[AINews] Not much happened piday • ButtondownTwitterTwitter

The section provides insights into recent developments in the AI field, including advancements in AI agents and environments, large language models and scaling, AI coding assistants and software engineering, AI safety and regulation, memes and humor, as well as other notable topics. It covers a variety of discussions, from the EU AI Act approval to the release of new AI models like Cerebras CS-3, Anthropic's Claude 3 Haiku, and Cohere's Command-R model. Additionally, it discusses AI safety vulnerabilities, advances in multimodal AI and interpretability, synthetic data generation for LLM pretraining, and fine-tuning, among other interesting topics. The section also highlights debates on decentralized AI model platforms and the introduction of new projects like Shoggoth.

The guild on Perplexity AI's Discord engages in vigorous debates comparing AI models like GPT-4 and Mistral. Users discuss the integration of Perplexity's API for various applications, noting limitations such as a 25MB upload limit. There are inquiries about the API's closed beta features like URL citations and discussions on models for real-time data access. Additionally, users explore running LM Studio's services outside the UI, workarounds for LM Studio limitations, model extensions, ROCm support challenges, and debates on hardware choices and AI startup investments.

"ComPromptMized" Exposes GenAI Weaknesses:

A new study titled "ComPromptMized: Unleashing Zero-click Worms that Target GenAI-Powered Applications" reveals prompt injection attacks on several AI models including Gemini Pro, ChatGPT 4.0, and LLaVA. The paper delves into susceptibilities in GenAI-powered applications, particularly in email assistants. Read the full paper

Quest for Code Assistant Supremacy:

A member is on the lookout for a comprehensive framework to measure and compare the efficacy of AI models like Mistral or L

Meet Devin, the Autonomous Code Whiz:

Cognition introduces Devin, touted as the world's first fully autonomous AI software engineer, capable of handling complex tasks and learning from its experiences as per Scott Wu's blog.

• Challenging AI's Social Skills: Participants are encouraged to showcase their creativity in the "The Most Interesting Bot In the World Contest" at the Voice + AI event. Contest details are available on the event’s Notion page.

Nous Research AI - Hermes 2.5 surpasses Hermes 2 with new context models and integrated features. Distinction made between function calling and JSON mode functionalities. Discussions on new models like Hermes 2 Pro and Genstruct 7B, clarifying JSON mode requirements.### Nous Research AI - TAO vs Hugging Face debate and discussions on centralized vs decentralized benchmarking, impact of crypto incentives, competition hindering collaboration, and the need for decentralized benchmarking.### Perplexity AI - New offerings like Claude 3 Haiku, local search enhancements, and use cases across coding and image recognition. Discussions on AI model comparisons, data uploads, and voice recognition features.### Perplexity AI - Sharing on topics like AI news controversies, medical information provision, image description capabilities, and script innovations for LM Studio.### LM Studio - Conversations on LM Studio features, API capabilities, deployment options, and impacts on employment.### LM Studio - Discussions on expanding model capabilities to 128k tokens for longer context comprehension.

The section discusses various topics related to hardware discussions within the LM Studio community. It includes discussions on expensive Nvidia links, VRAM hurdles on Mac OS, PC hardware upgrades, running multiple models in LM Studio, and monitor selection for high-end gaming and productivity. The conversations cover issues like compatibility, performance optimization, and user experiences with different hardware setups. Additionally, links to articles discussing advancements in GPU technology and AI chips are mentioned.

Synthetic Data for Finetuning Survey Presentation

• Reminder for the presentation on Synthetic Data for Finetuning at 12pm PT.
• Synthetic data is highlighted as a faster, cheaper, and often better-quality alternative to human annotations for pretraining and fine-tuning models.

Urgent Luma Invite for Paper Club Event

• Members urged to accept Luma invite for calendar reminders.

Corrected Synthetic Data Link Provided

• Include the corrected link to the survey on synthetic data.

