> AI Discords for 2/26/2024. We checked **22** guilds, **349** channels, and **12885** messages for you. Estimated reading time saved (at 200wpm): **1063 minutes**.

Not much happened today, so it’s a nice occasion to introduce 2 new Discords that have passed our quality bar: Interconnects (run by Nathan Lambert who we recently had on Latent Space) and OpenRouter (Alex Atallah who will surely join us at some point).

image.png

Table of Contents

[TOC]

PART 0: Summary of Summaries of Summaries

Model Comparisons and Optimizations: Discord users actively engaged in discussions about the performance and optimization of various AI models, including Mistral AI, Miqu, and GGUF quantized models. Key topics included the comparison of Mistral AI's new Mistral Large model with OpenAI's GPT-4, highlighting its cost-effectiveness and performance. Users also explored efficient loading practices and model quantization methods, such as GPTQ and QLORA, to enhance model performance and reduce VRAM usage​​​​​​.

Advanced Features and Applications: There was significant interest in leveraging AI for specific applications like role-playing and story-writing, emphasizing the need for models to manage consistent timelines and character emotions. Moreover, the potential of AI in code clarity and AI-assisted decompilation was discussed, with a particular focus on developing tools like an asynchronous summarization script for Mistral 7b, indicating a push towards making AI more accessible and practical for developers​​​​.

Quantum Computing and AI's Future: Discussions in various Discords touched upon the intersection of quantum computing and AI, speculating on how quantum advancements could revolutionize AI model processing. The discourse extended to the implications of DARPA-funded projects on AI development and the exploration of encoder-based diffusion techniques for stable diffusion in image processing, reflecting a deep interest in cutting-edge technologies that could shape the future of AI​​​​.

Community and Collaboration Initiatives: Various communities highlighted efforts to foster collaboration and share knowledge on AI model development, ranging from new Spanish LLM announcements to hardware experimentation for model efficiency. Platforms like Perplexity AI and LlamaIndex were noted for their innovative features and integration capabilities. There was a strong emphasis on open-source projects, as seen in discussions about Mistral's open-source commitment and the development of tools like R2R for rapid RAG system deployment, showcasing the vibrant collaborative spirit within the AI community​​​​​​.

PART 1: High level Discord summaries

TheBloke Discord Summary

  • Polymind’s Haunting Echoes: Engineers are troubleshooting Polymind’s UI bug where ghost messages linger until new tokens stream. Debugging involves frontend HTML variables, like currentmsg, to clear residual content.

  • Model Assessment Mania: Discussions highlight Miqu, a leaked Mistral medium equivalent, performing admirably despite being slower due to higher parameters compared to Mixtral. Users are also actively exchanging efficient loading practices for GGUF quantized models, including tweaks like disabling shared video memory and GPU settings to prevent VRAM overflow.

  • LangChain’s Loop: LangChain is under fire for its claimed context reordering feature, which @capt.genius labeled a “fake implementation”, and users suggest it might just be a basic looping mechanism.

  • Mistral’s Mysterious Moves: There’s a buzz around Mistral AI’s change in language on their website regarding open-source models, fueling speculation about the company’s commitment to open models, in light of models from other AI providers like Qwen.

  • Optimizing Prompt Engineering: Engineers in the roleplay and story-writing channels express a desire for models adept at role-playing and story-writing with consistent timelines, character emotions, and mood management.

  • Quantizing Questions and Training Troubles: There’s debate around model quantization methods, with @orel1212 considering GPTQ over QLORA for a small dataset training. Questions loom about Mistral’s PDF image text extraction capabilities and the hardware needed for Mixtral 8x7b without clear answers. Meanwhile, @dzgxxamine seeks advice on teaching LLMs to understand and use newly released Python libraries.

  • Merge Meltdown Mystery: A single message from @jsarnecki reports an unsuccessful model merge between Orca-2-13b and WhiteRabbitNeo-13b, resulting in incomprehensible output, with details of the merge process shared via a readme from mergekit.

