[AINews] Nothing much happened today • ButtondownTwitterTwitter

This section provides recaps of discussions and updates from AI-related Twitter and Reddit channels. The AI Twitter Recap includes updates on AI models like the Yi model on GitHub and discussions on cognitive computing. On the other hand, the AI Reddit Recap covers model releases such as Meta's Chameleon models, Salesforce's xLAM-1b, and Anole's text-image-video generation, as well as use cases like Xiaomi's automated factory and AI agents collaborating in Minecraft. These recaps offer insights into recent developments in the AI community across various platforms.

New Language Model Releases

• Ghost 8B Beta Debuts with Multilingual Prowess: The Ghost 8B Beta large language model promises robust multilingual capabilities and cost-efficiency, available in 8k and 128k versions, with comprehensive documentation detailing its architecture and techniques. Excitement surrounds the model's debut, though some express concern over its knowledge capabilities compared to more specialized models.

• Anole: First Open-Source Autoregressive LMM: Anole is introduced as the first open-source, autoregressive native Large Multimodal Model (LMM), built on Chameleon by @AIatMeta and promising multimodal generation capabilities. However, efforts to fine-tune Anole to reintroduce image capabilities removed from Chameleon have faced backlash, with concerns over undoing explicit design choices.

AI Model Benchmarking and Evaluation

• Rapid Theorem Proving Progress Showcased: HarmonicMath announced achieving a remarkable 90% state-of-the-art on the challenging MiniF2F benchmark, a significant leap from their 83% result just a month prior, as shared in their update. The AI community lauded the blistering pace of progress in theorem proving, considering the benchmark's simpler version stood at only 50% earlier this year.

• Scrutinizing VLM Performance on Basic Tasks: A new paper highlights state-of-the-art Vision Language Models (VLMs) like GPT-4o and Gemini 1.5 Pro struggling with rudimentary visual tasks such as identifying overlapping shapes and object counting, despite high scores on conventional benchmarks. The findings raise concerns about the real-world applicability of VLMs and question the validity of existing evaluation metrics.

Synthetic Data Generation and Feedback Loops

• Preventing Model Collapse with Reinforced Synthetic Data: New research explores using feedback on synthesized data to prevent model collapse in large language models, as detailed in this paper. The study illustrates how naïve synthetic data usage leads to performance degradation, advocating for feedback-augmented synthesized data to maintain high performance on practical tasks like matrix eigenvalue computation and news summarization.

• Exponential Integrator Accelerates Diffusion Sampling: A member sought clarification on the term "marginal distributions as p̂∗_t" from the paper FAST SAMPLING OF DIFFUSION MODELS WITH EXPONENTIAL INTEGRATOR, which proposes a method to accelerate the notoriously slow sampling process of diffusion models. The paper's approach promises to enhance the sampling efficiency of diffusion models while preserving their capability to generate high-fidelity samples across various generative modeling tasks.

Hackathon Hoopla: AGI's Weekend Code Rally

A hackathon is being hosted by AGI House this Saturday 7/13, featuring collaborations with @togethercompute, @SambaNovaAI, and others, with a call for participants to apply here.

Llama-Agents recently launched has already surpassed 1100 stars on GitHub, with @MervinPraison providing a thorough walkthrough available on this link.

Challenges in setting up Stable Diffusion on macOS sparked dialogues, with recommendations for Python file solutions. Enthusiasts unraveled that Adetailer sidesteps VAE encoding for direct full-resolution outputs. A community-contributed guide simplified the setup process for Stable Diffusion, hinting at operational costs. Curiosity flared around running Stable Diffusion on AMD GPUs, with the AMD RX6800 taking the spotlight. The high-resolution fix button showed enhancements in skin textures and facial characteristics.

• Google TPUs are now available on Hugging Face for building, training, and deploying Generative AI models.
• Dataset filtering feature is added to filter almost 200,000 datasets by modality, size, and format.
• Gemini Nano can now run in the browser using Chrome's window.ai feature.
• Rust-Based Inference is now possible with Kyutai Labs' Candle framework, supporting CPU, CUDA, and Metal for inference.
• New depth estimation models, Depth Anything v2, and ZoeDepth, are released providing relative and absolute distance estimations.

The HuggingFace Discord server is buzzing with discussions and updates in various channels. From debates on pursuing Data Science vs. Machine Learning, issues with Gemma model usage, to tips on debugging and optimizing Python code snippets, the community is actively engaging and sharing valuable insights. New developments like the release of qdurllm demo, exploration of early exit in LLM with Branchy-phi-2, and a live workshop on knowledge graphs are also being discussed. Additionally, exciting news about AMD's acquisition of Finnish AI start-up Silo AI for $665mn to enhance their AI services and compete with Nvidia is making headlines in Europe.

