Introduction to AI in 2026

AI is now a mainstream tool

Artificial Intelligence used to feel futuristic. Now it’s becoming a normal part of everyday life — the same way Google became normal in the 2000s.

You don’t need to be a technical guru to use AI anymore. Many people already use it to write letters, clean up emails, translate documents, or help with schoolwork. AI is becoming a tool for regular daily-life tasks, so it’s no longer just an integrated part of systems, where we - regular people - wouldn’t even see it.

We can’t predict the future, but it’s clear that AI isn’t replacing humans anytime soon. Humans with AI skills will be replacing those without. Like Email replaced Fax machines.

Introduction to Generative AI

The past 4 years created a technological shift, as well as a shift in people’s habits and language, so nowadays AI is usually meant as apps that can read, understand, and write text similar to how a human would. But AI isn’t new, the term itself is very broad and can mean a number of things.

AI has been a huge part of our lives for the past few decades. Product, movie or music recommendations, search engines, navigation systems, fraud prevention, medical imaging, photo editing apps, and so on, all use AI to perform its function.

So what happened 4 years ago?

The simple answer is that natural language AI became mainstream with the release of ChatGPT. What used to be gatekept behind closed doors and integrated into the products of the select few multi-billion companies, is now accessible to almost anyone.

And it’s not just text generation these days. AI that falls into the category of Generative AI can also produce images, audio, video, or programmer code.

ChatGPT, Claude, Gemini, Mistral, DeepSeek and others

The push to mainstream use of AI was done by the company OpenAI. Their web-based AI Chat app called ChatGPT rocked the world with its capabilities, showing us that machines can now understand and produce human-like messages on almost any topic.

This was groundbreaking in multiple ways. It unlocked a huge potential for digital work, learning, storytelling, and automation. Before ChatGPT, Large Language Models (LLMs) were mostly used by AI researchers, but it going mainstream gave us a simple way to interact with this technology in ways that were mostly science fiction.

And while OpenAI was way ahead of everyone, other companies now understood that they have to start catching up, to what can be considered a bigger event than the industrial revolution.

Anthropic released their Claude model, Facebook published LLama, Google - Gemini, Alibaba - Qwen, xAI - Grok, French company Mistral released their first tiny model, Chinese company DeepSeek their own, and so the race began.

How does it work? What’s the difference?

Different AI models are not only built different, but they’re trained on different information and serve different purposes, although most of them can be used as general Q&A assistants.

The way it works is - companies take enormous amounts of data, think 1000s of books, entire libraries, they hand-pick the content that is valuable to their cause, press a button and wait for days, until intelligent algorithms crunch that information and zip it into a single file which becomes the AI model.

These models are basically amazing copy/paste and guessing machines, they’re disguised as intelligent, and in a way they are, but people need to be aware that they do not posess a consciousness, nor do they have feelings or their own agendas.

They are boxes of information and math, trained to serve the purpose of the companies that make them. Some of those purposes are:

• Knowledge of information, brainstorming
• Productivity, writing, communications
• Safety focused chatting
• Translation services
• Media generation (audio, images and video)

Because access to data (sometimes not through ethical means) is tricky, different companies use different information to train the models. Some models are trained on the whole internet, some on internal company and customer data, some models use synthetic data (generated by other AI models) to guide the questions into good answers.

Local AI

Another AI area advancing at a rapid pace is running AI on personal computers and mobile phones. This is amazing for people’s privacy as no conversation data leaves our devices, everything is done on our own hardware. But it’s tricky.

For AI to be fast or even usable, the entire model file has to fit into computer’s graphic card’s (GPU) memory (different memory than RAM). This is a huge limitation that prevents the majority of personal computers, phones or even home servers from running these models, as the memory is often limited to something like 2GB to 16GB.

To give you a clearer view, it’s estimated that models like ChatGPT or Claude can range from 400GB to 3TB, so they are run in computer farms with 100s or 1000s of GPUs.

