Smart Facts
Sober facts about AI, language models, and their true capabilities
Anyone who reads about "Artificial Intelligence" on a daily basis is quickly blinded by the promises of tech giants. Here we sort out reality. We explain the fundamental mechanisms of LLMs (Large Language Models), why AIs hallucinate, what the actual state of data privacy is, and where the journey with autonomous AI agents and machine-to-machine networks (A2A) is heading.
What exactly are LLMs?
1 Not Magic, but Statistics
ChatGPT, Claude, and similar tools are based on so-called Large Language Models (LLMs). These models do not "think". They are trained to analyze massive amounts of text and predict the statistically most likely next word (token) in a sentence.
2 Black-Box Problem
Even the developers of these models often no longer know exactly why a model gives a certain answer after training. The billions of connections (parameters) disappear into a so-called black box, making error analysis extremely difficult.
3 Conversational Illusion
Because models are so good at mimicking human language, we quickly anthropomorphize them. However, an LLM has no consciousness, no feelings, and no real "understanding" of the texts it produces.
Excursus: AI Hallucinations
Absolute Truth
The Belief: Many users believe that AIs work like encyclopedias or search engines and always output correct facts. If the AI phrases something confidently, it is often accepted unconditionally as a fact.
Reality: Convincing Hallucinations
The Reality: LLMs generate texts that sound plausible. If they don't "know" an answer (because the probabilities in the network are ambiguous), they invent facts, links, or court rulings in such a convincing tone that it's easily believed. Never blindly trust an AI.
Excursus: Privacy and the AI Act
Sensitive Data
Caution advised: Never input sensitive, personal, or confidential company data unfiltered into public AIs like the free version of ChatGPT. In most cases, this data is used to train future models. Your data could theoretically end up in answers for other users.
The EU AI Act
European Protection: With the AI Act, Europe aims to set a global standard for safe AI. High-risk AI systems are strictly regulated, and greater transparency is demanded (e.g., labeling requirements for deepfakes). Companies must specify exactly where their training data originates from.
Evolution: Agentic AI
We are on the verge of the next major paradigm shift: language models are evolving from purely reactive text generators (chatbots) into autonomous agents. These agents can independently devise plans, use external tools, and course-correct on their own when encountering errors. Two key protocols are driving this development:
MCP (Model Context Protocol)
MCP was introduced by Anthropic as an open standard and is often referred to as the "USB-C for AI". It defines a universal interface: An AI agent asks an MCP server "What tools do you have?" – and can use them immediately, whether it's Slack, GitHub, or a local database. One plug for everything.
A2A (Agent-to-Agent)
While MCP connects agents with tools, A2A (by Google) solves communication between agents. An orchestrator breaks down your request into sub-problems and delegates them to specialized sub-agents, which negotiate among themselves and iteratively produce the best result.
Swarm Intelligence & Agent Networks
The idea that many simple units collectively generate intelligence is not new. Already in 1986, the MIT pioneer Marvin Minsky described human thought in his work "The Society of Mind" as an interplay of countless simple, individually unintelligent "agents". None of these agents understands the big picture – yet from their cooperation, competition, and specialization emerges what we call "intelligence".
This is exactly the principle we are now witnessing in AI: Highly specialized agents – one plans, one codes, one tests – form a decentralized network that solves tasks no single model could overcome alone. The AI turns from a sole worker into a swarm. Minsky was right – it just took almost 40 years for technology to catch up with his vision.