Dusseldorf At the German start-up Aleph Alpha, artificial intelligence (AI) should now have to explain itself: Like the US company OpenAI, the Heidelberg-based company has developed a language model that can create text and answer questions. In the future, users of “Luminous” should be able to understand exactly where the AI gets its information from – and whether it might be completely wrong.
Founder Jonas Andrulis and his 50-strong team are addressing a major problem: language models such as the recently famous GPT-3 and its successor GPT-4 have made progress. For example, they can summarize texts, write business letters and prepare presentations. But the models still produce hair-raising errors again and again.
Aleph Alpha is now hoping that the new functionality will help it differentiate itself from the competition. “This transparency will enable the use of generative AI for critical tasks in the legal, healthcare and banking sectors,” says Jonas Andrulis, who has relied on business customers from the start.
As the Handelsblatt learned exclusively, Aleph Alpha will activate an “explanation function” for all of the company’s customers on Thursday. This should enable them to understand how reliable the information provided is based on texts and images.
In an example prepared by the company itself, Luminous is presented with a text about Aleph Alpha. If the user asks the AI where the headquarters of the start-up is and whether “Berlin” could be the right answer, it marks all passages in the text that could provide information on this question: passages that make the connection between Berlin and the Occupy headquarters, contradictory passages are highlighted in purple.
Neuroscientist Bethge: “Important step in the right direction”
In this specific case, the case for the current five-language model is clear: Aleph Alpha’s headquarters are in Heidelberg, as can be seen from the sentence “The German startup, located in the picturesque German city Heidelberg”. Since Aleph Alpha relied on multimodality right from the start, Luminous can also describe images – and in future show which image sections it is referring to.
Handelsblatt has already been able to test the new function itself. Interim conclusion: It doesn’t always work as well as in the showcase example. Sometimes the user is more reminded of the Greek philosopher Socrates, who questioned what he thought he knew.
If the AI is completely wrong with its text references, at least the realization remains: “I know that the AI doesn’t know anything.” However, this can also be an advantage when you consider that language models sometimes represent absolute nonsense credibly.
“The fact that the model is self-explanatory is an important step in the right direction,” says neuroscientist Matthias Bethge from the University of Tübingen. About the current reliability of language models, he says: “Sometimes GPT also reminds me of a student who didn’t understand the content exactly, but tolerated his knowledge enough to pass the exam.”
The background is that the language models always try to give an answer – even if they only have limited information on a specific issue. For example, Luminous could claim that Aleph Alpha’s headquarters are in New York – for example because the density of companies is particularly high there. The problem: For users, this derivation has so far been difficult to understand.
Difficult race with OpenAI and Co.
The Darmstadt computer science professor Kristian Kersting, who researches together with Aleph Alpha and has invested in the company himself, is also critical of this. “Mathematically, I can tell you exactly what the models are doing and how they’re learning, at a level of abstraction,” he says. But the calculations are so complex that they can no longer be understood. “We know exactly what a model is doing, but we don’t know why it has now specifically recognized that an image shows a cat.”
With a joint scientific work, which is the basis for the new functionality in Luminous, the question of this traceability would now be advanced. Basic research and marketable innovations are sometimes close together in artificial intelligence than in other scientific fields.
To what extent this progress also represents a competitive advantage for Aleph Alpha cannot be answered unequivocally. It is clear that no competitor can currently offer such a powerful language model as OpenAI with GPT-4, in which Microsoft is investing tens of billions. That’s why it’s also about focus.
>> Now also read: How ChatGPT and Sam Altman will impact your life
The experts agree that OpenAI could – at least theoretically – quickly catch up when it comes to the explainability of their model. “If OpenAI wanted to introduce source references, it certainly could,” says Matthias Bethge. The question is, however, whether competitors want to set the same focus and invest money and time in explainability. He therefore confirms Aleph Alpha: “It is a smart strategy to position yourself as a German start-up that can be explained.” He also sees opportunities for Aleph Alpha in the fact that the company has understood how sensitive data processing issues are for corporate customers.
Jonas Andrulis does not talk around it when it comes to the powerful and financially strong competition. “Everyone I know who is involved in AI is under a lot of pressure right now,” he says. The former Apple manager is also well connected in the US tech scene.
But he also draws something positive from the competition with Microsoft and OpenAI, for example when looking for new employees who would come to Heidelberg from all over the world. “We’re taking on the world’s largest monopolist, who is certainly also known for being rather aggressive,” he says. This “mission” inspires people who “want and can do something special” to come to Aleph Alpha.
At the moment, however, the main thing is to convince investors of this mission. For the further development of Luminous, Andrulis needs money. So far, 28 million euros have flowed into the company – a fraction of the funds at OpenAI. The aim of an upcoming financing round should probably be a three-digit million amount.
More: Tech investor Peter Thiel: “Google is completely in panic mode”