Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
How many P’s are in Google? According to Google, there are two.
There’s also “1 ‘r’ in the word ‘poop’,” Google’s AI Overview says, and ‘d’s in the word journalism, but they’re still spelled: journalism. Google noticed that there is one P in the US President’s last name, but instead wrote it as trpum.
You didn’t have to be a prophet to say that Google’s revamp of AI-forward Search it was supposed to go well. We’ve done this before. The first time Google added AI Overviews to search, the feature was gone citing offensive posts from The Onion and Redditadvising people to eat rocks and put glue on their pizza.
This time around, as Google ramps up its commitment to making AI a bigger part of its 29-year-old product line, it’s not surprising to see it stumble.
“Speech reading has been a challenge for LLMs, and we’re working to fix that problem,” Google told TechCrunch in an emailed statement.
These basic mistakes may seem familiar. LLMs, the type of artificial intelligence that powers chatbots and other generators, are not designed to understand style. It’s been a joke for years that when a company unveils a new AI model, you should ask. how many ‘r’s are in the word strawberry. These types of AI — which can write a program in seconds, or solve problems that have stumped mathematicians for years — are as good as a preschooler learning to spell.
Google’s AI problems go beyond typos. Google already posted an article from last week when searching for the term “to ignore” would give what appears to be a dictionary definition, only the definition is expressed as, “To be understood. Let me know any time you have a question or a new question! ” But these spelling mistakes are funny because they are difficult to solve.
As researchers have done we have already explained when we ask about this syntax, the AI doesn’t see sentences as parts of language made up of words and letters. Most LLMs are built on conversion models, which break words into tokens, which can be whole words, syllables, or letters, depending on the model. Instead of “reading” as a human would, the AI converts the text into its own numbers, which are then processed to help the AI come up with a logical answer.
“LLMs are based on a transformer architecture, which is not really reading. What happens when you input quickly is translated into encoding,” said Matthew Guzdial, an AI researcher and assistant professor at the University of Alberta. he told TechCrunch. “When it sees the word ‘the,’ it has ‘the’ meaning, but it doesn’t know about ‘T,’ ‘H,’ ‘E.’
The token-based architecture that powers LLMs such as Google’s AI is in short supply, and researchers have not been optimistic about solving the problem.
“It’s difficult to get around the question of what exactly ‘words’ should be for a language model, and even if we have human experts to agree on a clear word, models can be useful to ‘chunk’ things to go,” said Sheridan Feucht, a PhD student who studies large-scale linguistic interpretation at Northeastern University. he told TechCrunch. “I guess there’s no such thing as a good sign for this kind of stupidity.”
This is not an urgent problem in the opinion of researchers, because the use of LLMs cannot be specified. But these glaring failures help us remember that AI is not perfect, even if it can sometimes seem more omniscient than our understanding. We cannot blindly trust AI results without double-checking their accuracy.
When you purchase through links in our articles, we can get a little work. This does not affect our right to repair.