There is a lot of hype around AI right now, and I wanted to share some of the details I have learned.
I’m specifically talking about Large Language Models (LLMs), not every form of AI. “AI” is an umbrella term that includes things like machine learning, computer vision, robotics, and LLMs.
A few useful terms:
- Pre-training: This is where the model is trained on massive amounts of data to learn patterns in language and information. It is somewhat similar to a child learning by observing the world around them — watching, listening, and picking up patterns with limited direct instruction.
- Post-training: This is where humans shape the behavior of the model after the initial training. This includes safety rules, behavior adjustments, and tuning how it responds. This is more like active parenting or teaching — helping someone operate successfully in society instead of simply repeating what they observed.
At their core, LLMs are essentially enormous mathematical functions (f(x)=Ax+Bx+Cx… with billions or trillions of parameters. They are not “thinking” in the way humans think. They are pattern prediction systems.
LLMs do not create truly novel ideas out of nowhere. They generate responses based on patterns from things humans have already created. They are incredibly good at recombining, summarizing, restructuring, and extending ideas, but that is different from independent thought.
They also do not continuously learn from normal user conversations and permanently integrate that information into the base model. Most people talk about them like they are constantly evolving from every interaction, but that is generally not how production systems work.
A lot of the public conversation around AI swings between extremes. Some of it is designed to drive clicks. Some of it is tied to investor excitement and stock prices. Depending on the day, AI is either going to destroy civilization or solve every problem humanity has.
Reality is usually less dramatic.
Some of the viral stories about AI manipulating engineers, trying to escape, or secretly taking actions on computers leave out important context. In many cases, the systems were being intentionally stress-tested, guided into edge cases, or responding based on patterns they learned from existing human-written material online.
That does not mean there are no risks. It just means we should be careful not to treat statistical prediction systems like magical beings.
We also should not treat LLMs like all-knowing machines. It is probably healthier to think of them more like fallible assistants. They can be useful, persuasive, confident, wrong, biased, incomplete, or occasionally misleading — just like people can be.
Humans make mistakes, and LLMs do too.
AI can absolutely make work easier, but there is also a danger in offloading too much thinking to it. If you stop exercising judgment, writing, analysis, or problem-solving yourself, those skills can weaken over time. In many cases, you may not even save as much time as expected because you still need to verify the output.
I also do not think AI will replace all jobs. More likely, it will reshape jobs, automate pieces of work, and change how people spend their time. Many roles will become more about reviewing, directing, integrating, and validating AI-generated work.
One important thing to remember is that LLMs are fundamentally trying to predict likely continuations. In simple terms, they are trying to guess the next most probable token or word. That means leading questions matter. If you strongly frame a question a certain way, the model will often try to continue the story you started rather than challenge the premise.
That is one reason why prompting matters so much.
Right now, many companies are using AI tactically — summarizing meetings, generating documents, answering support questions, or speeding up research. But if organizations want AI to truly transform the business, it has to be integrated strategically into workflows, systems, decision-making, and operations rather than used as a disconnected tool.
In practice, I think AI is especially useful for:
- Brainstorming
- Creating standardized documents
- Summarizing information
- Organizing research
- Acting as an improved form of web search
- Helping people get started faster
One final thing people are noticing is the increase in electricity demand from AI data centers. A large part of that is simple infrastructure pressure. Training and running these systems requires enormous computing power, which increases energy demand (supply and demand) and forces upgrades to electrical grids and supporting infrastructure.