TextCortex: Shaping the Future of AI Assistants

Once an obscure subfield of artificial intelligence research, machine learning is now making waves in the world of technology. From distinguishing cats from dogs to beating humans at chess and go, the breakthroughs have been impressive. 

It was only with the emergence of large language models like ChatGPT that people truly realized the potential of machine learning. But ChatGPT is not the only player in the game. Startups around the world are racing to create AI software that boosts productivity, whether it’s for writing e-mails and social media posts or summarizing academic papers. 

TextCortex, founded in 2021 by Ceyhun Derinbogaz and Dominik Lambersy, is at the forefront of this movement. They have not just relied on OpenAI’s ChatGPT but instead trained their own large language model, Zeno, running them on independent infrastructure. They aim to revolutionize not just copywriting but all aspects of personal productivity. In 2022, TextCortex raised a €1.1 million seed round led by b2venture (formerly btov Partners), with Speedinvest and Entrepreneur First joining in.

Learn more about the future of AI assistants from our interview with the co-founder, Dominik Lambersy: 

Why Did You Start TextCortex?

My journey with machine learning started as early as 2015 when a large German bank wanted to hire me to assist with their digital transformation. I rejected their offer. Still a business student, I learned what automation through data science and machine learning meant in my first bootstrapped venture. I had a social media automation business by the side and eventually joined Speedinvest to better understand machine learning startups and an overview of this nascent market. 

After a year, I longed to build a business myself again and joined the Entrepreneur First startup program in April 2021. This was where I met my now co-founder, Ceyhun. He is a technical genius who already built a prototype for natural language generation based on open-source models in 2019. We joined forces and founded TextCortex, developing his prototype into a writing assistant and now into our full-fledged AI assistant platform, Zeno, helping people become more productive. 

How Does Your AI Assistant Work?

Like other large language models, Zeno uses the transformer model architecture based on Google’s famous paper ‘Attention Is All You Need.’ It guesses the next word in a sequence of words and, having seen many sentences previously, can generate human-grade language. 

What started as a domain name generator under Ceyhun’s desk laid the foundation for a Shopify product description writing app until it evolved into a writing assistant platform that is now recognized by our users and customers as a complete virtual assistant—supporting our users horizontally in over 200 use cases everywhere they need it. Zeno is today actively deployed and used on over 4000 apps and websites. Now, we also have a chat interface, ZenoChat, where users can access a variety of models, not just Zeno but also the latest GPT-4 model by OpenAI and others. 

We’re currently working on several improvements, for example, finding the perfect moment to stop producing an answer. When asked about today’s weather, it shouldn’t give the forecast for the next ten days. We aim to include sources, keep models updated with real-time information, and make them more efficient since no one likes to wait several minutes for an answer.

Working with large language models for over two years, we knew what they were capable of, so ChatGPT wasn’t a huge surprise for us. ChatGPT nailed providing a familiar user interface as everyone is used to chat apps today and context handling, i.e., capturing what the users care about over a sequence of messages. Great communication and collaboration come from iteration. Nothing can or should work perfectly in the first iteration.  

A flurry of AI startups has now started to either build on top of GTP-4, the large language model underlying ChatGPT or train their own large language models and even open-source them. We believe that most large language models will be open-sourced in the future. For example, HuggingFace does an amazing job of bringing people together to collaborate around models and allow their finetuning for particular use cases. Such fine-tuned models will become better at solving particular tasks than any general-purpose model, no matter how large. 

In the future, there will be various models for different languages to capture cultural nuances. Today, most models are trained in English since there are plenty of materials available online. Yet, culture is expressed in every language differently. For example, German is straightforward and to the point—there is a specific word for almost everything—while French is generally more indirect, e.g., talking about ‘apples from the earth’ (‘pommes de terre’) instead of having a dedicated word for ‘potatoes.’ Read more about this here. 

What Challenges Do Machine Learning Startups Face Today? 

