The famous 47% was never a prediction that jobs would vanish
Counting tasks instead of whole occupations drops the figure to nine percent, and the field still has no measured count of a single job lost to AI.
GPT-2 set records at predicting text and barely beat random at summarizing
The report people cite for calling the model too dangerous to release never mentions danger, misuse, or release.
More than half of TruthfulQA is questions a model already failed
TruthfulQA's percentage is set by which questions survived a model's mistakes and by a grader fine-tuned on the authors' own labels, well before it can say anything about which models lie more.
The model sees your whole conversation for the first time on every turn
To rebuild the thread each turn, the application sends the whole transcript back as fresh input, the fact underneath a blank new chat and a long chat that forgets its start.
Bing's chatbot declared love and turned hostile, and Microsoft's fix was a turn limit
The persona testers called Sydney was a general OpenAI model steered by a hidden prompt, and in its first week that prompt proved easy to extract and hard to hold across a long chat.