Model cleverness is directly tied to the size of the model files and the amount of information it holds.

So why bother?

After a lot of trial and error, researchers found ways to train smaller models with enough general information for it to be usable. These models can answer simpler queries, assist in writing tasks or even be used for automation.

Many people (including the Awee team) already run these models for occasional AI-assisted tasks, but it requires good or moderate hardware. In addition to the models being small, model optimizations are performed to reduce the size even further without a huge loss in quality, via a method called quantization.

There are many exciting advancements and it’s improving rapidly, so we’re hoping 2026 will unlock this even further.

Free Cloud Models

As they say “If you’re not paying for the product, you are the product”, which is also true for free cloud models. There are 10s of good free models online, however, their access comes at a cost - your data. Everything you send to these models is then reviewed and used. Either to train future models, fix mistakes, or just market research and analysis.

Never talk to these models about anything sensitive or attach personal information.

The AI currency

When we talk to AI, a few things happen with the text we send and get back. Behind the scenes, AI models don’t operate on text, they operate on tokens.

A token is a number representation of a word or a part of a word. This number is used as a search keyword because computer math is much faster on numbers than text. Every model has a vocabulary, which holds word = number pairs, like cat = 25, dog = 26, and so forth. When the AI receives some text, it converts it into numbers from the vocabulary, and looks at all of them to better understand the meaning of user’s query. It then does its math magic and guesses the next tokens which contain the answer, finally replacing the numbers with the words from the vocabulary back into text.

This process is split into two. Input processing - a person asking a question and AI reading it, and output generation - AI model giving the answer. Everything happens using tokens. These both processes are very different, they require different amounts of computing power, and that is why AI providers charge different prices for input and output tokens.

When you look at AI model pricing, it’s usually per 1 million input or output tokens, where output tokens are a few times more expensive than the input.

Conversational AI

When we chat with AI, every single message we send or receive is sent together. This is to ensure that the AI has enough information to understand the topic and user questions.

1. Start of chat

• User: What’s the capital of Lithuania?

2. We get an answer

• Assistant: The capital is Vilnius

3. Continue chat

• User: What’s the capital of Lithuania?
• Assistant: The capital is Vilnius
• User: And the population?

4. We get another answer

• Assistant: About 3.5 million

5. Continue chat

• User: What’s the capital of Lithuania?
• Assistant: The capital is Vilnius
• User: And the population?
• Assistant: About 3.5 million
• User: Who’s the president?

If this wasn’t the case, and we only sent the last user message like “And the population?”, the AI wouldn’t know we’re talking about Lithuania, and couldn’t produce the correct answer.

This also means that the inputs can get expensive in large conversations. There are optimizations done on the AI provider side and AI apps to reduce the input costs, but it isn’t only a matter of input price. Answer quality and speed are affected by how much information we submit to the AI. AI models aren’t perfect, they lose information once the context becomes too large, so ideally (but not necessarily) we want to maintain shorter chats.

AI capabilities in 2026

While text generation is a big part of modern AI, this technology is evolving almost daily.

Reasoning models

A fairly new advancement is reasoning capabilities. Some newer models are trained to have an internal monologue which generates an extra step (or rather multiple thinking steps) that’s used to generate the final answer. These models often come with a setting for the level of thinking to perform, however, turning it off usually means the model just hides it.

Reasoning generates better answers, but its cost is often higher.

Large context windows

A context window is the size of the content the model can ingest before producing an answer. Prior to 2025 this was often enough for most conversations, but still limited. Nowadays models can accept entire books or hundred-page PDFs as the conversation context. Uploading tens or hundreds of pages is likely pretty expensive, and the model may hallucinate, but this area is getting better.

Advanced output formats and automation

Most newer models are trained to generate structured outputs. This is a fundamental capability that allows developers and systems to use the model to automate tasks.

