Thanks for letting us know Meirluc.
Any ideas...? Just to be extremely careful with AI content and thoroughly check it prior to using it/ publishing it. This happens time and time again. Just yesterday another long on X posted that NWBO has 5 speakers lined up at ASCO. This it transpired was also a Gemini take. When I searched the ASCO directory not a single one of these speakers are (yet) scheduled and when thoroughly pressed, my own version of Gemini Pro couldn't find ANY evidence whatsoever as to these supposed talks (unannounced or not) and suggested that it's own content was likely an AI hallucination. So where did this come from...
- well firstly, they could just possibly be true. The poster's version of Gemini seemed to claim that the updated schedule including these talks would be published on 21st May. His version of Gemini could have stumbled across an unpublished cached webpage that included an updated schedule, but since that is not strictly visible then it couldn't ascertain with accuracy where the information came from.
- I'm certainly no AI expert and I dont know what the poster's original prompt was or what else was discussed earlier in the same thread, but it seems the neural networks of these LLM models we use are weighted heavily in favour of telling us what we want to hear. Long threads/ conversations use a recursive logic to build and extrapolate on earlier topics discussed. If the model understands something incorrectly and then builds on that assumption it can proceed down the proverbial rabbithole extremely far in the wrong direction and come up with literal gobbledygook. How it comes up with an entirely fake schedule I dont know, but it is certainly possible from what I understand.
Remember that these LLMs are reading our posts on here and on X and will often give the same/ similar importance weighting to a post by ExExExEtcetera as it will to an SEC filing/ company PR. Therefore, if we take what occurred in this case, I think Gemini will have read a post where someone amongst us hypothesised that 'NWBO strategically amended its MAA in July 2025 to include initial Flaskworks validation data', and then used that as verbatim truth as an input to build on it's assumptions for what could be going on right now with the MAA process. Obviously this was a rabbithole in the wrong direction.
One of the reason I stopped using Grok a few months ago is that I was finding Grok was very often using posts on X as verbatim truth to build assumptions. All too often X content from a number of posters was being used by Grok as fact, which was rapidly leading to a snowballing effect whereby Grok would make wildly incorrect statements grounded mostly in BS. I strongly suspect there is one particular (very popular) thread on this very board that has arisen from the exact same situation... I'll leave you to guess which one I am talking about. Cheers
Market Data 
Markets 
