AI Universe

Dive into the universe of artificial intelligence: the history of AI from its very beginnings, fun facts, benchmarks explained, comparisons and guides to finding the right AI for every kind of task. In-depth content to understand AI beyond the headlines.

6 published articles

The universe of artificial intelligence, beyond the headlines

The news tells you what happened this week. This category covers everything else: where artificial intelligence came from, how it got here, what benchmark scores actually mean and which AI makes sense for which kind of task. It's the place for anyone who wants a deeper understanding of AI, without the rush of the latest launch.

Here you'll find the history, fun facts, benchmarks, task-by-task comparisons and the ideas that explain why the technology works the way it does — including when it gets things wrong.

When AI was invented

It sounds like a simple question, but the answer depends on where you start counting. The idea of thinking machines is old; the scientific field has a date and a place.

1950 — Turing's question. Mathematician Alan Turing publishes a paper posing a provocative question: can machines think? To sidestep the philosophical debate, he proposes a practical test, which became known as the Turing Test.
1956 — the name and the field. A summer workshop at Dartmouth College brings researchers together around the term "artificial intelligence," coined by John McCarthy in the proposal for the event. It's considered the official birth of the field.
1958 — the perceptron. Frank Rosenblatt introduces one of the first neural network models, loosely inspired by neurons. The idea sat on the back burner for decades before becoming the foundation of modern AI.
1966 — ELIZA. An MIT program that mimicked a therapist by turning the user's statements back into questions. It was simple, but plenty of people felt understood by it.
The 1970s and late 1980s — the AI winters. The promises outran the results, funding dried up and the field went through stretches of skepticism.
1997 — Deep Blue. IBM's computer beats world chess champion Garry Kasparov in an official match.
2012 — the deep learning breakthrough. A neural network trained on graphics cards wins an image recognition competition by a wide margin, and deep learning takes over the field.
2016 — AlphaGo. Google DeepMind's system defeats Lee Sedol at Go, a game long considered too hard for computers.
2017 — the Transformer. Google researchers publish the architecture behind virtually every large language model in use today.
2022 — ChatGPT. Its launch brings conversational AI to the mainstream and kicks off today's race between the labs.

Fun facts about AI

The history of artificial intelligence is full of moments that explain the technology better than any definition:

The word "robot" came from the theater. It first appeared in a 1920 Czech play by Karel Čapek, derived from a word tied to forced labor.
The ELIZA effect. People tend to credit any system that responds in natural language with understanding and feelings. It was first noticed in the 1960s and still holds true for today's chatbots.
Neural networks were almost abandoned. The technique that powers AI today was written off as a dead end for years, until massive amounts of data and graphics cards changed the game.
Video games helped build modern AI. GPUs, designed for gaming graphics, turned out to be perfect for the repetitive math involved in training models.
Language models don't read words. They read tokens, chunks of words. That's why something like counting the letters in a word can trip up a model that writes an entire essay without missing a beat.

Benchmarks: how AI gets measured

Every launch comes with a table of numbers. Those numbers are benchmarks, standardized tests that let you compare models on the same tasks. Some of the most cited:

MMLU — multiple-choice questions across dozens of subjects, from law to medicine.
GPQA — graduate-level science questions designed so they can't be answered with a quick search.
SWE-bench — real issues pulled from software projects that the model has to fix in the code.
ARC-AGI — visual puzzles that test the ability to learn a new rule from just a few examples.
Preference leaderboards, like LMArena, where people compare answers from two models without knowing which is which.

Benchmarks are useful, but they have limits. Older tests become "saturated," with every model scoring near the top. Questions can leak into training data. And a high score on a test doesn't guarantee strong performance on your specific use case. That's why, here, the numbers always come with context: what the test measures, what it leaves out and how much it matters in real-world use.

Which AI to use for each task

No AI is the best at everything. There's the right tool for each kind of work, and choosing well saves you time and money. Broadly speaking:

Writing and long-form content: chat assistants like ChatGPT, Claude and Gemini, each with its own writing style.
Coding: models that are strong at code and tools that work inside your project, like coding agents.
Research and sources: tools that search the web and cite where the information came from.
Images, video and audio: specialized generative models, which work very differently from text models.
Document and data analysis: models with a large context window that can read long files in one go.
Automation: models connected to other tools through integrations and agents.

The guides in this category dig into each of these cases, comparing options by what they actually deliver, not just what they promise.

Frequently asked questions

How is this category different from News? News covers what just happened. AI Universe covers what still matters after the buzz fades: history, concepts, comparisons and fun facts.

Who invented artificial intelligence? No single person. Alan Turing posed the question in 1950, and John McCarthy named the field in 1956, alongside researchers like Marvin Minsky, Claude Shannon and Nathaniel Rochester.

Can you trust benchmarks? They're a good starting point, not the final word. The test that matters most is still using the model on your own task.

Does today's AI think? Not in the human sense. Current models predict the most likely continuation of a text, and they do it so well that it looks like understanding. Making sense of that difference is one of the goals of this category.