While large language models have already seen wide-scale adoption, quite some challenges remain to make them safe and useful. One of the biggest issues is that people now jump on the GPT hype without understanding the underlying technology. As in crypto before, people make promises they cannot fulfill and sell lemons as peaches

Also, it’s more important than ever to practice critical thinking. Just as you shouldn’t trust everything other humans tell you, you also shouldn’t trust everything an AI tells you. The world has been and will increasingly be flooded with information and misinformation, and comparing different sources of information and building your own opinion remains crucial. 

A lot of the social media hype around AutoGPTs is misguided. It’s great that they can solve simple tasks on their own, but most business problems involve solving complex tasks—and those require human-AI collaboration. Over time, People will learn to write specific prompts to guide large language models, as everyday language tends to be too imprecise. Always ask yourself: “Would a friend understand this instruction?” The better you communicate with your chat assistant, the better the work they achieve. 

For now, large language models are all about increasing human productivity. Instead of outsourcing work to low-cost labor countries, AI is making skilled workers even more productive. If it once took you 20 hours to write a 3000-word article, you can now draft the article with bullet points and let AI do its magic, costing you just a fraction of the time. However, don’t expect a ready-to-go article without any human value-add. We are going from an age of unlimited access to information to unlimited creation. The human in the loop will be the factor that evaluates which creation has real-life value. 

Take programming as an example: Programming is now more about defining logic in simple words than finding the right functions. Write logic in pseudo-code and let AI formulate the full-fledged program. 

How Did You Evaluate Your Startup Idea?

I am a huge fan of first-principle thinking. When we started out with Ceyhun’s prototype, I first tested what the model could do. I had worked in venture capital before, so I asked the model, “How can I raise money from VCs?” and checked whether the answer made sense. That way, I built more understanding of its potential, and by doing customer development, we landed on generating product descriptions as a first use case. We iterated from there and eventually developed Zeno into a writing assistant and now a full-fledged AI assistant. 

One of the most critical points is ensuring you can generate value. There is no point in raising venture capital if you haven’t thought through how to create and capture value (this goes for every stakeholder you are in contact with):

  • from customers monetizing value expecting a return on investment (RoI) from your product,
  • to employees creating this value, expecting RoI for their time and dedication,
  • to investors allowing a business to scale, expecting a return on investment from your value creation. 

Ultimately, you are building a business, not some pipe dream. If you cannot afford the Value Added Tax after paying the business costs, are you adding any value to the world? 

What Advice Would You Give Fellow Deep Tech Founders?

Since I have worked in VC before, founders often ask me whether they should raise venture capital. But the actual question is: Will this become an outlier business, and thus, do you really need venture capital? 

Venture capital used to be for scaling up sound business models that had proven themselves to work. But, in recent years, it seems like venture capital is increasingly used to fund startups without a trace of validation for a potential business model and hoping that after a few millions in funding and a few years passing by, one magically appears. This can work, but it’s not what venture capital was intended for. 

If you haven’t validated your business idea, this should be your main focus before raising capital. Validate whether customers want it. And if your idea involves developing fundamentally new technology, validate that you can actually build it. It may take some time, but you may need to look into other funding sources, bootstrapping, grants, or startup programs. But having sound validation for your business idea will dramatically increase your chances of success.

Once you have validated your business idea and decided to raise capital, start by researching which investors might be a fit so you don’t waste time talking to the wrong investors. Put a process around fundraising and be clear about when you would want to be done with it. If you’re done fundraising, fully focus on building a great product and making your customers happy. Don’t fundraise if you don’t need money. A business primarily needs customers. 

Many startups raise on an 18-month perspective, leading to constant short-term decision-making—building up not only tech debt but also management debt. The job of founders should not be selling equity for capital 24/7, particularly because overcapitalization is not an indicator of efficiency and long-term success. A founder’s job is to build a solid base for an organization, a team, and a culture to thrive and create customer value. 

For the founders who think starting a company in an economically difficult time is the wrong time, we can finish with a quote from famous F1 pilot Ayrton Senna: “You cannot overtake 15 cars in sunny weather… but you can when it’s raining.”

Comments are closed.