This blog represents a translation from a Japanese interview Anastassia gave to Warwick Matthew.
Warwick Matthews, CTO & CDO @CDL, Japan, recently sat down with Dr Anastassia Lauterbach, a world-renowned expert in AI and business, to discuss recent trends and developments in the AI space, and also how Japanese businesses, government and society might benefit from the AI revolution.
What are you working on at present?
I have three things I am working on at present:
First is “AI Edutainment,” a Switzerland-based company I founded to improve AI literacy. Our mission and commitment are to bring knowledge of AI, robotics, and quantum technologies to one million families around the globe. We do this through smart content distributed via social media and partnerships with companies and Universities. Everything is centred around explaining AI in simple language. In fact, the adults and children who read my stories and the associated background materials will be better equipped to understand the subject than most of the CIOs in Europe!
I also teach business ethics, cybersecurity, and foundations of AI at several universities. Teaching is an excellent way to test things with different audiences. I get a lot of requests to teach, so I have to be very picky about which faculty I work with.
The third area is that I am still doing a lot of corporate governance work, serving on advisory and supervisory boards of public and private companies. I have seen a lot of great governance and even more bad governance. I’m a member of a very prominent group called “Applied Governance” in Germany, which is dedicated to reforming the German governance system, which, in my eyes, is very poor.
Speaking of corporate governance and embracing AI, how can companies and governments surpass the hype and “noise” everywhere right now?
The first thing to understand is that the noise is mostly centred around Large Language Models (LLMs), and those are not the whole world of AI.
With all the hype around recent breakthroughs in transformers and NLP and the democratisation of AI applications, it is very important to understand that the roots are in technologies from the 1940s, 50s, and 60s.
Of course, we had Google’s major paper of 2017, “Attention is All You Need.”
We have then seen the “transformer architecture” kicking in, along with “generative AI” (GenAI), which was coined in its current meaning in 2014.
However, the foundational technology for AI is actually old, and this LLM-based technology has a couple of issues.
First, this technology will never stop making mistakes and “hallucinating,” despite what the AI companies claim now. An example: I am married to a trauma surgeon. We asked ChatGPT3 in December 2020 (just after it came out) about doing surgery in our kitchen. Its advice was that this was fine and that the smell of fresh cookies during Christmas would help keep the patient calm and asleep. Obviously, it was completely incorrect and the result of a major hallucination, which, of course, was not a problem for my husband as he is an expert. Imagine you rely on LLMs to gather knowledge without understanding a domain at all? There might be serious issues.
So why will those hallucinations never stop? It is complex, but fundamentally, those models don’t master what I call a “distribution shift” in math. In simple language, in terms of deep learning models, those neural network technologies are based upon the interactions of multiple layers, representing combinations of “points and clouds.”
If your environment is clearly defined, your data is in order, and you have computational resources to push this data through processing infrastructure, then you might generate fantastic results. For example, AIs are excellent at discovering X-ray anomalies and identifying cancer. They can be better than human radiologists, who finally have more time to communicate with patients. This use case is just processing this mass of points and clouds around a question.
Let’s start deviating from a specific question into more distant neighbourhoods of inquiry. In that case, AI might hallucinate because it can’t master the distribution shift. The problem has to do with how AI learns and the mere fact that AIs generally lack reasoning and true understanding.
Frankly, these LLMs are fundamentally statistically generated “token streams.” For an LLM, a token could be a word, a piece of a word, or even a sound. For example, a token stream-based model completed “Beethoven’s Tenth Symphony” in 2021.
Inside those models, so-called “synthetics ” are generated. Two competing networks in Gen AI won’t work without producing synthetic data based on input from real-world data. Today, the actual behaviour of these synthetics is not quite clear. I’m not a statistician, but my friends in statistics report that synthetics behave differently from real-world data.
Beethoven’s Tenth Symphony doesn’t hurt us because it’s ‘just’ music, and we might discuss taste, whether the output is flat or truly innovative. However, the stakes are much higher regarding medical trials or truly assessing, such as research based on health data.
Along the same lines, LLMs won’t ever bring us AGI (Artificial General Intelligence). I think it’s a nice claim of some LLM companies based in the United States because they pushed billions of dollars of funding into what they do. I would bet my cat (whom I love very much) that these funds won’t ever be returned to investors.
We are seduced by the perceived fluency of the models. But we must understand that we can’t deliver true value by just relying on prompting. You can’t prompt your way out of a problem. I love a quote attributed to Pablo Picasso:
Machines are stupid; they just generate answers.
The real issue is to focus on how to ask good questions.
So, how do we know what AI to choose and where to turn to?
Whenever I speak in front of countless businesses about implementing AI, not just LLMs but AIs in general, it could be something as simple as Bayesian networks or linear regression, which provides fantastic results for countless businesses.
I always say you’re moving into a magical triangle between your strategy and your response to competitiveness.
The first point is your strategy: how do you compete? What drives your business? Make a list of what is behind your abilities to win!
The second point concerns tools, infrastructure, and execution—the “How.” This will depend a lot on your IT landscape; for example, do you have what I call (with great love) “vintage architecture?” Suppose you are on tech that is 30 or 40 years old. In that case, it’s a very different situation from when a new company is born in Silicon Valley, benefiting from the newest and shiniest hardware.
