Last year I was heavily experimenting with Knowledge Graphs because it’s been clear that LLMs by themselves fall short because of the lack of knowledge. This paper by Chirag Shah and Ryen White (you can click the heading above) from Dec 2024 expands on those shortcomings by exploring not just knowledge but also value generation, personalization, and trust.
They open the paper by casting a very wide definition of an “agent” everything from thermostats to LLM tools. While this seems facetious at first, their next point is interesting. Agents by definition “remove agency from a user in order to do things on the user’s behalf and save them time and effort.”. I think this is an interesting way to injext an LLM flavored principal agent problem into the Agentic AI conversation.
Their broad suggestion is to expand the ecosystem of agents by including “Sims”. Sims are simulations of the user which address
privacy and security
automated interactions and
representing the interests of the user by holding intimate knowledge about the user
Infinite knowledge is available through the internet today. It is available trivially and, some, ahem, blogs make a performance of consuming it. Machiavelli used to
put on the garments of court and palace. Fitted out appropriately, I step inside the venerable courts of the ancients, where, solicitously received by them, I nourish myself on that food that alone is mine and for which I was born, where I am unashamed to converse with them and to question them about the motives for their actions, and they, in their humanity, answer me. And for four hours at a time I feel no boredom, I forget all my troubles, I do not dread poverty, and I am not terrified by death. I absorb myself into them completely.
Some folks have a private office, but an office is not a study. A study or, studiolo
in Italian, a precursor to the modern-day study — came to offer readers access to a different kind of chamber, a personal hideaway in which to converse with the dead. Cocooned within four walls, the studiolo was an aperture through which one could cultivate the self. After all, to know the world, one must begin with knowing the self, as ancient philosophy instructs. In order to know the self, one ought to study other selves too, preferably their ideas as recorded in texts. And since interior spaces shape the inward soul, the studiolo became a sanctuary and a microcosm. The study thus mediates the world, the word, and the self.
In the 1500s Michel de Montaigne writes:
We should have wife, children, goods, and above all health, if we can; but we must not bind ourselves to them so strongly that our happiness [tout de heur] depends on them. We must reserve a back room [une arriereboutique] all our own, entirely free, in which to establish our real liberty and our principal retreat and solitude.
A little later, Virginia Woolf points out what seems to be an eternal inequality by struggling to find “a room of one’s own”.
The enclosure of the study, for those of us lucky to have one, offers us a paradoxical sort of freedom. Conceptually, the studiolo is a pharmakon, a cure or poison for the soul. In its highest aspirations, the studiolo, as developed by humanists from Petrarch to Machiavelli to Montaigne, is a sanctuary for self-cultivation. Bookishness was elevated into a saintly virtue
The world today would perhaps be better off if more of us had our own studiolos.
Another one of Frank Harrell‘s posts. Given my day job, and the R&D background this one is quite close to home. As a team leader on the industry side one hopes to build a culture with the team that aligns scientific rigor with company goals. Any method, statistical or cultural (in this case both), that solves for this tension will get you the most bang for your buck.
Building a startup, doing research, or even just launching a moonshot project is dependent on emergent actions and decisions. Behind this madness, there is a kind of Science of “Muddling Through” and Bayesian methods are perhaps best equipped:
make all the assumptions you want, but allow for departures from those assumptions. If the model contains a parameter for everything we know we don’t know (e.g., a parameter for the ratio of variances in a two-sample t-test), the resulting posterior distribution for the parameter of interest will be flatter, credible intervals wider, and confidence intervals wider. This makes them more likely to lead to the correct interpretation, and makes the result more likely to be reproducible.
In an environment of limited resources (time, money, investor patience), being able to quantify your data and obtaining bankable evidence is critical.
only the Bayesian approach allows insertion of skepticism at precisely the right point in the logic flow, one can think of a full Bayesian solution (prior + model) as a way to “get the model right”, taking the design and context into account, to obtain reliable scientific evidence.
The post provides a much more in-depth view, including an 8-fold path to enhancing the scientific process.
Chris Arnade gives us a view into why he goes on insanely long (both distance and time) trips by foot and what he discovers there. I found this passage hilarious, in that money seems to converge on a version of living that is identical with different paint while the rich likely search for that unique way of living.
Every large global city has a few upscale neighborhoods that are effectively all the same. It is where the very rich have their apartments, the five and four star hotels are, the most famous museums, and a shopping district with the same stores you would find in Manhattan’s Upper East Side, or London’s Mayfair.
