NLP – Natural Language Processing – refers to the field of AI that deals with understanding and processing human language. This includes everything from speech recognition to text classification and translation.

Every time you talk to a voice assistant, receive an automatic summary, or run a sentiment analysis of your customer reviews, NLP is at work. LLMs are currently the most powerful form of NLP.

A Small Language Model is a more compact language model with fewer parameters – typically under 10 billion. They are faster, more cost-effective to operate, and can often run on standard hardware.

Examples: Phi-3 (Microsoft), Gemma (Google), Llama 3 8B (Meta). For many business applications, SLMs are perfectly sufficient – especially when the task is clearly defined (e.g., classification, extraction, simple summaries). Not every problem requires the biggest hammer.

AGI refers to a hypothetical AI that masters any intellectual task at or above human level – not just writing texts, but also abstract thinking, learning, planning, and adapting to completely new situations.

Current AI systems are far from this. They are “narrow” – excellent at what they were trained for, but incapable of demonstrating true understanding. Whether and when AGI will be achieved is hotly debated in the research community. Estimates range from “in 5 years” to “never.”

GPU (Graphics Processing Unit): Originally developed for graphics and gaming, GPUs are now the most important hardware for AI training and operation. Their strength: they can execute thousands of calculations simultaneously. The dominant manufacturers are NVIDIA (market leader) and AMD.

TPU (Tensor Processing Unit): Specialized chips developed by Google, optimized exclusively for AI computations. They are deployed in Google Cloud.

The shortage of powerful GPUs is one of the biggest bottlenecks in the AI industry – and a reason why cloud AI services often make more sense for SMEs than their own hardware.

RLHF stands for “Reinforcement Learning from Human Feedback” – a training method in which human evaluators assess a model’s outputs and the model is refined based on this feedback.

RLHF is the reason why ChatGPT feels so “natural.” Without RLHF, language models would produce correct but often unusable, confusing, or inappropriate responses. It is the step that transforms a raw model into a useful assistant.

Inference is the process by which an already trained model processes an input and generates an output. When you ask ChatGPT a question and receive an answer, that is inference.

This is an important distinction: training happens once (and costs millions). Inference happens with every single use. The ongoing costs for AI applications are primarily determined by inference – which is why more efficient models are directly cheaper to operate.

A benchmark is a standardized test used to compare the performance of different AI models. Similar to crash tests for cars, there are different benchmarks for different capabilities.

Well-known benchmarks include: MMLU (general knowledge), HumanEval (programming), GSM8K (mathematics), Arena Elo (human evaluation in direct comparison). Benchmarks are useful as guidance, but should be taken with caution – a model that performs well in tests is not automatically the best for your specific task.

The knowledge cutoff is the date up to which a model absorbed information during its training. Everything that happened after this date is not “known” to the model – unless it is connected to a web search or current data sources.

Example: A model with a cutoff of March 2024 does not know about events from April 2024 onward. This is why it is important for businesses to connect AI systems with current data sources (via RAG or web search) when up-to-date information is relevant.

Tokenization is the process by which text is broken down into tokens before being processed by the model. Different models use different tokenizers – which is why the same text can result in a different number of tokens depending on the model.

Interesting detail: Most tokenizers are optimized for English. German texts therefore often consume more tokens than English texts of the same length. This is relevant for cost calculation – and a reason why multilingual optimization is important in AI projects.