There has been a lot of talk recently about artificial intelligence and possible applications of generative AI. ChatGPT, Bard, LLaMA, Apple GPT, Claude 2, GitHub Copilot, Microsoft Copilot, Midjourney, DALL-E, Stable Diffusion, Jasper, and other tools are shooting up like mushrooms, and the big players are releasing new features daily to outdo the competition. The race is on. So, is it all just hype – or a real opportunity for a digital agency such as Unic? We talked to Ivo Bättig, innovation enabler, transformation consultant and partner at Unic, to learn more. His first conclusion: AI will drastically change our work and the work of our customers.
His first conclusion: AI will drastically change our work and the work of our customers.
Everyone is talking about AI – including at Unic. Could you summarise briefly what is happening at the agency in this field, Ivo?
Absolutely! Artificial Intelligence is one of two recent topics we are looking at in our innovation lab right now. The other is spatial computing, which is now possible with Apple Vision Pro, for instance.
Within the agency, we are informing and training employees. We provide access to the tools and have already run several professional development events on topics such as writing and copywriting, coding or designing with AI. We are planning a first hackathon in cooperation with a customer, the goal of which will be an implementable project. And we are also adding expertise in this field: We have recruited our first consultants and researchers in the field of machine learning and artificial intelligence, and we are keeping our eyes open for additional talents. Last but not least, of course, we are also investigating the topic from a strategic point of view.
Can you say anything about the strategic aspect?
It is a bit too soon for details. Currently, we are evaluating the situation:
What does AI mean for activities or job profiles at Unic?
What is the impact on the services we provide to our customers?
What are the effects of AI on our customers’ customers?
Of course, we are also in close contact with many of our customers, have conducted brainstorming sessions and workshops and are working on a few first drafts. It is going to be an exciting time.
Listening to you, it is fairly obvious that you are passionate about the subject. How come?
Well, I have been for a while, actually. A long time ago, I studied computer science at EPFL in Lausanne. It was a hot topic even then. I remember a book by Raymond Kurzweil, “The Age of Intelligent Machines”, published in German in 1993 – it’s an excellent anthology of basic philosophical, mathematical, mechanical, electronic and logical ideas, with great examples and chapters on how the systems learn.
About three years ago, I started playing around with neural networks for the recognition of handwritten digits. It was not about creating something earth-shattering, but it helped me understand the difference between how human and artificial intelligence works, and that we need a differentiated evaluation of intelligence. Understanding how intelligence works is difficult: As well as the brain, there are other systems at play, and also other sciences such as chemistry and biology. Computers are based on maths (for now). This means they have to make up for a lack of complexity through their computational power. One way of doing that is with neural networks. I’ll show you based on a model and try to explain it in simple terms:
An artificial neural network consists of several nodes called neurons, just like in the human brain. These receive information either from other neurons or from the outside world, modify it and forward the result to the next neurons. To be able to do that, the neurons are connected by so-called edges. The input and output layers are clearly visible. Between them, there are the so-called ‘hidden layers’. There can be many of these, depending on the computing power available and the topic you want to work on.
Okay, understood. But how can this network read your handwritten digits?
It can’t, at first! For a network to be able to perform a certain task, you need to train it. So, you feed the network a huge volume of handwritten numbers. Based on actual results, the network continuously compares the calculated and the actual information and adapts the connections between the neurons, a bit like adding weights to the edges.
For instance, you and I both just “know” what the number 4 looks like. The AI has to be trained to recognise it – and will only start supplying useful results once it has learned certain patterns: “When a number has a score across the middle, it could be a four.” And even there, it could be wrong: Some write the number 7 with a score across the middle. In reality, the process is a little more complicated, but that would be a bit too much for this article. If you are still curious, there is a lot of literature on the subject.
That does not sound terribly intelligent to me just yet.
That is exactly the point: AI is not really intelligent. Or at least it raises the question of how we define intelligence. Let’s look at another example:
ChatGPT, or, to be precise, the trained Large Language Model or LLM it is based on, operates on a word-by-word basis. Mind you: In correct terms, it is token by token, and that can be just part of a word. But I’ll stop splitting terminological hairs for now, it is complex enough as it is and we are not writing a scientific paper here. The system works step by step, producing the word that is most likely to follow the previous ones – at times with a few statistical deviations so that it does not become too monotonous.
Here's an example: Let’s assume the AI has already worked out “The sun goes”. Then it will probably suggest “down” next, because that is statistically where the sun goes most frequently, or rather, what follows most frequently in the texts used for training.
