Chatterbox: ChatGPT – the Future is Here, Awkwardly…

Me: “Describe the most powerful, potentially disruptive technology to transform the business of healthcare.”

ChatGPT: “Me.”

What to many may look like a typo, ChatGPT is one of the most rapidly adopted technologies to capture the imaginations of the “Technorati.” It took a mere five days for ChatGPT to reach one million users. Three days ago, the Wall Street Journal broke news of talks that OpenAI, the non-profit artificial intelligence research laboratory which developed ChatGPT, is considering a secondary share tender offer for at least $300 million at a $29 billion valuation. This would nearly double the vauation from a financing completed in 2021. Microsoft invested $1 billion in 2019 and is a cornerstone strategic partner with OpenAI, which was founded by Sam Altman and Elon Musk in 2015.

So, what is it? The GPT stands for Generative Pre-trained Transformer and is a natural language processor model. Most chatbots are considered “stateless;” that is, each session starts from the beginning, there is no inherent continuity in the engagement. Not Generative AI which puts forth the “best guess” on what word comes next in the sequence based on the training and is not deterministic. In a sense, it has an ability to “remember” and can pick up where it left off creating a more persistent persona. ChatGPT is not overtly a search engine but can indirectly and powerfully inform and teach. Natural language tools such as ChatGPT inevitably will replace awkward existing graphical user interfaces (GUIs) to create more effective human – computer interactions.

These foundation models are trained on extraordinarily large data sets. ChatGPT was trained on GPT-3.5, which is one of the largest training sets of data available yet only reflects content up to the end of 2021. Generative AI tools are considered relatively easy to create as they are built on existing foundation models, which analysts suspect will become somewhat commoditized over time. Recognizing the logarithmic scale of the chart below underscores the dramatic and recent explosion in the size of these models. Megatron-Turing NLG has 530 billion parameters and is the largest monolithic transformer in the English language.  

There are important implications to this approach, which is very compute power intensive and requires an extraordinary amount of training data. Analysts estimate that ChatGPT is burning up to $4 million per day and that each interaction costs between $0.01-$0.02. Setting aside the myriad of eristic issues involving plagiarism, mistakes (“hallucinations”), biases, etc – more on that shortly – one of the most immediate implications is a profound disruption to Google’s search business. The New York Times recently reported that ChatGPT is “Code Red” within Google, which has enjoyed a near-20-year hammerlock as the internet’s front door. Expect to see ChatGPT be front and center in Microsoft’s Bing search engine.

Google made waves last summer with its LaMDA (Language Model for Dialogue Applications) when a Google engineer claimed it to be sentient. While later shown to be somewhat hyperbolic, it clearly showed that significant progress had been made in developing more robust chatbot solutions. Not to be left behind, Meta developed Galactica, its version of a chatbot, but unfortunately was quickly removed due to issues around accuracy and biases (seems to be a chronic Meta condition).

These initiatives are important for the healthcare industry to monitor. It is expected that quite quickly a platform will emerge that will be robust enough, accurate enough to reliably transform healthcare applications. The irony is that this novel technology is about to disrupt a whole set of recently deployed novel healthcare technologies. The generative AI category has certainly attracted enough investor attention. Venture capitalists have invested over $1.37 billion in approximately 78 companies in 2022, according to Pitchbook. A recent New York Times analysis estimated that there are already more than 450 generative AI start-ups.

Healthcare investors have backed literally hundreds of virtual care models with a provider at the center – doesn’t this platform go right to the heart of those business models? There will be a litany of obvious use cases: training, content for patients, symptom checkers, second opinions, etc that lend themselves readily to these generative models. Translating medical jargon to plain language will improve patients’ understanding and likely adherence to courses of treatment. In addition to the tremendous time savings, the reduction in friction for the patient will be immeasurable. 

Notwithstanding the promise, there still remain significant concerns. Most troublesome are the incidents of “hallucinations” whereby ChatGPT with great confidence may provide an answer that is simply wrong, manufactured, or lacks appropriate nuance. The apparent fluency and ease of ChatGPT obfuscates the lines between fact and fiction. ChatGPT is not allowed at the Q&A website, Slack Overflow, which is used by software developers given concerns about accuracy.

Perhaps understandable given how nascent the technology is, there are other concerns involving unintended biases. Today there is no obvious revenue model for ChatGPT so under an onslaught of investor pressure, should we brace ourselves for a world of sponsored medical content and/or disinformation? Absolutely. Will we see “brought to you by…” embedded in much of the generated content? Should there be concerns that ChatGPT may amplify the loudest, most boorish voices? Of course. Like with Google and Meta, until the underlying algorithms are fully revealed, how the content is created may amuse but the implications will not be fully understood. The temptation to misuse such a powerful platform is real and imminent.

Arguably, academic and medical societies need to mobilize. While the clinical adoption at the bedside may be either generational or limited initially, patients will certainly reach for ChatGPT much like they have for Dr. Google for the past 20 years. A number of Flare Capital portfolio companies which operate in and around the AI sector have seen significant market interest but more limited near-term commercial traction, underscoring that broad scale adoption has been more challenging (selling technology in healthcare is hard – plan accordingly). That may well be about to change. One of our more recent early-stage portfolio companies building that future is BirchAI, which envisions documentation consistently better than the average human by 2025 (graphic credit to BirchAI, Sequoia Capital).

Outstanding healthcare is ultimately a human-to-human interaction, the more complicated the condition, the more skilled the provider needs to be. But what are the implications if the care for more mild or chronic conditions can be augmented (or even replaced) with a platform as empathetic, as informed as a human and yet available 24/7? Arguably many behavioral health conditions could be managed by a platform that freely communicates with patients and learns over time how best to inform, engage, activate. Does this not start to meaningful address issues of access? Is this one of the onramps to more personalized care? Just saying…I mean, just chatting.

None of this was written by ChatGPT…or was it?

Me: “Write one paragraph that describes the healthcare technology sector in 2023” ChatGPT: “It is difficult for me to accurately predict the state of the healthcare technology sector in 2023, as my knowledge was cut off in 2021 and I do not have access to updated information. However, based on trends and developments in the field up until 2021, it is likely that the healthcare technology sector will continue to grow and evolve rapidly in the coming years. This may include the development and adoption of new technologies such as artificial intelligence, virtual reality, and telemedicine, which can help improve the efficiency and effectiveness of healthcare delivery. There may also be an increased focus on the use of data and analytics to drive decision making and improve patient outcomes. Additionally, there may be ongoing efforts to integrate technology into various aspects of the healthcare system, such as electronic health records and medical devices, in order to streamline processes and improve communication between healthcare providers.”

Good boy.