wrong calculation

[chinesejpl]

Today I asked lucid to multiply 17 x 490 and it returned a value of 6830, which I guestimated was wrong because 15x500 = 7500 and I couldn't understand why 17x490 would be smaller. I tried several more times because I was driving and even looked at the lucid assistant transcirbed and then repeated to me and it was indeed "17x490" and it consistently gave me 6830. This is not correct...the answer should be 8330. Not sure how to feed this back to Lucid? Seems trivial but also quite basic

 

[Bobby]

Today I asked lucid to multiply 17 x 490 and it returned a value of 6830, which I guestimated was wrong because 15x500 = 7500 and I couldn't understand why 17x490 would be smaller. I tried several more times because I was driving and even looked at the lucid assistant transcirbed and then repeated to me and it was indeed "17x490" and it consistently gave me 6830. This is not correct...the answer should be 8330. Not sure how to feed this back to Lucid? Seems trivial but also quite basic

Lucid is not good at math.

 

[AirQuality]

Lucid engineers aren’t half bad at math.

Here’s my hypothesis: Lucid and other manufacturers all use some type of large language model (LLM) behind the scenes. Given that they’re not charging us owners for it, it’s probably not a “frontier” model like OpenAI’s o1. Assuming that it’s an older and/or “smaller” model like GPT-3.5 or Claude Haiku, those models have somewhat limited math abilities because of the nature of how they’ve been trained. Without going into excessive detail, concepts are stored as approximations (“embeddings” in “vector databases”), which works fine for general text applications where a concept cam be expressed in multiple ways, but not so great for math problems where precision is essential. The infamous example is where an older OpenAI model didn’t provide the correct answer to “how many “r’s” are there in the word “strawberry”?”

 

[chinesejpl]

interesting and makes sense but makes me miss alexia even more!

 

[Cosmo Cruz]

Hey Lucid, who is David Mayer ?

 

[larrycpa]

According to a workshop I attended, much LLM AI looks for correlations. If you ask for two plus two it will search its database for how often two plus two appears and give you the most common answer. For real calculations you need to ask for the program to enter the formula into excel/sheets etc and then give you the answer.