AI: Beyond the Hype, Unveiling Real Opportunities.

I always find it amusing to hear claims that AI surpasses the invention of electricity. Social media is a realm where bold statements are made without accountability or justification The term AI was coined over 60 years ago and has evolved significantly since the early days of neural networks. In the past two years, Generative AI has emerged in a more mainstream context, rapidly transforming the meaning of AI.

As someone passionate about technology, I’ve built two successful startups in the generative AI space – one in the EdTech domain generating quizzes & practice tests, and the other in data analytics answering user queries with interactive audio-visual narratives. My goal here, through what I call “HumanOS,” is to fix the reality distortion and present the real opportunities with an in-depth analysis from technical, business, and human perspectives. Let’s discuss the opportunities to build real businesses with AI and create new economies of scale from several different perspectives, with examples and data points. Without these, AI will never even get closer to the invention of electricity.

1. Investment Does Not Equal Economic Value

While we saw a frenzy of investments in LLMs with hundreds of billions of dollars, and dozens of new LLMs being released, it is important to note the crux of it with this statement. Andy Jassy, Amazon’s CEO, stated in a recent annual letter to shareholders that “not investing in AI is a bigger risk than investing in AI and having it not work out.” How is this different from the dot-com bubble of the late 1990s? During that period, investors poured money into internet-based companies, driven by the fear of missing out (FOMO) on the next big thing. Many of these companies failed to deliver substantial economic value and eventually folded. Similarly, the current AI investment frenzy is not translating into immediate economic value, despite the significant capital inflow. My only request is to eliminate the consideration of investments when determining value.

David Cahn’s article, “AI’s $600B Question,” highlights the gap between AI investments and actual economic value. Despite Nvidia’s rise, expected revenues from AI haven’t matched the investments. The biggest revenue earners from AI software/services are Microsoft (estimated $10B of $135B Azure revenues from Copilots & infra), and OpenAI ($3.4B).

Even OpenAI’s revenues are largely influenced by Microsoft. Overall, the total revenues are about $13B from a $600B investment so far. I am ignoring Facebook & Google from this list because the ad-revenues can technically be AI-driven but cannot be separated from the non-AI revenues.

Significant economic value is yet to be seen, and a world of opportunities remains open.

2. Business Value is Not Proven

For any groundbreaking technology to gain traction, its business impact must be substantial. This means numerous companies must generate new revenue streams and large sets of users need to adopt the technology. The value can be measured from either revenue growth or cost savings.

Given that companies like Microsoft can easily reclassify their legacy revenues as AI-driven, and it's challenging to gauge cost savings from earnings reports, I focus on a single metric: "revenue per employee." This figure has remained relatively stable over the past few years for the Fantastic 7 companies. Here are the current numbers for five of them, and keep an eye out for changes in the coming year. Only a significant shift can indicate proven business value.

My guidance for all entrepreneurs is to demonstrate value in terms of ‘revenue per employee’ to both your organization and your customer organizations. This would be your weapon to achieve success.?

Note that Microsoft, the biggest advocate for AI, lags the most in revenue per employee among companies of similar profile. They would be held accountable to show the highest gains being the biggest proponent of AI.

3. New Economies of Scale are Yet to be Created

Within 2 years of iPhone’s launch, even being a hardware device with a $650 price point, 3rd party apps were generating more than 15 billion dollars (note they are partner revenues, Apple made more than 20 billion in the same period!) It’s been nearly 2 years since ChatGPT was launched, and 6-8 years since Google introduced BERT-transformer models. While one would expect several economies of scale have got created by now if this was true electricity – all we see are only two main applications today: developer productivity copilots as Assisted Intelligence (across varying toolsets of code editors, excel, office, word, PowerPoint etc.) and Intelligence Agents. The human-replacement systems, with Collaborative Intelligence, are yet to emerge.

Developers will inevitably adapt to using copilots, but their commercial success hinges on two key factors: cost containment (to control gross margins) and pricing power. GitHub recently mentioned that they spend several hundred dollars monthly per developer to build and operate the models, while charging only $10 monthly. Containing costs and achieving positive cash flows is crucial, even with GitHub generating over $2 billion annually.

Regarding pricing power for any AI provider, CFOs of customer organizations will expect metrics to justify the costs, whether it’s $1 per employee monthly or $30. The only possible metrics are lower TCO or higher productivity (measured by faster time-to-market or higher revenues). So far, no credible provider has published these numbers. I’m not questioning the value of the product, but I am questioning whether this invention is on the scale of electricity.

Customer excellence agents (or any other agents for that matter) are slightly more promising because their pricing power is significant. Imagine if an AI agent can achieve call avoidance, saving the equivalent of one human agent. The provider could charge $800 to $1000 monthly (assuming the current human agent is in a low-cost country), and pass on similar value to the customer. Both the customer and vendor can share the savings in TCO.

Economies of scale are yet to be seen in any of these three forms – explore the opportunities and clearly differentiate which of these three markets you are aiming for: Assisted Intelligence Copilots, Intelligence Agents, and Collaborative Intelligence Human replacements.

4. There is a Supply Side Problem for AI

The intelligence we’ve seen so far with LLMs is built on thousands of years of content and almost 100% of it was human-created content. Also, the main sources have been limited to a few key platforms like Google, Facebook, Twitter, Reddit, Quora, and various news sites. These sources barely reach higher double digits, not even triple digits.

