It randomly works, but it tends to stop generating once it notices that the output is conflicting with its policy somehow (it won't disclose the actual policy).
PROMPT ENGINEERING
-
AI Prompt to Draw Portrait of 17th Century Physicist
By
–
"please draw a portrait of a famous physicist of the 17th century"
-
Building Megaprompt for Meeting Transcripts Analysis Across LLMs
By
–
Spent the last few days building a megaprompt for a client. This one takes in a 1-hour meeting transcript and outputs a report that becomes the basis for a paid consulting deliverable. Ran it in 4 LLMS:
– ChatGPT (GPT-4 Turbo)
– GPT-4 Turbo in OpenAI Playground
– Claude-2 via -
Gemini 1.5’s 10M Context Window Enables Personalization Without Fine-tuning
By
–
With Gemini 1.5's nearly perfect 10 million context you can:
– give entire code bases, 100+ papers, documents
– give it all the information about me (emails, writing examples etc) and have it act entirely like me WITH 0 fine tuning Your entire life can be in its context window -
Gemini 1.5 Pro: In-Context Learning Outperforms Fine-Tuning
By
–
I've been testing Gemini 1.5 Pro's reasoning capabilities this morning. It's near-GPT-4-level with regular prompts. BUT Gemini's performance improves as I add dozens of examples. There doesn't seem to be an upper limit. Many-example prompting is the new fine-tuning.
-
Gemini 1.5 Passes Complex RPG Character Generation Test
By
–
A hard test of Gemini 1.5 that every other AI has failed. I gave it a truly over-the-top RPG (the 352 page rulebook for 60 Years in Space) and asked it to role up a character The instructions are scattered across many pages, and are very complicated, but Gemini seemed to get it.
-
Gemini 1.5 Pro Self-Implements Support in Open Source Codebase
By
–
I showed Gemini 1.5 Pro the ENTIRE Self-Operating Computer codebase, and an example Gemini 1.5 API call. From there, it was able to perfectly explain how the codebase works… and then it implemented itself as a new supported model for the repo! Not perfect, but very close.
-
RAG Query Translation: Decomposition Approaches From Scratch
By
–
RAG From Scratch: Query Translation (Decomposition) Our RAG From Scratch video series walks through impt RAG concepts in short / focused videos w/ code. This is the third in our videos on Query Translation, focused on approaches for decomposition w/ ideas from Least-to-Most
-
LoRA Land: 25 Fine-Tuned Mistral Models Outperforming GPT-4
By
–
Introducing #LoRA Land: 25 fine-tuned #mistral-7b models that outperform #gpt4 for specific tasks.
— Predibase by Rubrik (@predibase) 20 février 2024
You can prompt all of the fine-tuned #LLMs and compare their results to mistral-7b-instruct in real time!
Check out LoRA Land: https://t.co/QsBrFGAIry pic.twitter.com/qZYZcLQvvgIntroducing #LoRA Land: 25 fine-tuned #mistral-7b models that outperform #gpt4 for specific tasks. You can prompt all of the fine-tuned #LLMs and compare their results to mistral-7b-instruct in real time! Check out LoRA Land: https://
pbase.ai/3ORqcwa -
Groq’s 500 Tokens/sec Enables Production Self-Reflection for LLMs
By
–
@GroqInc at almost 500 tokens/sec makes repeated self-reflection usable in production, changing everything.
— Jiquan Ngiam (@JiquanNgiam) 20 février 2024
A key technique with LLMs is to use self-reflection (Reflexion) to get the model to do better. The problem used to be that it is slow – no longer.
We can now use… pic.twitter.com/d4QSRU9a6u@GroqInc at almost 500 tokens/sec makes repeated self-reflection usable in production, changing everything. A key technique with LLMs is to use self-reflection (Reflexion) to get the model to do better. The problem used to be that it is slow – no longer. We can now use