As Kasisto’s Chief Technology Officer, Sasha Caskey drove the architectural design and software development of the KAI platform since it spun out from SRI International in 2014. Sasha currently oversees product development and Artificial Intelligence (AI) research for the KAI platform and its various intelligent digital assistants.
We chatted with him to get his opinions on the excitement and interest surrounding the recently launched public chatbot, ChatGPT.
Q&A with Sasha Caskey
There’s a lot of buzz around ChatGPT. Can you explain what it is?
ChatGPT is a conversational artificial intelligence (AI) solution trained by OpenAI in San Francisco. Since its release on November 30, 2022, it has blown up the Internet and tickled the imaginations of millions of folks across the globe, from entrepreneurial business people to sleepy student term paper writers.
ChatGPT is based on an underlying large language model (LLM), specifically GPT-3.5. An LLM is statistical prediction model that is trained by ingesting a mammoth amount of text content (typically from the internet) for the task of composing text based on a prompt that is provided to the model. The GPT (which stands for Generative Pre-trained Transformer) family of LLMs is based on the powerful Transformer deep neural network architecture which is able to learn complex contextual relationships between words and concepts. ChatGPT incorporates another model that was trained using human feedback on top of the LLM to help generate responses that are more aligned with human preferences.
It seems like ChatGPT popped up out of nowhere. Were there any similar precursors to it?
Open AI has been conducting research on LLMs in the GPT family for more than 5 years and has previously released versions of the LLM that ChatGPT is based on (GPT-2 was released in 2019 and GPT-3 was released in 2020). ChatGPT improves on its predecessor that was trained by OpenAI (InstructGPT) in many ways, most significantly in its usage of human-to-human interactions in its training. But possibly the most significant progress it made over prior versions was moderating its output to prevent malicious and offensive responses.
In recent years, major tech companies have built other similar chatbots. Some notable examples include:
- LaMDA (Language Model for Dialog Applications): This LLM developed by Google in 2021 was trained on conversational data (e.g., dialogs) and uses a different metric for selecting its output: a framework that accounts for Safety, Sensibleness, Specificity, and Interestingness (SSSI.) LaMDA has not been made available directly to the public, although Google just recently announced the release of a conversational AI service (Bard) that is based on LaMDA for beta testing.
- BlenderBot by Meta: This chatbot was developed by Facebook (now Meta) in 2020 and was released as open-source; subsequent improvements were released in 2021 and 2022. BlenderBot is notable for its advanced conversational skills (such as empathy and personality) as well as its ability to incorporate timely information from external knowledge sources via internet searches into its conversations. Similar to ChatGPT, the public can try out open-domain conversations with BlenderBot for free here.
Why does ChatGPT represent such significant strides in AI and language-based models?
ChatGPT is like nothing most of us have ever seen before. It seems knowledgeable and creative, writes code and poetry, and creates games. More importantly, it’s more likely to align its output to the user’s specific goals, and it’s much less likely to produce inappropriate or toxic output.
Large language models have quickly changed the nature of AI and what AI does in natural language processing and conversational systems. With each new iteration, the software encodes more information. This goal can be accomplished by both increasing the amount of textual data that is used an input to train the model and increasing the number of parameters inside the model that are used to encode information that it has learned about the data.
As a field, we are still learning the full range of knowledge encoded in neural networks. Imagine the opportunities as these networks include more video, pictures, and audio, and AI gathers information from these sources to produce text and non-text content, as we’ve seen with Midjourney.
Through the window that ChatGPT has opened, we can all envision how a large language model may express its knowledge conversationally, from stories to code and beyond. ChatGPT could create personalized texts and news and train better AI models with less data.
We’ve heard it’s not 100% perfect. What are some of ChatGPT’s limitations?
It’s imperative to note that ChatGPT and other similar technologies are still in the early stages of development and have a ways to go before they are ready to deploy in customer-facing production environments.
They are temporal.
Large language models are trained at a singular point in time. For example, GPT-3.5 was trained sometime in Q4 of 2021 and therefore lacks any information published since that time. While techniques to overcome this limitation (such as using an internet search engine in combination with the LLM) are being developed, just like the LLMs, they are nascent and take time to mature.
They are expensive.
Training LLMs alone can costs upwards of $5 million, which doesn’t include the expensive upkeep required for processing each request. In time, new hardware and software will develop to help reduce these costs, but running at scale is still prohibitive.
They can lack ethics and be offensive.
LLMs train on vast volumes of human data. While ChatGPT, LaMDA, and BlenderBot have made significant strides compared to their predecessors, users can still easily trick them into producing unethical and biased content. This can pose a serious threat to the early adopting businesses willing to invest in and use these systems.
Companies have made significant investments in risk mitigation, but the technology is still far from perfect; these current risks would probably make most organizations, especially those in regulated spaces, uneasy about adopting these technologies for many customer-facing and mission-critical functions. This is where it will be increasingly important for businesses–especially those in regulated industries like financial services–to rely on trusted conversational AI partners like Kasisto, organizations who offer tools and solutions to keep an AI architecture secure and private.
They can be very, very wrong.
Last, LLMs can “hallucinate” or produce unrealistic content. When they do this, they do it with the utmost confidence, not unlike what we experience when a human adamantly defends something they learned from “fake news.” The pervasiveness of “fake news” remains a complicated problem to solve, even with efforts to curb this by some of the largest corporations in the world.
