Models suck at behaving like a human or even higher order life forms like mammals or birds. Humans can plan, think, memorize, act and make decisions. We learn by experiencing or doing and also our brains consume less energy (~20W) to perform all sorts of complex tasks and also think about other things while doing a complex task like driving, swimming or playing an instrument. Humans are also good at memorizing heaps of information and recalling quite a bit in real time with little to no context.

We also have nearly unlimited experiences of the world around us, and our experiences shape our reality. This is possible because of the functions of our brain (or mind) happen to be quick and super-efficient.

Somehow, models have not been built around the paradigm of how the human brain or mind works and efficiently computes things. Existing LLMs in contrast consume a lot of compute, capital, energy and need heavy governance for safety purposes. Most models are also centralized, stifling innovation.

We need to move on from models and find new paradigms or architectures to solve for the Agentic future.

An agentic paradigm is simple. It consumes less energy, less compute, less capital, learns by experience continually, can be governed, is safe and reliable. We also believe Agents can be decentralized or democratized rather than preferring a centralized approach.

The new approach is a type of Experiential AI and not Generative AI. The architecture can be hybrid and use Generative components but not restricted in any manner to it.

At AFCP AI we are working on a new approach for building our Agentic commerce OS based on Experiential AI. We are figuring out an artificial shopper’s mind with this approach. We believe solving an artificial shoppers’ mind will add tremendous value to commerce, consumers and businesses. It will transform commerce from the ground up, making commerce seamless and delightful with Agents doing the work for both consumers and businesses.

Instead of models, we need an artificial mind specific for cracking shopping intelligence. We believe in a divide and conquer approach to cracking AI.

We believe we need fewer chips to run our artificial shoppers’ minds. We aim to make this intelligence scalable and distributed. We want to repurpose old, discarded, or second-hand chips and servers to build and run our AI. We are also committed to powering our servers with sustainable electricity while consuming 10x less energy than model providers. By building an efficient and sustainable business, we can offer the world shopper agents that are not only affordable but also remarkably easy to build and use.

The last revolution was Generative, the new Agentic future needs an Experiential AI revolution. The way to build Agents is through Experiential AI

The human brain or mind also has adaptive agency to suddenly react quickly to an unexpected situation on the road like a car from a side lane suddenly entering the road we are on and causing us to (typically) put our foot on the brakes to either stop the car or just slow down depending on the distance from the hazard causing vehicle. So, our brains can autocorrect or adapt to our environments and take corrective actions in real-time. (Note: I am making a general point here and not talking about AI being used for self-driving vehicles in particular. In my opinion, the technology is still not quite there to accurately judge situations like these, be it pedestrians, other automobiles, road or weather conditions and make safe timely judgments.)

Now, our existing large language models (LLMs) are the opposite of this. They are highly inefficient since they consume tonnes of energy, compute and capital even if they have to do an operation which they have done before, let alone new complicated tasks or situations they would be faced with. Moreover, they are not able to autocorrect or adapt i.e. show agency to be adaptive in the face of unexpected situations they are not trained in. This would either mean updating the model by freezing it with more post-training (a very expensive operation) or doing inference (cheaper but recurring and often results in 70-90% of actual costs even after post-training).

Now why do LLMs have this problem?

Let’s begin with the Von Nuemann Transformer Architecture that LLMs are based on. This architecture requires the matrix operations for LLMs requiring to move data and memory constantly to perform compute. This wastes energy and compute, because the architecture of PCs is not grounded in how the human brain works.

Secondly, the “intelligence” generated by these transformer models is based on statistical pattern matching and prediction (passive), not based on how the human brain (or mind) operates (active, and for the most part, spontaneous). Moreover, we also need some form of adaptive agency, just like human beings correcting and improving on the fly.

Lastly, intelligence cannot emerge only from statistical pattern matching approaches which employ next-token-based prediction and are compute heavy. But this thesis comes at a cost, i.e. more better chips, more compute, more cash and more energy. But none of this is grounded in reality.

This is neither beneficial to businesses, nor consumers nor the environment.

We need better AI solutions or architectures for the future. What seems to be counted as a given “Artificial Intelligence” from probabilistic approaches like statistical pattern patching to predict the next token is still a far cry away from the way humans or even complex mammals like cats, dogs or lions think or reason.

Now, I am not saying that LLMs or approaches that employ statistical pattern matching do not have a place in the AI world. Of course, they do with regard to certain domains and use-cases. But they are not the “be all and end all” approach to determine a universal benchmark for artificial intelligence. There are many domain specific use-cases which would rely more on deterministic approaches of artificial intelligence over probabilistic ones. This is more so in the case of critical applications with mission-critical use-cases.

There are many ways to bring forth artificial intelligence. That’s what we at AFCP AI are excited to solve. We are developing artificial shopping intelligence for the world.

We believe the world is headed towards an Agentic paradigm. We would need AI agents to be scalable, compute efficient and affordable. We also need AI agents to consume 10x less energy just like a human brain. We also do not want Agents reasoning to depend on LLM alone. This is possible if we run Agents on a new and efficient architecture and not bigger token-based statistical prediction models.

We also want AI agents to be accessible, affordable and safe for everyone; agents that benefit businesses, consumers and the environment at large.

We believe in the existence of such agents as not only being a possibility but a necessity for the future.

That’s what we at AFCP.ai are bringing into existence, an AI operating system for figuring out a shopper’s mind. This can help us create AI shopper agents for businesses which are efficient, scalable and affordable for all.

Only a few like Satya Nadella or Mike Volpi were paying attention to the writing on the wall. They were putting that across. Yet much of the world was and is still hypnotized with LLM-first AGI narratives and whatever they want to call it these days.

And yet still none of the “leading players in AI” today want to talk about the ROI for all this infrastructure investments pitted against recurringly high inference costs. Or say the lack of enterprise wide mass adoption as Sarah Guo puts it. She has stated this time and again on her posts besides the need for AI companies that orchestrate complex workflows leveraging both AI and human expertise.

Moreover, we also need new architectures for more efficient compute as we head towards an agentic world. Existing Von Neumann architectures aren’t going to solve our compute and inference problems.

At afcp.ai we are on a mission to help save businesses and enterprises from dying out due to these problems. These problems keep them from embracing AI at a wider scale.