The only difference is the branding. So you get the Upper East Side with Turkish affectations, or a Peruvian themed Mayfair. The residents of these neighborhoods are also pretty comfortable in any global city. As long as it is the right neighborhood.
Having traveled a fair bit, the designation of tourist brings with it multiple horrors. Chris uses walking as a mini residing and side-steps much of it. :
Walking also changes how the city sees you, and consequently, how you see the city. As a pedestrian you are fully immersed in what is around you, literally one of the crowd. It allows for an anonymity, that if used right, breaks down barriers and expectations. It forces you to deal with and interact with things and people as a resident does. You’re another person going about your day, rather than a tourist looking to buy or be sold whatever stuff and image a place wants to sell you.
This particular experience reminded of the book Shantaram, a book about a foreigner being adsorbed to and absorbed by the local.
While interesting the research uses LLMs to play both patient and doctor. I like that there is research happening in this area and at least we are moving in the right direction with the testing of AI i.e. not just structured tests. I would perhaps not judge the “failure” too harshly as the source of the data as also an LLM and would suffer from deficiencies which an actual patient suffering from symptoms would not.
This paper introduces the Conversational Reasoning Assessment Framework for Testing in Medicine (CRAFT-MD) approach for evaluating clinical LLMs. Unlike traditional methods that rely on structured medical examinations, CRAFT-MD focuses on natural dialogues, using simulated artificial intelligence agents to interact with LLMs in a controlled environment.
While the paper has a negative result we do come away with a good set of recommendations for future evaluations.
Traversing through human history, even in the last two decades, we see a rapid increase in the accessibility of knowledge. The purpose of language, and of course all communication is to transfer a concept from one system to another. For humans this ability to transfer concepts has been driven by advancements in technology, communication, and social structures and norms.
This evolution has made knowledge increasingly composable, where individual pieces of information can be combined and recombined to create new understanding and innovation. Ten years ago I would have said being able to read a research paper and having the knowledge to repeat that experiment in my lab was strong evidence of this composability (reproducibility issues not withstanding).
Now, composability itself is getting an upgrade.
In the next essay I’ll be exploring the implications of the arrival of composable knowledge. This post is a light stroll to remind ourselves of how we got here.
In ancient times, knowledge was primarily transmitted orally. Stories, traditions, and teachings were passed down through generations by word of mouth. This method, while rich in cultural context, was limited in scope and permanence. The invention of writing systems around 3400 BCE in Mesopotamia marked a significant leap. Written records allowed for the preservation and dissemination of knowledge across time and space, enabling more complex compositions of ideas (Renn, 2018).
Shelves, Sheaves, and Smart Friends
The establishment of libraries, such as the Library of Alexandria in the 3rd century BCE, and scholarly communities in ancient Greece and Rome, further advanced the composability of knowledge. These institutions gathered diverse texts and fostered intellectual exchanges, allowing scholars to build upon existing works and integrate multiple sources of information into cohesive theories and philosophies (Elliott & Jacobson, 2002).
Scribes, Senpai, and Scholarship
During the Middle Ages, knowledge preservation and composition were largely the domain of monastic scribes who meticulously copied and studied manuscripts. The development of universities in the 12th century, such as those in Bologna and Paris, created centers for higher learning where scholars could debate and synthesize knowledge from various disciplines. This was probably when humans shifted perspective and started to view themselves as apart from nature (Grumbach & van der Leeuw, 2021).
Systems, Scripts and the Scientific Method
The invention of the printing press by Johannes Gutenberg in the 15th century revolutionized knowledge dissemination. Printed books became widely available, drastically reducing the cost and time required to share information. This democratization of knowledge fueled the Renaissance, a period marked by the synthesis of classical and contemporary ideas, and the Enlightenment, which emphasized empirical research and the scientific method as means to build, refine, share knowledge systematically (Ganguly, 2013).
Silicon, Servers, and Sharing
The 20th and 21st centuries have seen an exponential increase in the composability of knowledge due to digital technologies. The internet, open access journals, and digital libraries have made vast amounts of information accessible to anyone with an internet connection. Tools like online databases, search engines, and collaborative platforms enable individuals and organizations to gather, analyze, and integrate knowledge from a multitude of sources rapidly and efficiently. There have even been studies which allow, weirdly, future knowledge prediction (Liu et al., 2019).
