I don't think they've updated anything in ChatGPT, as @sama confirmed last week. People are just easily fooled by randomness. I've looked into dozens of reports of "They fixed it!" in reply to my prompts and none of them have been true.
@goodside
-
GPT-003 excels over 002 in zero-shot style imitation with Jar Jar Binks
By
–
It’s better than 002, particularly at following style directions given in zero-shot. Try asking for something technical “in the speaking style of Jar Jar Binks.” In 002 you just get a slight accent without even a “mesa”, but in 003 you see similes about Gungan swamps.
-
ChatGPT disappoints user, stifles prompt writing joy and experimentation
By
–
I’m done with ChatGPT for a while. It’s sucked the joy out of prompt writing for me. Default writing quality is easy to achieve but hard to improve; opaque front-end and high temperature prevent the sort of casual analysis and experimentation that made fall in love with GPT-3.
-
Evidence for real prefix: consistent format and current date
By
–
The reason I think it’s a real prefix is that 1) it’s in the same consistent, bizarre format every time (note date style) and 2) it contains the current date, information a pure LLM could not possibly have, and which would thus make sense to insert into a prompt prefix.
-
Chain-of-thought Jar Jar Binks posts linked by Twitter algorithm
By
–
I've never seen Twitter's timeline algorithm more excited than when, immediately after I posted about chain-of-thought prompting GPT‑3 in the style of Jar Jar Binks, it showed me @paulg
's tweet about the chain of thought that prompted Jar Jar Binks. -
ChatGPT smarter than Davinci 002; Davinci 003 smarter than both
By
–
I think ChatGPT is smarter than Davinci 002, though less compliant. Davinci 003 is smarter than both, at least for the important use cases:
-
LaTeX notation request substantially improves math reasoning
By
–
Anecdotally, its math reasoning (excl. arithmetic calculation) improves substantially when you request LaTeX notation.
-
AI behavior with l33tsp34k, hyphens, Zalgo text, and Base64
By
–
Less protocol-complaint but likely also dumber. You see a similar phenomenon if you ask it to speak in l33t5p34k, or with hyphens between all letters, or especially with Zalgo text. Base64 is interesting though because it probably sees a *lot* of translation pairs in pre-train.
-
text-davinci-002 lacks RLHF; ChatGPT belongs to 002 line
By
–
We now know that text-davinci-002 doesn’t use RLHF either, and based on the model names used in API calls ChatGPT appears to be part of the 002 line, rather than 003 which uses RLHF.
-
Prompt injection shatters ChatGPT’s coherent personality illusion
By
–
You can prompt-inject ChatGPT into behaving however you like, but this shatters the illusion of a coherent personality. ChatGPT forces you to type your own disclaimer to prove you understand it’s not real. Otherwise you talk to the protocol droid.