This is a lot more than app has bug - it effectively demonstrates that all the hype about LLMs being "almost AGI" and having real understanding is complete bullshit. You couldn't ask for a better demo that LLMs use statistics, not understanding.

While I agree that Marcus's tone has gotten a little too breathless lately, I think we need all the critiques we can get of the balogna coming from Open AI right now.

No, this doesn't show anything of the sort. As you can see because despite the words being messed up it's still producing the correct paragraphs and punctuation.

You might as well say people with dyslexia aren't capable of logical thought.

You worded his unstated assumptions beautifully. I completely disagree, though: this demonstrates the exact opposite, that LLMs are using statistical methods to mimic the universal grammars that govern human linguistic faculties (which, IMO, is the core of all our higher faculties). Like, why did it break like that instead of more clear gibberish? I’d say it’s because it’s still following linguistic structures — incorrectly, in this case, but it’s not random. See https://en.m.wikipedia.org/wiki/Colorless_green_ideas_sleep_...

Marcus’s big idea is that LLMs aren’t symbolic so they’ll never be enough for AGI. His huge mistake is staying in the scruffies vs neat dichotomy, when a win for either side is a win for both; symbolic techniques had been stuck for decades waiting for exactly this kind of breakthrough.

IMO :) Gary if you’re reading this we love you, please consider being a little less polemic lol