Bandung, IndonesiaSentinel.com — AI struggle to determine the number of “r” letters in the word “strawberry” shows the limitations of AI in grasping basic concepts like letters and syllables, highlighting that, despite its sophistication, AI still does not think like humans.
While advanced AI technologies like GPT-4 can write essays and solve mathematical equations in seconds, they actually cannot answer a simple task, such as counting the number of “r” letters in the word “strawberry.” These errors highlight that AI lacks human-like understanding and operates in a fundamentally different way.
As reported by TechCrunch on Friday, August 30, most large language models (LLMs) are based on transformer architecture, a type of deep learning that breaks down text into tokens, which can represent words, syllables, or letters, depending on the model.
According to Matthew Guzdial, an AI researcher at the University of Alberta, AI does not “read” text in the human sense but rather converts text into numerical representations. This means that while an AI might recognize that “strawberry” consists of the tokens “straw” and “berry,” it does not understand that the word is made up of the letters “s,” “t,” “r,” and so on.
This limitation comes from the fundamental workings of transformer architecture, which transforms text into numerical representations to understand context. This results in AI often struggling to grasp the detailed arrangement of letters and syllables. The challenge becomes even more complex when AI has to work with various languages that have different structures. Some languages, like Chinese, Japanese, and Thai, do not use spaces to separate words, which adds to the difficulties in the tokenization process.
Breakthrough in Atomic Physics, Scientists Aim to Add New Element 120 to Periodic Table
A recent study by Yennie Jun, an AI researcher at Google DeepMind, found that some languages require up to 10 times more tokens than English to convey the same meaning. This highlights the significant challenges AI faces in managing the linguistic diversity around the world.
Similarly, AI models used for generating images, such as Midjourney and DALL-E, face analogous issues. Diffusion models, which are used to create images, tend to excel at representing large, clear objects like cars or human faces. However, these models often struggle with finer details, such as fingers or handwriting, indicating that the challenges in AI extend beyond text to the visual realm as well.
Experts note that efforts are underway to address these shortcomings. For instance, OpenAI is developing a new AI product called “Strawberry” that can generate accurate synthetic data to improve their AI models. Meanwhile, Google DeepMind has launched a new system for solving formal mathematical problems called AlphaProof and AlphaGeometry 2, which has successfully solved several problems from the International Mathematical Olympiad.
Despite advancements in AI technology, the inability to solve simple tasks like spelling words demonstrates that AI still has a long way to go before it can truly think and understand like a human.
(Raidi/Agung)
