The Turning Point in AI

Artificial Intelligence (AI) has been evolving for decades, but we’ve now hit an inflection point where AI is moving from research labs into everyday tools, and everyone now talks about AI. Just as electricity transformed every industry a century ago, AI is in a position to have a similar sweeping impact.

In particular, the emergence of Large Language Models (LLMs) and the release of ChatGPT mark a turning point. AI is no longer limited to narrow, specialized tasks handled by big tech companies, but is becoming a thing that is available to everyone.

I had several discussions with tech leaders in the industry about this topic, and in this post, I want to share my perspective on why AI became the hype and the topic of everyone recently, while it has existed for multiple decades, how we got here, and what it means.

Note: The insights in this post reflect my personal views and experiences in exploring AI and machine learning.

AI and Machine Learning in Simple Terms

AI isn’t new. In fact, computer scientists have chased this goal for decades, the term “Artificial Intelligence” was first introduced in 50s. The core concept is that AI involves computers performing tasks typically requiring human intelligence.

Machine Learning (ML), a subset of AI, is about finding patterns in data and using them to make predictions. Machine Learning techniques broadly fall into three categories: Supervised Learning, Unsupervised Learning, and Reinforcement Learning. I won’t deep dive into explaining all those concepts because it’s not the main topic and I am not an expert in this area, but I will go through the basics so that if you don’t know, you can have a rough idea about what it is.

For example, one of the simplest ML techniques is linear regression, which fits a straight line through data points. If you have data about house prices and their sizes, you’ll likely see that bigger houses generally cost more. Linear regression finds the best-fit line that quantifies this relationship.

In the figure above, each data point represents a house, and the line is the model’s prediction: given a house’s size, it predicts the price. This shows how linear regression works. The model draws a “line of best fit” through the scattered data points. This illustrates how ML models learn from historical data.

Another major approach in ML is Reinforcement Learning (RL). Rather than learning from examples of correct answers, it learns by trial and error. It’s like training a dog: you reward it with a treat when it performs the right action. Similarly, an RL algorithm gets “rewards” for good decisions. Over time, it figures out how to act to maximize rewards.

Over the years, several key milestones have shaped the evolution of AI and ML. Initially, most models were based on statistical methods, such as logistic regression and decision trees, which provided simple but effective solutions. The early 2010s saw a major shift with the rise of deep learning, driven by the revival of “neural networks”, which significantly outperformed traditional methods in tasks like image and speech recognition. In 2015, the release of TensorFlow by Google made it easier for researchers and developers to build and deploy deep learning models at scale. Then, in 2017, the introduction of the Transformer architecture revolutionized natural language processing by enabling models to understand and generate language with contextual awareness. This innovation laid the foundation for today’s large language models (LLMs), such as GPT.

Whether it’s Supervised Learning (like linear regression) or Reinforcement Learning or any other method, many AI techniques boil down to recognizing patterns and making predictions. In the case of today’s AI models, that prediction is often literally “what comes next”. A large language model like GPT is trained by looking at trillions of sentences on the internet and learning to predict the next token in a sequence. A token can be as small as a character or as large as a word or subword unit, depending on how the model was trained. In other types of generative models, like those for images or video, the system predicts the next pixel values, often represented as arrays of numbers, rather than tokens of text.

So, it’s all probability-driven next-token prediction under the hood.

Before Large Language Models (LLMs)

To appreciate today’s AI, let’s look at its past.

Not long ago, most AI systems were narrow and domain-specific, each was an expert in one thing, like fraud detection or language translation, etc. These models required significant expertise and resources, limiting access mostly to large organizations. Google DeepMind’s AlphaGo beat the world champion in Go in 2016. OpenAI trained a system to play Dota 2. But each of these AIs could only perform one task. They were not general.

For the most part, if you weren’t a big company or research lab with access to huge datasets and specialized talent, your ability to use AI was limited. AI was powerful but locked behind narrow applications or APIs. However, advances in hardware, cloud computing, and model architectures allowed models to scale massively.

