How Did AI Suddenly Become Semi-Sentient?

Less than 2 years?

Photo by Luke Jones on Unsplash

In 2018 AI chatbots got stuck in repetitive loops if you asked anything complex.

Today they autonomously write software and manage multi-step plans.

Here’s exactly how that happened. Not the hype version. The actual technical story.

The foundation — transformers

Before 2017 language models read text one word at a time.

Painfully slow. Fundamentally limited.

Google researchers published a paper called Attention Is All You Need and introduced the transformer architecture.

The key innovation — separate encoder and decoder stacks that could process entire sequences of text in parallel simultaneously.

This unlocked GPU acceleration for language modeling. GPUs are built for parallel computation. Transformers needed parallel computation. The match changed everything.

Scaling laws

The popular narrative says engineers just kept building bigger versions until the systems woke up.

The reality is more mathematical than that.

In 2020 researchers published scaling laws — precise equations predicting exactly how much a model would improve based on three variables.

Parameter count. Training data volume. Raw compute power.

Increase all three together in the right ratio and performance improves predictably. Not magically. Not emergently. Mathematically.

This turned AI development into an industrial process. Massive financial investment.

Physical warehouses of specialized chips running for weeks. Brute force compression of a large chunk of human knowledge into what amounts to a very sophisticated pattern-matching system.

But raw pattern matching isn’t a useful assistant.

RLHF — the behavioral layer

Early massive language models had one objective.

Autocomplete internet text.

The internet is chaotic, contradictory, and frequently toxic. Raw model outputs reflected that. Frequently unhelpful. Sometimes dangerous.

Researchers introduced Reinforcement Learning from Human Feedback — RLHF.

Human labelers reviewed model outputs and rewarded accurate, safe, helpful responses. That feedback loop shifted the model’s internal structure. Turned an unpredictable autocomplete experiment into a polite, structured assistant.

Something interesting happened as this behavioral alignment layer stacked on top of massive pre-training.

Emergent abilities appeared.

Without being explicitly programmed to translate foreign languages — the models started translating foreign languages accurately.

Without being explicitly programmed to write code — they started generating complex functioning software.

RLHF is what made these systems commercially viable.

Test time scaling — the reasoning breakthrough

Even a perfectly aligned model had a critical limitation.

It relied entirely on fast instinctive text generation. What psychologists call System 1 thinking. When asked a complex logic question it was forced to guess instantly without checking its work.

The solution — test time scaling.

Instead of spending all compute during initial training, allocate extra computational time at the moment a user submits a prompt.

Using a chain of thought approach the model explores multiple branching paths. Flawed logic gets discarded. Correct alternative routes get found.

This is System 2 thinking — deliberate, slow, analytical reasoning rather than fast pattern matching.

Test time scaling broke the logic bottleneck. Hard problem-solving requires not just data but computational time to actually process it.

Tool use — escaping the chat box

A perfectly reasoned plan is useless if the system is trapped inside a text box with no connection to the outside world.

Engineers developed tool use.

Models can now emit specific text tokens that trigger actions in external software environments.

Neural networks are inherently bad at literal arithmetic — they’re pattern matchers, not calculators. So when faced with a complex equation the model writes a command to open a calculator. Guaranteed correct result.

It can write Python scripts. Execute them. Read the error messages. Correct its own logic based on what it finds.

External tools transform the model from a text generator into an active agent in the digital world — capable of verifying its own reasoning through live software interaction.

What an AI agent actually is

Combine test time reasoning to map out a plan with tool use to execute it autonomously.

That’s an agent.

The rise of agents isn’t a single breakthrough. It’s a compounding effect where specific technological layers stack together and multiply each other’s capabilities.

Five years ago the industry was struggling to generate coherent paragraphs.

Today it’s deploying systems that debug entire codebases, analyze legal contracts, and manage supply chains.

What comes next

Engineering focus is already shifting.

Instead of pouring everything into raw pre-training, labs are scaling inference compute — more thinking at query time rather than just more training upfront.

They’re also using synthetic data — AI generating training material for the next iteration of AI.

If stacking transformer architecture, scaling laws, RLHF, test time compute, and tool use created reasoning agents in five years — the next layer added to this stack will define how humans interact with technology for the decade after that.

Nobody knows exactly what that layer is yet.

That’s what makes this moment genuinely interesting rather than just hype.

Want more stuff like this? Even more detailed? I go into deep dives every single day on my 📌Substack📌

At the price of a coffee I can give you enough knowledge to replace a bachelor’s degree in computer science or maybe even a masters, who knows.

💜 If you like my work and would like to support to me financially. Even a small donation would help a lot!

My PayPal — Please Support