Control AI: Aleph Alpha Introduces New Generation of Explainable AI Models

Explainability and trustworthiness are weak points of large language models. Aleph Alpha's AI system is supposed to remedy this with a new control version.

In Pocket speichern vorlesen Druckansicht
Server,Racks,In,Server,Room,Data,Center.,3d,Render

(Bild: Sashkin/Shutterstock.com)

Lesezeit: 7 Min.
Von
  • Silke Hahn
Inhaltsverzeichnis


(Diesen Artikel gibt es auch auf Deutsch.)

The Heidelberg-based AI start-up Aleph Alpha has introduced a new generation of its AI models. The previous Luminous LLM family, which consists of three Foundation Models of different sizes, now includes a Control version. Technically, according to the provider, these are Instruct models that have been readjusted in a process similar to OpenAI and its InstructGPT for processing natural language and solving complex computational linguistic tasks, in the case of Aleph Alpha in an automated process without human feedback.

According to the announcement, what is new in the now presented model generation is the explainability and reproducibility of the output. According to company founder and CEO Jonas Andrulis, Aleph Alpha with its control models is "the only option on the market that solves tasks with explainability, reproducibility and trust". To this end, Aleph Alpha had published research contributions together with academic partners in Spring 2023, such as Attention Manipulation (AtMan) in cooperation with the German Research Center for Artificial Intelligence (DFKI) and TU Darmstadt in January and a new method for containing the distribution bias in large language models (QDAIF: Quality through Diversity in AI Feedback) in cooperation with Stability AI and Carper AI at the end of May.

According to the Heidelberg researchers, the research findings gained will be successively incorporated into the models and should make their output comprehensible. In April, for example, Aleph Alpha added a fact check to its models with AtMan Explain, which is intended to make the results comprehensible with source information. The origin of factual outputs can be traced back to sources in texts and images, up to counterfactual, i.e. contradictory observations in data.

The implemented functions are steps towards the upcoming regulatory requirements: According to the current state of the draft, the upcoming AI Act will impose clear requirements on the explainability and traceability of the claims of generative AI. Apparently, Aleph Alpha, as a provider of generative and general purpose AI, is having a close look at the upcoming regulation and is aligning its AI systems accordingly.

The previous three model versions have been fine-tuned to execute commands and instructions for the controlled model generation, as press spokesman Tim-André Thomas explains when asked. For this purpose, the research department of Aleph Alpha collected new data sets and aligned the output of the model more closely to the input in a supervised fine-tuning process, thus bringing the results closer to the input prompts and instructions.

The new model generation is not intended to replace the existing base models, but to complement them. Those who already work with the previous Foundation Models via API or Playground can continue to do so. The new models are supposed to be easier to use because they can be addressed in Zero-Shot Prompting, i.e. they require less context to execute tasks. Previously, it was apparently necessary to show the model more cues or examples in Few-Shot Prompting to get the desired results, such as a summary, answers to questions or a picture description in combined multimodal prompting.

The control models are supposed to deliver the desired result with a single instruction, as users are now familiar with and used to from systems like ChatGPT, and are supposed to be able to handle five languages as before (in addition to German and English, the basic models have also been trained in French, Italian and Spanish). According to the vendor, the new models enable a more streamlined development cycle of AI chatbots and digital assistants, the creation of which is arguably part of Aleph Alpha's core offering to its clients and partners.

According to Andrulis, his technology is particularly targeted at companies in the legal, health and banking sectors, as well as in public administration – Aleph Alpha specializes in corporate clients and public institutions, working for example with the Cyberhub of the German Armed Forces (Bundeswehr), a Federal Ministry, the city of Heidelberg and the State of Baden-Württemberg in the German Southwest. The Innovation Atlas of the German regions just named this state "the most innovative" among the 16 German states, according to its Finance Minister Danyal Bayaz. Andrulis mentioned pilot customer projects in which evaluations took place before the market launch. According to him, the control models can provide more accurate answers than before. However, benchmarks are not available. The pilot customers were not named upon request "for reasons of confidentiality".

AI researcher Jonas Andrulis in his own data centre: Andrulis heads Aleph Alpha, the Heidelberg-based company he founded. With his team, he received the German AI Award in 2021 – for a technology comparable to OpenAI and Google DeepMind. Previously, the German industrial engineer had led secret AI projects at Apple in California (Special Projects Group) and further developed the Siri voice system.

(Bild: © Manuel Schlüter)

The IT consulting firms ONTEC, adesso, the msg Group and the French consulting firm Sopra Steria are known partners. Apparently, a partnership with SAP is also in the offing, as could be read in WELT and Handelsblatt in the run-up to its annual conference SAPPHIRE in Orlando, Florida. SAP is said to be participating in the start-up's current financing round with a three-digit million sum. No official announcement has yet been made by SAP or Aleph Alpha – presumably because the funding round has not yet been completed.

However, Heise editors noticed the Aleph Alpha logo in SAP CEO Christian Klein's presentation at the virtual edition of SAPPHIRE. Klein had mentioned the Heidelberg-based company alongside Microsoft, Databricks and other possible AI partners in a keynote. In the press briefing, when asked, he did not mention any of the pilot partners by name, but he did mention a number of projects on AI applications that are now being worked on internally with partners.

The research report on "Quality Diversity through AI Feedback" can be read at Carper AI (long) and at Aleph Alpha (streamlined) respectively in the blog. The three new model types of the Control generation are now available to all users in the Aleph Alpha Playground. This playground can be tested as a demo with a free credit.

(sih)