From Algorithms to Autonomy: How AI Is Rewriting the Architecture of Software and Mobility

Jul 19, 2026

Artificial intelligence is not simply a new tool being added to the developer’s toolkit. It is rewriting the foundational contract between humans and machines.

Tl;dr: The evolution of problem-solving using computers. The tectonic shift happening around us that is resetting the tech industry. The breakthrough of AI: how it all started, the limitations of AI, and what’s coming up next.

I came up with this story line while I was trying to connect all the dots on the recent developments in the field of AI: where it all started and where it is now and what’s next? I’d like to thank Mr. Andrej Karpathy, Ms. Fei-Fei Li, and Mr. Benedict Evans for their perspectives and deep insights, which gave me a lens to see AI in a different way.

Artificial intelligence is not simply a new tool being added to the developer’s toolkit. It is rewriting the foundational contract between humans and machines. Recently I was listening to Mr. Karpathys talk “SW in the era of AI” where he spoke about Software 1.0, 2.0, and 3.0. It was quite insightful. In the first era, humans solved their problems using computers by programming the logic themselves. In the second era, the computer learned the logic on its own from data and interfaces fed by humans. And the current era is defined by a natural language interface, where humans converse with the machine, and the machine understands the intent, handles ambiguity, and produces solutions with fluency.

Software 1.0, 2.0, 3.0: Abstraction levels

Software 1.0 is the era where most engineers were trained to define explicit rules, write deterministic logic in various languages like assembly, C++, Python etc. telling machines exactly what to do. The developer has handcrafted the logic to take care of all possible “if-then-else.” For the problems we planned to solve, this is powerful, predictable, and is inherently limited by what a human can specify.

Software 2.0 era shifted the locus of logic from human-written instructions to learned weights & biases. As the complexity of problems to be solved increased, hand-crafted logic no longer worked effectively. For example, to achieve highly accurate traffic signs, traffic light, or pedestrian predictions for ADAS and Autonomous driving, mere 80-90% accuracies are not enough. Accuracy must be > 99%. In this era, developers merely curated the data and designed loss functions for the network to find the logic on its own and solve problems with better accuracy. One side effect is that the system became more probabilistic and not deterministic. “Mind = blown” moment. This is what I call Cambrian explosion of AI. This is when “AlexNet” from Alex, Ilya & Hinton has cracked “ImageNet challenge” using labeled images, hand crafted loss functions and using NVidia GPUs. Many new network deep learning architectures exploded after this defining moment. The infamous “AI winters” has never come after that.

Software 3.0 era is where we stand today. The interface is natural language. The source of logic is emergent comprehension i.e., it is not pre specified, not trained on a fixed tasks , but generalized across the domains. The developer’s role here changes to architecting the intent and giving precise context. The system here is more generative, not merely classifying or predicting, but synthesizing. While generative way which we call LLM (Large language Models) brings in meta learning and reasoning by learning the patterns across domains. But one draw back of this approach is hallucinations.

─────────────────────────────────────

“LLMs are kind of like these fallible people spirits that we have to learn to work with.” - Andrej Karpathy

─────────────────────────────────────

LLMs as General-Purpose Technology(GPT)

As Benedict Evans puts it, Large language models are a general-purpose technology (GPT) i.e.., a category that includes electricity, the internet, and GPS. It is characterized by broad applicability, scalability, and the capacity to spawn entirely new industries at the application layer.

If we zoom out and see technological evolution like an archeologist, we see that every platform shifts like an S-curve. Every S-curve displaces the old one and a new foundational layer emerges. The companies that recognized it early and built on top of it capture the value of the next decade for example Microsoft won the PC era., Google won the web era, Apple, Samsung, Google won smartphone era and now in GenAI era we see stratospheric valuations for the power houses like Meta, Anthropic, NVidia, Google, OpenAI etc.…and is popularly known as MANGO 😊 .

As LLMs commoditize and become infrastructure, the differentiated value will move to the application and experience layers. The products built on top of the models, the proprietary data that feeds them, and the domain expertise that shapes their deployment are going to be the game changers.

Anatomy of General purpose technology and the “Technology Diffusion”

If we closely observe, general purpose technologies, did not bring minor changes, but transformed the society. Electricity did not just replace oil or gas lamps, it restructured everything: enabled new working hours, organized societies, created new industries altogether.

