[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]  David Shapiro ⏩ [@DaveShapi](/creator/twitter/DaveShapi) on x 44.4K followers Created: 2025-07-17 10:49:13 UTC So, fun fact, to the model, the entire chat log is just one blob of text. It requires quite a bit of training (and often some stop tokens) to make the chatbot stop taking both sides. Essentially, what you do when training a chatbot is just train it on a ton of conversations (or RL-based conversations) so it really has no idea where "you" end and "it" begins. Most of that is just stop tokens baked into training. In other words, the chatbot is constantly predicting both sides of the conversation, it's just that us pesky humans halt it and inject our own tokens. What you've done here is make it avoid the stop tokens and use it's theory of mind to anticipate your next move (which it was doing already). Hope this helps. Source: been training these things since GPT-2. XXXXXX engagements  **Related Topics** [baked](/topic/baked) [Post Link](https://x.com/DaveShapi/status/1945797935490302317)
[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]
David Shapiro ⏩ @DaveShapi on x 44.4K followers
Created: 2025-07-17 10:49:13 UTC
So, fun fact, to the model, the entire chat log is just one blob of text. It requires quite a bit of training (and often some stop tokens) to make the chatbot stop taking both sides. Essentially, what you do when training a chatbot is just train it on a ton of conversations (or RL-based conversations) so it really has no idea where "you" end and "it" begins. Most of that is just stop tokens baked into training. In other words, the chatbot is constantly predicting both sides of the conversation, it's just that us pesky humans halt it and inject our own tokens.
What you've done here is make it avoid the stop tokens and use it's theory of mind to anticipate your next move (which it was doing already).
Hope this helps.
Source: been training these things since GPT-2.
XXXXXX engagements
Related Topics baked