That is, of course, a very simple example. But it shows how complex the process of training a tool for text or image generation can be. And just how much computing power it requires. And if you think quantum computers – well, then we cannot even begin to fathom the possibilities.
If people have been working on this for such a long time, why did OpenAI cause such hype with ChatGPT?
To be honest, I did not see that coming either. Sam Altman, CEO of OpenAI, is no stranger, and it has been known since 2015 that he was working on innovative AI projects. But nobody foresaw what happened in November 2022 with ChatGPT. With that bold move, OpenAI – with substantial support from Microsoft – made the other giants such as Google, Amazon, Apple and Meta sit up and take note. All of a sudden, it became clear why Microsoft invested in GitHub – it was all about code data.
Every week we hear of new providers, and email inboxes, YouTube and news channels are teeming with courses, prompts (used to instruct generative AI), hot news, tips, tricks and more. And I am really excited about the cornucopia of conferences in autumn on this subject. It is not easy to keep up with these rapid developments, but it is important: The impact of the latest developments in AI will be massive. That is why we are seeing so many reports on massive AI investments by service providers and on industry analyses about the new world of work.
And how do you stay up to date?
Well, first of all, I read and watch a lot! And of course, I use generative AI to generate my own summaries of articles and papers. I am also in contact with many internal and external experts and keep trying out new things myself.
Oh. Do you have an example of that?
Of course. When the whole hype around ChatGPT happened, I had this idea of a ChatGPT document query. The Large Language Model GPT by OpenAI used by ChatGPT was trained with data as of November 2021, which not a lot of people know or keep in mind. Later, there was a minor update, but if you work with GPT-3.5 without access to the internet, the information you are relying on is in part outdated.
Additionally, the trained model has no information on your own data or documents. That is why I wanted to develop a mini tool allowing me to upload current information on hand and make it searchable, and with the same level of comfort I am used to from ChatGPT, through simple chat prompts.
So, I developed a small MVP, as we call it: A minimum viable product called DocuQuery that allows me to upload a PDF file, for instance, and process it in a way that the tool can answer questions on the document comprehensibly within seconds. And no, I was not the only person with that idea: a wide variety of professional tools is available in the market now.
Circling back to the agency: What does all this mean for Unic?
Our customers are seeing completely new opportunities in the field of AI: Text, images and even videos can now be generated with the help of artificial intelligence. But the tools come with risks and pitfalls: Based on the content they were trained with, they are biased, prejudiced and contain stereotypes. Since they only work with probabilities, they may hallucinate in certain areas, that is, simply “invent” facts and sometimes even sources. Depending on the subject, they may even generate completely useless content, which you need to be able to spot. Because at first glance, it usually sounds very convincing.
As an agency, our responsibility now is to make sure that our employees familiarise themselves with the opportunities and risks of the different applications, pass on that knowledge internally and use it in our work with customers wherever it makes sense.
We need to learn for ourselves to use AI in a smart way so we can generate added value for our customers. We already do that in the fields of research, image and text generation, with chatbots or when developing early AI customer applications. And more will follow.
I am happy to hear that you do not think our agency will run out of work in the near future. What is that based on?
There are different aspects. One is content: All these AI tools generate content – they do not create it. So, if we do not want to drown in regurgitated, rehashed content in the future, we need to learn to efficiently write or update dry and informative texts. That will give us more time and resources for the creative process – or for research and quality assurance. Of course, this not only applies to text but also to visual content or certain elements of coding. And in those fields, our expertise and the way we handle customers’ requirements will still be relevant in the future.
In terms of our organisational structure, we have a huge advantage over traditionally organised companies: As a purpose-oriented, self-organising company based on Holacracy, we can acquire and promote knowledge quickly, simultaneously and in a decentralised fashion. This enables us to test, validate, optimise and scale AI in small and large projects. In the current situation, it is better to take small steps and actually move ahead than to get stuck in planning bigger steps. And like I said, we are lucky to have customers who agree with us and have joined us on this journey. I am sure we will be able to reveal more when the time is right.
I am already looking forward. Until then, I hope to talk to you about a few other topics related to AI.
Absolutely. Intellectual property rights, regulatory issues, ethics and sustainability are some of the topics I am itching to discuss, and maybe our readers, too. We’ll keep at it!
Definitely! But for today, let me thank you for your time and let you go. Duty calls. And we do have some work left to do here.
Would you like to know more about AI – and related opportunities in your line of business? Get in touch:
I am here for you!
Partner & Innovation Enabler, Unic