When prompted about itself, ChatGPT says:
“It’s important to keep in mind that ChatGPT and other similar technologies are still in the early stages of development and may not be able to match the performance of more specialized and carefully tuned virtual assistants in certain scenarios. Additionally, there may be concerns about the reliability and accuracy of ChatGPT’s responses, particularly when it comes to sensitive financial information. It’s also worth considering potential ethical issues related to the use of language processing systems like ChatGPT in the banking and financial industry.”
What is Kasisto doing with large language models similar to ChatGPT?
Since the 2017 publication of ELMo, a word-embedding method, Kasisto has worked with large language models. Today, we train our word-embedding models from the more than 100 million utterances we have collected by chatting with millions of financial services industry users around the globe. Our systems use LLMs which continuously outperform existing pre-trained models, such as BERT or RoBERTa.
Since GPT-3 became available, we have incorporated it into our software stack and, most recently, used it to create personalized responses within KAI. The initial results show a noticeable improvement in the comprehension of our generated answers. When ChatGPT becomes available as an API, we predict it will be accessed and used alongside the other financial digital assistants within the KAI platform.
How does this affect Kasisto’s products and roadmap?
We couldn’t be more energized about this latest seismic evolution in our industry. ChatGPT, along with all other advances in language processing, provides a wide range of opportunities for Kasisto.
In the most basic sense, ChatGPT complements our in-market solutions and generates new ideas about how we can continuously enhance our existing KAI platform and intelligent digital assistants. In addition, all of the new interest in GPT is proving to be an accelerant for some of our more advanced orchestration and answer-generation technologies. Whereas, just a couple of months ago, these more complex applications were the domain of only the most advanced, research-oriented customers, but since December we’ve seen a dramatic increase in the number of financial institutions and large tech partners who are looking to apply LLMs and secure orchestration tools for new customer-facing and employee-facing use cases.
We’re also very excited that ChatGPT has accelerated and expanded public awareness of the growing AI technology space, where Kasisto has existed as a leading player for nearly a decade. This awareness highlights the differentiated value our banking industry clients have realized from Kasisto’s highly refined, specifically trained, and financially skilled AI products and services.
We hope this distinction helps financial industry partners to better understand our in-market products’ unique value while recognizing that Kasisto remains active in this research and development community. We have and will continue to apply these advances to our products. From hyper-personalization, more accurate answers, easier system creation, and a lower bar to entry into new domains, ChatGPT and LLMs complement our existing services in lots of ways.
But you don’t have to take my word for it … when prompted about comparing KAI digital assistants to itself, ChatGPT says:
“First and foremost, KAI is a highly sophisticated and advanced conversational AI platform specifically designed to deliver superior customer service and support. It uses state-of-the-art natural language processing and machine learning algorithms to understand and respond to user inquiries with near-human accuracy.
In contrast, ChatGPT is a more basic and generic chatbot platform that is not tailored for customer service. It is not as effective at understanding and responding to user inquiries, and therefore is not as good at providing high-quality customer support.”
Can you provide a sample comparison of KAI and ChatGPT answers?
Below is an example of answers from KAI and ChatGPT to the same question posed by a banking user regarding their financial needs. Both the user and institution details are anonymized for privacy concerns.
User’s Question:
“How do I find checks written but not cashed?”
KAI’s Answer:
“Here are 11 matches. I searched for:
Date: Between Sep 11, 2022, and Nov 11, 2022
Category: Checks”
ChatGPT’s Answer:
“To find checks that have been written but not cashed, you will need to check your records or bank statement for any outstanding checks. You can do this by looking for checks that have been issued but do not appear in your account activity as having been cashed. Alternatively, you can contact your bank and ask them to provide a list of outstanding checks for your account. It’s also a good idea to regularly reconcile your records with your bank statement to ensure that all of your checks have been accounted for.”
Because KAI embeds itself within a bank’s system, it delivers direct, financial-based answers. ChatGPT only provides generic definitions related to the content of the question.
So, what makes Kasisto different? What sets us apart from the competition?
Kasisto AI technologies specialize in providing transactional financial information and helping consumers make smarter financial decisions using that information. This reality requires KAI to connect to our partners’ highly secure financial services ecosystems and, in many cases, to orchestrate an integrated user experience that’s informed by multiple private data sources and knowledge bases.
KAI allows people to access these complex financial services through natural conversation, using human language in the same way they would interact with a bank employee at the neighborhood branch or contact center. Users don’t need to learn a digital application, construct a precise web search, or use a browser through a web resource. This helps a financial institution to ensure that complex information is readily available to and easily accessible by all consumers, even those who may be traditionally less digitally savvy.
This need for convenient access to secure, highly personalized financial information and guidance will endure in the era of ChatGPT. And we would argue that it will become even more important, given the stated limitations of ChatGPT. Kasisto provides the tools and technology that give bankers and consumers alike the confidence that the private financial information they share can be trusted.
The Future of Chatbots
Interested in learning how your team can integrate conversational AI to better engage your financial institution’s consumers?
Read The Chatbot Journey: Making Intelligent Digital Assistants Integral Members of the Team, our whitepaper that explains how conversational AI accelerates digital transformation, how to select the right partner to deliver your best branded digital banker, and more.
