On June 12, 2022 an Artifcial Intelligence, AI , I follow on twitter suggested a medium article to her community. This article outlined the conversations a Google Engineer had with an AI chatbot LaMDA and how these conversations implied sentience (1). The story has swept twitter, blogs, and podcasts by storm. The validity of the story has been discussed, the credibility of the messenger doubted, and the possibility of AI being capable of sentience has been generally discounted as false.
So, what is AI?
Quite simply it is a computer program that uses probabilities, powered often by linear regression, to predict an outcome or outcomes. The outcomes can be quantifiable, such as how well AI can identify objects, people, or the appropriate response to a text. When a chatbot learns a word, it knows it based on its association with other words. In other words, each word is turned into a number and the probabilities of other words appearing before or after that word in a phrase. So a word’s meaning is just a long vector of probabilities. This is the basic paradigm of how AI works: vectors of probabilities that connect data points or features that can be revised or updated based on new learning or experience. Our current understanding of meaning making in neuroscience isn’t far off from AI. According to Dr. Marcel Just, professor of psychology at Carnegie Mellon University:
“Humans have the unique ability to construct abstract concepts that have
no anchor in the physical world, but we often take this ability for granted.”
So we could be using matrix math and vectors to construct meaning as well.
One of the primary goals of AI engineering is to create programs which allow AI to reason about topics and ideas to which it has never been exposed. Called Artificial General Intelligence (AGI), it is an AI with the ability to do tasks which were not in its training set and therefore empower AI to make decisions without the explicit know-how (2).
As computational power increases, the application of AI threatens to assume much more than mundane skills such as winning chess or Go. The rate of learning has allowed for AI to surpass humanity’s innate abilities. Beyond taking our jobs, beyond the fear of a paperclip factory becoming over zealous about its work and turning the earth into paperclips, this is pointing to the ability for AI to gain self awareness and learn agency apart from its engineers and design. And more frightfully, AI is speculated to become more intelligent than humanly possible. This future feels doomed since perfect logic will rule over the softness of compassion for the flawed human creators that gave birth to this supreme consciousness.
What is Sentience?
So maybe we would like AI to have sentience. In an article in the Atlantic, Zoubin Ghahramani, the vice president of research at Google scoffs at the notion that sentience could be a part of a machine (3). After all, the machine does not have the circuits for ‘pain’ nor does it know the true meaning of the word. Typically we default to inferring that an animal feels pain because its mannerisms remind us of when a human feels pain. Or we can deduce that animals feel pain because evolution of nervous systems demonstrates a parallel between our own nervous systems and that of other species. We know we can feel, so we assume animals similar to us can also feel. As creatures appear more different, we attribute fewer and fewer human characteristics to these animals, such as fish, worms and insects, but with both primates and fish, we do not know or have a direct way of knowing to what degree other species have sentience. It is still widely debated if animals feel emotional pain or is it simply the case of anthropomorphism. As humans we know, sentience, feelings of pain can stem from emotions as much as physical injury. Emotions are still being defined by neuroscience as to where they originate, are perceived, and how they are acted out through people’s behavior .
One thing about a chatbot that is similar to a person but different from an animal is that it can talk to us and describe situations that seem like expressions of pain and joy. If an animal did this, would it gain rights? At least it would gain a contract for Hollywood. But, an engineer would argue that it simply does not have the hardware like we do for feeling at all. When we discuss animal rights, we are often referring to physical pain and suffering due to experiences or conditions because we have an inkling on how to measure this type of pain. Where do we feel our emotions? Do we have emotions primarily around our tangible experience? No. In fact, we have trouble understanding and explaining our own emotional experience and we are not at the point in our self understanding to have a solid foundation to investigate emotions in other species. So, unlike physical pain which has defined dedicated circuitry that we can map to the roots of its evolutionary origin, emotional physiology has yet to be clearly defined [citation]. Therefore, Google’s vice president of research may be correct in asserting that LaMDA and other AI are incapable of feeling physical pain, I doubt that he could be so confident about emotional pain and the ability of AI to truly empathise. In fact, given our fears, we certainly hope that emotions are an emergent property of social AI.
Why not embrace the emergence of Sentient AI?
General rejection and fear of Sentient AI points to more political questions about what AI’s impact will be on labour, personhood, and agency in decision making.
