleted responses.
What is FactSet’s vision concerning AI and LLMs?
FactSet envisions success in this landscape for entities possessing comprehensive, well-connected, trusted data that can be traced back to its source.
We are reimagining the FactSet experience and actively exploring innovative solutions, including:
- Introducing a conversational user interface for bankers to inquire, source information, and initiate tasks.
- Enhancing the portfolio manager bot to engage in conversational interactions with asset managers on the buy side.
- Leveraging generative AI in the front office to generate code within FactSet’s programmatic environment, eliminating the necessity for Python proficiency and democratizing access to this powerful tool.
- Developing solutions for wealth managers to drive optimal actions, create portfolio summaries for proposal generation and client engagement, and establishing generative-AI-ready data bundles to augment clients’ Large Language Models with our interconnected, auditable data.
The initial set of solutions is undergoing client testing as part of FactSet Explorer, our product preview initiative, underscoring our commitment to prioritizing client-centric product development.
For a deeper insight into our AI perspective, explore the FactSet AI Blueprint, our product roadmap designed to harness AI’s potential to enhance our leading solutions, delivering unparalleled levels of personalization, discoverability, and productivity to our clientele.
For further insights on AI, refer to the select articles available from FactSet.