The Rise of Large Language Models (LLMs)

The rise of LLMs as a transformative force in AI was not driven by a single breakthrough, but by the convergence of multiple technological and methodological advancements.

First, the evolution of hardware, Graphics Processing Units (GPUs), initially designed for rendering high-performance graphics, turned out to be effective for deep learning workloads due to their ability to handle large-scale matrix and tensor operations in parallel. This parallelism significantly accelerated the training of neural networks compared to conventional CPUs.

Second, the availability of large-scale, diverse datasets enabled training at unprecedented scales. OpenAI’s GPT series exemplified this shift by training on massive text data on the internet, including datasets like websites, digital books, encyclopedias, public code repositories, etc. This provided models with wide-ranging exposure to human language, logic, and domain knowledge.

GPT-3 released by OpenAI in 2020, showcased an impressive leap in AI capabilities, performing tasks it wasn’t directly trained to do. But the launch of ChatGPT in late 2022 truly changed the game. Built on a fine-tuned version of GPT-3, ChatGPT made AI feel accessible, helpful, and fun. Within just two months, ChatGPT hit a hundred million users, faster than any app before it.

What made it special wasn’t just the technology, but how easily anyone could use it. Whether you were a student, a teacher, a marketer, a software engineer, or someone with no tech background at all, you could ask it a question or describe a task and see instant results. For the first time, people could personally experience what modern AI could do, without needing to know how it worked under the hood. This shift brought AI out of the hands of experts and into the everyday lives of millions. It marked the moment when AI stopped being a research buzzword and started becoming something you could use to get real things done.

Where do we go from here?

Today’s LLMs are impressive, but they’re not perfect. One common issue is that they sometimes generate information that sounds correct but is actually wrong, which people call “hallucination”. This happens because LLMs predict what comes next in a sentence based on patterns in data, not real understanding or a live database. Their training data also has a cutoff, so they might not know about recent events. LLMs can also pick up biases from the data they were trained on, which can lead to unfair or stereotypical responses. They can struggle with tasks that require complex, step-by-step thinking or deep knowledge of specialized topics unless you guide them carefully, which is also the reason behind the recent experimentations, such as chain-of-thought prompting.

AI development has gone through a cycle, from building many small models for specific tasks, to creating giant general-purpose LLMs, and now back to building specialized models again. But this time, we’re starting with a strong foundation: the general-purpose models. These specialized models are more powerful because they build on broad language and reasoning abilities, then get fine-tuned for specific tasks. To make these specialized models effective, researchers use techniques like

• Fine-tuning: A general model is trained further on a smaller, domain-specific dataset
• Knowledge distillation: Compressing a large model into a smaller, faster one while retaining much of its performance.

These techniques help adapt general models to specific use cases efficiently, without requiring massive resources or retraining from scratch.

This shift is also about practicality. Training huge models for every single task is expensive and overkill for many use cases. Instead, we’re creating smaller, faster models that are tailored for specific jobs. This helps reduce cost, speed up performance, and make AI more useful in everyday settings. In the future, we’ll likely use a mix of models, big and small, general and specialized, depending on what the job requires.

We will reach the point where having a data center running a network of thousands of AI models, each trained to be an expert in a different area, such as coding, math, physics, and more. These models work together like teams, combining their strengths to solve problems. For example, a coding model might team up with math and physics models to solve a scientific problem and then write the code for it. This setup can be replicated and scaled, allowing these expert AIs to work 24/7 to tackle challenges from all angles.

AI or LLMs specifically, like electricity, only becomes valuable when people build useful things with it. A powerful model by itself doesn’t change much, just like electricity needs light bulbs, refrigerators, or fans to be useful, AI needs real applications to make a difference.

What made ChatGPT and similar tools so impactful was that they put AI in the hands of everyone, not just researchers and big companies. Now, developers, startups, and even everyday users can build AI-powered tools to solve real problems.

We saw something similar when electricity was first introduced. It didn’t transform factories right away. At first, people just replaced steam engines with electric motors, which didn’t help much. The real change happened when they redesigned factories to use electricity.