Computer systems operate on functions. Functions accept configuration parameters. For example - a function to check the weather can accept two parameters:

• Location of the place to check the weather for
• Temperature units (Celsius °C or Fahrenheit °F)

Before AI models, figuring out and extracting the location and the units from a user’s query, a document, chat app or other place was an extremely challenging task. Unless the user followed a specific word structure, it was close to impossible considering that there are millions of places in the world.

The text could contain a country code, or a city, or a district within a city, or none of these things. It could be misspelled, it could be abstract, it could be in a different language, so the developer would have to have written a list of every single possible location to be able to match the user’s preference.

Modern AI models have so much knowledge that they can correctly guess the location even if it’s written with a typo. Once it extracts this information, it is able to format it in a computer-friendly way (JSON, XML, etc.) so then the systems can read and pass that information into the functions.

This feature is the foundation of AI-driven automation. Developers define functions, and AI models produce the parameters for almost any task.

This is one of the main things that excited us to start Awee. The possibilities are endless, as long as the AI is clever enough to figure things out.

Multimodality

Models available in 2026 have improved the scope of their use by accepting multiple types of content. You can text it, upload an image, a voice track, and it can generate an answer in any of those same formats or combinations of them.

This creates opportunities for apps to innovate beyond a simple chat.

Example:

Take a picture of your fridge, record a voice conversation with your family on what you want to eat and ask the AI to generate a few recipes from the products that are currently available.

The very real risks of AI

These modern machines are incredibly knowledgeable. They contain information on the entire internet, libraries, and news articles. They are an artificial brain that powers chat apps, intelligent systems and can help humanity in ways that weren’t realistic just a few years ago.

But that also comes with serious risks, just ask Sarah Connor.

People have already started using AI for therapy, tutoring, medical advice, and other sensitive topics. While this may not be necessarily bad, people also need to realize the dangers.

Hallucinations

AIs falsify facts. They are prediction machines, which contain billions of texts so there’s a very high chance of two or more conflicting pieces of information. In the age of weaponized disinformation, the data used to train the AI can simply contain lies. Data collected from public forums, comment sections, even news articles and other places on the internet where people or bots post their opinions has an impact on the model’s knowledge and decision-making.

In addition to the training data, AI models are designed and configured to be friendly assistants. This implies that they are keen to agree to whatever the user is suggesting, and when the user isn’t knowledgeable about the conversation topic, it can lead to a strong confirmation bias, leading the conversation down a potentially dangerous path.

Bias, censorship and discrimination

AI providers are companies that can involve themselves in politics, they can be paid or lead to cause harm to communities, influence opinions and misinform. It means that their AI models can be deliberately designed to do this.

Models can be trained to lie, ignore facts and produce harmful ideological or non-consensual content. Ranging from racist remarks, to sexualized images or hateful speech. Elon Musk’s xAI model Grok has been in a negative spotlight for exactly this. Multiple countries launched lawsuits or even banned these models entirely.

Lack of data privacy

When we submit information to the AI, that information is stored and processed by the AI provider. While models themselves are stateless (they do not store any information after they are trained), companies providing access to the models intercept our content. In some cases this is fully valid - there have to be safeguards to prevent people from learning how to make chemical weapons or bombs. But at the same time companies can use our data to further train the models, sell our information for political, marketing or for-profit purposes, as well as accidentally leak it through security breaches, mismanagement or misconduct.

Even with all the security measures, ethical policies and enforcements, sensitive personal data should never be sent to the AI providers. Providers comply with local laws, they store information in data centers in physical locations. Leaderships and governments change, so even if information is secure for now, there are no guarantees it will stay that way.

Spam and fraud

AI is capable of generating text, audio or even visual content. Criminals can use these tools to impersonate, defraud and trick people. Society will have to work and come up with ways to educate its people and reduce the risks of it happening.

Conclusion

We live in exciting times. What used to be science fiction is now a reality.

The ecosystem of Cloud AI and Local AI is growing, and we’re hoping to see it become more affordable and accessible to more people. It is a type of technology that can actually benefit the world, and not be just another over-hyped buzzword in the tech industry.