The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions

OpenAI Blog2024-04-20 03:000 views

Today's LLMs are susceptible to prompt injections, jailbreaks, and other attacks that allow adversaries to overwrite a model's original instructions with their own malicious prompts.

Source：https://openai.com/index/the-instruction-hierarchy

My favorites

PreviousOpenAI’s commitment to child safety: adopting safety by design principles NextIntroducing OpenAI Japan

How I Built an LLM Inference Engine From Scratch in CUDA (235 tokens/s on a 6 GB Laptop GPU)

My laptop's RTX 3050 now generates 235 tokens per second with TinyLlama-1.1B, using an inference engine where every GPU

0dev.to

I Replaced Every Cable in My Travel Bag with MagSafe & Qi2 for 30 Days

A few weeks ago, I made an aggressive bet with myself: I packed up every single USB cable, braided wire, and multi-port

0dev.to Watercooler

What Does 95520# Do on OPay? How to Check N50,000 Relief Eligibility

What Does *955*20# Do on OPay? How to Check N50,000 Relief Eligibility The OPay USSD code *955*20# allows customers to c

0dev.to News

DeepSeek V3 en 190 líneas de JAX puro

Una implementación de 190 líneas en JAX puro reproduce toda la arquitectura de DeepSeek V3: atención latente multi-cabez

0dev.to Discuss

Saving Telegram Web Media in Batches: How ExtStack's Downloader Works

Disclosure: this is a product guide from the ExtStack team. Telegram chats can contain a mixture of videos, pictures, vo

0dev.to Productivity

AI literacy training is now a paperwork problem, not a training problem

The EU AI Act's Article 4 — AI literacy — is the obligation almost everyone overlooks, because it doesn't ban anything.