New Episode with Suno AI Released

• Announcement of a new podcast episode featuring Suno AI.
• Includes a link to the Twitter announcement and a YouTube video titled 'Making Transformers Sing - with Mikey Shulman of Suno'.

Seeking Token Probability Visualization

• Inquiry about visualizing token probability in a sentence.
• Suggestions on using lm_head's output and softmax.

Rapid AI Evolution

• Discussions on upcoming releases like Elon Musk's open Grok model.
• Speculations and rumors circulating within the community.

Eleuther - AI Interpretability

• Frustration is expressed over limited access to research due to publisher restrictions, prompting a member to share links to available papers.
• Discussions focus on the training dynamics of neural networks and the potential of combining multiple architectures for problem-solving.
• Members engage in debates regarding the efficacy of AI content detectors, concerns over watermarking AI outputs, and challenges in model agnosticism.
• An interest in evaluating large language models in competitive gaming environments is discussed, with doubts raised about AI's current ability to compete at human levels in certain games.
• Members contemplate strategies for model integration and performance improvements, as well as adaptations for generative models and logit handling in the LM evaluation feature.

Visualize LLM Leaderboard with Ease:

• The Open LLM Leaderboard Viz update allows users to change metrics order and plot up to 3 models for easy visual comparison.

Storytelling Gets Visual with GPT:

• A new space called Kosmos-2 by Tonic1 brings GPT-based visual storytelling to users.

ARC Dataset Augmented with Reasoning:

• Augmented ARC-Challenge Dataset now incorporates Chain-of-Thought reasoning, offering more depth in answers to common questions.

Python Package for Vertex AI Inference:

• A new Python package, vertex-ai-huggingface-inference, is available to streamline running HuggingFace models on Google Cloud's Vertex AI.

Rich Portuguese Pretrained Model Debuts:

• Introducing Mambarim-110M, a Portuguese LLM with over 119 million parameters trained on a 6.2B token dataset.

HuggingFace Reading Group Highlights

• No Show This Week: This week's reading group session was cancelled, with a presentation planned for the following week.
• MNIST Digit Classification Question: A member asked about the number of units in the first layer for an MNIST digit classification given 20x20 pixel images from Andrew Ng's neural network course.
• Exploring Neural Network Architecture: Members discussed determining the number of neurons and hidden layers in neural networks, emphasizing the need for experimentation and leveraging past successful configurations balancing processing power, speed, and accuracy.

In this section, various discussions on model training and selection were highlighted. The considerations ranged from the choice of open-source chat-ready models like Mistral and Mixtral to the challenges of training large models like PyTorch's Metal Performance Shaders (MPS) backend. Participants also debated between Mixtral and Qwen 70B for medical training and discussed best practices for training formats. Additionally, there were suggestions to enhance performance by integrating NVIDIA's GPUDirect® Storage technology into the Axolotl system. The section also included links to relevant resources such as PyTorch's Metal Performance Shaders (MPS) backend, a paper on enabling and accelerating model fine-tuning on a single GPU, and the cuFile API Reference Guide from NVIDIA.

Interconnects (Nathan Lambert)

This section features various discussions and exchanges among members in different channels related to topics such as language applications, collaborations, LLMs, model performance, and more. Members engage in conversations about polyglot projects in European academia, the support of German language by LLMs, and the performance of Aleph. Additionally, discussions touch on GPT-4's dominance, model inquiries, safety filtering, Claude models, the challenges in AI literature survey assistants, and more. Links are shared regarding tweets, updates on models, and further details on discussed topics.

This section provides insights into different discussions happening in various AI-related Discord channels. From inquiries about Munich meetups, AI events in Berlin, fine-tuning language models, seeking benchmarks for German embeddings, successful creative writing benchmark tests, to debates on visual document processing tools like Haiku and content filtering challenges. Additionally, announcements of a zero-click worm paper targeting GenAI-powered apps, a voice + AI event bot contest, and the unveiling of Devin, claimed to be the world's first fully autonomous AI software engineer, are highlighted.