GPT Store is Live, Persuasive LLM Jailbreaks, Mistral's Router Network, & ChatGPT SKU for Small Teams
The GPT Store is officially live, researchers found that persuasive prompts can jailbreak advanced LLMs, Mistral details their router network in a new white paper, and ChatGPT has a new SKU
OpenAI Launches GPT Store
We mentioned this was likely coming last week, and it's now official. OpenAI officially launched the GPT Store, a platform for sharing user-created chatbots, for paying users
Over 3 million custom GPTs have been developed since the GPT Builder program was introduced in November, but the feedback online has been fairly muted.
OpenAI stated that US builders will be paid based on user engagement with their GPTs, and will share specifics later in Q1.
Advanced LLMs Are More Vulnerable to Persuasive Prompt Jailbreaks
Researchers have developed a method to jailbreak language models using Persuasive Adversarial Prompts (PAPs), which humanize LLMs through persuasive communication techniques. This method, tested on models like GPT-4, achieved a 92% success rate in eliciting harmful outputs, without needing specialized optimization.
Interestingly, they found that more advanced models like GPT-4 were more vulnerable to these prompts, possibly because advanced models can grasp context better. The researchers also noted the effectiveness of adaptive defenses against these attacks, underscoring the importance of continually improving AI safety measures.
Mistral Unveils Cutting-Edge Router Network in Mixtral White Paper
Mistral AI's Mixtral 8x7B model has been out for about a month now, and Mixtral's white paper was released this week. The big takeaway is their router network, which selects two experts that process the current state and combine their outputs.
Even though each token only sees two experts, the selected experts can be different at each timestep. As a result, each token has access to 47B parameters, but only uses 13B active parameters during inference.
Mixtral was trained with a context size of 32k tokens, and it outperforms or matches Llama 2 70B and GPT-3.5 across all evaluated benchmarks. In particular, Mixtral vastly outperforms Llama 2 70B on mathematics, code generation, and multilingual benchmarks.
OpenAI Launches ChatGPT Team Subscription for Small Teams
OpenAI also introduced ChatGPT Team this week, a new SKU specifically designed for small to medium-sized teams, offering enhanced collaboration and management tools.
The enhanced collaboration and management tools are broken down in their pricing.
OpenAI is positioning ChatGPT Team as an affordable alternative to the higher-tier 'ChatGPT Enterprise' plan, hoping this caters to the growing demand for AI-powered tools for various applications such as coding, content creation, and document summarization.