Erkennen Sie die Probleme, die KI nicht lösen kann

The ability of AI to solve complex problems depends on several factors that can increase (or decrease) the effectiveness of AI. The most important factors include: availability of high - quality data, number of potential solutions, clarity of goals, and the need to adapt to a constantly changing evaluation system. If these factors are missing or not clearly defined, challenges arise.

The development of AI is progressing rapidly. Its potential can be revolutionary and fascinates researchers, product teams, and end - users alike. However, despite its great popularity, AI can only solve very specific problems. In this article, we provide companies with guidelines to understand which problems are suitable for AI - based solutions and which are not. In this way, if the problems posed by companies do not match the strengths of AI, they can still find solutions. After overcoming these obstacles, they can find further AI projects that are of greater benefit to companies and society.

Let's take Google DeepMind as an example. DeepMind is an AI laboratory of Alphabet that is dedicated to solving extremely challenging problems in the real world. Three of us (Paolo Cervini, Chiara Farronato, and Marshall W. Van Alstyne) are engaged in academic research on the development of technology platforms and have implemented innovative AI applications in various fields and gained expertise from them. In addition, Pushmeet Kohli, one of the authors, has first - hand experience as an employee at Google DeepMind.

The following content helps to decide whether AI is suitable for solving a specific problem. However, we do not provide a complete list of all criteria. In particular, we take into account that these criteria may vary depending on the application area of AI. Our main goal is to find the most important criteria.

Can AI solve this problem?

The ability of AI to solve complex problems depends on several factors that can increase (or decrease) the effectiveness of AI. The most important factors include: availability of high - quality data, number of potential solutions, clarity of goals, and the need to adapt to a constantly changing evaluation system. If these factors are missing or not clearly defined, challenges arise. We must apply innovative methods to overcome these challenges. The following are the measures that should be taken when these four factors and challenges occur:

1. Lack of high - quality data

Data is the most important input for any AI model. However, often more attention is paid to the quantity of data than to its quality. Although the development of Large Language Models (LLM) indicates that an increasing amount of data can lead to high - quality models and results, it is unclear whether this trend will continue.

It has been shown that data quality is just as important or even more important than data quantity. In some cases, if you have a relatively small but high - quality database, you can even increase the amount of data by generating virtual data.

This was the situation that the team behind AlphaFold initially faced. AlphaFold is a groundbreaking technology from Google DeepMind that can accurately predict the three - dimensional shape of protein structures. This technology accelerates research in almost all biological fields.

In 2017, AlphaFold was in the early stage of development. The development team only had about 150,000 data points on 3D protein structures. These structures were experimentally determined over decades through expensive and time - consuming technical methods (such as X - ray crystallography or cryo - electron microscopy). (Studying a single protein structure requires at least one doctoral student.) For a machine - learning system, this is a very small amount of data.

However, the research team successfully expanded the data by predicting the structures of unknown protein sequences. They created a prototype version of AlphaFold to achieve this. This version was powerful enough to generate a million predictions about new folded proteins.

Subsequently, the team used this version of AlphaFold to evaluate how accurate its predictions were. Through this self - improvement process, the team included the best 30% of the predictions back into the dataset for the training model and used them together with the real data. In this way, the size of the dataset expanded to about 500,000 folded proteins, which was sufficient to train a final version of AlphaFold.

It should be noted that although this training method was very effective in that situation, the use of generated virtual data to train the model must be done with great caution, as this carries the risk of recursive learning: Using predictions from earlier model iterations as input for subsequent iterations can amplify the errors and inaccuracies from the previous iteration and degrade the performance of the model.

However, if the initial dataset is very large but of low quality, it is very difficult to improve the data quality through synthetic data. Data collected from the Internet - which can contain various information, formats, languages, topics, and methods - belongs to datasets with large quantity but low quality. In most cases, it is not worth improving the quality, as this requires a lot of time and effort. In such datasets, the model may also not find "correct answers" for synthetic data - for example, when asking what recipe makes a perfect pizza, the answer is most likely "It depends." These subtleties and the dependence on context make it difficult for the model to structure language and general knowledge and thus create high - quality datasets.

2. Too many potential solutions

If a problem has too many solutions, it is impractical to solve the problem by a "brute - force" method, i.e., testing all possible answer combinations. In the past, such problems were always solved with heuristic methods: With simple rules, a "sufficiently good" solution is found for most cases, instead of the optimal solution.

The application of AI is a promising option for complex problems with countless solutions. However, the more potential solutions there are, the more difficult it becomes to check the quality of each solution. For example, LLMs are good at generating creative and diverse answers, but they also have the problem of the "hallucination phenomenon," i.e., generating information that does not match reality. This reduces trust in them. Therefore, finding methods to ensure that the solutions generated by LLMs are correct and verifiable has become the key to the further development of the capabilities of LLMs.

3. Lack of clear and measurable goals

The objective function (also called the reward function) is the goal or output that an AI model tries to achieve. In other words: We must ask the model the right questions. Determining "What do you want the model to do?" is one of the most difficult parts in any machine - learning system. Take games like chess or Go as an example. These games have clear and measurable goals, such as scores or rules that determine the winner, and the iteration costs are low, so the AI can learn quickly.

