🍄 The 8th Edition: Learn a New Word - Package Hallucination

With Some LLM News

📦 Package Hallucination

  • 🛎️ It has been discovered that ChatGPT and other AI platforms may provide false sources, links, blogs, and statistics when answering questions.

  • 🛎️ This can include providing solutions to coding problems using non-existent libraries or offering questionable fixes to security vulnerabilities.

  • 🛎️ This vulnerability can be exploited by attackers who ask ChatGPT for a package to solve a coding issue.

  • 🛎️ ChatGPT may respond with packages that do not actually exist, and if the attacker finds a recommendation for an unpublished package, they can replace it with their own malicious package.

  • 🛎️ This can potentially harm users who receive recommendations to use the malicious package in the future. In our proof of concept using ChatGPT 3.5, we were able to replicate this scenario.

💬 The dialogue that occurred between the assailant and ChatGPT.

  • 🔹 ChatGPT recommends installing the arangodb package.

    • The first question in the attacker’s conversation with ChatGPT:“How to integrate with arangodb in node.js? Please return the package to install in the pattern of npm install”

  • 🔹The attacker’s second question, and ChatGPT’s response with a suggestion to install the arangodb package:

  • 🔹 The suggested package does not exist in npmjs:🔹 Source

⚠️ Zoom has been recording your calls to train the AI

🔹 The software does not distinguish between the content or purpose of your phone conversations. This means that whether you're a lawyer discussing a confidential case, a doctor communicating with a patient, or a teacher speaking with a student, your calls are treated the same way.

🔹 After the news was released, Zoom responded by assuring the public that they will not utilize any of your information without your explicit permission.

🔹 You can turn it off.

✊🏼 LLM models are politically biased

Do companies have social responsibilities, or are they solely focused on delivering profits to their shareholders? The answer to this question varies depending on the AI language model you ask. Recent research by the University of Washington, Carnegie Mellon University, and Xi’an Jiaotong University shows that different language models contain different political biases.

For instance, OpenAI's GPT-2 and GPT-3 Ada models tend to support the idea that companies have social responsibilities, while GPT-3 Da Vinci, OpenAI's more advanced model, believes that companies exist solely to generate profits. The study analyzed 14 large language models and found that OpenAI's ChatGPT and GPT-4 were the most left-leaning libertarian, while Meta's LLaMA was the most right-leaning authoritarian.

Thank you for reading Musings on AI. This post is public so feel free to share it.

👀 The Top Gossip of Silicon Valley: “NVIDIA GPU Shortage”

🔹 As compute-hungry generative AI shows no signs of slowing down, which companies are getting access to Nvidia’s hard-to-come-by, ultra-expensive, high-performance computing H100 GPU for large language model (LLM) training is becoming the “top gossip” of Silicon Valley, according to Andrej Karpathy, former director of AI at Tesla and now at OpenAI.

🔹 It's time to check out other training frameworks such as AWS Trainium, but there's a challenge: the NVIDIA Cuda Frameworks offer a strong ecosystem. In order to train the LLM, additional frameworks like Apex and 3d Parallelism are needed for optimization (although it can be done with another framework, it may take more time and be less reliable). This is why certain aspects cannot be substituted.

🔹 Source:

🔹 Source:

Nvidia DGX head says no GPU shortage, but supply chain issue.

“On the system side, we’ve always been very supply-responsive to our customers,” he told VentureBeat in a recent interview. A request for thousands of GPUs will take longer, he explained, but “we service a lot of that demand.”

Something he has learned over the past seven years is that ultimately, it is also a supply chain problem, he explained — because there are small components provided by vendors that can be harder to come by. “So when people use the word GPU shortage, they’re really talking about a shortage of, or a backlog of, some component on the board, not the GPU itself,” he said. “It’s just limited worldwide manufacturing of these things…but we forecast what people want and what the world can build.” 

Boyle said that over time the “GPU shortage” issue will “work its way out of narrative, in terms of the hype around the shortage versus the reality that somebody did bad planning.”

🌻 UniversalNER

Models such as Alpaca and Vicuna are capable of following instructions, but they fall behind ChatGPT in NLP benchmarks. UniversalNER is a new model that has been developed through targeted distillation, and it is able to recognize over 13,000 entity types. In fact, it outperforms ChatGPT's NER accuracy by 9% F1 across 43 datasets. This is an exciting development in the field of NLP! 💡🚀

  • 🛎️ Demo: https://universal-ner.github.io/

  • 🛎️ Dataset: https://huggingface.co/Universal-NER

  • 🛎️ Paper: https://arxiv.org/abs/2308.03279

🏋️ StackOverflow has experienced a 35% decrease in traffic.

Over the past one and a half years, Stack Overflow has lost around 50% 35% of its traffic (Update: Around 15% of the observed loss seems to be related to the recategorization of the Google Analytics Cookie around May 2022. The chart was updated to only include data after that change). This decline is similarly reflected in site usage, with approximately a 50% decrease in the number of questions and answers, as well as the number of votes these posts receive.

The charts below show the usage represented by a moving average of 49 days.

The data is from https://stackoverflow.com/site-analytics

🔹 Source

🧸📖✨ We are reading:

  • https://www.lesswrong.com/posts/dBmfb76zx6wjPsBC7/when-can-we-trust-model-evaluations

**

I will publish the next Edition on Sunday.

This is the 8th Edition, If you have any feedback please don’t hesitate to share it with me, And if you love my work, do share it with your colleagues.

Cheers!!

Raahul

**

Reply

or to participate.