Title: AI Startups Face Uncertain Future as Hype Surrounding Generative AI Grows
The field of generative artificial intelligence (AI) has been flooded with hype recently, leading to concerns that numerous AI startups receiving venture capital (VC) funding may meet an untimely demise. Drawing parallels to the cryptocurrency bubble, experts predict that a significant portion of these startups, approximately 70-80%, will fail, leaving only a selective few with viable and sustainable business models.
Even industry heavyweight OpenAI, known for its developments in generative AI chatbots, has experienced a decline in usage. This decline suggests that the promised world domination by these AI chatbots may not come to fruition as anticipated.
As an AI software engineer working for an AI legal startup, I have witnessed similar patterns before in the self-driving car industry. Despite advancements in AI technology and the impressive abilities of programs like ChatGPT and Dall-E, it is critical to understand that these systems merely mimic information they have learned from the past.
The current capabilities of AI fall short when tasked with reliable prediction. Throughout the years, we have witnessed three distinct waves of machine learning within AI development: supervised learning, unsupervised learning, and reinforced learning.
Supervised learning involves training AI models to perform specific tasks, such as identifying objects or answering queries. Startups emerged to fulfill the manual labeling and training needs for these models. Unsupervised learning, on the other hand, focuses on algorithmic rule setting to detect and classify objects based on specific attributes, such as color recognition. Reinforced learning takes the training process a step further by enabling the model to learn from feedback and reinforcement, affirming correct predictions or identifying errors.
These three waves laid the foundation for AI engineering, leveraging a combination of reinforced and unsupervised learning techniques. However, it is crucial to recognize that AI technology today is essentially a probability game. It relies on determining the highest probability of future occurrences, making predictions based on vast amounts of data ingested by the system.
Take self-driving cars as an example. Deep learning models analyze various factors, such as a person’s posture or actions, to assess the likelihood of certain outcomes. However, accidents still happen due to the inherent probabilistic nature of deep learning. Therefore, while AI systems can be powerful when used correctly, they heavily depend on the quality and bias-free nature of the information provided.
To determine the potential success or failure of an AI startup, it is essential to assess whether their technology relies on static information replication or prediction of future outcomes. Startups focusing on reusing static data, such as warehouse robot routing in controlled environments or call center triage, have more favorable prospects. These tasks involve consistent and predictable patterns that AI can handle effectively.
However, startups requiring heavy prediction capabilities, mirroring the self-driving car industry’s challenges, are likely to face significant hurdles. Virtual human apps aiming to replace human managerial roles, AI bots substituting salespeople, or technologies demanding strategic defense development fall into this category. Additionally, AI’s ability to understand and predict human emotions, as required in concierge services, presents further difficulties.
While AI remains limited in its capabilities, an intermediate solution involves using AI to perform consistent and repetitive prework, while human oversight keeps processes in check. However, even this approach comes with its true cost. The Amazon Go stores, for example, aimed to provide an improved user experience by replacing human cashiers with smart technology. Nonetheless, the hefty expenses associated with engineering talent, proprietary technology development, ongoing support costs, and model maintenance have challenged the cost-effectiveness of such endeavors. Retailers opting for self-checkout have achieved similar efficiency gains without the need for costly AI implementations.
In conclusion, the current hype surrounding generative AI has prompted concerns that many AI startups will struggle or fail. Startups focusing on tasks requiring static replication of information are more likely to succeed. Conversely, those relying on accurate prediction of future outcomes, such as virtual human apps or AI-assisted legal defense strategies, face challenging prospects. Until AI technology evolves further, leveraging AI for consistent prework tasks alongside human oversight remains the most viable approach.