Conclusion
From oral traditions to digital repositories, the composability of knowledge has continually evolved, breaking down barriers to information and enabling more sophisticated and collaborative forms of understanding. Today, the ease with which we can access, combine, and build upon knowledge drives innovation and fosters a more informed and connected global society.
In Newton’s era it was rare to say things like “if I have seen further, it is by standing on the shoulders of giants” and actually mean it. Now it’s trivial. With education, training, and experience, professionals always stand “on shoulders of giants” (OSOG). Experts readily solve complex problems but the truly difficult ones aren’t solved through training. Instead, a combination of muddling through and the dancer style of curiosity is deployed, more on this later. We have industries like semiconductors, solar, and gene sequencing with such high learning rates that the whole field seems to ascend OSOG levels daily.
These fast moving industries follow Wright’s Law. Most industries don’t follow Wright’s law due to friction against discovering and distributing efficiencies. In healthcare regulatory barriers, high upfront research costs, and resistance to change keeps learning rates low. Of course, individuals have expert level proficiencies, many with private hacks to make life easier. Unfortunately, the broader field does not benefit from individual gains and progress is made only when knowledge trickles down to the level of education, training, and regulation.
This makes me rather unhappy, and I wonder if even highly recalcitrant fields like healthcare could be nudged into the Wright’s law regime.
No surprise that I view AI being central, but it’s a specific cocktail of intelligence that has my attention. Even before silicon, scaling computation has advanced intelligence. However, we will soon run into limits of scaling compute and the next stage of intelligence will need a mixed (or massed, as proposed by Venkatesh Rao). Expertise + AI Agents + Knowledge Graphs will be the composite material that will enable us not just to see further, but to bus entire domains across what I think of as the Giant’s Causeway of Intelligence.
Lets explore the properties of this composite material a little deeper, starting with expertise and it’s effects.
An individual’s motivation and drive are touted as being the reason behind high levels of expertise and achievement. At best, motivation is an emergent phenomenon, a layer that people add to understand their own behavior and subjective experience (ref,ref). Meanwhile, curiosity is a fundamental force. Building knowledge networks, compressing them, and then applying them in flexible ways is a core drive. Everyday, all of us (not just the “motivated”) cluster similar concepts under an identity then use that identity in highly composable ways (ref).
There are a few architectural styles of curiosity that are deployed. ‘Architecture’ is the network structure of concepts and connections uncovered during exploration. STEM fields have a “hunter” style of curiosity, tight clusters and goal directed. While great for answers, the hunter style has difficulty making novel connections. Echoing Feyerabend’s ‘anything goes’ philosophy, novel connections require what is formally termed as high forward flow. An exploration mode where there is significant distance between previous thoughts and new thoughts (ref). Experts don’t make random wild connections when at the edge of their field but control risk by picking between options likely to succeed, what has been termed as ‘muddling through’.
Stepping back, if you consider that even experts are muddling at the edges then the only difference between low and high expertise is their knowledge network. The book Accelerated Expertise, summarized here, explores methods of rapidly extracting and transmitting expertise in the context of the US military. Through the process of Cognitive Task Analysis expertise can be extracted and used in simulations to induce the same knowledge networks in the minds of trainees. From this exercise we can take away that expertise can be accelerated by giving people with base training access to new networks of knowledge.
Another way to build a great knowledge network is through process repetition, you know… experience. These experience/learning curves predict success in industries that follow Wright’s Law. Wright’s Law is the observation that every time output doubles the cost of production falls by a certain percentage. This rate of cost reduction is termed as the learning rate. As a reference point, solar energy drops in price by 20% every time the installed solar capacity doubles. While most industries benefit from things like economies of scale they can’t compete with these steady efficiency gains. Wright’s Law isn’t flipped on through some single lever but emerges through the culture right from the factory floor all the way up to strategy.
Labor efficiency – where workers are more confident, learn shortcuts and design improvements.
Methods improvement, specialization, and standardization – through repeated use the tools and protocols of work improve.
Technology-driven learning – better ways of accessing information and automated production increases rates of production.
Better use of equipment – machinery is used as full capacity as experience grows
Network effects – a shared culture of work allows people to work across companies with little training
Shared experience effects – two or more products following a similar design philosophy means little retraining is needed.
Each of these is essentially a creation, compression, and application of knowledge networks. In fields like healthcare efficiency gain is difficult because skill and knowledge diffusion is slow.
