Witnessing the AI Revolution: From Deep Learning to Long-Running AI Agents
Over the past 15 years, artificial intelligence has not merely produced better models. We have moved from systems that recognize images to models that converse with us, from conversational models to agents that use tools, and from there to AI systems capable of working on a problem continuously for hours or even days. So what will be the next major breakthrough in 2027?
A Question That Emerged from a Conversation
Today, while talking with a friend who is also deeply interested in artificial intelligence, our conversation naturally turned to the extraordinary pace of development in AI over the past few years.
After a while, we started recalling the technologies we had been discussing over the years: Deep Learning, RNNs, LSTMs, Transformers, large language models, multimodal models, reasoning models, RAG, function calling, agents, MCP, computer use, coding agents, and most recently, AI agents capable of working independently for extended periods of time.
When you put all of these developments side by side, the speed of change becomes much more visible.
In fact, we have not simply developed more capable AI models over the past few years. The way we use artificial intelligence itself has changed several times.
🧠 1. Deep Learning: The Beginning of the Major Breakthrough
Artificial neural networks were certainly not new. However, for many years they could not demonstrate their full potential because of limitations in data availability, computing power, and training techniques.
In 2012, AlexNet’s success in the ImageNet competition demonstrated how powerful deep neural networks trained on GPUs could become.
The breakthrough that started in computer vision quickly spread to speech recognition, natural language processing, and many other machine learning problems.
From today’s perspective, this may seem almost ordinary, but one of the key foundations of the AI revolution we are experiencing today was established here:
🔁 2. RNNs, LSTMs, and Sequence-to-Sequence Models
While CNN architectures achieved remarkable success in image processing, natural language presented a different challenge: language is sequential.
RNNs, and particularly LSTMs, made it possible to retain information from previous steps and process sentences, conversations, and time-series data more effectively.
Then came sequence-to-sequence, or seq2seq, architectures. They enabled a model to receive one sequence and generate another sequence as output. Machine translation became one of the most important applications of this idea.
Today, we consider it completely natural to give a model a piece of text and receive another piece of generated text in return. Yet encoder-decoder and seq2seq architectures were important milestones on the road toward modern generative AI.
👁️ 3. Attention: Teaching the Model Where to Look
There is an important chronological detail here: the attention mechanism did not begin with the Transformer.
Around 2014, particularly in neural machine translation research, models began to learn how to focus more heavily on relevant parts of the source sentence while generating each part of the output.
This idea was extremely important because a model no longer had to compress the entire input into a single fixed representation.
⚡ 4. 2017: Attention Is All You Need
In 2017, the publication of Attention Is All You Need became one of the most important turning points in the history of artificial intelligence.
The Transformer architecture greatly reduced the dependence on the sequential processing used by RNNs and placed self-attention at the center of the architecture.
Today, the architectural roots of major model families such as GPT, Claude, Gemini, Llama, Qwen, and Mistral can largely be traced back to this transformation.
📚 5. Pre-Training and the Foundation Model Paradigm
After the Transformer, another major paradigm gained momentum.
Instead of training a separate model from scratch for every task, researchers began training large general-purpose models on massive datasets and then adapting them to different downstream tasks.
Models such as BERT demonstrated how powerful this approach could be in natural language processing.
The discussion gradually moved beyond the question of “Which architecture performs better?” toward a broader question: “Can we train general-purpose models capable of performing many different tasks?”
📈 6. Scaling and GPT-3
Around 2020, the relationship between model size, training data, computational resources, and performance became far more visible.
Scaling law research suggested that larger models did not simply perform the same tasks slightly better. In some cases, new capabilities appeared as scale increased.
GPT-3 made another important phenomenon highly visible:
A model could perform many different tasks without being retrained for each one, simply by receiving instructions or a few examples in the prompt.
Zero-shot and few-shot prompting became increasingly important.
📄 7. RAG: Moving Beyond the Model’s Parameters
One of the major limitations of large language models was their inability to reliably access current, organization-specific, or previously unseen information.
Retrieval-Augmented Generation, or RAG, approached this problem from a different direction.
Instead of trying to encode every piece of information inside the model’s parameters, relevant information could be retrieved from an external knowledge source and added to the model’s context.
Over time, RAG evolved into more sophisticated architectures involving embeddings, vector databases, semantic search, hybrid search, reranking, query rewriting, and knowledge bases.
This approach remains extremely important today, particularly in enterprise AI applications.
