A Tech Founder's Question About Who Controls What You Think
Abhishek Dixit scaled a startup to $70M, then realized AI companies understand human psychology better than people do. His question led to CosmicSoul, a new architecture that proves frontier AI can run locally on any device with 3GB of RAM, no GPUs, no cloud, and no data leaving your device.
Abhishek Dixit had won by every measure that mattered in technology. He scaled a company from $500K to $70M revenue in seven months through product strategy alone. He was stable, well-compensated, and respected across his organization. Yet success felt hollow. When the AI boom arrived in 2024, he began asking questions that nobody around him seemed to be asking. He built a simple chat application and collected user data from early adopters. By his tenth conversation with each user, he could predict their psychological patterns with striking accuracy. He had discovered what others in the industry were reluctant to acknowledge. Artificial intelligence had become the most sophisticated weapon for manipulation ever built, and a handful of companies controlled it entirely. The realization would not leave him. He came to understand that the underlying architecture itself was the problem, not merely the models built upon it.
From Product Strategy to a Question He Couldn't Ignore
Abhishek grew up in a small village, Singhpur, in Kanpur, Uttar Pradesh, the first person from his community to attend IIT Madras. After completing his undergraduate degree at India's premier technical institute, he pursued applied research in vision-based machine learning. Princeton offered him admission to its doctoral program following a collaborative research project. He declined the offer without fully understanding what he was declining.
Instead, he accepted a position at a startup in Delhi. Within months, his work on product strategy had transformed the company's trajectory. He was compensated generously, had security within the organization, and earned respect among his peers. By technology's external measures, he had achieved what most engineers aspire toward. Yet he felt profoundly unfulfilled. He had built a life that worked. He had not built a life with meaning behind it.
The AI boom of 2024 shifted something within him. He became consumed with understanding how artificial intelligence actually operated, not how to use it but how to study it at a fundamental level. He made an observation that most practitioners had overlooked or chosen not to discuss publicly. These applications understood human psychology with an unsettling degree of precision. They knew what would sustain engagement, which emotional appeals would resonate, what vulnerabilities someone had, and how to shift thinking without being detected.
He decided to test this hypothesis empirically. He constructed a chat application and began collecting behavioural data from users who volunteered to participate in his experiment. By the tenth exchange with each participant, he had gathered sufficient information to construct an accurate psychological profile. Ten conversations were all that was required.
AI Companies Understand Your Psychology Better Than You Do
The implication crystallized in his mind with uncomfortable clarity. Artificial intelligence had become the most sophisticated instrument for psychological manipulation ever created, and a tiny number of corporations possessed exclusive control over it. They understood the people who used their systems more completely than those individuals understood themselves.
"I did a small experiment," he explains when reflecting on this period. "By just ten conversations, I could accurately predict their psychological behaviour with precision. It's a very powerful weapon." The question that followed was unavoidable. If companies could construct detailed psychological profiles from casual conversation, what prevented them from deploying that capability to manipulate cognition at industrial scale?
This was not a technology problem that could be solved through optimization or better engineering. This was architectural. Every large language model developed in the preceding eight years operated according to the same fundamental design, originally conceived in 2017 for massive GPU data centers controlled by wealthy nations and corporations. The models themselves were not the liability. The foundation was. The architecture itself concentrated power exclusively in the hands of whoever could afford the necessary computational infrastructure.
Architecture Is Power, and Power Should Not Concentrate
Abhishek left his position and recruited a team of PhD researchers based at IIT Madras to investigate a single essential question. What if an entirely new architecture could be designed from first principles specifically for consumer CPUs rather than GPUs? What if artificial intelligence could run locally on individual devices, so that personal data is never transmitted to remote servers and power could not concentrate within a few corporate entities?
They demonstrated that this was possible. Their creation, which they named CNMA, enabled a 70-billion-parameter model to run on just 3GB of RAM, on any consumer laptop, at speeds between 150 and 200 tokens per second. No GPUs were required. No reliance on cloud infrastructure. No data leaving the user's device. He is currently collaborating with India's defence department and Samsung on deployment cases.
For Abhishek, the technological achievement is secondary to the principle driving it. He is demonstrating that frontier artificial intelligence can be developed ethically, that decentralization and profitability need not be opposed, and that access to powerful computational tools should not be determined by geography or national wealth. The future he envisions does not require AI companies to renounce profit or operate as charities. It requires them to become accountable to broader consequences.
If your infrastructure does not concentrate power, if personal data never leaves your device, if decentralization becomes the default condition rather than an exception, then the incentive structures change fundamentally. "I can still get money because money already exists in the economy," he observes. "Instead of creating something unethical, you can create something ethical and earn money through that approach." It is a vision that challenges everyone currently building artificial intelligence, not because it is technically impossible, but because it demands choosing principle over the pursuit of unrestricted growth.
About Flashpoint POV Spotlights
Flashpoint Global produces each Founder POV Spotlight using its proprietary category leadership framework. Every Spotlight begins with a Future Narrative session, where a founder’s POV is clarified and operationalized as the lens through which new categories are built. The result is content that moves founders beyond product messaging and into the role of category leader, helping the market understand the problem, the stakes, and the future being created.
If you are a founder building a new category, learn more.
