The ‘strawberrry’ drawback: How you can overcome AI’s limitations

By now, massive language fashions (LLMs) like ChatGPT and Claude have turn into an on a regular basis phrase throughout the globe. Many individuals have began worrying that AI is coming for his or her jobs, so it’s ironic to see nearly all LLM-based methods flounder at an easy job: Counting the variety of “r”s within the phrase “strawberry.” They don’t seem to be solely failing on the alphabet “r”; different examples embrace counting “m”s in “mammal”, and “p”s in “hippopotamus.” On this article, I’ll break down the explanation for these failures and supply a easy workaround.

LLMs are highly effective AI methods skilled on huge quantities of textual content to grasp and generate human-like language. They excel at duties like answering questions, translating languages, summarizing content material and even producing inventive writing by predicting and setting up coherent responses based mostly on the enter they obtain. LLMs are designed to acknowledge patterns in textual content, which permits them to deal with a variety of language-related duties with spectacular accuracy.

Regardless of their prowess, failing at counting the variety of “r”s within the phrase “strawberry” is a reminder that LLMs will not be able to “considering” like people. They don’t course of the knowledge we feed them like a human would.

Nearly all the present excessive efficiency LLMs are constructed on transformers. This deep studying structure doesn’t instantly ingest textual content as their enter. They use a course of referred to as tokenization, which transforms the textual content into numerical representations, or tokens. Some tokens is likely to be full phrases (like “monkey”), whereas others could possibly be components of a phrase (like “mon” and “key”). Every token is sort of a code that the mannequin understands. By breaking the whole lot down into tokens, the mannequin can higher predict the following token in a sentence. 

LLMs don’t memorize phrases; they attempt to perceive how these tokens match collectively in several methods, making them good at guessing what comes subsequent. Within the case of the phrase “hippopotamus,” the mannequin would possibly see the tokens of letters “hip,” “pop,” “o” and “tamus”, and never know that the phrase “hippopotamus” is made from the letters — “h”, “i”, “p”, “p”, “o”, “p”, “o”, “t”, “a”, “m”, “u”, “s”.

A mannequin structure that may instantly have a look at particular person letters with out tokenizing them could doubtlessly not have this drawback, however for at this time’s transformer architectures, it isn’t computationally possible.

Additional, how LLMs generate output textual content: They predict what the following phrase will probably be based mostly on the earlier enter and output tokens. Whereas this works for producing contextually conscious human-like textual content, it isn’t appropriate for easy duties like counting letters. When requested to reply the variety of “r”s within the phrase “strawberry”, LLMs are purely predicting the reply based mostly on the construction of the enter sentence.

Right here’s a workaround

Whereas LLMs may not have the ability to “suppose” or logically cause, they’re adept at understanding structured textual content. A splendid instance of structured textual content is pc code, of many many programming languages. If we ask ChatGPT to make use of Python to depend the variety of “r”s in “strawberry”, it’ll more than likely get the right reply. When there’s a want for LLMs to do counting or another job that will require logical reasoning or arithmetic computation, the broader software program will be designed such that the prompts embrace asking the LLM to make use of a programming language to course of the enter question.

Conclusion

A easy letter counting experiment exposes a basic limitation of LLMs like ChatGPT and Claude. Regardless of their spectacular capabilities in producing human-like textual content, writing code and answering any query thrown at them, these AI fashions can’t but “suppose” like a human. The experiment exhibits the fashions for what they’re, sample matching predictive algorithms, and never “intelligence” able to understanding or reasoning. Nonetheless, having a previous data of what kind of prompts work properly can alleviate the issue to some extent. As the mixing of AI in our lives will increase, recognizing its limitations is essential for accountable utilization and real looking expectations of those fashions.

 Chinmay Jog is a senior machine studying engineer at Pangiam.