Maybe, there could be an app for that…
Knowledge graphs (KGs) are databases but instead of a table they create a network graph, capturing relationships between entities where both the entities and the relationship have metadata. Much like the mental knowledge networks built during curious exploration, knowledge graphs don’t just capture information like Keanu → Matrix but more like Keanu -star of→ Matrix. And all three, Keanu, star of, and Matrix have associated properties. In a way KGs are crystalized expertise and have congruent advantages. They don’t hallucinate and are easy to audit, fix, and update. Data in KGs can be linked to real world evidence enabling them to serve as a source of truth and even causality, a critical feature for medicine (ref).
Medicine pulls from a wide array of domains to manage diseases. It’s impossible for all the information to be present in one mind, but knowledge graphs can visualise relationships across domains and help uncover novel solutions. Recently projects like PrimeKG have combined several knowledge graphs to integrate multimodal clinical knowledge. KGs have already shown great promise in fields like drug discovery and leading hospitals, like Mayo Clinic, think that they are the path to the future. The one drawback is poor interactivity.
LLMs meanwhile are easy to interact with and have wonderful expressivity. Due to their generative structure LLMs have zero explainability and completely lack credibility. LLMs are a powerful, their shortcomings make them risky in applications like disease diagnosis. The right research paper and textbooks trump generativity. Further, the way that AI is built today can’t fix these problems. Methods like fine-tuning and retraining exists, but they require massive compute which is difficult to access and quality isn’t guaranteed. The current ways of building AI, throwing in mountains of data into hot cauldrons of compute and stirred with the network of choice (mandatory xkcd), ignores the very accessible stores of expertise like KGs.
LLMs (and really LxMs) are the perfect complement to KGs. LLM can access and operate KGs in agentic ways making understanding network relationships easy through natural language. As a major benefit, retrieving an accurate answer from KGs is 50x cheaper than generating one. KGs make AI explainable “by structuring information, extracting features and relations, and performing reasoning” (ref). With easy update and audit abilities KGs can easily disseminate know-how. When combined with a formal process like expertise extraction, KGs could serve as a powerful store of knowledge for institutions and even whole domains. We will no longer have to wait a generation to apply breakthroughs.
Experts+LxMs+KGs are the composite material to accelerate innovation and lower costs of building the next generation of intelligence. We have seen how experts are always trying to have a more complete knowledge network with high compression and flexibility allowing better composability. The combination of knowledge graphs and LLMs provide the medium to stimulate dancer like exploration of options. This framework will allow high-proficiency but not-yet-experts to cross the barrier of experience with ease. Instead of climbing up a giant, one simply walks The Giant’s Causeway. Using a combination of modern tools and updated practices for expertise extraction we can accelerate proficiency even in domains which are resistant to Wright’s Law unlocking rapid progress.
****
Appendix
Diving a little deeper into my area of expertise, healthcare, a few ways where agents and KGs can help:
Application
Role of Intelligence
Outcomes
Efficiency in Data Management
KGs organize data in a way that reflects how entities are interconnected, which can significantly enhance data accessibility and usability
faster and more accurate diagnoses, streamlined patient care processes, and more personalized treatment plans
Predictive Analytics
AI can analyze vast amounts of healthcare data to predict disease outbreaks, patient admissions, and other important metrics
allows healthcare facilities to optimize their resource allocation and reduce wastage, potentially lowering the cost per unit of care provided
Automation of Routine Tasks
AI agents can automate administrative tasks such as scheduling, billing, and compliance tracking using institution specific KGs.
With widespread use, the cumulative cost savings could be in a similar range as Wright’s law
Improvement in Treatment Protocols
Refine treatment protocols using the knowledge graph of patient cases.
More effective treatments being identified faster, reducing the cost and duration of care.
Scalability of Telehealth Services
Agentic platforms rooted in strong Knowledge Graphs can handle more patients simultaneously, offering services like initial consultations, follow-up appointments, and routine check-ups with minimal human intervention.
Drive down costs of service delivery at high patient volumes
Enhanced Research and Development
Already in play, AI and KGs accelerate medical research by better utilizing existing data for new insights.
Decreases time and cost of developing new treatments
Customized Patient Care
AI can analyze multimodal KGs of individual patients integrating history, tests, and symptoms for highly customized care plans.
