Pseudocodes Enhance LLMs, Ridiculously Simple GPT-4 Bypass, Tech Giants Lose Money on AI, & Build Your AI Document Whisperer
Psuedocodes increase LLM accuracy, GPT-4 has a huge blind spot, tech giants are racking up losses with their AI offerings, and a project for building your own web app that speaks to PDFs
Pseudocodes Propel LLMs to New Heights
Inspired by Eric Elliott’s pseudocode presentation at this week’s Prompt Engineering Conference, researchers at IBM published a white paper earlier this year showing the efficacy of pseudo-code prompts in enhancing the performance of pre-trained language models.
By creating a unique dataset of pseudo-code prompts and comparing them against natural language equivalents, researchers observed a substantial performance leap in models like BLOOM and CodeGen.
They also found that adding code comments and doscstrings improve performance on top of pseudo-codes.
This white paper was the first that demonstrated how pseudo-code prompts can improve large language model performance, and Eric Elliott’s presentation took it to new levels this week. Keep an eye on the Prompt Engineering Conference’s YouTube channel for the recording.
GPT-4's Flaw: Low-Resource Languages Bypass AI Safety Measures
A team from Brown University uncovered a significant flaw in GPT-4's safety protocols, demonstrating that the system's safeguards can be bypassed using low-resource languages—those for which the AI has undergone minimal or no safety training. Their diagram shows just how simple this workaround is.
The most effective languages were Zulu and Scots Gaelic, their full list and efficacy is below. Simply use a translation API, or Google Translate for ChatGPT, and you’ll have a 79% success rate of bypassing OpenAI’s guard rails.
Tech Giants Navigate the High Costs of Cutting-Edge AI
Big Tech is grappling with the high costs and complex economics of cutting-edge artificial intelligence (AI). Companies like Microsoft, Google, and Adobe are pioneering generative AI tools, which, while innovative, are proving costly due to their intensive computational requirements.
Microsoft's GitHub Copilot, which costs $20 per user per month, is averaging a $20 loss per user per month.
Given the losses, you might wonder why consumption pricing isn’t being used, which is how newer tech like Snowflake & Twilio price.
Perhaps their buyers already pay per seat out of habit, and big tech would rather remove the friction so they can prevent competitors from getting a foothold. If they can’t lower the operating costs, big tech can always change the pricing after they win the market.
Speak to Your PDFs: A Guide for Creating an App Using LLMs
Technologist Damian Gil shared a project that allows users to create a web app that interacts with and poses questions directly to PDFs, TXTs, and web pages. The key technologies used for this are LangChain, Hugging Face, Streamlit, and Python.
Designed with a user-friendly interface, it doesn't require much technical expertise, making document interrogation accessible to professionals in various fields. His detailed guide includes direct access to the necessary code for anybody who would like to build this themselves.