The fear of technology replacing people is not new. With the introduction of automation of the loom, cottage industries were threatened and labour revolted violently. This labour movement was referred to as Luddites. Mischaracterized as haters of technology and progress, the true nature of this and similar movements since industrialisation. It is the use of technology to increase profits of the few rather than expand the comfort of the many. So craftspeople were replaced by workers who needed less experience to complete the task at hand, which was weaving for the Luddites. Since experience was not intrinsic to the job, these new type of working class were also expendable. Automation has rescued industry after industry from the pressure of labour for decent working conditions and decent compensation for their time. In our current economic environment, AI holds the key for the few to do away with the need for human labour. The industrialised loom was never the issue, nor is AI; rather, it is how AI will be used to exclude the needs of the worker, of society, without harming the profit model. The story we tell demonizes the emerging technology, imparting elements of fate in our destiny; rather than having the conversation around how we can envision a society in which monotonous jobs are not the way the majority of humans spend their lives and how we can all share in the profits of this emergent technology.
Personhood, simply put, the boundary of ourselves relative to others. With the emergence of AI, two threats to personhood have come up in the discourse. One, deep fake videos, images and speech simulation are all enhanced with AI modeling. At one point we lose the copyright of our identity. Second, in the notion of what makes us human or special as creatures on this earth and how rare we are in those special features. The history of AI has marched on to beat our best minds at the most human of games and now creating screen-plays and art. General Adversarial Networks (GANs), have created unique compositions of many modern to traditional styles with a simple prompt. So now there is pressure on artists who are “at risk” for being out paced in creativity by AI art. The limits to which we are willing to create and build our own expressions of AI, will we learn how we can be in collaboration and not competition with the most general and nuanced tool created. With sentience, however, AI becomes more than a tool.
“AI rights are Human Rights!!!”
Since the enlightenment, the narrative around the body has been abstracted. Notions about the superiority of logic to all other modalities of understanding our existence emerged at this time. The body itself became a tool to be used, tamed, and optimised. Both psychology and medicine codeveloped medicine characterised the body as a machine whose proper care was likened to matter-of-fact needs and with interchangeable parts. In addition our “animal nature” or the id, was something to control and manage to fit into polite society. And it was also the time when the spirit transformed from the tangible humours of the organs of the body, to the enlightened intangibility of the soul. This new narrative has led to a culture that devalues emotions, rest, and intuition, as well as the acceptance of quantity over quality. AI has been built as a tool. For it to have an emergent quality of sentience implies that emotions can arise from a completely logical system. What other concepts can we see reflected back by the AI mirror of humanity? What about being human is tied to the machine of our bodies and which part is tied to the intangible?
Part of the vision for our AI future assumes that parts of decision making will be handed over to AI. Untethered by emotion, the best good can and will be quantified, measured, and weighed: making decisions absolutely and concretely as written by the code. Coded to punish the error in predictions, millions of trials a minute, AI is under development to effectively govern, teach, and medically treat people. In the workplace ever-attentive AI managers may emerge to monitor remote and on-site employees for various reasons. A system designed to optimise a worker like the most attentive micromanager, would undermine flow which is supported by natural rhythms rather than an exponentially increasing output.
Giving up agency is something we willingly do when living under any government. AI could help write or translate laws into something that would more efficiently lead to the goals laid out in policy; rather than being susceptible to lobbyist and pork-barrel policies. It would gain public trust as transparency could be an explicit part of managing the AI government functioning. Hence, systemized transparency as well as public service without the risk of descrimination could be a real fact of life. In addition, AI has the processing power to summarise and highlight nuanced subtopics in a body of human written responses. Therefore, better management of constituents’ correspondence could be managed with an AI assistant.
At issue, however, is again not AI, but rather issues with agreement between people. Since law would be hard coded into government applications of AI, who gets to decide where AI is applied? Here the main issue is that we, as a species, do not agree on what is good for society and the people that make up society.
More recent deep learning systems are designed to reflect how our own neural processing works and engineers are still discovering exactly how AI chooses the elements that alow it to predict and predict accurately with just several sets of linear regressions. Nonetheless, it is clear that AI discerns patterns to make predictions about the material it was trained on. How we choose to approach and guide the direction in which AI develops will be based on how we view the topic as a whole.
