Die Anwendung von großen Sprachmodellen

Text generation, formally referred to as natural language generation, aims to produce reasonable and readable text in human language based on input data, e.g. a sequence and keywords. It’s also the foundation of the question and answering system (i.e. response generation, answer generation).

The automated generation and completion of text on sentence or paragraph level is an important Natural Language Processing (NLP) task and can be widely used in many industries with shared use cases for generating articles, news, marketing slogans, social posts, protocol, and all kinds of product descriptions. This list can be extended for many more text types, therefore, text generation is an important task across industries and use cases. We recommend you to get an impression of the state of the art technology by trying it out yourself, the Generation API , it is one of the most powerful and popular software offerings of OpenAI , based on GPT-3. It can generate all kinds of content, the popular example which you can test out below is to generate new ideas , business plans, marketing slogans or even support your interview questions based on your prompt (see explanation in workbook 1). Below is an example to get you motivated.

Taking text generation as a core NLP task, many startups and software have been initialized and generate a lot of usages. For example, CopyAI is well-known as a GPT-3 powered tool to automate creativity tools and generate marketing copies within a short period of time efficiently. Businesses can use this AI model for digital ad copy, social media content, website copy, eCommerce copy, blog content, and sales copy. And Snazzy AI is known for providing the simplest way to create content with GPT-3. Multiple services include creating landing pages, copywriting, Google Ads, and many more with just three clicks.

Next to text, the ability to generate artificial images based on pre-selected image examples or textual description is another important key application of AI models in general, and of latest large language models specifically. It’s also known as image synthesis.

Various different AI model architectures (i.e. deep learning methods) have been published in recent years, and are also known as generative AI models. Recently, LLM and diffusion models have arisen with a significant improvement of the ability to generate up to photo-realistic images with a broad range of content, style and variation based on a solely text-based prompt. In 2022, LLM models such as the DALL.E 2 of OpenAI, Stable Diffusion of stability.ai or Luminous by aleph alpha became very popular for this task.

DALL.E 2 is a 12-billion parameter version of GPT-3 trained to generate images from text descriptions, using a dataset of text–image pairs. OpenAI has released a first text to image model called DALL.E in 2021 and it has got a lot of attention in 2022 with its 2nd generation. The performance and robustness of DALL.E 2 also made it possible for a wide range of use. Microsoft is bringing  DALL.E 2 not only into its cloud AI services, but also directly integrating it to its software products and ecosystems, such as Teams APP Platform, new Designer platform.

You can also have a free test here , which is an open source playground to generate images from any text prompt using DALL-E Mini and based on OpenAI’s DALL-E. And luckily in this area, there’s also something happening closer to us in Germany, which the main research work has been done in LMU, Munich, e.g. the announcement of the stable diffusion model  from stability.ai . Stable Diffusion is a text-to-image model similar to DELL.E that will empower billions of people to create stunning art within seconds.

Different then DALL.E and Diffusion Models, Luminous from aleph-alpha focusing in a special and interesting direction. Luminous has a basic and extended model, both are Multilingual models trained on English, German, French, Spanish and Italian. Luminous-base can read, explain and generate text from existing images, such as c ontext question and answering based on given images; or to generate creative artwork titles based on image s ; or analyze and summarize based on image comparison . You can test them out in the aleph-alpha playground .

With this significant progress of LLM models for image generation in 2022, we observe a shift in the way we consider AI-based content generation at all. Coming from text generation only, various AI models now focus multi-modal generation of video, audio or even music. As a result, large language models have founded a new hype of Artificial General Intelligence (AGI) in research, tech companies as well as industry around the globe.

Code generation could mean translating natural language to code, or translating code between different programming languages. Since the LLM became capable of generating code, the Low code / No code tools has always been one of the hot topics in AI use cases and among all the developers. One of the major questions everybody has in mind is, how big could the market of AI software engineering grow, if it will really be able to replace some of the software developer roles?

Well the question shouldn’t be answered by us, we would like to encourage you to take a look at the various examples of CodeX which you can use your human language (see an example below and read more  ) or if you are developers, definitely try out first code generation product from Microsoft Copilot , then you will have your own answer 🙂

Also some start up companies showed the possibilities. One example is Debuild.co is a top GPT-3 driven app for developing website apps rapidly and effectively. With just one click, SQL code can be generated and an interface can be put together visually. It makes use of this AI model to develop an autonomous software writing system that is on par with the best developers.