When aggregated across the patient population healthcare systems can benefit from economies of scale and new efficiencies
While exploring the application of AI agents in healthcare we see that standard Retrieval-Augmented Generation (RAG) and fine-tuning methods often fall short in the interconnected realms of healthcare and research. These traditional methods struggle to leverage the structured knowledge available, such as knowledge graphs. Data approaches like Fast Healthcare Interoperability Resources (FHIR) used alongside advanced knowledge graphs can significantly enhance AI agents, providing more effective and context-aware solutions.
The Shortcomings of Standard RAG in Healthcare
Traditional RAG models, designed to pull information from external databases or texts, often disappoint in healthcare—a domain marked by complex, interlinked data. These models typically fail to utilize the nuanced relationships and detailed data essential for accurate medical insights (GitHub) (ar5iv).
Leveraging FHIR and Knowledge Graphs
FHIR offers a robust framework for electronic health records (EHR), enhancing data accessibility and interoperability. Integrated with knowledge graphs, FHIR transforms healthcare data into a format ideal for AI applications, enriching the AI’s ability to predict complex medical conditions through a dynamic use of real-time and historical data (ar5iv) (Mayo Clinic Platform).
Enhancing AI with Advanced RAG Techniques
Advanced RAG techniques utilize detailed knowledge graphs covering diseases, treatments, and patient histories. These graphs underpin AI models, enabling more accurate and relevant information retrieval and generation. This capability allows healthcare providers to offer personalized care based on a comprehensive understanding of patient health (Ethical AI Authority) (Microsoft Cloud).
Implementing AI Agents in Healthcare
AI agents enhanced with RAG and knowledge graphs can revolutionize diagnosis accuracy, patient outcome predictions, and treatment optimizations. These agents offer actionable insights derived from a deep understanding of individual and aggregated medical data (SpringerOpen).
A Novel Approach: RAG + FHIR Knowledge Graphs
Integrating RAG with FHIR-knowledge graphs to significantly enhance AI capabilities in healthcare. This method maps FHIR resources to a knowledge graph, augmenting the RAG model’s access to structured medical data, thus enriching AI responses with verified medical knowledge and patient-specific information. View the complete notebook in my AI Studio.
Challenges and Future Directions
While promising, integrating FHIR, knowledge graphs, and advanced RAG with AI agents in healthcare faces challenges such as data privacy, computational demands, and knowledge graph maintenance. These issues must be addressed to ensure ethical implementation and stakeholder consideration (MDPI).
Conclusion
Integrating FHIR, knowledge graphs, and advanced RAG techniques into AI agents represents a significant advancement in healthcare AI applications. These technologies enable a precision and understanding previously unattainable, promising to dramatically improve care delivery and management as they evolve.
If you’re in the field or exploring applying AI, do get in touch!
In this video, Stephen Fry narrates Nick Cave’s letter, which argues that using ChatGPT as a shortcut to creativity is detrimental. Surprisingly, I found myself in agreement. Having been involved in the AI industry for over a decade, I’ve always viewed AI positively. As an entrepreneur who pitches AI to investors and customers, I liken AI to technologies like spreadsheets: they eliminate tedious tasks, but you still need to understand what you’re doing.
I use various contemporary AI tools daily for tasks ranging from creating ISO template documents to drafting reference letters. However, when I’ve tried using AI as a thinking partner or advisor, it has fallen short, primarily due to its inability to discern or have taste.
Expertise involves having taste – the ability to distinguish good from bad, one decision from another. We depend on experts and advisors not just for their knowledge, but for their ability to guide us optimally, sometimes even questioning our intended goals. They speak confidently amid uncertainty, drawing on their experience.
ChatGPT/LLMs and their generative counterparts exhibit confidence but, by design, lack real-world experience.
My agreement with the video’s sentiment stems not from an inherent issue with computational tools, but from the understanding that taste develops through experience, however imperfect. Just as we learn arithmetic before using calculators, and calculators before spreadsheets, we need to cultivate a new culture around these emerging tools.
In mission-critical fields like healthcare, balancing exploration and regulation is crucial. Over-regulation can hinder society from benefiting from innovative uses and discoveries, while a lack of regulation places undue risk on vulnerable populations, as seen in historical clinical trials
I envision a path where doctors integrate AI into their workflows with enthusiasm yet maintain high standards. Unlike drug development and other biotech fields, AI and its software can be corrected relatively easily. Creating a selective environment will drive quality.
Users who recognize what is good will elevate the collective ability to achieve excellence.
Aneesh Sathe 