The current status of the AI hype cycle: what’s still standing/next to come?
In the last few years artificial intelligence has turned into a hype and been plastered on to every pitch deck remotely related to AI almost turning into a laughing stock in the VC community. Despite this hype we still believe the best opportunities are yet to come. Why? It’s different this time. We have many reasons to believe that the best is yet to come and that AI has the ability to accelerate transformation, solve significant problems and shape the industry of tomorrow.
IT IS DIFFERENT THIS TIME
AI has been around since the mid 1950’s and has gone through several hype-cycles. When AI was first introduced computer scientists overestimated the speed by which AI would yield significant scientific and commercial advancements. Despite several later breakthroughs in hard but well defined areas, such as AI being able to beat the world champion of Go, the long-lasting goal of reaching “general” artificial intelligence is still far from being reached, if ever. So why is it different this time and not just a temporary hype with a new AI winter waiting?
- First of all, the fact that we have 60+ years of research within AI means we have more, and more mature, algorithms as well as better theoretical understanding of the underlying mathematics behind AI. There have also been significant advances in the past 10 years showing unprecedented progress for example within natural language processing and computer vision.
- Secondly, there is an immense amount of data available today: 90% of all data today was created in the past two years and the trend is expected to continue with exponential growth. Furthermore, increased capacity and lower prices of data infrastructure and storage make the vast troves of data accessible to more people and organizations. The first Gigabyte hard drive was created by IBM in 1980. It was the size of a refrigerator and cost $40,000. Today you can store one terabyte of data with AWS for $20 per month.
- Thirdly, the progress within computing power, for example the use of graphic processors as AI accelerators in the form of GPUs or lately ASICS, such as Google’s TPU’s, have significantly improved our ability to process large data sets and train complex algorithms – and at lower price points.
- Finally, access to open source libraries and frameworks for development and deployment of AI has lowered the knowledge barrier significantly as well as increased development and iteration speed. Here, Google with Tensorflow, LinkedIn with Apache Kafka, and others have contributed a lot with open source systems that can be used both in research and production. In Sweden we have several useful initiatives as well, e.g. the National Library making the BERT-model for natural language processing (NLP) available, ICE in Luleå – one of the leading data center research facility in Europe and initiatives by AI Sweden, such as the data factory.
RESEARCH GENERATING THE FUTURE IN AI – LOOKING AT SWEDEN
Industrifonden’s mission is to enable the industry of tomorrow in Sweden. As an investor in sometimes research-heavy technologies we have talked to several researchers, professors, non-profit organizations and research institutes around Sweden to get a picture of the ongoing research in Sweden within different fields of AI. Our mapping shows several areas where Sweden has a competitive advantage with excellent research ongoing. Based on these finding we predict that Sweden has the potential to generate unique and transformative applications using AI for example in:
- Data driven life science:
- AI used in for example drug discovery, precision medicine or diagnostics are areas where much significant research has been performed for example at Chalmers University, Royal Institute of Technology (KTH), Karolinska Institutet (KI) and Lund University. One company emerging from this science is our portfolio company AMRA medical.
- An example of interesting research is from the team lead by professor Emma Lundberg (Stanford & KTH) working at the interface between bioimaging, proteomics and artificial intelligence to understand fundamental aspects of human cell biology.
- Google’s deep learning program DeepMind’s recent gigantic leap in solving protein structures paired with the state-of-the-art research in the Swedish life science sector, may also hold the key to solving several really hard problems. For example, cures to diseases such as Alzheimer’s and Parkinson’s.
- Natural Language Processing (NLP):
- Sweden has historically generated – and is generating – global state-of-the-art research at several universities such as Uppsala University. One example is research on understanding the grammar and structure of language by Professor Joakim Nivre and his team
- This research and progress in related fields such as computational linguistics, will continue to increase our ability to generate accurate information analytics, which can give rise to products accelerating research in itself.
- Example applications include: turning human-readable text into knowledge structures suitable for machine processing, but also next level of conversational AI in voice first systems for example in industrial applications.
- Sweden has several strong research teams in robotics, for example at KTH. At KTH there is also interesting progress in spatial computing being used to allow human machine interaction in novel ways. Worth highlighting are for example research projects led by Hedvig Kellström or Danica Kragic at KTH.
- Example applications we expect to evolve further include: robots used in healthcare, everything from assisting in elderly care to performing more advanced surgery; products making interactions with robots or consumer electronics just a matter of thinking.
- Computer vision and image recognition:
- There is a lot of great research already conducted and ongoing, for example in Lund and at Linköping University.
- Applications include: further adoption of image recognition within diagnostics, interpreting x-ray or ultrasound images, for example able to assist faster diagnosis of several cancer types but also applications for autonomous vehicles and several applications in agricultures and farming.
- Real time distributed processing is another area where research in Sweden is showing great progress.
- Examples include: Boosting AI deployment by enabling real-time scalability, which can be critical for automated business decisions that are based on real time computation and fast response.
