After foundation models have taken research by storm since 2018, commercial applications have been emerging at high speed in the last 12 months. Due to the low entry barriers described above, there are great opportunities for companies to use these methods to optimise internal processes. However, companies face a number of challenges before they can use the methods in practice:
- The content generated with foundation models is impressive. However, in detail, the results can be misleading or wrong, as the models rely on training data, some of which is outdated or incorrect (e.g. product descriptions or prices). The models recycle the content of this training data, which essentially takes place on the “common crawl” dataset (includes all websites that have been available online since 2011). In order to achieve optimal results for one’s own application purpose, fine-tuning is usually required in which the models are trained on company-specific terminology, imagery or tasks. Companies need a sufficient amount of training data and knowledge to create their own proprietary content from the publicly available models.
- Even if the models are addressed via a simple text, the formulation of the prompts is not trivial. To achieve optimal results, a deep understanding of the individual models is necessary, which is expressed in the form of correct prompt engineering. This is about using specific terms and phrases to achieve the desired output.
- With the pay-per-use model, the leading AI providers have created a low-threshold, but in the long run expensive offer. In addition, the rights to the models do not lie with the companies; the use and proprietary training of open source models are associated with technical hurdles and usually with extremely high innovation speeds.
- Ethical aspects/ bias/ privacy: When using foundation models, other aspects also need to be considered: Data protection and privacy: Foundation models often use large amounts of data to learn and improve. The use of personal data for training the models and the storage of the data itself may raise data protection and privacy concerns.
-
- Bias and discrimination: Foundation models may tend to match the biases present in the training data. If these biases are based on gender, race, ethnicity or other factors, this can lead to discrimination.
- Responsibility: Because foundation models can generate large amounts of text, it can be difficult to take responsibility for the output of these models. It can also be difficult to trace the source of the generated texts.
- Security: Foundation models can be used to create fake texts and deepfakes. If this technology falls into the wrong hands, it can be used to spread false information and manipulated content.
- Data protection: GPT-3 or ChatGPT from OpenAI are not open source technology. If companies, for example, adapt the models with customer-specific knowledge, they are also training the language model of OpenAI or Microsoft in a certain way. What the companies will ultimately do with the sensitive data is still unclear.
