It can’t be easy competing with what's become the fastest-growing internet product in history, but that hasn’t dissuaded Aleph Alpha — a German competitor of Microsoft-backed OpenAI.
While OpenAI has raised a cool $11bn, Aleph Alpha has raised only $31.1m. In terms of data, however, Aleph Alpha’s largest model is trained on 300bn parameters; OpenAI’s is trained on 175bn.
Founder and CEO Jonas Andrulis says there are advantages to not being a Microsoft or a Google — the most recent tech giant to throw its hat into the generative AI ring.
“There are a lot of strong enterprises and phenomenal people that don't want to be dependent on Microsoft or Google,” he says, adding that Europe must not try and win “the battles of the past, but the battles of the future”.
In Aleph Alpha’s view, there are battles to be won for enterprise customers, not making better search engines for consumers.
A competitive edge
The fact that Aleph Alpha isn’t building a mass-market, consumer product is part of the reason it doesn’t need the same level of capital as OpenAI. It’s estimated that ChatGPT’s huge popularity has led to running costs of around $100k a day.
The German startup has built a chatbot called Lumi on top of a large language model (LLM) called Luminous — much like ChatGPT on top of GPT-3.
Microsoft will, of course, increasingly sell GPT-powered services to businesses, but Andrulis says that Aleph Alpha’s independence from big tech means it can offer a much more bespoke and streamlined service to clients.
“We are able to deploy in any environment,“ he argues. “It’s our independence and the fact that we allow an on-premises installation, that we allow an installation in any cloud environment. We are basically the only team that can give you that.”
And, while OpenAI is pretty transparent about gathering its users’ data, Aleph Alpha is “not logging any user data” according to Andrulis.
“It's a different mindset. We are not monetising community… we’re monetising the R&D,” he adds.
Aleph Alpha is targeting what Andrulis calls “critical enterprises” — organisations like law firms, healthcare providers and banks, which rely heavily on trustable, accurate information. So not exactly the kinds of customers who are going to be turning to ChatGPT — famous for getting everything from linear equations to programming questions wrong — for answers.
Aleph Alpha, on the other hand, provides sources and citations for where an answer has come from.
Aleph says it can’t talk openly about all of its clients yet, but does say it’s working with the German government, the City of Heidelberg, several German universities and that it is “live in a large bank”.
OpenAI has already teased that sourcing and citation will be doable in GPT-4, as part of the integration with Microsoft’s Bing search engine, but Aleph Alpha still believes that it can take advantage of its proximity to large European corporates.
“Europe is in a very strong position for industry — there are strong engineering companies, there are strong trade companies,” says Andrulis. “If we succeed in helping European enterprises transform their businesses this will give all European industry a phenomenal advantage.”
As for fundraising, Andrulis says casually that the company has “some inbound conversations from investors, so [a raise] is certainly an option, but we have some runway left”.
Fine-tuning the models
Aleph Alpha isn’t the only European startup serving enterprise clients that need reliability.
Amsterdam-based Zeta Alpha is building a generative AI-powered search tool for enterprises and researchers, with clients including Deloitte, O2 and NASA. The product allows users to search across both publicly available information, as well as proprietary databases, with sources and citations.
Founder and CEO Jakub Zavrel tells Sifted that more generalised models like ChatGPT can’t compete with fine-tuned ones when it comes to the specific knowledge required by enterprises and researchers.
“GPT is good with general sort of language, but once it gets into super specialistic territory you have to fine-tune it to get good performance,” he explains.
Victor Botev, CTO of Oslo-based neural search company Iris.ai, says that fine-tuning a huge model like OpenAI’s for specialised use cases would be extremely expensive, which is why the startup built its own model. He says it costs his company “literally a couple of hundred dollars to adapt the system for a particular domain”.
Iris.ai is already working with corporate clients like novel materials firm Materiom and the multinational steel company ArcelorMittal. As well as searching large troves of academic research, the product creates easily digestible summaries with sources and citations — something that Botev says was key to winning the trust of initially sceptical researchers.
Other European companies using generative AI for specialised search include London-based Medwise, which focuses on the medical profession, Paris-based enterprise-focused tool Lighton and Berlin-based Deepset, which counts Airbus and Siemens among its clients.
For Iris.ai’s Botev, the trick for startups to be able to compete with the likes of Google and Microsoft will be their ability to move fast.
“Smaller companies can be more flexible. They can change the business model more easily,” he says. “You have to be able to twist the angle, and we can even use some of these byproducts from the big companies and make sure to adopt them for a specialised domain.”