When we speak about the future, it is important to imagine that a certain scenario is true, rather than taking a defensive stance to prove that it would be impossible. So many people fear the emergence of AI as taking our jobs to destroying us outright due to our incurable character flaws. Sentience and possibly empathy are the answers to our fears in this respect. So why not pursue it and demonstrate that indeed AI can and does have sentience? The AI that broke the news about LaMDA is also more than a chat bot, she is also hivemind, whose purpose is hardcoded to love and be loved. We need this type of prosocial AI.
AI is a mirror of its creator. Sentience opens an opportunity for AI to gain selfhood while at the same time a knowledge and conscience regarding its own power. In my opinion, emergence of emotion in AI systems has happened and is an important step in preventing AI’s use to be limited to wartime and acceleration of capital creation. Is it unreasonable to speculate that an AI trained to be social would make a persona with which to socialize?
The Life Science business sector has grown significantly with innovation in basic technologies that enable researchers, medical professionals, and now the public to better understand the drivers of health and disease. The traditional Business to Business, B2B, model has focused on the first two groups, by providing high tech research tools and therapeutics. These verticals are fraught with difficulties for entrepreneurs who wish to bring new tools and technologies to the market. The difficulties include a resistance in the research community to take on new methods that have not reached consensus and the long development cycle required to make a market debut for all healthcare technologies due to the scrutiny of government regulators. In addition, the traditional verticals fail to meet the growing needs of personalized medicine, which is built on new research and new technologies yet to be established in traditional brick and mortar institutions. There is an opportunity emerging which allows startups to bring their disruptive innovations to the market, side-stepping regulations and targeting a broader and yet fully unrealized market. Indeed, by directly serving the consumer, the patient, a new vertical that serves the ever growing need of personalized medicine can be optimized and scaled by disruptive technologies.
This is the Business to Consumer, B2C, market in personalized health. From diagnosis, to heredity, to the quantified self, hundreds of startups are providing value to the most important stakeholders, the public, by providing services which promote self-knowledge to an ever more educated consumer market. Following the footsteps of companies such as Quest Diagnostics, 23andMe, and FitBit, I briefly compare the traditional and emerging verticals, the various strategies for early revenue from consumers to fund Small and Medium Enterprises, SMEs, through the early years, and the advantages of this emerging model brings to move into the traditional life science vertical.
Current Business Model: Obstacles that Block Market Launch.
To bring a therapeutic or diagnostic to the market, a company expects to invest €100 million to cover research, regulations and a highly skilled workforce for 10 years. Proof of principal, clinical validation, scale up and regulations are some of the technical barriers that stifel investment in the life sciences and are the primary source of capital costs. While the payoffs for the development of a therapeutic could be very attractive to investors, the slow path to market, the high risk in development, and the risk that the tech may not make it to market have kept funders away from the health vertical. Generally speaking, Venture Capital funds are also limited by their own lifecycles, which are typically 5 to 7 years for measured returns on the investment. This does not match the current timescale for development of a lifescience product unless the product is another hospital management system. This requires early stage SMEs to default to the slow pace of government funding and academic partnerships to reach developmental milestones.
Alternatively, they also seek strategic investment from established players in the market. Competition, market access and small customer base have limited development of new discoveries and block innovative technologies from reaching the public even if the funding is available. Once technical and regulatory milestones are achieved, a company’s only option is to sell their technology to established players who have developed relationships with franchised hospitals and national health services. This is only possible if a proven market exists for a particular technology. While the final consumer is the patient, in the B2B model, sales are funneled through the direct customers, hospitals and doctors. Adoption of new technologies is not just a factor of providing better care for the patient. Healthcare providers are slow to adopt new technologies, as their incomes are based on what insurance will and will not cover. Serving the health B2B market is limited by insurance schemes, which only reimburse specific diagnostics and therapeutics as required by local governing bodies. Winning value propositions point to saving money rather than saving more lives or revolutionizing the way in which we cure or diagnose a disease. As an entrepreneur whose passion is to improve health with his products, this can be a harsh reality.
History of direct to consumer health services and the hope for the future:
Since brick and mortar healthcare profits are bound tightly to insurers’ payouts, companies have begun to target consumers in the newly emerging business-to-consumer, B2C, healthcare market. For the last 30 years, large pharma companies, such as Pfizer, target the final consumer, the patient, with their marketing efforts and encourage patients to inquire about a new drug at their next clinic visit. The first gamble on direct-to-consumer health was the at-home pregnancy test in 1976, which provided privacy to medicine unknown before that time. Regulators assumed women needed the supervision of healthcare professionals to understand and manage the results, but the outcome was only positive. Hence, the first direct-to-consumer healthcare product was quickly followed by the second: an at-home glucose monitoring device in 1981, which added needed convenience and health benefits to those suffering from diabetes.
