Image from Aleph Alpha’s Medium blog
Hey Everyone,
Aleph Alpha is the most important foundational LLM startup you’ve never heard of. Situated in Germany, they are aligned with the EU’s higher standards of privacy, ethics and accountability. In terms of AI governance some startups and firms have taken more care than others.
For this reason I believe enterprise and governments clients will work with them. I asked my chief European guest contributing writer to take us for a deeper look into them.
Aleph Alpha is worth checking out. The founder and CEO is Jonas Andrulis. I’m sure I’ll be covering this AI startup again sometime soon.
📶 From our sponsor: 📶
Build a custom ChatGPT for your website
Instantly build AI chatbots from a link to your knowledge base and your files.
ChatGPT-powered bots that effectively automate up to 70% of customer and sales support for your company.
Spend 5 minutes to create a bot that saves you 5000 minutes spent working on email support.
How to Sign-up to A.I. Supremacy 📑
To get access to all of my deep dives consider becoming a paid supporter of this channel. Join 550+ others globally.
Discounts on Premium – to get Full Access
If before March, 2024:
❄️ Special Offer to Go Paid 🌨️🧊☕️⛄️
This special 2-year Anniversary offer expires after February, 2024. It is only available for yearly plans. This amounts to $53.60 a year, or $4.16 a month, or $0.26 an article.
By of Futuristic Lawyer. ~ Subscribe
In today’s post, we will take a look at Aleph Alpha, from the company’s funding, origin, and market position, to its flagship product Luminous.
Aleph Alpha’s Series B Funding
The German AI company, Aleph Alpha, became a serious competitor in the global corporate AI race, namely after it secured $500 million in Series B funding in November of last year. Among its biggest investors is the owner of the Lidl supermarket chain, Schwarz Group, who co-led the funding round with Bosch Ventures and a local AI hub in south-west Germany, Innovation Park Artificial Intelligence (Ipai).
Other investors included Berlin-based Christ&Company Consulting, Hewlett Packard Enterprise (HPE), SAP, and Burda Principal Investments. All of these investors, including the lead investors, were new. The already-oversubscribed funding round included additional participation from existing institutional investors from Aleph Alpha’s Series A funding in July 2021, where it raised $27 million.
According to data from Crunchbase, Aleph Alpha’s Series B round was the seventh-biggest deal for all AI companies in 2023. Yet, in terms of funding amounts it’s still far behind its competitors across the Atlantic. Notably, it’s behind OpenAI’s $10 billion investment from Microsoft, Anthropic’s three deals in 2023 of respectively $4 billion, $2 billion, and $450 million, and Inflection AI’s fresh funding round of $1.3 billion. Established enterprises like Google and Meta have an advantage over Aleph Alpha with their wide access to computing power, cloud resources, data, and cash.
One advantage the German AI startup holds in the international market may be that big AI companies in the US tend to work in silos. Most of their knowledge is heavily guarded behind NDAs, and model architectures are tightly held trade secrets as if they were the secret ingredient in Coca-Cola. A highly competitive and uncollaborative environment like this is bad for innovation and terrible for addressing AI safety risks.
Aleph Alpha may have an edge in this regard, since the European AI startup space seems a bit more open and collaborative. Legal frameworks like GDPR and the upcoming AI Act push companies to be transparent in the first place. Aleph Alpha is engaging in several partnerships with universities, governments, and other tech companies, fostering a collegial and open research environment that could be beneficial for addressing dangers, for Aleph Alpha’s global position in the market, and even for Europe’s AI sovereignty.
Is Aleph Alpha EU’s OpenAI?
Contrary to its US counterparts, Aleph Alpha is not focused on the consumer product side but sells generative AI as a service to enterprises and government institutions. It boasts of having more than 10.000 customers.
The company was founded in 2019 in Heidelberg, a city in the southwestern part of Germany by Jonas Andrulis, current CEO, and Samuel Weinbach, now VP of Technology. The two likely met at Deloitte where they both worked as managers a few years prior to establishing Aleph Alpha. The 41-year-old Andrulis was working as a senior AI researcher at Apple when he realized that no company in Europe was working towards developing an AGI model. That became his calling.
Aleph Alpha was hailed by Wired as Europe’s answer to OpenAI. Both companies are on a focused march towards developing AGI. Here, from “Mission” on the company’s website:
“This new generation of AI has the power to shape history’s most impactful development – the transformation of society’s very fabric. Our responsibility here is to forge a future that upholds our values and principles, where the bedrock of liberal democracy and aligned institutions ensure prosperity and fairness for all.”
Aleph Alpha’s view on the transformational, future effect of AI on society somewhat mirrors OpenAI’s mission statement to “ensure that artificial general intelligence—AI systems that are generally smarter than humans—benefits all of humanity”. But while it’s very important for OpenAI to be a leader and a first-mover in AGI, Aleph Alpha’s focus is on “sovereignty first”. “Sovereignty first” means a strict prioritization of data privacy, security, transparency, and ownership for its customers.
