It appears WebGPT is doing what the GOOGLE TRILLION TOKENS thing James Bowery posted recently, but using the web, not a dataset.
Also it's doing what I said to do, apparently, which is to predict words and those link to motor actions like notepad and browser etc controls, or a robot body even. Using vision would work too. Also they seem to be making it "grab" (motor command) a preferred passage. So it therefore maybe is predicting GrabIt when it sees/predicts a preferred condition/state it wants....A>B....like I mean they fine tune it strongly on ex. "lots o food button" > "grab it!" - - - - - "such motor action".....so then when it sees that similar passage, it predicts to grab it strongly. So they essentially can/ are/ will be making it predict food a lot for example let's say, and also making it predict grab_it but ONLY if sees food first, conditionally, i.e. you could fine tune it by shoeing it lots of (or strongly raise) "food" token and "food, grab it" token. That way, when it sees food, it predict that token "food, grab it" in full because it's not finished activating the node yet. At least...in theory...this makes sense.... ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T2451a036e13f64be-M271665662201c98268afe5fa Delivery options: https://agi.topicbox.com/groups/agi/subscription