Machine translation has a very long history starting from 1954 with the first rule-based solution of IBM and is a fundamental task of NLP and AI in general. A major milestone in the history of machine translation was the introduction of Deep Learning starting in 2016. It was a paradigm shift by outperforming statistical machine translation ( SMT ) which was the common approach at this time. Today, machine translation solutions are only using deep learning and, therefore, we call it neural machine translation ( NMT ) .

One of the companies that took the chance is German MT specialist DeepL . Its product is currently even becoming one of the most popular websites globally with more than a billion of people using their services. Because MT is very easy and necessary to integrate into different end customers‘ products, specially for international markets.

Nowadays, we observe more specific large language models that are also being integrated into NMT solutions. One important reason for this is to improve the translation quality of so-called low resource languages by the use of LLMs, mostly even multi-lingual LLMs. Low resource languages can be understood as less studied, resource scarce, less computerized, less privileged, less commonly taught, or low density, among other denominations languages, in the MT topic specially for NMT, those human languages with low amounts of digitally available corpora data (translation pairs) to train NMT models, are all low resource languages. For example, Korean, Vietnamese and many other languages in Asia, Africa. In Europe, except English, German, French, Spanish, most other languages are also from NMT perspective as low resource languages.

Various technology companies do heavily invest in research on different LLM architectures and methods optimized for machine translation.

For instance Amazon Web Services (AWS) has recently released the Bloom of Alexa Teacher Model  which showed state-of-the-art performance on machine translation focusing on in-context learning. These results are important for AWS as it will improve key offerings such as its own machine translation product AWS Translate and its virtual assistant Alexa.

And a traditional research company like OpenAI provides its MT services  with its Chinchilla model which focuses more on optimal model size and number of tokens under a given compute budget. Open source projects such as BLOOM  are trained to continue text from a prompt on vast amounts of text data using industrial-scale computational resources. As such, it is able to output coherent text in 46 languages and 13 programming languages that is hardly distinguishable from text written by humans. It is not only powerful but also more focusing on collaborative and supporting future AI applications such as chatbots for the open source communities and try to bring more human value such as avoid discrimination and bias.

Text summarization is aiming to convert a large body of text to a few sentences without losing the key themes of the text. Very often we need to grasp the context of text and rephrase it in different ways. This can be used widely from news and

summarization, reports and briefings preparation, even

. In digital times, information is exploding every second, people always want to get the most important information in even a few seconds. Therefore text summarization has drawn attention to researchers as well as tech companies for a long time with many

And this is a typical task where it is hard to agree on a best method, since the performance highly depends on the user scenarios. meeting  notes generate mindmap from text documents different approaches. 

Here our Open GPT-X partner aleph-alpha also provided their Luminous-extended Model, which is a Multilingual model trained on English, German, French, Spanish and Italian for

of text and document summarization. If you know any of the languages above, it would be fun to test it out! testing 

Speech recognition, also known as automatic speech recognition (ASR) or speech-to-text, is a capability which  enables a program to process human speech into a written format . It’s commonly confused with voice recognition, the simplest explanation of the  differences between speech and voice recognition is, speech recognition translates anyone’s voice; voice recognition understands a specific user’s voice. Speech recognition enabled all hands-free user experience, Siri, Alexa, Google Assistant. All kinds of voice assistants are already a must-have for smart devices from mobile phones to smart speakers and even in cars and robots. Except for complicated applications such as voice assistants, speech recognition is also widely used in use cases such as call centers or dictation related applications, e.g. in healthcare supporting doctors and manufacturings. But normally, models behind those famous products are block boxes for the public, Alexa or Google; they both provide their voice assistant as a development platform, to encourage worldwide developers to help build 3rd party functions into their voice assistant, such as  Alexa skills and  Google Actions , which your product also becomes part of their big ecosystem. However, in the past, not only as users but also for researchers, ASR systems have not really achieved human level performance. Last month openAI released their result to try to overcome this challenge, a new ASR model called Whisper  that shows human levels of accuracy and robustness outperforming both supervised and unsupervised models in the space, you can now give a test  here .