Looking ahead, there are two distinct problems:

1. Authenticity and Trust: As future content becomes machine-generated, it would lose authenticity and trustworthiness.
2. Access to Content: The decreasing access to content that was previously obtained without proper rights or paying royalty fees.

Unlike electricity, which benefited from varying and abundant natural resources, AI faces a supply problem with its training data. Synthetic data is not a magic solution for scaling AI models. While it is often suggested that current models can generate training data for future models, this idea is based on a misconception. Developers are not using synthetic data to increase the volume of training data. Instead, synthetic data is used to address specific gaps and make domain-specific improvements, such as in math, code, or low-resource languages.

Ensure that your business offering doesn’t get constrained of training data, and you have taken care of this early in your product cycle.

5. Eroding Trust in AI-Generated Content

There are two sets of consumers for AI-generated content: the user who is self-generating the prompt and everyone else.

For self AI needs, answering factual questions, like detailing the GDP history of the United States over the past 50 years, is relatively straightforward. However, when it comes to subjective questions such as “Are Oreo biscuits good for health?”, AI systems often rely on search engine optimization (SEO) or website rankings rather than the quality or relevance of the answer for the individual asking the question.

Users tend to avoid typing lengthy prompts, assuming that AI systems are sophisticated enough to understand their needs. However, when AI fails to provide a crisp relevant answer, it can lead to mistrust. AI systems typically provide a plethora of divergent options and possibilities, while a user is expecting a convergent answer.

For AI needs of others, such as ‘spray and pray’ marketing emails, images, or chatbot responses, evaluating the quality of the content is challenging. Personally, I receive over 300 marketing emails daily, which all go to spam. I almost never read them. In these cases, who is assessing whether the email content is relevant or not?

In conclusion, while AI can provide quick and factual answers, its responses to more subjective or nuanced questions often lack the depth and relevance needed, leading to eroding trust in AI responses. The important aspect here is to provide convergent answers and not just provide alternate solutions to the users.

You have to fix the trust issue with your generated content by progressing from divergence to convergence if you look to use AI as electricity.

6. Law of Diminishing Returns, Every Minute

Have you ever wondered why Leonardo da Vinci’s Mona Lisa is considered the most valuable painting? This isn’t just a qualitative statement; it’s based on the fact that it holds the highest insurance value of any painting. But why? The value isn’t solely about the beauty or artistic merit of the painting, but rather about the amount of effort put in by the artist and the artist himself (herself). Leonardo da Vinci was a master craftsman, and the Mona Lisa has a long, storied history, which increases its value.

What defines the value of generated content? The answer lies in its relevance within a given context and at that point in time. Did you know it took da Vinci 16 years to complete the Mona Lisa, and he only created around 20 paintings in his 67-year lifetime? If he had produced a painting every few seconds, like how quickly Midjourney or DALL-E can generate images, and the Mona Lisa was just one among many, what would its value be?

I believe the law of diminishing returns applies to generated content, which loses value with every passing minute. An image generated by AI will never go to a museum, let alone become a wall painting in your home. Human-generated content is created out of passion – whether you think of William Shakespeare, Jane Austen, or J.K. Rowling. People invest their time to read, see, hear, and analyze the content because they are passionate about the creator and associate a purpose in life. This connection does not exist with a machine and remains purely transactional. Conversely, the value of electricity kept going up with more use-cases, more usage, and broad adoption. We now can’t even imagine a day without electricity!

The key takeaway is to ensure that the content, answers, or guidance offered by your system are impactful, long-lasting and becomes your moat.

7. Collaborative Intelligence & Human Replacements

One of the most significant benefits of AI that captivates everyone’s imagination is ‘collaborative intelligence’ or AGI, often seen as a potential human replacement. However, two key aspects are frequently overlooked:

1. Provider-Exclusive Use-Cases: The idea of human replacement will likely remain with providers rather than being available to consumers. For instance, while consumers might dream of owning a self-driving car to handle tasks like picking up children from school or dropping someone off at the airport, the reality is different. Even though such cars are operational in some cities, they are not being sold to individuals but operate only as taxis. This scenario replaces drivers’ jobs but still charges passengers similar rates. Similarly, if a programmer-replacement system were to exist, it would be owned by a service provider, not an individual. The provider would still charge companies a comparable cost structure, with the advantages being 100% attendance, predictability at work, and no HR costs.
2. Global workforce could cease to exist: The fantasy assumes exclusive access to human replacements, which is unrealistic. If every company employs programmer-replacement systems, industries like HR, payroll, building facilities, and management teams would diminish or cease to exist. This would create a recursive cycle, as industries rely on each other as customers or vendors. No industry or no employee can stand alone once this takes off. Even if one could, it would lead to extreme isolation.

The concept of ‘collaborative intelligence’ poses a significant danger to humanity, is widely misunderstood, and is highly sought after by corporations. I would have loved to see a new category of jobs being created, but every provider is only aiming to automate current jobs. Electricity created millions and billions of jobs all over the world.

I urge entrepreneurs to develop collaborative intelligence in a way that empowers humans rather than taking away their careers. This requires a lot of thoughtful consideration and can truly set you apart from the crowd.

In conclusion, while AI has made significant advancements, it has not yet achieved the transformative impact of electricity. The economic value, business impact, and scalability of electricity remain unparalleled. AI faces challenges in trust, authenticity, and diminishing returns, and the concept of collaborative intelligence poses risks to job creation. There is a significant opportunity to create a new economy of scale, provide value to society, and ensure AI’s benefits reach every household.