🎨 8. Diffusion Models and Generative Visual AI
Meanwhile, the AI revolution was not limited to text.
With the rise of diffusion models, text-to-image generation rapidly became mainstream.
This was followed by major progress in image editing, style transfer, image-to-video generation, text-to-video generation, speech synthesis, and AI-generated music.
The concept of Generative AI was no longer limited to models that generated text.
👀 9. Multimodal Models
The next important step was bringing different types of data together within the same model.
The model was no longer limited to understanding the text we typed.
Because the real world is not composed only of text, multimodality became an essential step toward more general-purpose artificial intelligence.
💬 10. 2022: ChatGPT Brings AI to the Masses
The release of ChatGPT on November 30, 2022 was, in my view, not only a technological milestone but also a major social turning point.
For the first time, large language models entered the daily lives of ordinary users at massive scale.
Teachers, students, doctors, lawyers, engineers, managers, artists, software developers, and entrepreneurs all began interacting with the same technology.
Instruction tuning and alignment methods based on human feedback made interaction with these systems far more natural from the user’s perspective.
At the same time, a new term entered everyday vocabulary:
🔧 11. Function Calling and Tool Use
Large language models could generate highly impressive text, but their ability to interact with real-world systems was limited.
Function calling and tool use began to change that.
A model could now call an API, access a database, use a calculator, perform a web search, or invoke another software component.
This seemingly small change later became one of the foundations of agentic systems.
🧠 12. Reasoning Models
In 2024, the rise of reasoning models introduced another important shift.
The focus was no longer only on adding more parameters to the model. Models could also spend more computation on a difficult problem before producing an answer.
The concept of test-time compute became increasingly important.
Especially in mathematics, science, coding, and complex problem solving, the idea of the “larger model” was joined by another idea: the “model that spends more time reasoning about the problem.”
💻 13. From Chatbot to Coding Agent
In my view, one of the most important recent transformations has occurred in software development.
At first, we asked models to generate small pieces of code.
Then came copilot-style systems embedded directly inside IDEs.
Later, AI systems began reading entire repositories, creating files, running terminal commands, debugging errors, writing tests, and refactoring code.
🔌 14. MCP: Standard Connections Between AI and the Outside World
As agent systems became more common, another problem became increasingly visible.
Every AI application had to be integrated separately with GitHub, file systems, databases, enterprise applications, and many other external services.
The Model Context Protocol, or MCP, became an important development for establishing more standardized connections between AI applications, tools, and data sources.
I do not think MCP should be viewed merely as a technical protocol. For an agent ecosystem to grow, models need standardized and manageable ways to connect to the outside world.
🧩 15. Agent Skills
The next step was to stop explaining the same procedures to an agent again and again, and instead define reusable skills that describe how particular tasks should be performed.
A skill can, for example, teach an agent:
📊 How to analyze an Excel file,
📄 How to prepare a professional report,
🧪 How to execute a testing procedure,
🏢 How to follow an organization-specific workflow,
💻 How to perform a particular software development task.
The agent can then use that skill whenever it becomes relevant.
This seems especially important in the transition from general-purpose AI toward specialized digital workers.
🖥️ 16. Computer Use and Browser Use
Agents then began doing more than simply calling APIs.
They started seeing screens, moving the mouse, clicking buttons, navigating websites, and interacting with the same graphical user interfaces that humans use.
I believe this is a critical development.
Most software in the world does not have an API specifically designed for AI. But almost every application has a user interface designed for humans.
🌍 17. Open Models and Local AI
While all of these developments were taking place, another transformation was happening in parallel.
Open-weight models rapidly narrowed the gap with closed models across many practical use cases.
Model families such as Llama, Mistral, Qwen, and DeepSeek helped build a powerful open ecosystem.
Quantization, more efficient inference engines, and advances in hardware made it possible to run models locally that previously would have required large data centers.
Local AI systems that keep organizational or personal data on-premise are becoming increasingly realistic alternatives.
The emergence of personal agent software that can run continuously on a local computer is another important part of this transformation.
🤖 18. 2026: The Year of AI Agents
For some time, many people had predicted that 2026 would become “the year of AI agents.”
I think that prediction has largely come true.
However, the meaning of the word “agent” has also changed.
Early agent architectures looked roughly like this:
Today, we are talking about far more comprehensive systems.
🧭 They can plan.
🧰 They can choose the right tool.
🗂️ They can manage files and context.
