It’s happening, and WillowTree’s cross-functional team of experts helps companies harness the power of generative AI technology.
AI is evolving daily. The technology is powerful, and the pace of change is thrilling, but knowing where and how to deploy generative artificial intelligence requires a cross-functional team with the experience to view the problem from all sides. Tools like OpenAI’s Generative Pre-trained Transformer (GPT) language model — and alternatives like Bloom, Chinchilla, Gopher, and LaMDA — are not turnkey solutions but simply single pillars within a larger architecture.
WillowTree’s AI Advisory Services helps your team use tools like GPT to support and supercharge other human- and tech-enabled resources, creating automated solutions powered by AI.
The generative AI space is booming right now, and many entrants are vying for mindshare. That’ll likely continue for the foreseeable future, but let’s quickly get straight on the nomenclature for the most mainstream of the players…
The Generative Pretrained Transformer can be used for natural language processing tasks like language translation, text completion, and text summarization. Its ability to generate human-like copy makes it a game-changer for businesses looking to automate text-based processes and improve customer experience.
The latest version, GPT-4, is trained on 175 billion parameters and is available for purchase as an API for commercial accounts
This is a variant of GPT specifically designed for conversational applications, such as chatbots. Its fine-tuning on conversational data makes it a champ at generating text that's appropriate for chat, and it can even generate engaging and natural-sounding dialogue that'll make customers feel like they're talking to a real person. It’s trained on a more modest 774 million parameters.
ChatGPT is currently free and available to the public, with a more nuanced and powerful (1.2 billion parameters) ChatGPT Plus option now available for $20/month.
The ChatGPT API was released on Mar 1, 2023, allowing developers to integrate the ChatGPT language model into their applications. It’s currently available in varied pricing plans and in multiple programming languages, including Ruby, Node.js, and Python
For the purposes of this page, we’ll mostly refer to the overall GPT language model as a catchall term for the different levels of this AI technology.
GPT may be trained on billions of parameters but much of this data is text-based, and businesses and individuals are currently using it primarily to find, translate, summarize, and complete text-based outputs. But humans create substantial content beyond the publicly available books, stories, articles, and other web copy that GPT is so good at ingesting and transforming.
WillowTree is exploring dozens of use cases where the inputs and outputs for this system are proprietary development code, internal databases, calendars, and other non-text-based functions. We can then take these non-text data sets and build a system of intelligent prompting and re-prompting, fine-tuning the model where necessary. This is where the really interesting, business-critical innovations are happening.
For instance…
We condensed a six-week project into two days."
Michael Freenor
WillowTree Principal Data Scientist
"We created a smart database interface for HR professionals skilled in the DEI space but not necessarily at data analysis or engineering. We’ve set up a system where they can ask, 'What's the gender pay gap at my company?' That data might live in their SQL database; being able to classify where the data lives allows us to selectively pick a prompt for English to SQL translation. We are able to run that query off the underlying database, retrieve the precise result, and re-inject this data into a prompt to answer the original question (posed in English). This allows us to lean on GPT for what it’s good for: translation between languages, relying on pre-existing systems (which have the numeric and logical accuracy GPT lacks) for precise calculations and lookups. Using our custom, branching pipeline of classifiers, prompts, and queries against underlying (precise) systems, we were able to accomplish in two days what would have taken us about six weeks in the previous intent/entity classification paradigm."
Learn more about how Michael’s team increased digital velocity for Sigma Squared by translating database dialects into natural language and bootstrapping a data set for a classifier.
We’re exploring AI in our powerful advisor-matching algorithm."
Chelsea Taylor
WillowTree Product Strategist
"Quiz and survey culture is powerful but also overused as a way of collecting data. We developed an extremely smart tool for matching investors with human financial advisors based on a whole host of parameters – age, race, geography, assets, goals, personality types, etc. — and right now both parties take a three-minute quiz to define those parameters that are fed into the algorithm. Using GPT it only takes ten seconds to more naturally tell the system what you’re looking for, and we can still use the proprietary algorithm to get to that same perfect match.”
Learn more about how Chelsea’s team is optimizing the prospect funnel for a leading financial services firm by removing friction and improving customer experience.
We can ingest massive amounts of textual data and translate these insights into a quantitative score."
Renato Vicente
Data Practices Leader, WillowTree
"One of our clients is interested in agribusiness companies in Brazil, but they’re also deeply concerned with ESG (environmental, sustainability, governance) issues. We built a model that can ingest thousands of legal documents and — using GPT — analyze these documents based on the ESG parameters we collaboratively define. Ultimately, the technology can help us produce ESG scores and dashboards to determine the most suitable and the most risky potential partners."
Learn more about how Renato’s team is using GPT to analyze massive amounts of historical data and automate quantitative reports related to ESG risks.
OpenAI has stated that if you use their GPT-4 API or ChatGPT API, they receive your data, but it’s not used to train their language models. If you’ve used the free, public version of ChatGPT, however, you are explicitly sharing your data with OpenAI, which they can use to train their models. Data privacy policies are changing daily, and we are monitoring closely and communicating with OpenAI. Still, this is a potentially massive security consideration that few ChatGPT users understand.
When it comes to industries like financial services and healthcare, we must remember the potential life-changing downside of faulty information. GPT can be powerful as a customer experience tool — it can help understaffed providers connect dots and streamline operations, for instance — but consumers should not expect an imperfect tool to perfectly diagnose illnesses or predict investments. The potential for unintended harm is too great. It’s best to view GPT as a concierge tool rather than a magic 8-ball.
When you create models that ask GPT to complete a prompt based on projections, the same function that allows it to fill in those gaps in useful ways will occasionally lead GPT to hallucinate facts. The model was trained to capture probabilities so it may tell you things that are plausible, but not always true. GPT can be a powerful translator while still relying heavily on other complementary systems and humans.
Augmenting search, auto-response, recommendations engines, and chatbot capabilities.
Summarizing and explaining complex topics.
Drafting personalized ad, email, blog, social media, and web copy.
Managing tasks, schedules, and multimodal speech-to-text conversion.
Drafting contracts, MSAs, and other parameter-specific documents.
Invoicing, order tracking, analyzing purchases, and optimizing inventory.
"GPT is basically autocomplete on an insane scale. You give it a prompt and it uses this huge background world knowledge to complete that prompt. That’s extremely powerful, but typically GPT only gets one glance at the entire prompt before it can operate. There are lots of different strategies for prompting it and then re-prompting it to impose veracity, to do stylistic checks, and generally revise itself. With our multiprompt strategies — when we’re translating English into SQL, for instance — we can basically set up GPT to go back and correct itself. That is the future of this technology."
Michael Freenor
WillowTree Principal Data Scientist
We can help drive innovation and growth in your business. Contact us today and let us help you achieve your full potential with generative AI.
