Investment Views On Artificial Intelligence

We at PanCosmic Capital take pride in coming up with granular and detailed investment thesis in spaces where the fund intends to invest in. This investment thesis not only helps us create a targeted filter to source the right set of companies but also acts as a base for establishing a close relationship with entrepreneurs. This article gives a glimpse of our nuanced views on Artificial Intelligence, a space the fund will look to closely invest in. We begin by dividing all startup AI investment opportunities into Application AI and Infrastructure AI and then go deep into what would work in each spaces while emphasizing our views on Infrastructure AI, giving the example of a company – that our Managing Partner is an investor in and the fund intends to invest in. And in conclusion, we push the envelope further and express our views on AI for the future, where we believe current use of AI in mainstream consumer applications utilizes negative cognitive bias amplification to fulfill objective function of maximizing engagement leading to suboptimal societal outcomes where the opportunity for the future lies in developing AI solutions that do increase engagement but by freeing users from cognitive biases and by putting them on cognitive expansion streams. A simple explanation of not letting AI treat us as robots but “humanizing” AI.

In our world view, high potential, investable opportunities in Artificial Intelligence can be broadly be divided into 2 categories: Application AI and Infrastructure AI.

Application AI: This category in our dictionary is defined as all frontend employee and customer facing applications and use cases that are powered by AI algorithms resulting in intangible – productivity / efficiency gains of employees or tangible gains resulting from either better ROI (resulting from increased automation and better margins) or uplift in sales by significantly improving the process of tracking, understanding, monitoring and fulfilling unrealized / unmet demands of customers.

In our opinion, an AI first approach for enterprise applications is fraught with danger, meaning if you start with AI at the center of your application strategy and then build a product around it, you run the risk of shoveling down a technology down the throat of customers where no real tangible improvement is achieved over current status quo and as such the value proposition of adopting AI is often unclear. Plenty of such AI first solutions are frequently rejected by enterprise clients. And as such, we think there are 2 approaches that could work in Application AI: 1) Vertical approach – Go deep into a vertical, unearth industry pain points, dig out inefficient industry specific workflows and mundane compliance and regulatory requirement, build an application to solve these issues and then see if AI can fit into the picture; 2) Horizontal approach – Build industry defining horizontal SaaS solutions in legacy product categories, where current incumbents are way behind the innovation curve and then use AI to improve product automation / functionality / efficacy, scope, scale, user experience / efficiency and productivity.

With the vertical / industry specific approach, we believe the verticals where the use of AI makes a lot of sense are financial services, retail / commerce, healthcare / pharma, energy, insurance and manufacturing, while under the horizontal approach, the product categories where incumbent industry leaders are behind the innovation curve, and startups can be disruptive and also effectively explore use of AI are workforce management, supply chain management, enterprise risk management, IT service management, IT spend management and product lifecycle management. Over subsequent posts and newsletters we will address the nuances of how exactly can Application AI be disruptive in these vertical and horizontal product categories. But for this post, we want to focus and emphasize on Infrastructure AI.

Infrastructure AI: This category in our dictionary is defined as the backend nuts, bolts and tools that are used by the data science / machine learning teams that create algorithms and ML solutions which power Application AI. To understand this category, a high-level understanding of the workflow that data scientists, machine learning experts adhere to, to build these algorithms is necessary and also an appreciation of the fact that no matter what the use case, all enterprise scale AI/ML algorithms go through the same workflow / process.

The data science / machine learning workflow has 4 stages in a closed iterative loop: Stage 1: Data Operations – Data scientists take a look at the data and perform cleaning, transformation, enrichment and make it machine learning ready; Stage 2: Model Selection and Building – Data scientists then take a look at the cleaned data, notice a trend, build a model hypothesis and then select + feature engineer (i.e. build) the model; Stage 3: Training – Models are then fine-tuned, trained, tested and then retrained using this data; Stage 4: Deployment – Finally, models are deployed in a production environment.

