These days, you can’t read a newspaper or newsletter without being flooded with articles about AI and how big and important these developments are. Huge investments, geopolitical interests, the impact on our lives and on the planet: sometimes it seems as if everything revolves around AI these days. That’s why, as part of Digital Independence Day, we’re turning our sights this Sunday to the world of AI and language models in the workplace.
Because in our daily work, too, AI is increasingly being pushed on us. If your organisation uses Microsoft 365, you can now ‘oh-so-easily’ use Copilot, Microsoft’s AI assistant, in almost all its products. And if you haven’t tried it yourself yet, you probably have colleagues who eagerly use ChatGPT or Claude. To write a summary or a translation, for example, or perhaps a first draft of a presentation? And do we still need that expensive illustrator if we can just ‘generate’ an image? (Editor’s note: the illustrations for this series are made by a real human being!)
A veritable battle is currently raging between the big American technology companies, such as OpenAI, Anthropic, Microsoft, xAI and Meta. And in China, too, companies like Alibaba and Baidu are investing on a massive scale. These are the companies at the forefront of developing large language models, with the enormous computing power needed to develop and run them (the so-called frontier labs). But these are also the companies that will be looking to recoup their investments. By convincing people and organisations that they can’t do without them, they make people dependent on their products and services. That’s how Big Tech tries to embed itself in every sector and extract value.
So it’s time to explore the alternatives: how can you use AI without big tech?
And yes, AI is an umbrella term that actually covers a great deal. From machine learning to image recognition, and from semantic search engines to image generators. That’s why the overview below is limited to services you can use for everyday office tasks.
Vanish
The Dutch search engine Startpage.com has, just like its American competitor DuckDuckGo, launched a privacy-friendly AI tool: Vanish.
With Vanish, you can anonymously use a number of AI models from the big players. Chats aren’t stored on servers, but are kept in the app. And user data should not be sent along to the underlying servers.
So yes, it’s a Dutch product, and yes, it’s better for your privacy, but of course you still remain dependent on what the big tech companies have to offer. And that’s exactly what we don’t want…
Vibe
Fancy chatting with an AI that isn’t of American or Chinese origin? Then Vibe (formerly Le Chat), from the French company Mistral, is a serious European alternative.
Mistral has developed its own language models and focuses mainly on the business market. You can start for free, but if you want to use it more, you can also take out a paid subscription for yourself, your team or your department.
Chatting with Vibe is relatively safe. Mistral uses an opt-in system, which means you have to explicitly consent to your chat history being used for training purposes. They also don’t store data for advertising purposes or for data collection by third parties. As a genuinely European company, they comply with the GDPR, and as a user you can easily have your data deleted.
Vibe is very versatile. Besides chatting, you can connect the service to various tools, such as your email, chat or your own databases, and they also offer a coding assistant that lets you develop software together with the chatbot.
Lumo
We mainly know the Swiss company Proton for its privacy-friendly email and VPN services. But Proton recently launched its own chat service too: Lumo. Lumo presents itself as a privacy-friendly AI chat service.
Lumo is a paid service. You can try it for free, but with limited features.
Unlike Mistral, Proton hasn’t developed its own models. Lumo uses so-called open weights models. These are models that AI research labs make available online, and which users can often fine-tune a little further. For example, Lumo uses an open model from Mistral, mentioned above, but also the Qwen model developed by Chinese company Alibaba.
Because Proton can run these models on its own servers, it can guarantee that its users’ data doesn’t leak to external parties. And chat history is stored encrypted at Lumo, so that even the company’s own employees can’t see your conversations with the bot.
Lumo can search the internet on its own and do a bit of coding (although there’s no dedicated environment for that), but it’s especially good at working with text. You can easily use it to summarise long texts, restructure or correct your own writing, or brainstorm. Lumo works well as an all-round writing or search assistant.
Open, open and more open
The first value in the PublicSpaces manifesto is ‘Open’, so it’s worth taking a moment to consider what open means in the context of language models and AI.
When it comes to language models, there are really three different kinds of open: open weights, open source and being an open book.