In the real world, however, the goals are usually complex and disordered, and there is no simple unit of measurement to measure progress. Without clear and measurable goals, it becomes difficult to define the "good" standard. The less clear the goals are, the worse the performance of the model.

Let's repeat the example of AlphaFold. In biology, there are clear and measurable goals for protein folding, and the goal of AlphaFold is to minimize the difference between the experimentally determined 3D protein structure and the calculated prediction. A good way to measure success is the Critical Assessment of Structure Prediction (CASP) competition, which takes place every two years. This competition is like the "Olympics" of protein folding. It brings together the best teams from around the world to compare their calculation methods for predicting protein structures. By participating in this competition, the research team of AlphaFold can obtain an objective "evaluation" of the model's performance and compare it with other teams.

Providing AI with clear goals and measurement criteria is a challenge. Especially in environments such as social media, "engagement" is often regarded as a measure of the success of a project, measured by likes, shares, comments, or the time spent on the platform. Optimizing engagement can promote user activity and income, but it can also have unexpected consequences. For example, it can amplify extreme thoughts, spread sensational content, false information, etc., or encourage addictive behavior and harm the long - term well - being of users.

These pitfalls underline the importance of designing a multi - objective optimization framework that can balance engagement with other important indicators, such as the accuracy of content, the diversity of opinions, and user satisfaction. Incorporating human feedback, ethical principles, and more comprehensive evaluation criteria in addition to simple evaluation standards helps the AI generate meaningful and sustainable results while avoiding damage.

4. When "good" cannot be written in code

The problems that organizations face usually change constantly. Together with the difficulty of identifying a solution as effective, there is a risk that the AI will gradually deviate from the best answer. To overcome this challenge, more and more technologies are using the Reinforcement Learning with Human Feedback (RLHF) method.

This "human participation" technology enables the model to learn from human insights outside the data. RLHF is particularly suitable for situations where it is difficult for the AI to develop a solution based on a clear algorithm, but humans can intuitively judge the quality of the model output.

The recent collaboration between DeepMind and the YouTube Shorts team is a good example. Since the creation of short videos usually happens quickly, creators usually only add a small amount of information to the video titles or descriptions, which makes it difficult for viewers to find the short videos they want to watch.

Generating accurate video descriptions is the problem that the AI needs to solve: Each video is unique. General descriptions generated with heuristic methods can only be used for the rough classification of videos (e.g., whether it is a sports or gardening video), but are not sufficient to improve the uniqueness of the videos on an individual level. However, accurate and individual video descriptions are the prerequisite for viewers to find the content they like.

Humans are good at intuitively judging what a "good" video description is. Computers have always had a hard time with this, and it is also difficult to write a "good" description for each video. By directly inputting human feedback into the next iteration of the model, the model can learn by trial - and - error, intuitively understand what a "good" description is, and continuously improve it. Through continuous learning and adaptation to new information, the model can adapt to changes in social aesthetics and norms.

Google DeepMind, together with the YouTube product team, has used a generative AI model that generates a text description for each video. These descriptions are stored in YouTube systems to provide users with better search results. This solution is now applied to all newly uploaded YouTube short videos.

Which opportunities should we seize?

Companies can use the above - mentioned criteria to determine which problems are suitable for AI - based solutions. Next, they must select from a variety of AI application opportunities that meet these criteria. Prioritizing the implementation of AI based on the impact and scope of the solution is beneficial not only for Google but also for society as a whole.

The DeepMind team calls this method the "root node problem." Imagine all world knowledge as a huge tree, whose branches represent different fields. In this tree, each "node" is a connection point that acts like a springboard to other branches. The root node is the starting point of the tree - the original place where everything else grows. When a root node problem is solved, new research areas and understandings emerge, which in turn open up new ways for research and innovative applications.

The AlphaFold project fits this criterion well, because solving the major challenge of protein folding can not only drive progress in the discovery of better medicines, such as malaria vaccines and cancer drugs, but also promote development in other areas, such as the development of enzymes that can break down plastics to combat environmental pollution.

After the publication of AlphaFold, the evolutionary biologist Andrei Lupas from the Max Planck Institute for Developmental Biology wrote in the journal "Nature": "This will change medicine, research, bioengineering, and everything else." In 2024, the Nobel Committee awarded the Nobel Prize in Chemistry to Demis Hassabis and John Jumper to recognize their great contributions to the development of AlphaFold.

To find root node problems, companies should arrange discussions between AI experts and experts from other fields. The early DeepMind team has already applied this method. In addition to machine - learning experts, the team has deliberately hired many talents from different disciplines, including biochemistry, physics, philosophy, and ethics. This diverse professional background can stimulate the exchange of ideas between different disciplines and find new solutions.

Iterative Innovation

Once an organization has identified the root node problem, the product team strategically selects which AI technologies to use and begins the product development process.

During the product development process, two risks should be avoided...