In our opinion, an approach that will not work in this space for startups is you try and do it all i.e. provide machine learning as a service platform where you tackle all the four stages and there are a couple of reasons behind it – control and specialization. First reason is control, i.e., if you are a large enterprise and are basing a business-critical function (for e.g. revenue generating or critical to security) on AI such as Netflix / Youtube doing video recommendations, Airbnb doing home recommendation or JP Morgan doing fraud detection, then you are not going to outsource this to a third party machine learning as a service solution, but would rather build your own platform and subsequently your own ML solution and have control over this business-critical function. Second reason is specialization. It is near impossible for a startup to create a cookie cutter, one size fits all solution that caters to all 4 stages of the data science-machine learning workflow and that works for clients across multiple industries and verticals. There is just too much wood to chop and every data science-machine learning workflow stage has umpteen-nuanced complexities that requires a dedicated-specialized solution.

So, enterprises want to and are creating their own ML platforms for control but want to benefit from innovation in this space and are open to adopt best of breed, pluggable and “specialized” solutions that that can plug into their platform.

Thus, in our opinion, a startup looking to succeed in this space, to start with, should go deep into one of the stages of the data science machine learning workflow to maximize possibility of enterprise adoption. We personally pursued the path of finding an investment opportunity in this space about a couple of years ago and after extensive research, diligence, multiple company and corporate level conversations found out that the first 3 stages of the data science / machine learning workflow had plenty of startups and solutions. For e.g., Stage 1 (Data Operations) has startups such as Trifacta, Paxata, Nexla as well as Confluent; Stage 2 (Model Selection / Building) has startups such as Domino Data Labs and DataRobot; and in Stage 3 (Training) every machine learning framework such as Tensorflow, Café, Torche comes with its own hyperparameter tuning (i.e., training) solution.

But stage 4 deployment was a white space opportunity and enterprises were facing issues related to portability, heterogeneity of production environment, scalability, frequent post production updates, machine learning explainanbility, feature transformation, auditing / tracking / monitoring among a host of other issues. And thus, our Managing Partner decided to invest in machine learning deployment solution build on top of containers and Kubernetes called ( Without disclosing confidential company information, the company since the initial pre-seed round in 2018 has performed well and has gone on to raise a seed, series A and is on track for a successful Series B. Seldon has already returned an attractive multiple to it’s pre seed investors such. For detailed analysis and views on the space, white space opportunities and Seldon, please check out these articles on Vishal Arora’s Social Media feed:

Article 1: What is infrastructure AI?: 

Article 2: Where are the white space opportunities in infrastructure AI?:

Article 3: How is Seldon well positioned to capitalize on this white space opportunity?:

Thoughts for the future: Humanizing AI

The magnitude of impact that AI can and is already having in our lives is not up for questioning. As digital platforms control a disproportionately large portion of our existence (media, shopping, social engagement, working) and as AI becomes the technology of choice that drives functional efficiency of these digital platforms, what begs the question is whether this high impact technology is leading to an optimal outcome collectively for the society as a whole.

If on one hand AI is helping social media, ecommerce, digital media platforms maximize their revenues, and on the other is causing negative externalities – such as capitol hill riots, altering political opinions based on psychological manipulation and not facts thereby unfairly affecting public elections, excess digital media exposure leading to poor health, social and family outcomes, then one has to wonder whether this technology is to be encouraged or not. Well in our opinion, the answer is a resounding – Yes, but with a caveat. “The key to changing an outcome is to not break the algorithm but to be inside the algorithm.”

In order to fix these negative outcomes and tweak this powerful technology to further optimize its use, we need to closely look at how it works. Every consumer facing AI algorithm (the ones in social media, digital media, ecommerce platforms) has an objective function. And without getting into complications, the objective function of every consumer facing AI algorithm is “maximizing engagement”. Higher the engagement, higher is the amount of time spend, higher are the advertising dollars and higher are associated sales (if products are being sold such as on ecommerce). Now the issue is that the algorithm is not trained on “how” to maximize engagement, but only to maximize engagement at any cost. What this almost inevitably leads to is “negative cognitive bias amplification”. Cognitive biases are subconscious perceptions baked in our consciousness via a process of natural evolution to maximize the probability of the survival of the species. Such as tribal mentality – which creates a positive bias towards people of same color, nationality, gender, social / income status and negative bias to all who are not like you. For e.g., in the olden days, the arrival of someone who did not look like you / did not belong to your land, would be perceived as acute danger, potentially leading to death, as a result of disease or war. And as a result, our brains are attuned to avoid (and these biases are being reshaped but are far from being reversed) people who are not like us in terms of cast, creed, religion, gender, color and nationality.