In a sense, Aleph Alpha’s approach is more similar to that of Anthropic, the world’s second best-funded AI startup that was founded by former OpenAI executives who grew concerned about safety, after OpenAI engaged in a commercial partnership with Microsoft. However, Aleph Alpha cannot really be compared with any company in the US. Anthropic is dedicated to understanding the long-term effects of superintelligent AI, whereas Aleph Alpha is more focused on the pragmatic, immediate, and near-term effects of AI and complying with the heavy regulatory framework of EU data laws.
Aleph Alpha’s collaborative and “sovereignty first”-approach is reflected in its partnerships with universities such as the University of Duisburg-Essen and the Technical University of Darmstadt, with the open-source AI organization EleutherAI, and with its customer base of governmental institutions.
For example, Aleph Alpha is cooperating with its local government in Baden-Württemberg to implement an AI system that can support overburdened administrative workers. The system is called F13 and is being developed in collaboration with the state’s innovation lab. Right now, it is being tested and feedback from the state officials so far has been positive. F-13 eases their workload by assisting in tasks related to research and text creation. Simultaneously, the system adheres to the state’s internal form specifications, and data protection policies, maintains full confidentiality, and the system even provides sources for all information it generates.
Luminous
Aleph Alpha’s flagship product is a family of language models called Luminous. Luminous is available in three models, the smallest Luminous-base (13 million parameters), the medium-sized Luminous extended (30 billion parameters), and the largest Luminous-supreme (70 billion parameters). The three models are trained in the five most commonly spoken European languages: English, Spanish, German, French, and Italian.
Luminous-base and Luminous-extended are both multimodal, meaning they can process both text and images . All models are available in an instruction-finetuned version, called Luminous-control. These versions can be instructed by the user to respond to prompts in certain ways, for example, “use Australian slang”, or “be polite”.
You can try out the Luminous for free via the playground here.
Compared to let’s say ChatGPT, the user interface is a bit confusing and probably requires some technical introduction to use, at least for the non-AI-aficionado. I am not very technical myself but used this guide as a starting point.
Once the basics of the interface are understood, I see at least some usefulness to the complexity. Luminous provides the user with more options than other AI tools I have tried to adjust the model’s output and experiment. Although there is a bit of a learning curve, I think that playing with advanced parameters like Top P, Tok K, frequency penalty, presence penalty, stop sequences, maximum length, and temperature, is a healthy exercise, even for the casual user, since it kind of demystifies the whole experience of interacting with a language model. And we are dealing with a language model after all, not HAL 9000.
The Luminous family of models cannot currently compete with its American competitors on raw model capabilities or performance on benchmarks. However, the Luminous family of models can do something that other foundation models cannot: explain the source of its outputs.
Samuel Weinbech and another researcher at Aleph Alpha co-authored a paper from November 2023 with researchers from Technical University Darmstadt, Hessian Center for Artificial Intelligence, German Center for Artificial Intelligence, and LAION concerning an algorithm called AtMan “that provides explanations of generative transformer models at almost no extra cost”. Aleph Alpha integrated AtMan in Luminous last year. All the company’s models now have the ability to explain where its outputs came from. Or at least to some extent.
The Luminous models can “explain” how it arrived at an output based on the input data. But it cannot refer to its underlying training data. For example: I provided Luminous Supreme with a prompt (“The best things in life are free” said the man as he gazed out onto the ocean) and the model returned with a target. Hereafter, Luminous Supreme can “explain” its target by marking those parts of the prompt with a color code that was either positively or negatively correlated with the target. It looks like this:
The feature can be used for verifying that internal knowledge in a company is reflected in the output’s model and provides some context to what information the model had emphasized in its answer. Or as one of the co-developers of AtMan, Björn Deiseroth explains to Scientific America:
“If a bank uses an algorithm to calculate a person’s creditworthiness, for example, it is possible to check which personal data led to the result: Did the AI use discriminatory characteristics such as skin color, gender, and so on?”
The feature is a step towards explainable AI, but there is still a long way to go. Aleph Alpha is dedicated to research in this area, and it aligns very well with the upcoming AI Acts’ requirements in terms of transparency.
Wrapping up
I hope this post gave you a better understanding of Aleph Alpha and the Luminous models. It certainly did for me writing it. Aleph Alpha is an exciting company to look out for in the future. Particularly, I am excited to see how the company and its approach develops in comparison with its competitors across the Atlantic.
Theoretically, Aleph Alpha could be a key part of the infrastructure for future EU sovereignty in AI. It will be hard, if not impossible for Aleph Alpha to keep up performance-wise with the big players in the US, short of a major technological breakthrough. But the emphasis on security and ethicalness is refreshing and aligns well with European values.
Read More in AI Supremacy
