How AI Superintelligence Will Transform Our Products

In the last couple of weeks, the word “superintelligence” has been quite trending. All thanks to Mark Zuckerberg, who poached some of the greatest AI minds from OpenAI, Anthropic, Google, Apple, and others to build his own Superintelligence lab at Meta.

Well, this was not the beginning of superintelligence. 

The term was coined by Nick Bostrom in 1997 in the paper titled “How Long Before Superintelligence”. The paper discusses what superintelligence is, how it can be implemented, and what the possible hardware requirements could be to achieve it.  

With each hardware generation delivering cheaper petaflops—and frontier models leaping benchmarks every quarter—the path to human‑level or beyond no longer feels speculative. Investors and researchers are acting accordingly.

For instance, Ilya Sutskever, Daniel Gross (who was poached by Mark Zuckerberg recently), and Daniel Levy founded their own company, Safe Superintelligence. Not only that, but already established AI startups such as OpenAI, Anthropic, XAI, and Google DeepMind have already shared their thoughts on “Superintelligence” and taken steps to build it.

But what is superintelligence? And why are teams, top AI researchers, and Engineers obsessed with it? 

In this article, I have explained the origin of AI Superintelligence, why it is an engineering problem, and how our products are being transformed as we march towards AI Superintelligence.

What is Superintelligence?

As per the original definition, according to Nick Bostrom, Superintelligence is a system that is “much smarter than the best human brains in practically every field, including scientific creativity, general wisdom, and social skills”. In essence, it is a general intelligent system that excels in all fields, including science, engineering, socioeconomics, pathology, and the arts. 

Today, we are witnessing the power that AI, especially LLM with reasoning capabilities, holds to crush the most difficult benchmarks out there. You talk about ARC-AGI, HLE, AIME, SWE, or even the toughest mathematical, medical, engineering questions, or any competitive exam. LLMs are crushing them as they come. 

Just take a look at the recent benchmarks from Grok 4.

Source: XAI

It is INSANE to think how fast we are progressing toward achieving superintelligence with every passing month. 

These reasoning models are leveraging computing power to scale their thinking capabilities and distribute user queries or prompts to multiple agents, thereby achieving the best results.

This is the path to building superintelligence. It is safe to say that a superintelligent system will gradually become an expert in all disciplines.

And it is worth noting that building it will be an iterative process. I guess that's why Sam Altman tweeted, Building superintelligence will be an engineering problem rather than a scientific problem.

Engineering Problem Compared to a Scientific Problem

Now, tell me, expand a little on this phrase: “building superintelligence is an engineering problem”. 

To achieve superintelligence, initially, most teams will be working with the following aspects:

1. 1
   Data
2. 2
   Architecture
3. 3
   Learning algorithm
4. 4
   Loss function
5. 5
   Optimizing function

Above the five, the three most essential components that interest me are the data, architecture, and learning algorithm. 

When it comes to the data, the most efficient form is text. It is cheaper to process and even produce synthetically compared to other modalities. What we are looking at in the near future is that LLMs will be the first type of model to achieve superintelligence. 

Now, the architecture. 

Since the release of the transformer architecture, it is evident that this architecture will be one to bring superintelligence. Why? Because it can handle long-term dependencies and contextual awareness. Even though it is really expensive to train a larger and denser transformer, it is the architecture that has stood the test of time and has continuously improved, making it robust. 

Now comes my favourite part of all, the learning algorithm. 

There is no doubt that reinforcement learning will be the main algorithm for training these models. Why? 

RL is extremely good at learning from feedback, i.e., alignment. Keep in mind that it is not efficient, and scaling RL is extremely difficult compared to supervised learning. But it can simulate a system very well. All it requires is a lot of time and computing power. If you have these two factors already settled, then you can build a terrific model. 

What I have learned from the Grok-4 release is that spending more time training the model with RL yields significantly better performance. 

However, although we have all the major ingredients to create a superintelligent system, a significant amount of engineering work is required to integrate all the components perfectly. Meaning, 

1. 1
   Does superintelligence mean developing models that are significantly larger? If yes, then how much computing power is required to run a single task?  Is scaling the real path to move forward?
2. 2
   A superintelligent model will definitely be an AI agent. So, what stops the AI agent from “reward hacking” by chasing easy reward loopholes?
3. 3
   How can we find more data to train efficient models? Or we just train the model entirely on RL and let it figure out the best path to intelligence. This would mean that the model will lack safety alignment. 
4. 4
   Do we need to use supervised learning at all during post-training to counter cold-start? 
5. 5
   How do we feed the model fresh experience fast enough to keep it learning?

Once we find the answers to these questions and a lot more, I’m sure we will create products that will change our lives. 

But are we not creating products that are transforming our experience? 

Short answer. Yes, we are. 

Transforming Our Product Experience

Humans have always been eager to produce products that replicate human thinking. But why? I think the answer is to provide leverage to think faster and come up with results quickly, take an abacus, for example. 

Humans are fighting the battle against time. We aim to accomplish tasks quickly and efficiently before we go beyond the curtain of time. And the thing that takes the most time, effort, and energy is “thinking”. So why not outsource it? 

What we are doing with AI is to develop products that can improve our lives across a wide spectrum of disciplines. 

Essentially, we want AI to handle the research (finding appropriate resources), do the cognitive heavy lifting (thinking), and develop products that align with human preferences. That’s what we are doing at present as well. 

But we need more power and a holistic research approach to develop products. 

Today, AI agents are not capable of perfectly aligning with preferences, emotions, and style. There is still a lack. Today’s AI can only help us connect the dots or place the information that we already have correctly. If we don’t have subject matter expertise or a proper plan to execute a task, then AI will definitely fail us.  

Bottom line: we still need to do our reading and grasp all the information. 

