AI in 10 minutes

Artificial intelligence is a set of techniques that let a program perform tasks we associate with human intelligence: understanding text, recognising images, planning, solving problems. It is not a single machine, and it is not a person. It is a family of very different systems, from a spam filter to an assistant you can chat with. Today the word most often means systems that learned from a large mass of examples, rather than systems where every rule was written by hand. These systems are useful and sometimes impressive. They also make mistakes, and they do not understand the world the way we do. Throughout this path, we keep one simple question in mind: what has been shown, and by whom?

In machine learning, you do not tell the program how to recognise a cat. You show it many photos of cats and other animals, with the right answer, and it adjusts its internal settings to be wrong less often. These settings are numbers, called parameters, stored in a neural network: a stack of small computing units linked together. This phase is called training. It takes a lot of data and computing power. Once training is done, the model is frozen. When it is used to answer a request, that is called inference. One point to keep: a model partly reflects what it saw during training, including the biases and mistakes in the data.

A language model is a neural network trained on very large amounts of text to predict what comes next. From this simple task come useful abilities: answering a question, summarising, translating, writing code. Most rely on an architecture called the transformer. One large model can serve as a base for many uses: this is called a foundation model. Some also handle images or sound: they are multimodal. Some models publish their parameters: they are said to have open weights, which lets others examine them and run them. It is not always open source, because the licence and the data can stay closed.

Models are increasingly asked to chain steps: this is called reasoning. Some first produce intermediate steps, then an answer. An agent goes further: it uses tools (search, code, files), looks at the result of its actions and adapts to reach a goal. The more an agent acts alone, the more human oversight matters. Finally, in robotics, so-called VLA models link what a robot sees and an instruction in plain language to a movement to carry out. Systems of this kind appear in the DiscoverIA registry, for example agent platforms or autonomous research systems. Each entry says what is established and what is not.

A language model can produce an answer that is well written and confident, but wrong: a source that does not exist, a proof with a gap. This is called a hallucination. The tone says nothing about accuracy. This is why systems are evaluated on benchmarks, standardised tests that allow comparison. These tests have limits: a good score does not guarantee success elsewhere. In mathematics, some AI claims turned out to be invalid or incomplete after review, and the registry keeps a record of them. An announcement, even from a large lab, remains an announcement until others have checked it.

In science, a result goes through several filters. First, a claim: someone says they have achieved something. It is often posted as a preprint, before any review. Next comes peer review: experts independent of the authors examine the paper, and a journal then publishes it. Another check is reproduction: someone else redoes the computation and sees whether the same thing comes out. In mathematics there is an even stricter check: writing the proof in a proof assistant such as Lean, which validates each step. Even then, it remains to check that the written statement says what the source claims. This is statement fidelity, and DiscoverIA assesses it separately.

Each registry entry carries a status, sourced and dated. Formally verified: a named third party has checked the proof by machine. Peer reviewed: published in a journal or reviewed by named experts. Under review: a public review is under way, with no verdict. Disputed: a sourced public objection exists, with no established refutation. Withdrawn: abandoned by its author or classified incorrect by the reference source. Not evaluated: no third-party evaluation has been found. This last status does not mean “false”. It means “not yet examined, as far as we know”. Many recent results are in this case, simply because the time needed for review is longer than the time needed to announce.

Each source in an entry also says how it was read. PAGE: the page was read and the quotation taken word for word. PAGE-WF: the page was read through a model summary, the quotation is not re-checked. Search-engine snippet: the information was seen only in a search engine's snippet, the weakest level. An entry reads on three levels: “In one sentence”, “To understand” (an analogy, what changes, what remains unclear), then “For experts” (method, limits, sources). Every entry says what is not demonstrated. If you spot a mistake, you can report it: the correction stays visible on the entry, with its date.