So far, we have mainly been delighted with AI software from the USA. DeepSeek-V2 Alpha, a public AI chat that responds to normal language input like ChatGPT, was recently launched. According to our tests, the time horizon extends to the year 2023, so it was trained with training material up to the year 2023. It was reluctant to respond to certain questions in our tests, such as questions about the Tianmen massacre. Other questions relating to global politics were also answered from a clearly Chinese perspective during our initial tests.
You can ask the AI chat DeepSeek-V2 a lot of questions, but the Chinese AI answers some of them cautiously or from a Chinese perspective.
The DeepSeek AI chat is in the alpha stage and is probably the most accessible part of the DeepSeek VL universe for end users. This is a series of multimodal large-scale language models developed by DeepSeek-AI from the Chinese metropolis of Huangzhou. The company was only founded in 2023.
Object recognition in images
The models can process both visual and linguistic information in screenshots, PDFs, scans, graphics and texts. DeepSeek-VL has already shown interesting performance in visually centered scenarios in some demo cases. For example, the AI was able to recognize from a complicated image whether a certain object, named in text form, was to the left or right of a barely recognizable bicycle.
DeepSeek-VL can answer questions about an uploaded photo and recognizes, for example, where objects are located relative to others.
However, DeepSeek Chat can only recognize text. References to other websites, for example for translating and summarizing certain passages from them, are currently still too much for the AI.
The DeepSeek terms of use reveal that the substructure is based on Transformer and that DeepSeek was trained with PyTorch. In comparison to ChatGPT-4, DeepSeek-V2 performs slightly worse according to our initial impression. However, the DeepSeek models are open source and currently available on GitHub and HuggingFace. DeepSeek-AI plans to further scale DeepSeek-VL and integrate MoE architectures (MoE = Mixture of Experts). DeepSeek offers SDKs and APIs for programmers. The latter is OpenAI-compatible.
(mil)
Don’t miss any news – follow us on
Facebook,
LinkedIn or
Mastodon.
This article was originally published in
German.
It was translated with technical assistance and editorially reviewed before publication.
Dieser Link ist leider nicht mehr gültig.
Links zu verschenkten Artikeln werden ungültig,
wenn diese älter als 7 Tage sind oder zu oft aufgerufen wurden.
Sie benötigen ein heise+ Paket, um diesen Artikel zu lesen. Jetzt eine Woche unverbindlich testen – ohne Verpflichtung!
