Distinguishing Hype from Reality, Superlatives from Optimism - Innospark's Approach to AI Investing
Updated: May 17
We see unbounded potential for AI to positively impact how we live and work. We seek opportunities where computational intelligence intersects with deep understanding of a specialized domain or a challenging cross-functional problem.
2023 started with an explosion of excitement around Artificial Intelligence, driven by the public launch of ChatGPT. Many called this a crucible moment for AI as the general populous became aware of what modern large language models (LLMs) are capable of, and were initially spellbound and let their imaginations run wild. Some of the prognostications where truly galactic in scale as the general media pushed the theme that this could change everything. Just two months later, some of that excitement has since been tempered as the limitations or flaws of these still developing models have also been more broadly surfaced. The oscillations of attitudes towards AI of late are actually nothing new — they are just happening on a more public stage and at a faster pace. As a team that has been building, investing in and working with data science driven companies for decades, we felt it was timely to share more about how we approach investing in AI, or using the term we prefer, computational intelligence, at Innospark.
The term Artificial Intelligence is generally attributed to John McCarthy and dates back to 1956 and a conference he assembled at Dartmouth College. Since then, many academics and practitioners have sought to further advance the “intelligence” of computers with progress coming in fits and spurts and opposing ideological camps forming and dissolving. There have been numerous AI “winters” where most believed AI could never deliver on the original visions and was a lost cause, and then periods of exuberance, each more grand than the last, such as what we have been experiencing of late.
Throughout this time, computing resources have become dramatically cheaper, faster and more accessible. Processing power and storage are at levels today that were likely inconceivable 67 years ago, and yet, even with all this advancement, we are still a long way from replicating human-like intelligence in a computer. Perhaps this is because we still have a very limited understanding of how the human brain truly works, perhaps this is because even with all the data and processing power we have it still poses a constraint. Regardless, a machine that can understand concepts and reason like a human is still well over the visible horizon.
And yet, there is much to be excited about.
There exists tremendous potential in applying what constitutes modern AI to address a vast myriad of challenges across industry and society. This is where we live and operate at Innospark. The pace of technological change continues to accelerate and the opportunity set is truly only governed by our collective imaginations. So, with all of this potential, what is it that we look for? Stated simply we look for innovative applications of computational intelligence at the intersection with deep domain expertise in fields such as business, life sciences, health care and education. Done properly, we believe computational intelligence based solutions can provide substantial advantage over current manual or automation based offerings. We also prioritize architectures and approaches that are transparent and where the benefits and risks are understandable and clear. We seek to support and promote solutions that will have positive impacts on our society at large.
Not new models per se, but clever application of them
Seeking truly novel AI is a very high bar, and not one we feel needs to be met in order to find compelling opportunities. While there are some startups seeking to “invent” new AI, much of the recent advancement has come from large tech companies with massive research budgets and resources, or from leading academic institutions. In many instances, there is no need to recreate the wheel so to speak. Clever application of existing innovative models and their continuous improvements is often sufficient for many use cases. Teams that have a strong working knowledge of the plethora of modern computational intelligence tools at their disposal are positioned to build well architected solutions.
We expect these solutions to be systems of models, where neural networks (such as LLMs), perhaps many of them, are combined with symbolic systems and other methodologies in modularized ways. Some call this “chaining”, where the output of one model becomes the input of the next. Each section of the system is distinct, performs its needed function and operates in coordination with the other sections or modules. These models can be used to govern each other, or improve each other, as we see in GANs as one example. Another approach is called ensembles, where multiple models are used in parallel with their output combined to reach a more accurate prediction than any one model could produce operating alone. Combining these different models via thoughtful architectures can produce very capable and advanced prediction machines, or artificial intelligence. In many cases, these combined systems, leveraging deep domain expertise and AI, may engender patentable innovations, even if entirely novel AI is not present.
Much has been written about explainability or interpretability. For solutions oriented towards fields such as business, health care, life science, education, government etc., we believe this is paramount. There must be logic that connects the output/predictions to the inputs. While “black box” elements may exist in parts of the system if neural networks are present, overall the system output must be interpretable. For example, it is of little use to help diagnose a medical condition if the reasons for the diagnosis are entirely obfuscated.
Innovative architecture also implies that these systems have been built to be flexible, meaning modules can be improved or switched out without impacting the other parts. Other characteristics such as scalability, resiliency and ease of deployment are also important and are well documented in current literature.
Computational intelligence & data — the fuel for compelling AI
When pursing an AI based solution, the right data is a requirement. Such data is most valuable when it is hard to assemble and specialized, yet high quality, and can be combined with other data sources to enhance or augment its depth and breadth. It will be needed in sufficient quantities to train and test models to make accurate predictions. The required volume for accuracy will vary based on specific use cases and types of model deployed against the specific problem.
Given the availability of powerful computational models today, applying them to general problems with commonly available data is not especially appealing to us. While potentially fun and relatively easy to do, we do not see this as delivering sustainable value. Solving the hard problems requires deeper insights only gained through extensive experience or research. People that have put in the work to reach such levels of knowledge are well positioned to best understand how to address these problems or needs.
Computational intelligence is gained from applying domain expertise on data. Domain expertise often enables the creation of distinctly unique model features. Such features are constructed based on rules, parameters, or behaviors that again, are only well understood through experience. Over time these features can come to represent important trade secrets that help the system perform in superior ways to alternative more general AI based options.
AI: The Knock on Effects
Much has been written recently about the potential risks and downsides of AI. These are critically important topics and as a society we will likely need regulations and frameworks to help govern and guide AI developments to ensure we enable the benefits while controlling and minimizing negative impacts and bias.
As noted, at Innospark we give considerable thought as to the societal impacts of the companies we evaluate. Examples of challenges and problems that companies we support are working to solve range from accelerating the development and delivery of therapeutics, to reducing the impacts of Essential Tremor or Parkinson’s, to compressing the time to match patients to clinical trials, to enabling wireless networks to self-heal, and to improving literacy at scale among underserved populations. These are specific solutions to specific problems, with broad ramifications. We also work with AI-driven solutions with more horizontal applications including a non-invasive brain computer interface, a robotics platform that can be adapted to different enterprise uses cases and a no-code platform that brings computer vision to anyone while maintaining data privacy.
There remains much to be done to bring more stability, scalability, equity and tractability to the popular AI models of the day. We expect these models will also continue to be significantly improved upon in terms of the quality and accuracy of their output. In considering what the future may hold, we anticipate entirely new types of AI models that will leap beyond the capabilities of current approaches as we learn more about how the human brain functions and new generations of researchers seek to outdo what already exists.
Computational Intelligence is both Real & Ready
Working with new companies formed by founders who together innovate at this intersection, where AI skills merge with real domain expertise, is what Innospark is all about. The strength of these teams, based on their experience, skills, personal characteristics and drive, is paramount to our engagement, and the most important ingredient when formulating a successful venture. When we find teams that bring this mixture of talent, AI expertise and domain knowledge, we are compelled to learn more and investigate ways to work together to help them transform their vision into reality. Looking into the future, it is possible that AI will develop to the point where AGI may come to resemble what today we can only witness in science fiction novels or films. As we collectively strive towards that goal, there remains a massive opportunity to creatively deploy today’s cutting edge data science tools against the vast array of industry and societal challenges. We seek to work with those innovators leading the charge. Let’s get to work!
And just for good measure, we want to note that NONE of this was written by artificial intelligence powered text generators. Just humans.