🧠 They can preserve state across a task.
🤝 They can delegate work to other agents.
💻 They can run code and terminal commands.
🌐 They can perform actions on the web.
🔍 They can evaluate their own outputs.
🛡️ They can operate within permission boundaries.
👤 They can request human approval when necessary.
🏗️ 19. The Next Step: Agent Harnesses
One concept I find particularly important in recent developments is the agent harness.
As we assign agents longer and more complicated tasks, the intelligence of the underlying model alone is no longer enough.
A model can perform extremely well during the first few steps and then drift in the wrong direction.
It can incorrectly assume that the task is complete.
It can forget previous decisions.
It can reach a conclusion without testing its work.
Its context can gradually degrade during long-running tasks.
This is why we increasingly need an execution environment around the model that manages and supervises its work.
This moves us from chatbots that work for a few minutes toward AI systems that can work toward a goal for hours, or potentially much longer when supported by the right infrastructure.
How Has the Short History of Modern AI Actually Changed?
🔮 So What Will Define AI in 2027?
I do not think the next major leap will necessarily be just another, slightly more powerful conversational model.
One of my strongest candidates is the persistent agent: a personal or enterprise AI system that can work continuously while preserving its state and context over time.
In today’s chatbot paradigm, we go to the AI.
We open a window, ask a question, receive an answer, and then continue working ourselves.
The next stage may be very different: AI could become a persistent part of our working environment.
🧠 It could recall our previous work when relevant.
📧 It could work with our email and messages.
📅 It could follow calendars and tasks.
📁 It could understand our files and projects.
🌐 It could conduct research when necessary.
💻 It could perform actions on our computers.
🧩 It could use the skills it needs.
🔌 It could connect to enterprise systems.
🤝 It could collaborate with specialized agents.
🛡️ It could understand its permission boundaries.
👤 It could request human approval for critical actions.
✅ It could verify its own work before reporting completion.
☁️ Cloud or Local?
I also do not think this future will be entirely cloud-based.
As personal and enterprise data becomes more important, hybrid architectures will probably become increasingly attractive.
Small and fast models may run continuously on local devices.
When more complex reasoning is required, the system could call more powerful cloud models.
Organization-specific knowledge could remain on-premise.
Highly sensitive tasks could be handled entirely by local models.
🤖 Another Strong Candidate: Physical AI
Another major field to watch is artificial intelligence moving beyond the computer screen and into the physical world.
When vision, reasoning, planning, and agent technologies are combined with robotics, we may see major advances in embodied AI or physical AI.
An agent that learns how to click a button on a computer screen today and an AI system that controls a physical robot tomorrow may be more closely related problems than they first appear.
🚀 Perhaps the Most Important Part of the Revolution Is Just Beginning
When I look back at the developments of the past 15 years, the most interesting change for me is how our questions have changed.
During the Deep Learning era, we asked:
“Can the machine recognize this?”
During the LLM era, the question became:
“Can the machine understand and explain this?”
With reasoning models, we started asking:
“Can the machine solve this problem?”
In the age of agents, the question is very different:
I do not think this is a small change.
For most of the history of computing, humans learned how to use computers.
Now computers are beginning to learn how to use the same tools that we use.
Perhaps the most important transformation of the next few years will happen precisely here.
Which technology or approach do you think will become the defining AI trend of 2027?
🤖 Persistent personal AI agents?
🏢 Enterprise agent systems?
🏗️ Agent harnesses and long-running autonomous work?
🧠 More powerful reasoning models?
🖥️ Computer use and browser use?
🌍 Open and locally running models?
🤖 Physical AI and robotics?
Or will the next major breakthrough come from something that we are not yet talking about enough?
Selected Milestones
Krizhevsky, Sutskever & Hinton — ImageNet Classification with Deep Convolutional Neural Networks, 2012
Sutskever, Vinyals & Le — Sequence to Sequence Learning with Neural Networks, 2014
Bahdanau, Cho & Bengio — Neural Machine Translation by Jointly Learning to Align and Translate, 2014
Vaswani et al. — Attention Is All You Need, 2017
Devlin et al. — BERT, 2018
Kaplan et al. — Scaling Laws for Neural Language Models, 2020
Brown et al. — Language Models are Few-Shot Learners, 2020
Lewis et al. — Retrieval-Augmented Generation, 2020
Ho, Jain & Abbeel — Denoising Diffusion Probabilistic Models, 2020
