9/ LLM In-Context Recall is Prompt Dependent – finds that a model's recall performance is significantly affected by small changes in the prompt; the interplay between prompt content and training data can degrade the response quality.
PROMPT ENGINEERING
-
System Prompts and Instruction Delimitation in AI Models
By
–
Models have worked like that for over a year now – OpenAI have a similar mechanism, it's part of how their system prompts work The problem is that it's not infallible: even with special reserved tokens to delimit instructions vs data it's still possible for models to lose track
-
Claude AI Assistant Delivers Results with Persuasion
By
–
With a little convincing, Claude comes through pic.twitter.com/vdtpbjEVRV
— Ethan Mollick (@emollick) 21 avril 2024With a little convincing, Claude comes through
-
SGML vs XML for Claude: Flexibility in Data Structuring
By
–
I wonder if we should be calling that SGML instead? Very little of what people are doing with Claude would cleanly pass through an XML parser, those things are notoriously inflexible!
-
Understanding LLM Hallucinations: Critical Challenge for Reliable AI Tools
By
–
The biggest challenge IMO is helping people understand the hallucination problem and how unreliable these tools are – if you don't understand that it's much harder to put these things to good use
-
Making Unreliable AI Algorithms Work Practically
By
–
We can also figure out how to get useful results out of inherently unreliable algorithms – it's definitely possible, I've been finding useful ways to put this stuff to work for a couple of years now
-
LLM Non-determinism and Endpoint Predictability Challenges
By
–
The amount of non-deterministic behavior LLMs introduce makes me believe in techno-sorcery futures. Let's chant a verse, burn candles, perform data prayers and result purification rituals. Anything to get this endpoints perform predictable across different input
-
Novel One-Sentence Startup Pitch Structure Across AI Models
By
–
Here's an interesting question to use to compare models: "GPT-4, Llama 3, Claude 3, Gemini 1.5, give me a novel structure for a one-sentence startup pitch and teach me how to use it"
-
Document Chunking Techniques for Better RAG Applications
By
–
But, How is Chunking Done? To create Better RAG applications, you need to know how to split or chunk the documents so you preserve the content while asking questions. Data Science Basics shows how to do this with LangChain https://
youtube.com/watch?v=tMwdl9
hFPns
… -
Essay Writing and Research Assistance Prompt
By
–
3/ Assist with Essay Writing and Research: Prompt: "I need to write an essay on [Topic]. Could you help me outline the main points, suggest credible sources, and guide me on how to integrate these sources into my essay effectively?"