@avi_eisen

@avi_eisen Switch strawberry for carrots, I asked gpt5.5 fast , cow many rs in strawberry and it got it right, then asked how many cs in carrots and it said 0. I'm guessing the strawberry meme is now in its training data but they have not fixed the underlying problem.

@avi_eisen

@avi_eisen Looks like Deepseek v4 flash is AGI too.

@avi_eisen Now each letter is a token so this question will cost like 5x what it used to 😎

@avi_eisen even llms have brainrot now

@avi_eisen It feels like they created a different system to identify the letters in a word, and used it as a patch. Just to make sure people won't complain and stick with this issue for measuring model's capability.

@Woa8814 @avi_eisen overused ass midwit meme. What are you even trying to claim?

@morgangilroy @avi_eisen Well the underlying problem is that llms don't actually understand anything, they just predict tokens. Important to remember this, regardless of what the tech bros would have you believe.

@miler_zak

@avi_eisen Don’t worry, Google is still batting 1000 on this.

@avi_eisen in a roundabout way, yeah. make it smart + patch the gaps as they come = functional general purpose system

@Woa8814 @avi_eisen You feel smart?

@avi_eisen It’s uwu-Ing words? If that’s not agI idk what is

@adyingdeath @avi_eisen No personality bro.

@avi_eisen someone doesn't understand how tokenization works

@avi_eisen It feels like the AI labs spent ungodly amounts of time preparing the newest models for this kind of 'test' 🤣

@avi_eisen A 27B model can do the same

@avi_eisen No Rs in Stwabewwy.

But one huge fat L in Linkedin.

@avi_eisen Those two prompts burned more watts than a G Wagon in 100 miles