Now most external stimuli – what we hear, see, touch, feel, smell, pass through our senses without warranting any attention, but when a stimulus hits the radar of one of our biases, we immediately attend to it. This on digital platform is measured by the AI algorithm as a success because it leads to higher engagement and higher time spent on the platform. Again, the algorithm does not care, what that bias is but as long as it exists it will make sure, it is utilized to the tee to maximize engagement. When we developed these cognitive biases, the purpose was to ward against one off danger and not to put the human body in a constant state of stress and anxiety. And so, AI algorithms based on the objective function of maximizing engagement achieved through negative cognitive bias amplification are creating chronic stress and anxiety in the lives of billions of people worldwide leading to optimal business outcomes but suboptimal societal outcomes related to mental/physical health, community/family and law/order. In layman’s terms, these algorithms, “boil the blood”, “rouse the emotions”, “revv up neurotic tendencies” to make more money for companies.

The key to correcting this trend is to make the AI work for us and not the other way around and this can be achieved by tweaking objective functions of these algorithms, and this is where the next big opportunity in technology lies. Why can’t engagement be “meaningfully” maximized? Is our attention attuned only by our neurotic tendencies based on negative cognitive biases or is there a more natural, subtler, more permeable, more settled and more consistent attention based on positive emotions? If the answer is yes, then what are those positive emotions? Can those positive emotions, be intertwined with optimal societal goals while also optimizing value for businesses?

How about an algorithm that frees us from our biases, leads to cognitive expansion, broadens our field of view, improves empathy and helps individuals develop better non-digital social / community ties. If an AI algorithm has the capability to decipher the code of our negative emotions and use it to maximize our attention, why can’t it decipher the code of our positive emotions / tendencies and nudge us towards further development in areas which would make us better rounded / multi-dimensional human beings and why wouldn’t this also lead to higher engagement, such that companies are incentivized to make more and better of such algorithms? How about a music service that not only recommends what we have listened to or what people similar to us have listened to, but puts us on a subtle track of expansion of our taste in music. Why can’t an algorithm through acute careful understanding of the frequencies emanating from our favorite rock songs, gently nudge us over a period of time to also appreciate a completely new genre – for eg classical, where every new song is only subtly different than the previous song but these subtle differences accumulated over a period of times truly transform and expand our taste in music. Or how about a restaurant recommendation engine that not only recommends eating in an Indian restaurant if that is the propensity you have shown, but nudges you to try other cuisines, through discounts and reward mechanisms? Why would social media engines only show us feeds that are so negative that they boil our blood or are so addictive that they act like cocaine where you neglect other aspects of your existence? Why would they not challenge us to expand our consciousness, broaden our world view, develop new hobbies / interests, exude empathy and spend more time in the non-digital realm? Why if such a platform exists, would it not earn the appreciation and loyalty of its consumers?

Central to all of the above thoughts is the notion of AI being used to create “Cognitive Expansion Streams”. But the danger arises in the knowledge that such a technology could be used as a knob for mass control. So, let’s say if we figure out through AI on how to make someone more empathetic, more trusting or who develops a taste in fine arts or an interest in world politics, or someone who begins to appreciate classical music, or someone who becomes aware of environmental issues, then who decides which cognitive expansion stream or streams to put a user on? This is a highly important issue, because if not thought about, then such a tool would be used by governments for mass control or businesses again to sway consumers to buy more of what they have to offer by altering their consciousness. I think the answer lies in giving intellectually inclined users the option to choose their own cognitive expansion streams and those who are curious and want to opt in, but are not inclined to make these choices, make the allocation of cognitive expansion streams “truly random”, which eliminates the possibility of any individual’s consciousness being controlled by an external agent.

The possibilities of the potential of AI are endless but the key lies in making the algorithm work for us and not the other way around and the key to changing the algorithm is to not break it but to be in it. The key to transforming AI is “humanizing” it.