AI is on the same path. Its biggest impact won’t come from just speeding up what we already do for the general model, but from rethinking how we work, create, and solve problems using AI as a core part of the process.

Why did coding become an early prime target?

Interestingly, coding emerged as an early AI success. Tools like Bolt, Cursor, and Lovable saw quick adoption, while big AI players released their own coding agents, such as OpenAI’s Codex, Anthropic’s Claude Code, and GitHub Copilot Agent. They are pushing AI-powered development into the mainstream. But why coding?

Personally, I think there are 2 main reasons:

• Code has rules and structure, which makes it easier for models to generate syntactically correct output.
• Code can be tested, if it runs, it’s probably doing something right.

Additionally, abundant training data (public code repositories) accelerated learning.

But that doesn’t mean the AI-generated code is good. Much of the training data is from open-source code, which varies in quality. It’s likely that only a small portion of code available on the internet is good code. What’s considered “good” code today might not hold up tomorrow as standards, technologies, and best practices evolve. The model may generate code that works, but it might not be secure, efficient, or maintainable. That’s where human engineers still play an important role in reviewing, improving, and validating the output.

Implications for Engineers and Builders

AI isn’t just a tool for completing your sentences or suggesting the next line of code. We need to see it as a new way of building software. Think bigger:

• Instead of writing every line of logic, you prompt the model to fetch and format data.
• You can build apps that orchestrate LLMs to handle user queries, data processing, and even reasoning.

The real potential of LLMs isn’t just making old workflows faster, but it’s about changing how we build things completely.

Let’s use a simple example: a weather app.

• Before AI: An engineer finds a weather API, reads the docs, writes code to connect to it, pulls the data, and builds a UI.
• With basic AI help: The engineer uses an LLM to search the API, then uses it to write parts of the code faster, like parsing the API response. It saves time, but the process is still mostly the same.
• With a new mindset: The engineer can ask the LLM to gather and organize weather data directly. For example: “Give me temperature, humidity, wind speed, and rain chances for the inputted location in a structured format”. The LLM could return a clean JSON with all that info, skipping the need to write low-level integration code, and focus more on making a good user experience.

This shows how LLMs can take on bigger roles, not just writing code, but more. That frees up engineers to focus on what makes their app unique. Treat AI as an engine you can build around. Prompting becomes programming. Human creativity and context still matter, but the AI gives us a new layer of abstraction and power. This is the new infrastructure for innovation.

Companies also need to rethink how they invest in machine learning. In the past, building a custom AI system and fitting it in the business context could take years. It required collecting data, training a model from scratch, fine-tuning it, and setting up complex infrastructure.

Today, engineers can move much faster by using a technique called Retrieval-Augmented Generation (RAG). Instead of training a new model, RAG uses a general-purpose language model and connects it to a data source such as a company’s knowledge base. When someone asks a question, the model pulls in the right information and gives a more accurate response based on that context.

This not only reduces the chances of the model making things up, but it also makes it easy to keep the system updated, just update the data, not the model itself. With this approach, companies don’t need large teams to build everything from scratch. They can start quickly by building on top of existing foundation models and benefit as those models continue to improve.

Conclusion

The turning point in AI isn’t just about smarter models or faster hardware. It’s the accessibility. It marks a shift in how we build and use technology. AI is now a tool for everyone, not just experts.

We’ve moved from narrow use cases to general-purpose models, and now we’re entering a new phase where these big models are fine-tuned to solve specific, real-world problems more efficiently and effectively. This cycle is not a step backward, it’s evolution with a stronger foundation.

Think of AI like electricity. Its true power wasn’t just in generating energy, but in enabling new tools, light bulbs, refrigerators, and everything else that changed daily life. AI is similar. Its value shows up when people build useful things with it, apps, tools, and services that solve problems in new ways. This will be the next turning point.

For engineers, this shift means thinking differently. Now, it’s about designing systems that work with AI as the engine. This is where engineers shine by applying creativity, domain knowledge, and critical thinking to unlock what AI can really do.

If you’re an engineer, a builder, or just a curious mind, this is your moment. The models are here. The tools are improving. The canvas is wide open. Let’s build!