GPTZero: Facebook LLama Leaks 🦙 & the Future

What happens when GPT4 comes out? We shed some light with llama.

Core Updates:

• We’ve built out GPTZero Pro features for educators, available to use today!

• The internal model for Meta’s version of generative AI leaked last week. Our team is the first to leverage it to improve AI detection.

• GPTZero’s chrome extension is launching! Let us know if you’re interested

Educator Plan

We’ve received countless requests from teachers for new features and access to more runs beyond our rate limit. Now we’re releasing a plan with sky-high limits, that allows you to run up to 1 million words per month through GPTZero + premium access to our new fine-tuned models for education.

We anticipate that our free plan will still satisfy the needs of the majority of educators and users, but are incredible excited to now support our super users with additional features.

To sum up, our classic/original plan hasn’t changed at all. It still has the same limits and features we’ve been providing the entire past month. In addition, we’re committed to maintaining the classic version of GPTZero for all users free forever.

Llama Leaks 🦙💦

On Thursday, Facebook’s version of generative AI, Llama, had its model weights leaked publicly.

Llama, a model in between GPT3¹ and unreleased future language models in maturity and performance, helps us anticipate the question, what happens when new models like GPT4 comes out?

Our research team was keen to shed some answers, quickly downloading LLama’s model weights and generating large swaths of Llama text over the weekend.

We tested and compared our detection performance on Llama outputs with one other open source method (Roberta-Large) and two other AI detection services. As our ML researcher put it, “we definitely blow the competitors out of the water” in terms of detection accuracy.

Comparison results from testing AI detectors from 10000 and 300 generations as tested by GPTZero ML researchers Jacob Tian and Alex Cui

Even for pre-llama datasets, GPTZero performed far superior to other AI detection competitors, as validated by independent sources, including TechCrunch deeming GPTZero the ‘best and most reliable’ AI detector.

Screenshot of Techcrunch analysis of popular AI detectors

Finally, and excitingly, our team² was able to adapt GPTZero with Llama data to successfully improve performance and accuracy on Llama generations. Whether GPT4 or the next big thing, the Llama leaks has given our team evidence-based confidence that we’ll be able to upkeep detection for language models far into the future.

Open GPTZero 🔍 🌍

Open GPTZero is launching on the chrome store — for us, it’s not just a chrome extension. It’s a new window for educators, students, and anyone browsing the web to interact with an AI polluted internet.

In addition to detecting and highlighting AI, Open GPTZero also will

• detect and flag biases and toxicity in information consumed

• highlight the most original and human portions of writing, flipping our technology around to promote and identify awesome writing and unique thoughts

Interested in becoming a beta tester for Open GPTZero? Please let us know here.

Thanks for following us on everything in the works at GPTZero!

- Edward

1. https://aibusiness.com/meta/meta-s-llama-language-model-outperforms-openai-s-gpt-3

2. Our awesome and growing team :)