BET ON A VERTICAL AND HOPE FOR HORIZONTAL EXPANSION
We are actively looking for investment opportunities in applied AI and in the near future we believe the most promising AI start-ups are the ones solving significant problems within different vertical niches in Industry and society. AI companies aiming directly towards the broader, more horizontal, applications of AI will have a much fiercer competition as this market will likely be dominated by the global tech giants such as Google, Facebook and Amazon.
As an example Deep Mind has an R&D budget that is many times larger than the research grants for natural sciences and technology issued by the Swedish Research Council. That, combined with enormous amounts of data, will make competing with a horizontal application really challenging.
We believe companies with a vertical application, solving a significant problem, with proprietary algorithms or data, that in the future might be used more horizontally – could be the biggest winners.
INTERESTING APPLICATIONS WE LOOK FOR
The progress within machine learning and deep learning is of particular interest. For commercial applications of deep learning it is still early days, there are challenges connected to lack of interpretability and it being computationally heavy. Over time though we believe deep learning could offer some of the most transformative applications of AI and hope for many interesting vertical applications using deep learning, e.g. within speech recognition and medical image analysis.
Some other applications we especially encourage and look for since attempting to solve significant problems and benefit society, are for example;
- AI being used to minimize negative climate impact within different verticals: optimizing the use of energy in different sectors, improving efficiency in manufacturing or in agriculture.
- AI being used to improve our health e.g. diagnostics or AI used to predict negative health indicators in everyday life or AI used to accelerate drug development
- AI used to improve education by creating great personalized education tools
- AI improving our industry – Fault detection, predicative maintenance and assistance in solving errors, increasing efficiency and productivity.
Another really important part of the AI chain is companies working on making AI work more practicable, in terms of data quality, validation, literacy and code deployment. This is today still a really inefficient area and with better tools, increased efficiency and automation, this could enable more innovation and acceleration of applied AI in many verticals.
OUR LITMUS TEST FOR AI
Many companies now have AI to some extent in their software and are happily positioning themselves as an AI company. Therefore, as an investor we want to make sure that these areas are closely evaluated:
- Are they actually solving a significant problem that needs AI?
- Do they possess access to quality data and especially proprietary data? Can they even build a data moat?
- Do they have the right competence and people?
- Is the use of AI meeting necessary ethical criteria? Is the company considering this aspect and closely monitoring?
We always begin evaluating a potential investment based on their insight into a problem that the company is trying to solve. Einstein once said the formulation of a problem is just as important as its solution. We therefore evaluate if the insight is unique and transformative and if this problem is significant and widespread enough. With companies in applied AI we also ask ourselves whether the problem actually needs AI as part of the solution. We have seen many examples of companies trying to solve problems using AI without adding significant customer value in a unique way. At Industrifonden, we look for AI-companies that are truly transformative, contributing with more than just incremental improvement but also benefit society.
When evaluating what AI application to invest in, another important aspect is the company access to (preferably proprietary) data. AI relies heavily on data; garbage in – garbage out. Without sufficient access to reliable, correctly labeled, up-to-date high quality data, the algorithms cannot function reliably and the solution will falter. Today too much of the time spent in AI project both in research and commercial are spent on structuring and qualifying data. Automating this process will be an important enabler for a wider application of AI. We look for companies ensuring:
- High quality and reliability of the data.
- Cost and ease of access to necessary data and timeliness of the data.
- Training data closely resembling the real data used in the actual AI application.
- Able to create a sizable data differentiator. Using data assets to protect from competitors for example by building a Data Moat could be an important differentiator. This is also of importance due to the diminishing return effect over time when training AI-algorithms
We also review the team and make sure they have relevant competence in relation to data and artificial intelligence. Although, number one is unique insight into a problem and how it can be solved using AI and a deep understanding of customer pain points. If founders get this part right they do not always have to have a deep AI competence rather be able to attract relevant talent in this area.
A final aspect to keep in mind is ethical considerations. Ensuring ethically aligned AI systems requires more than designing systems whose result can be trusted. It is about the way they are designed and the purpose of the design. Furthermore, the AI generated is only as good as its data. Considering, monitoring and systematically avoiding bias or being aware of bias and adjusting accordingly. Transparency and explainability of data and algorithms are of increased importance with potential regulation coming from the EU. Companies will need to keep track of data and algorithms in a very systematic way. This is a challenge with many ML systems being black boxes. All this also highlights the extra care needed when looking to use ML in applications related to medicine, law and security. We recommend EU’s Ethical guidelines for trustworthy AI for further understanding.
SOME FINAL THOUGHTS
Just like electric power was overly hyped in the beginning and it took almost 100 years before factories changed from steam and could utilize it fully, leading to the second industrial revolution; AI has been overly hyped, many companies have not been ready and are still not ready. This is changing and we look forward to investing in companies using AI to solve significant problems and shaping the industry of tomorrow.