AI Newsletter #100 - Adversarial prompting exposes API-integrated LLM vulnerabilities

Welcome to Nural's newsletter focusing on how AI is being used to tackle global grand challenges.

Packed inside we have

Meta open sources LLaMA - 10x smaller than GPT-3 with improved performance
Adversarial prompting exposes API-integrated LLM vulnerabilities
and ChatGPT allowed to be used in International Baccalaureate student's essays

If you would like to support our continued work from £1 then click here!

Marcel Hedman

Key Recent Developments

ChatGPT allowed in International Baccalaureate essays

What: The international Baccalaureate (IB), has allowed Schoolchildren to quote from content created by ChatGPT in their essays.

Matt Glanville, the IB’s head of assessment principles and practice, said "The clear line between using ChatGPT and providing original work is exactly the same as using ideas taken from other people or the internet. As with any quote or material adapted from another source, it must be credited in the body of the text and appropriately referenced in the bibliography"

He added essay writing would feature less prominently in the qualifications process in the future because of the rise of chatbot technology

Key Takeaway: Since the release of ChatGPT, there has been significant concern on the tools ability to faciliate cheating by students. Given the tool's power to create fully formed essays. While some see the International Baccalaureate's decision to encourage the tool's use as a progressive step in adapting to technological advancements, others worry about the lack of standardization in the assessment process and the potential long-term impact on the evaluation of students.

ChatGPT as a marketing tool - the non obvious

What: OpenCage, an API for geocoding, recently found it had an influx of users who were quickly unhappy and complaining. The source... ChatGPT, which had recommended it to users that the service "offer an API to turn a mobile phone number into the location of the phone"

They do not!

Key takeaway: This is a case where the training data held incorrrect information and began blasting it as truth at scale. However, beyond the obvious questions around how to mitigate factual errors (& hallucinations) in LLMs with millions of users. Another interesting concept is the idea of ChatGPT as a marketing engine based on what's contained in its training data. Will we see a world where companies pay for certain "facts" to be included and overweighted in a training corpus?

EU's AI Act faces delay with lawmakers deadlocked after crunch meeting

AI Ethics

🚀 Facebook "opensource" LLaMA LLM

Today we release LLaMA, 4 foundation models ranging from 7B to 65B parameters.
LLaMA-13B outperforms OPT and GPT-3 175B on most benchmarks. LLaMA-65B is competitive with Chinchilla 70B and PaLM 540B.
The weights for all models are open and available at https://t.co/q51f2oPZlE
1/n pic.twitter.com/DPyJFBfWEq
— Guillaume Lample (@GuillaumeLample) February 24, 2023

🚀 Novel adversarial prompting techniques

Additional prompt engineering guide

🚀 Responsible practices with synthetic media framework

🚀 Elon Musk and Tesla face a fresh lawsuit alleging his self-driving tech is a fraud

🚀 OpenAI founder blog: "Planning for AGI"

Here's a critique

Cool companies found this week

Data privacy

Metomic - Protect sensitive data in SaaS applications through detection of sensitive info, remediation of critical issues and employee real time coaching. Recently raised $20m

Talent

HireLogic - AI-based candidate insights by Listening to any Job Interview

Developers

Nebuly - Plug & play open-source AI modules

Best,

Marcel Hedman
Nural Research Founder
www.nural.cc

If this has been interesting, share it with a friend who will find it equally valuable. If you are not already a subscriber, then subscribe here.

If you are enjoying this content and would like to support the work financially then you can amend your plan here from £1/month!