The third point is talent. It’s not just about your internal employees; it’s also about how you pick vendors and advisors. For me, there’s one basic rule that should be understood by every single executive:
If someone who is pitching you AI has more sales and marketing personnel than data scientists, be very cautious about working with this company.
I mentor CISOs and CIOs as part of my work because I want technical people to speak in understandable language to the rest of us, e.g., including the C-Suite. Technical experts can sometimes be very introverted and struggle to express themselves clearly. Normal (non-technical) people need “stories” to navigate these complex topics.
When I advise the C-Suites and the boards I am working with, I will tell them: you don’t need to take all those great technical minds into your board of directors, but instead establish a functioning and vital “Advisory Board” around what you need. So, if your issue is (for example) cybersecurity and infrastructure, please look for people who have truly implemented and delivered in real life. Don’t hire consultants with slides from big firms. Look for real-life experience in digital transformation. Learn from those people.
Digital transformation is a long, arduous process (we jokingly call it the “trail of tears”), not a considerable bang effort, but a party at the end. Suppose you believe that you might benefit from unstructured data for your competitiveness. This is a requirement for your entire IT architecture, not just some fancy new AI models.
An old joke is, “How long does implementing an ERP / SAP system take?” The answer: “3 CIOs.” AI will hopefully not be this slow, but we should not underestimate the fundamental digital transformation that must precede successful AI adoption.
And this is for all business types, not just technology companies?
In my eyes, every single business has a “Data Shadow.” And this digital shadow drives the value of every company on this planet, whether in agriculture, healthcare, or banking. However, to release this value into the world means you must first invest. If you are dominated by fear of failure, you won’t succeed. You must invest, take risks, be courageous, make mistakes, and learn from them. As a rule of thumb, the best company that is really AI-dedicated would spend 25% of the revenue line on the Cloud because they don’t have proprietary systems or cloud systems with GPUs. And this company will pay 15% of the revenue line on cleansing data because, without cleansing data, you can’t really release any useful model. Of course, this is easier said than done…sometimes it’s tedious to sit down, go file by file, apply some algorithms to aggregate certain files into clusters, try to understand what is happening there, and just clean and repeat, clean and repeat.
Significant investment is becoming very important. What about the claims that DeepSeek was created for USD 6 million?
Frankly, this US$6 million figure behind Deep Seek is entirely misleading. The $6 million figure refers to the computing cost of a single training run for Deep Seek V3 or R1 models using “2.048 Nvidia H800 GPUs,” the version of Nvidia H100 designed for the Chinese market. Chinese people don’t have access to the most advanced AI hardware, which is actually created in the United States.
However, this 6 million does not include the cost of multiple training runs and failed experiments. How many of those were there is a question mark.
It does not include data collection, data cleansing, and data curation, which are skills that must exist in every company. It does not include research and development expenditures. For example, I discovered some pieces of Deep Seek technology using research from a Jürgen Schmidhuber paper from 2015 and 2018. Jürgen Schmidhuber is one of the most outstanding computer scientists around the globe (who does not get the same recognition as Hinton or Bengio in the US, but absolutely deserves it). Still, he is one of the greatest, and he is the father of long- and short-term memory, which sits in every single smartphone around the globe, among many other things. So this cost does not include Human Resources costs, like salaries for researchers, engineers, and support staff, and it does not include hardware acquisition, maintenance and infrastructure.
What do you think about AI for Japan specifically?
The first thing is really the property of the Japanese as a language and a culture. We don’t have a model, an LLM designed for Japan. And I think that in Japan, there is sufficient scale to support the building up of such a model.
You require this country’s whole treasure as inputs, including cultural nuances and, of course, more mundane things like regulation, technical specifics in infrastructure, etc.
For example, I’m working with a Silicon Valley-based deep tech company that is reforming how radio access networks (RANs) are architected as an open RAN environment. Semiconductor veterans built this company. It is very technical, with specialised hardware and software. So, they built this LLM, which they trained based on their domain of expertise to do things for them, provide analysis for operators, give summaries, provide recommendations, and just communicate.
If we see Japan as a specialised domain and not just hand-waving about adapting models trained in Silicon Valley in English, in American culture, then there is a case for creating a truly Japanese AI.
What should traditional companies consider when looking to embrace AI? What skills will they need to lead these transformations?
The most basic skill is understanding how data plays into strategy. The second thing is, where does this data sit?
Data Warehouse? Data Lake? Some hybrid setup? Next, what touches the data, and who is using the data? What is a piece of data in your business? A file? A database record?
For example, when I work with boards on cybersecurity, sometimes I’m literally asking them, “Are you aware how many files are required to run your business successfully?“
The way this question is answered tells us a lot. But this is a very strategic question because if you want your business to be lean and agile (these days, we are detoxing our bodies, and perhaps we need the same approach to our business your systems, but in order to detox, you need to understand where the “crown jewels” are located).
Talk to your people, go through the systems, identify those crown jewels, and then talk to specialists about how to give more power to those crown jewels to amplify value, improve your products, and possibly open up new third-party channels and partnerships.
I often hear this claim: “Amazon and Google have all the data for AI; we don’t.” You do have data, and your competitors do, so perhaps creating a “data trust” to increase the collective volume of data is possible. It is not about providing trading secrets but developing better value propositions for your respective businesses.
[There will be a future article on data democratisation and data sharing best practices.]