Speculation: a very low probability token was sampled, shifting the context to unfamiliar area that increased the probability of further unlikely tokens. This c

Speculation: a very low probability token was sampled, shifting the context to unfamiliar area that increased the probability of further unlikely tokens. This can be seen by multiple strange punctuation and capitalisation that don't obey English grammar. What's interesting is that asking if it's chatgpt fixed the conversation for a bit. This suggests that chatgpt is nearly (or is) hard coded to respond with "No, I am an AI developed by OpenAI known as GPT-3. How may I assist you?". Another interesting point is the grammar is fixed in its responses after first breaking, suggesting that chatgpt handles previous responses very differently than the current response it's writing. This likely results from instruction tuning for dialog applications, I suspect gpt3 (not chat version) would have not recovered from broken grammar.