First, there are so-called open weights models. These are models developed by AI labs that can often be downloaded for free from websites such as huggingface.co. These models contain the end result of all the training they’ve gone through. Developers can often fine-tune such a model a little further to make it better at a specific task, but the foundation of the neural network has already been laid.
Open weights models may often be free and self-hostable, but you shouldn’t confuse them with open source models. With genuinely open source models, not only the result of the training is public, but also how the training was done and the software used to do it. In Europe, there are several collaborative projects that work in an open source way, such as the OpenEuroLLM consortium and the Swiss Apertus project. EuroLLM and MiniLingua are European language models that aim to lower language barriers in Europe.
Openness also plays an important role in model development when you look at where the data used to train the models comes from. Not every AI lab is an open book about this, or transparent about the origin of its training data. They often say their training data comes from the ‘open web’, but take no responsibility for the fact that the open web also contains a great deal of illegal, harmful or copyrighted material that gets absorbed into the models without a second thought. Mistral (which presents itself as fairly open), for example, has been accused of using a large collection of illegal ebooks to train its open weights models.
The autonomous options
Want to stay completely independent, without running the risk of a language model provider reading along, shutting down the service at some point or suddenly raising its prices? Then you can also host open weights models yourself.
Ollama
If you have a powerful computer or a recent Apple laptop with an M-series processor, you can use Ollama to help you. Ollama is an open source tool that makes it fairly easy to download a large number of open weights models and talk to them through a simple chat interface.
Models come in all shapes and sizes. Depending on how powerful your computer is and how much memory the graphics card (GPU) can use, you can try out different models. Because the models can be quite large and need a fair amount of computing power to query, you shouldn’t expect too much from Ollama on your desktop or laptop. It may run locally, making you completely independent of the cloud, but you’ll often be waiting a long time, or the results will be mediocre.
It might work for a simple summary or classification task, but definitely don’t use a local model like this as a source of answers. Through Ollama, local models can’t simply consult external sources, which makes their hallucination rate very high.
Murmure
Specifically for converting speech to text, you can also use Murmure on your own computer, an open source tool that runs on a model from Nvidia. You can download it for free and use it offline on your own computer. For example, if you want to make a transcript of an interview or meeting, or if you’d simply rather talk to your computer than type on your keyboard.
But be careful: always read it back and don’t blindly rely on it.
There are also European models from Mistral that can convert speech to text. You can download them, but there isn’t a handy tool available yet to use them with directly.
GPT-NL
In the Netherlands, too, there’s an initiative to develop a homegrown language model: https://gpt-nl.nl/.
The GPT-NL model is mainly trained on Dutch data and is intended as a ‘responsible’ model of our own. Responsible here means looking not only at what the model can do, but also at how it was made and how to ensure that all stakeholders can benefit from such an initiative. For example, revenues will flow back to the rights holders of the data used for training, and the methods used will be accounted for.
To make sure revenues can reach the rights holders, it was decided not to make the model freely available as open weights. So you can’t download it with Ollama and get started with it today. GPT-NL is, however, very transparent (the open book) about the data used (much of which is also accessible to other developers) and about how the system works, by releasing the software under open source licences. They’re also transparent about the energy used to train the model.
Want to get started with GPT-NL? You can currently still sign up as a launching customer if you want to start a pilot or project with your organisation.
To wrap up
There are many pitfalls to using AI in work processes, and at PublicSpaces we’re very aware of them. The global impact of the AI industry (energy use, data centres, data and capital extraction, to name but a few) is huge, and the social consequences (employment, learning outcomes) are still hard to grasp. The dependence on just a handful of large companies is also worrying. No wonder, then, that more and more organisations and companies are looking for alternatives.
At the moment, there’s no alternative yet that fully meets the PublicSpaces manifesto. So a good alternative to using big tech language models remains… not using AI! Ask a colleague, write that summary yourself.
So how can we design and govern AI entirely in the public interest? That’s what PublicSpaces will be exploring in the coming period through our ‘Public AI’ programme. Keep an eye on our website and newsletter for updates.
Still in doubt after reading this about whether or not to use a chatbot? Ask a human! <– Listening tip (in Dutch)!
