ChatGPT has become more powerful throughout its first year of iterations. It responds better to most prompts (questions and directions), and new capabilities help reduce hallucinations.
Reflecting on the progress since the popular LLM was released to the public Nov. 30, 2022, we chatted with FactSet’s Lucy Tancredi, Head of Strategic Technology Initiatives, and Ruggero Scorcioni, Director of Machine Learning. Below they highlight the key developments and what’s next for artificial intelligence at FactSet. (Note: This article does not address news of OpenAI’s leadership and governance changes.)
Addressing a major criticism that the original release of ChatGPT was unaware of recent events due to a September 2021 training data cutoff, training data up to April 2023 is reflected in the newest model, GPT-4 Turbo.
The model also allows longer text prompts—more than 300 pages at a time—with faster response times and a lower price tag for input and output tokens. The expanded prompt capability is a game changer for anyone using it to summarize, analyze, or translate substantial texts such as research documents, books, code bases, or meeting transcripts. Instead of passing in their content in multiple batches—which was time-consuming and may not have yielded the best results—users can finish in just one step now.
Individuals who want ChatGPT Plus to use up-to-the-minute data no longer need to enable a web plugin. Unlike a search tool that returns multiple pages of results to click through, the LLM will format internet-sourced responses as the user directs—in a bulleted list or brief paragraphs, for example.
Plugins don’t eliminate hallucinations, but they do a much better job reducing them. Plugins also provide access to third-party data, such as grocery orders, email systems, and air/car/hotel reservations.
Custom versions of ChatGPT enable users to build, without coding, single-purpose GPTs (generative pre-trained transformers, a type of Large Language Model). Think of GPTs as virtual assistants in your daily life—technical support, an editor for personal emails, or a coach for prompt engineering. Businesses can also build (or provide employees access to build) GPTs for internal uses that keep their proprietary information secure. (The data will still travel to OpenAI or Azure servers.)
ChatGPT Enterprise was released in August to offer enterprise-grade security and privacy to businesses. It now also offers financial/legal protection called Copyright Shield for claims of copyright infringement, a potential catalyst for businesses wary about using AI technologies due to legal risks.
You can speak to ChatGPT and show it pictures through an intuitive interface with voice and image capabilities in GPT-4V. Individuals who are blind or visually impaired can prompt the tool to read a subway map, summarize a restaurant menu, or give directions to a nearby product in the grocery. Travelers can ask it to translate signs and some other foreign-language content. And here’s an impressive example of the model not only creating working code from a visual mockup but doing it from a rough sketch whiteboard session and properly interpreting the in-the-moment design decisions.
Generating images and designs became easier when the DALL-E 3 image generation model was incorporated in ChatGPT and ChatGPT Plus. The latter is automated to search the web, understand images, and generate a combination of text and images—without the need to select specific details.
Software developers can save significant time by asking ChatGPT to explain poorly written code, locate and repair buggy code, and perform quality-assurance testing in software. It can also help stem the tedium of writing documentation while developing code.
Users can apply customized settings so that ChatGPT provides personalized assistance. For example, an individual could reflect their expertise in machine learning (but not marketing) and their preference for brief, bulleted responses.
We believe the winners in this environment will have the broadest suite of data, the most well-connected data, the most trusted data, and data that can be source-linked back to where the answers came from.
We are reimaging the FactSet experience and actively exploring innovative solutions. For example:
A conversational user interface that allows bankers to ask questions, discover and source information, and initiate tasks.
On the buy side, we are enhancing our portfolio manager bot to answer questions in conversation with asset managers.
In the front office, we are harnessing generative AI to create code in FactSet's programmatic environment, reducing the need to know Python. This will make the power of that programmatic environment available to more users.
For wealth managers, we are developing solutions to drive the next best action and to create portfolio summaries for proposal generation and client engagement. And we are establishing generative-AI-ready data bundles, allowing clients to augment their own Large Language Models with our connected, auditable data.
The first wave of solutions is undergoing client testing as part of FactSet Explorer, our product preview program, as we remain committed to our client-first product development strategy.
Learn more in the FactSet AI Blueprint, our product roadmap designed to leverage the power of AI to further enhance our industry-leading solutions and deliver unparalleled levels of personalization, discoverability, and productivity to our clients.
For more of our perspective on AI, check out these select articles:
This blog post is for informational purposes only. The information contained in this blog post is not legal, tax, or investment advice. FactSet does not endorse or recommend any investments and assumes no liability for any consequence relating directly or indirectly to any action or inaction taken based on the information contained in this article.