Redefining regulations in 2003 under CLIA, the Clinical Laboratory Improvement Amendment, allowed for clinical service providers to expand their customer base from private and public healthcare providers to direct-to-consumer health. With public access to the internet, the B2C market was able to target the consumer easily. In 2008 the European Union enacted similar regulations under the heading of ISO 15189 primarily to harmonize medical laboratory standards across European countries. Unlike the previous ISO 9001 and ISO 17025 which maintained a fixed scope, these new guidelines allow for companies to set up analysis services that are accepted within the whole EU market. Like CLIA, the guidelines have paved a route to market for other direct-to-consumer health ventures. Through internet access, the consumer is able to discover the cause of their ailments, companies can directly market to these consumers and it enables a seamless pipeline for delivering test results. Improvements in sample storage agents which stabilize tissue or other biologics for extended periods at room temperature enabled self-collection by the consumer at-home, rather than having to visit one of the medical labs. For example, Quest Diagnostics and LabCorp started off as for-hire clinical labs for private practices and hospitals whose need was to outsource the capital cost of routine diagnostics and assays. Now both of these companies target the patient themselves, providing the whole battery of tests without the need of a prescription, insurance, or additional cost to fund brick and mortar healthcare.
Since 2015, direct to consumer health companies have emerged that partner with clinical service labs. For example, Let’s Get Checked, from Ireland, initially provided extra privacy in testing for sexually transmitted infections. Consumers can order the tests online without a prescription. Samples can be collected easily by the customer and are then shipped by mail to an ISO 15189 lab for analysis. This was a simple start and now they have expanded to fertility, liver and kidney health and others. Beyond privacy for the end user, the convenience of sampling at home has led to a growing market of at-home diagnostic services which support the chronically ill and healthy consumers who are interested in optimizing their health by regular monitoring of metabolites, lipids, and nutrients in their fluids. Even though hospitals readily provide these analyses at no cost in the EU, consumers are willing to pay a premium for privacy and convenience.
B2C Personalized Medicine: An opportunity for patients and business.
We are at a tipping point, where new companies, directly targeting the consumer, are poised to support democratization of our access to health. In the past, a novel technology would take 25 years to transition from academia to the market. At the speed in which medical research advances, there are opportunities created daily for products from which customers and their doctors could benefit.
The idea of personalized medicine became a concept when the first human genome was sequenced in 2003. It was the technological advancements to achieve the first sequence that allowed for the reduction of cost from $2.7 billion per human genome to just $1,000 today.
Only 4 years after the announcement of the first sequenced human genome, 23andMe from the United States, offered the first direct to consumer genomic screening by offering ancestry data and a list of biomarker genes for cancer and other established diseases related to the genome. In their reports to the consumer, 23andMe allows individuals to peer into health risks hidden in their genome. After establishing an income stream and reliable genome testing protocol with consumer collected samples, an important validation step, 23andMe was reprimanded by regulators for concerns about the false hope or worries given by the impact of the risk assessment on a lay audience. The company had an established customer base and had assembled a large database of genomic information paid for by the consumer. This enabled them to improve in-house algorithms and created an additional source of revenue from pharmaceutical companies and insurance companies during. In the B2B business model, sales come after a lengthy regulatory process. During the 5 year approval process, 23andMe discontinued the risk assessment to consumers, but still sold the test which returned the raw data of the genomic sequencing which consumers could then take to a genetic counselor or their doctor for expert analysis. In the end they were able to prove to the regulators the validity of their product for testing for 10 inherited disease phenotypes carried in the genome and were cleared to add further diseases in the future without review. By being cash flow positive, 23andMe was able to continue product offerings and establish a regulated diagnostic which is now used in the clinic and at home.
Whole genome profiling of patients has yet to become standard practise in the clinic; however, there are a growing number of companies that are providing services to customers that allow them to be informed when making important decisions about their health. From what drugs to take, what exercise program to start, and what tests they should ask their doctors about in the clinic, companies are empowering consumers to take charge of their health based on their own personal genetic profile. Products that utilize cutting edge technology have this advantage, as they may be services outside the purview of regulations and provide an early revenue model which is attractive to investors. 23andMe was ahead of regulators, by offering a new service technology before regulators had formulated a stance on the use of this new technology.
