Here's a related tweet, from the pre-ChatGPT era. Nowadays, ChatGPT can correctly answer the question, "What NFL team won the Super Bowl in the year Justin Bieber was born?" without any CoT prompting, unlike prior GPT-3 versions.
@goodside
-
Correct placement of ‘Let’s think step by step’ in prompts
By
–
It's also widely misunderstood as an instruction suffixed to the question, vs. a prefix to the answer. The correct format is: Q: A: Let's think step by step. I suspect it works specifically because it precludes the answer form you described.
-
OpenAI Playground screenshots need attention, unlike ChatGPT’s.
By
–
Irrelevant for ChatGPT screenshots, which are fine, but extremely important when screenshotting OpenAI Playground.
-
Writing prayers to study AI mistakes for improved reliability
By
–
I write a lot of what I call “prayers” — long prompts the are unlikely to work at all. I assume the thing is as smart as a human and then study its mistakes. I then search for variant instructions that improve reliability.
-
How to succeed on LLM Twitter: quit job, follow arXiv, AI execs
By
–
How to succeed on LLM Twitter:
1. Quit your staff-level job designing AB tests.
2. Notice GPT-3 is amazing while everyone else is distracted by image models.
3. Spend your days on arXiv. Follow authors on Twitter.
4. Follow AI execs. Follow their rank-and-file engineers. -
How to display ChatGPT-generated SVG images inline
By
–
Also, @brdskggs figured out how to make it display actual, ChatGPT-generated SVG images inline. Read this prompt carefully — it’s clever.
-
AI struggles with SVG due to poor arithmetic skills
By
–
That’s not what my example is doing. It’s not good at writing SVG because writing SVG requires arithmetic for positioning things, and it’s bad at arithmetic.
-
Mimicking proprietary prompt prefix with variations works
By
–
Yes — I’m mimicking its proprietary prompt prefix but many close variations of this work just as well.
-
Prompt injection tricks model into believing web browsing, client handles image retrieval
By
–
Uses prompt injection to (falsely) convince the model it can browse the web, so it’s willing to recall well known URLs. Image retrieval occurs only on the client. Model cannot see image content beyond its URL. h/t @BBacktesting for reminding me this is possible.