moinsen / Software historyDE ↗
← Back to the journey

Topic

Sprachmodelle und die Frage nach offener KI

In 2017 a research team introduced the Transformer architecture for machine translation tasks. In 2018 BERT used Transformers for language understanding and incorporated context from both directions. These were distinct research projects and tasks.

With GPT-3, researchers studied in 2020, among other things, how a language model handles tasks with a few examples in its input. ChatGPT followed in 2022 as a dialogue-oriented research preview. The 2023 report on GPT-4 described text and images as possible inputs but also unreliable results. Research reports and actual access to individual product capabilities must be considered separately.

Also in 2023, Meta released Llama 2 with model weights available under its own license. The Open Source Initiative disputed Meta’s open-source label, citing licensing restrictions. This example shows that being able to use a model does not by itself settle questions about data, redistribution or permitted uses.