Railways did the same thing a generation earlier. People thought railways were faster horses. They were wrong. Railways did not just move people faster. They made geography elastic.

LLMs are now considered GPT, not only generative pretrained but they are General purpose technology . This is mainly because of 3 reasons

1. Emergent reasoning: Unlike previous AI systems, LLMs can solve problems they were never explicitly trained to solve. They generalize across domains in ways that task-specific models cannot

1. In context learning: The program adapts to new tasks at inference time. It learns from examples you show it in the conversation. This makes LLMs radically flexible

1. Generalization at scale: The more compute and data you throw at these models, the more capable they become, in ways that are not linear and are not fully understood. Hence the race to be the top foundation model with more data centers, more compute & memory needs, if you observe industry closely.

Andrej Karpathy observes that large language models (LLMs) have flipped the traditional direction of technology diffusion.

In the past, General purpose technologies like electricity, computing, cryptography, flight, the internet etc. followed a specific path: Governments and large corporations first adopted them because they were new and expensive and only later diffused down to the general consumer. For example, early computers were primarily used by governments for military & other research.

However, LLMs have reversed this pattern in several ways: Unlike early computers used for military science, LLMs are used by everyday people for mundane tasks

Because LLMs are software-based, they are not restricted to a few elite entities. ChatGPT, for instance, immediately accessible to billions of people’s computers overnight before the world’s major institutions had fully integrated it.

The Limits of Language: Why Physical AI Is the Next Frontier

LLMs are extraordinarily good at reasoning over text. They compress and recombine patterns from vast corpus of human knowledge. But they have a fundamental limitation: they understand the world through descriptions of it, not through direct experience of it.

A human looking at a night sky experiences it. Reason humans say the night sky is beautiful is because they see that it is, whereas an LLM says it because it has been said enough times in its training data.

─────────────────────────────────────

“Current LLMs have mastered abstract knowledge, but they remain wordsmiths in the dark.” - Fei-Fei Li

─────────────────────────────────────

This becomes an existential constraint as AI moves from language tasks into the physical world. The path forward runs through spatial intelligence: the ability to perceive three-dimensional environments, reason about geometry and depth, assign semantic meaning to objects in space, and act on that understanding.

In her TED talk by Ms. Fei Fei Li explains that in biological evolution, the development of eyes catalyzed the Cambrian explosion where there was a rapid diversification of life enabled by the ability to perceive and navigate physical space and various other factors. Spatial intelligence may represent an analogous inflection point for AI.

This connects directly to Moravec’s Paradox the counterintuitive observation that tasks humans find trivially easy (walking, recognizing faces, catching a ball) require enormous computational resources for machines. These are the tasks that evolved over billions of years before it got genetically coded in humans. While other tasks that humans find challenging like chess, logic, art, reasoning etc. are straightforward for machines. These are the tasks that evolved over few thousands of years in human evolution.

According to Ms. Fei Fei Li, while current Large Language Models (LLMs) have mastered abstract knowledge, they remain “wordsmiths in the dark,” lacking the physical grounding necessary to navigate or manipulate the real world. These models fail at basic spatial tasks such as estimating dimensions, mentally rotating objects or maintaining the temporal coherence as they ignore the holistic nature of physical reality. To overcome these limits, AI must transition toward a deeper understanding of the geometric, physical, and dynamic rules that govern our world. Next frontier hence would be towards Physical AI or Spatial intelligence.

Towards Physical AI: The Software-Defined Vehicle & Humanoids

In AI, spatial intelligence is the missing link that pulls models out of text and into the physical world. It needs enablement of physical sensors to perceive the environment. For example Humanoids and Autonomous vehicles employ a strategy called “Sense – Plan – Act ” where sensors perceive the environment, Compute does the planning ( path & maneuvers) and Actuators respond according to decisions. Adding intelligence layer that implements feedback loop and enhances the thin “veneer of intelligence” is going to be the next frontier in Physical AI, may not be towards “conscious” AI.

Conclusion: Architects of an Intelligent World

Engineers and leaders who internalize the Software 3.0 paradigm, who deploy AI along with guard rails and who build the bridge between language intelligence and spatial intelligence are going to be the authors of next transformation.

(And a quick note: while the storyline, insights, and flow are mine, I used AI to help refine and polish this article.)