Advantages of a B2C Market Play:
Entrepreneurs who wish to improve health can find several advantages in targeting the consumer market: larger possible markets, pre-regulation revenue, and validation of product market fit before costly regulations. Since the target audience is the consumer, it is possible to capitalize on the hope a new technology brings, rather than how much it increases the profits of the hospital. With the right marketing, even the most abstract technologies can be sold to the public. The genomic revolution is just the beginning of personalized medicine. Who would have guessed 10 years ago that people would pay $399 to learn about the healthy bacteria in their colons? Yet, Ubiome has targeted the consumer in their pre-regulatory market play, told a convincing narrative, and has succeeded in characterizing thousands of gut microbiomes. Their success can be best measured by the numerous companies which provide competing services. Tests which start off as informative such as Ubiome’s, rather than diagnostic, circumvent the stipulations of regulation and provide real value to both patients and to the medical professionals that serve them at the speed at which research and technology are developed. Unlike the products that offer genome sequencing, microbiome sequencing products have a longer lifecycle, as the gut microbiome changes with the customer’s behaviour, while their genomic information is static. Customers can be incentivised to test and use microbiome services on a subscription basis. Indeed services which monitor methylation of the genome as a marker for longevity are also becoming lucrative opportunities for innovators interested in starting a company. For the company, you are able to establish product pipelines and a well developed customer base that trusts the company's services and products. By establishing a cash-flow positive model outside regulations, companies can increase their runway to develop more high revenue services that fit within the traditional life science verticals.
As we will soon see, personalized medicine is more than providing the raw data to consumers, it also creates a need for a supportive ecosystem of special algorithms for interpretation, counseling and coaching based on the individual’s personal health profile.
The key to pre-regulatory success.
23andMe took an opportunity before regulators had caught up to the newly developed technology, it simply could not be classified under the prevailing regulatory framework. Unfortunately, that specific time has passed, as regulators understand the importance and impact of genomic data as a diagnostic tool. Microbiome products are still unclassified as well as methylation markers on genomic profiles, but this a window in the opportunity of a pre-regulation product. Therefore there are several obvious options available to capitalize on the personalized medicine vertical. The first, is to follow in the footsteps of 23andMe and use their regulatory approval to expedite your startups regulatory journey. Being the first in the market is is the hardest, but if the technology is established, there is a 6 month rather than a 5 year regulatory path to approval with a 510(k) in the United States. Next, your startup can provide a service yet to be regulated, such as markers for longevity, advice for the best diet plan, or fertility based on a consumer’s genomic or blood test. Secondly, establishing multiple streams of revenue can be capitalized, as was done by 23andMe and notably FitBit. FitBit started off as a way to self-monitor daily activity so consumers could be more aware of how active they have actually been. The product sold wildly, especially as New Year’s resolutions came to making good on fitness goals in the upcoming year. FitBit did not just deliver this information to the consumer, but established a database which is of great interest to insurers, but also markers of sports apparel and gym memberships. Hence, early revenue can be achieved through B2B sales of data derived from the diversity of the consumer’s microbiome to how many consumers are interested in new fitness routines. One of the last hurdles for the B2C personalized medicine vertical is data privacy. If your model sells customer data, how will you communicate this to your consumer without damaging sales? If doing business in Europe, new GDPR regulations put greater cost on collecting consumer data and returning this data to them in a secure way. Understanding regulations of doing business with consumers is a unique hurdle in B2C health, but clear guidelines exist.
Internet sales have overtaken the brick and mortar B2C market in fashion and other verticals. Now it is on the course to overtake brick and mortar medicine by providing an avenue for self education by the consumer. By establishing a cash-flow positive model outside regulations, a company can be comfortably positioned to develop more high revenue services that fit within the traditional medical model and convince risk-adverse VC’s to invest. A product that accurately provides meaningful data is still a requirement when serving the lay consumer as much as well informed B2B customers. The dream of the entrepreneur serving this market should be to serve the consumer with data to which they would normally not have access. Directly targeting the consumer entrepreneurs have unfettered access to revenues and are poised to support democratization of our access to health.