In the latest developments, AMD is set to buy Finnish start-up Silo AI for $665mn, marking the largest all-cash acquisition in Europe in a decade. Additionally, discussions in the GPU development community include topics such as comparing the performance of GPUs like 3090 and 4090 for AI tasks, issues related to electricity costs in Australia due to high-power GPU setups, and challenges in building multi-GPU setups with considerations for case compatibility and power supply requirements.

Modular highlights the challenges enterprises face in deploying PyTorch models in production despite its popularity for development and research. The flexibility and ease-of-use of PyTorch in development can lead to complications like resource management and latency issues in full-scale production settings. Modular also addresses the difficulties in creating streamlined AI development workflows that are both locally manageable and scalable for cloud deployment. Developers often face fragmented AI tooling that complicates the end-to-end workflow for effective AI development and deployment. Moreover, Modular encourages enterprises to adopt and integrate AI to enhance productivity and maintain a competitive edge in their services. According to a Bain & Company survey, 87% of companies are already developing, piloting, or deploying generative AI, mainly in software development, customer service, marketing, and product differentiation.

• Brain Size Not Sole Indicator of Intelligence: Brain size is only part of the picture - structure and neuronal density are important within clades.
  • In birds and lizards, density of all neuron types matters more, unless you dig in and differentiate by structures, but data is sparse.
• Cortical Neuron Count Maps Intelligence in Mammals: In mammals, overall cortical neuron count gives a reliable map of the intelligence distribution, owing to the similar brain structure across species.
• Link Between Bigger Brains and Intelligence is Complicated: Discussion on whether bigger brains imply higher intelligence highlighted that bigger animals have bigger brains mainly to control their larger body size.
  • One point raised was that there are uncomfortable ideas around genetics and IQ, particularly concerning human intelligence.

Perplexity teams up with AWS:

Perplexity announced a collaboration with Amazon Web Services to bring Perplexity Enterprise Pro to all AWS customers. More details can be found in the official announcement.

AWS Marketplace Enhances Offerings:

AWS Marketplace expands its catalog with the inclusion of Perplexity Enterprise Pro for its customers. This move aims to provide enhanced AI capabilities to businesses leveraging AWS services.

Misunderstandings about Gemini 1.5 and Claude 3 Pricing:

A user criticized an AI for inaccurately stating that Claude 3 Haiku is considerably cheaper than Gemini 1.5 Flash, pointing out it's actually Gemini 1.5 Flash that's slightly cheaper. Further confusion arose when the AI compared Haiku with Gemini 1.5 Pro instead, despite them being different models entirely.

Perplexity AI's Confusing Image Generation:

A new user was confused about image generation on Perplexity's web and mobile platforms, noting inconsistent capabilities and unclear instructions. Other users explained that while image generation is possible on the web, it is more complicated and limited on mobile devices.

Context Window Limits on LLM Responses:

A user highlighted that LLMs tend to stop generating code after a certain number of lines, necessitating multiple segments for long outputs. Another member explained that this is to prevent excessive token consumption, impacting both usability and cost.

Perplexity AI Lacks Comprehensive Pharmacy Pricing:

A pharmacist noted that Perplexity AI does not initially include CostPlusDrugs.com in its search results for drug prices. While manually prompting the tool to include CostPlusDrugs.com works, the user hopes Perplexity AI will include this by default in the future.

Anticipation for Claude 3.5 Opus:

Users inquired about the release timeline for Claude 3.5 Opus, expressing confusion about its existence. Another member clarified that while Anthropic has announced its upcoming release, a specific date has not been given yet.

A member inquired about LangChain's ConversationSummaryMemory and its capability to support multiple humans for summarizing conversations efficiently. Another member detailed a use case involving the creation of Agents and subagents in LangGraph. A member noted recurring retrieval issues with Chroma. A user described integrating sequential algorithms in a chatbot within the LangChain framework. Lastly, a member asked about adding costs to LangSmith using the traceable decorator for gemini-1.5-flash model via httpx calls.

Aiman1993 hosted a live online workshop on Knowledge Graphs using Video Game Sales as a case study for RAG. The workshop heavily utilized the Langchain library, and Aiman1993 sought feedback on its content. A participant found Aiman1993's workshop helpful and inquired about additional LangChain use cases.