Let’s examine how the products have evolved since AI was integrated into product development and how they will continue to evolve as we approach superintelligence. 

2023-2024 (Q1-Q3)

By this time, AI has been greatly involved in product development. LLMs that are capable of summarizing large amounts of text produce long and contextual responses, such as writing a 1500+ word essay, assisting in writing small codebases, and composing emails using bullet points, among other tasks. 

The era of serious automation had begun. This time frame proved that AI could follow instructions and could produce responses that were safe and trustworthy. This was also where search capabilities were largely integrated into LLMs to avoid hallucination and produce accurate results. 

Late 2024 (Q4)

This was the period when reasoning models were introduced to the general public. With the release of o1, OpenAI claimed that, given enough test-time compute, the model could think and produce multiple chains-of-thought (CoT) internally. The model also uses reinforcement learning to shape and structure its internal thinking or CoT using constant feed or reward before answering. 

After the release, a lot of companies integrated o1 to build their products. Here are some insights:

Early-Mid 2025: Release of Research Agent

The early 2025 saw the release of AI agents, especially in the first quarter. Perplexity AI and OpenAI released a web-based “Deep Research” tool to the general public, which changed the way people research. 

Although Google had already released its own web-based Deep Research tool in December 2023, the research tool gained significant traction when OpenAI launched it, building on the success of their “o1” release. 

Following that, XAI and Anthropic also launched their own research tool.

This quarter also saw the release of Agentic tools like OpenAI “Operators” and Agent-SDKs. Additionally, a flood of Agentic tools was released in this period. 

Providers like CrewAI, SuperAGI, and AWS Strands SDK provide agentic platforms and toolkits for custom agent orchestration and automation in enterprise environments.

So far, we are seeing that teams and businesses are heavily relying on AI agents to build their products. Here are some examples taken from this report:

1. 1
   BMW Group, in collaboration with Monkeyway, developed the AI solution SORDI.ai to optimize industrial planning processes and supply chains with gen AI. This involves scanning assets and using Vertex AI to create 3D models that act as digital twins that perform thousands of simulations to optimize distribution efficiency.
2. 2
   Capgemini has been using Code Assist to improve software engineering productivity, quality, security, and developer experience, with early results showing workload gains for coding and more stable code quality.
3. 3
   HDFC ERGO, India's leading non-life insurance company, built a pair of insurance "superapps" for the Indian market. On the 1Up app, the insurer leverages Vertex AI to give insurance agents context-sensitive "nudges" through different scenarios to facilitate the customer onboarding experience.
4. 4
   PwC uses AI agent technology, powered by Google Cloud, to help oncology clinics to streamline administrative work so that doctors can better optimize the time they spend with patients.

Mid 2025-till 2030

It turns out that the rest of 2025 will be the year of two things: 

1. 1
   More expensive AI like GPT-5.
2. 2
   More AI Agents.

One thing is for sure: the scaling of AI models will not stop any sooner. Labs are in a race to build larger and larger models. 

I would to quote Ilya Sutskever,

The release of Grok-4 has already given us a hint of how the development of superintelligent AI will happen – multiple AI agents. But I could be wrong as well. There are two new startups that are still working and have not shown us any of their products yet – Safe Superintelligence and Thinking Machines.

Also, let’s not forget about Meta. They will definitely do something that will blow our minds. The late 2025 is the quarter to be mindful of. 

However, during this period, the primary focus will be on developing AI that can code effectively. Once AI can accomplish this, other disciplines can be integrated. 

So far, AI is not good at “alignment” coding. It codes, but it is still gibberish. 

But, just imagine the size of the neural network with all the information and skills of every unique individual. It will be huge and expensive to run a single job. 

However, there will also be an algorithmic breakthrough as well. Meaning, the efficiency of neural networks will improve such that they will consume less FLOPs to process information. This also means that cost of running a single job will reduce as we continue to reach superintelligence. 

Here are the remaining insights:

1. 1
   By 2027, we will see that products are more tailored to human preferences. 
2. 2
   The development cycle will reduce from quarters to weeks. 
3. 3
   Products will be more thoughtfully developed. Roles inside product teams will shift from building features to curating data and reward functions. They will closely monitor the design of a reward function to produce better features for users. 
4. 4
   “AI safety and alignment engineering” will improve and be more robust.
5. 5
   Significant breakthroughs will occur in medicine and healthcare.
6. 6
   Better materials will be invented and tested within a span of months. 
7. 7
   Safety protocols will be enhanced, and cyberterrorism will be reduced. 
8. 8
   There can be financial stability. Maybe?

By 2030, products will be shipped faster and with greater confidence. There will be continuous iteration and development. AI will be more involved in team conversations, and there will be more continuous feedback.  

If we manage to develop a superintelligent AI, then our products will have a holistic understanding of what humans prefer, and it will be aligned with our emotions and goals, leveraging scientific creativity, general wisdom, and social skills. 

Conclusion

The journey toward superintelligence—AI that surpasses human capabilities across all domains including scientific creativity, wisdom, and social skills—is no longer a distant vision but an engineering challenge unfolding before us. As Nick Bostrom originally defined it, we're witnessing this transformation through reasoning models that crush complex benchmarks and leverage massive compute power.

Currently, we're in the agent era. Companies are integrating AI into everything from GitHub Copilot's coding assistance to BMW's industrial planning optimization. The shift from basic automation to sophisticated reasoning models like o1 has already begun reshaping how we build products.

As we develop better AI we will fundamentally transform product development. 

By 2030, we’ll see products shipped faster with greater confidence, development cycles compressed from quarters to weeks. AI handles the cognitive heavy lifting while humans focus on curating data and reward functions. Products will become more thoughtfully designed, deeply aligned with human preferences, and continuously refined through AI’s holistic understanding of our needs and emotions.