  • AI’s Role in Code Clarity: Within the coding discussions, there’s excitement about AI’s potential in AI-assisted decompilation, as @mrjackspade anticipates a future without manual code reconstruction. Additionally, @wolfsauge introduces an asynchronous summarization script tested on Mistral 7b, with an invite for peers to experiment with it, available on GitHub.

LM Studio Discord Summary

  • New Spanish LLM Announced: Spain’s President Pedro Sánchez announced the creation of a Spanish-language AI model at Mobile World Congress; it’s significant for Spanish language applications in AI.

  • Hardware Hijinks: RTX 2060 Action: Discussions in the hardware channel have users experimenting with multi-GPU setups without a bridge connector and seeing success, even with mismatched graphics cards, driving efficiency for models as large as 70b, pointing to changing norms in hardware compatibility.

  • WSL Networking Fix for AI Tools: Users resolved an issue connecting to a LM Studio endpoint from within WSL (Windows Subsystem for Linux) on Windows 11 by bypassing local host and using the network IP (http://192.168.x.y:5000/v1/completions), addressing unique WSL network behavior outlined in this guide.

  • Tech for Text-to-Talk: Piper, a neural text-to-speech system, has been highlighted for its local processing capabilities and support for different languages, indicating a trend towards efficient, offline solutions for language model applications.

  • Techies Taming Language Models: In the model discussions, users highlighted the importance of specific quantization parameters like mirostat_mode and mirostat_tau for optimizing LLM performance, which needs to be manually configured. Conversations also acknowledged the role of human evaluation in assessing models’ performance.

Nous Research AI Discord Summary

  • RWKV/Mamba Models Primed for Context-Length Flexibility: @vatsadev commented on the viability of rwkv/mamba models being extendable through fine-tuning for longer context, mitigating the need to pretrain from scratch. @blackl1ght achieved a 32k token context length on Solar 10.7B using a Tesla V100 and Q8 quantization without hitting memory limits, indicating the potential for extending these models even further.

  • Ring Attention Sparks 1 Million Token Model Feasibility: Discussions highlighted the application of ring attention in models managing up to 1 million tokens, @bloc97 revealed that this could be applied to 7b models, illustrating advancements in attention mechanisms for larger scale models.

  • Quantization Techniques Promise Efficient Inference: Efforts to optimize inference efficiencies involving key-value cache and potential use of fp8 were mentioned; @stefangliga introduced the concept of quantizing the kvcache which could pave the way for handling longer contexts more effectively.

  • Benchmarks Point to Selective Reporting: In a conversation about AI model benchmarks, @orabazes pointed out that a tweet on AI performance omitted certain model comparisons, hinting at the need for transparency in benchmarks.

  • Mistral TOS Controversy Settles: After intense debate regarding Mistral’s Terms of Service, a tweet from co-founder @arthurmensch resolved issues by removing a contentious clause, reaffirming the model’s open-use for training competitive LLMs.

  • Structured Data Gets Dedicated LLM Attention: New insights from @_akhaliq’s tweet into Google’s StructLM project indicate a focus on enhancing LLMs’ handling of structured data, despite the absence of shared training approaches or model frameworks.

  • Curiosity in Mistral Loop Dilemmas: @.ben.com theorized that repeated text looping in Mistral could be attributed to classic feedback system issues, indicating a technical curiosity in understanding and resolving model oscillations.

  • High-Demand for Self-Extend Model Practices: The community’s enthusiasm for local model improvements was evident after @blackl1ght’s successful experiments with self-extension, leading to a collective interest in replicating the configurations that allow for improved memory management and feature utilization.

Links mentioned:

OpenAI Discord Summary

EBDI Agent Challenges and Solutions: @.braydie explored EBDI frameworks for agent goal determination, but encountered thinking loops after integrating the ReAct framework. They examined decision-making models from a JASSS paper to address the issue.

Mistral Steps Up to Rival GPT-4: A TechCrunch article reported that Mistral Large, a new model from Mistral AI, is positioned to compete with OpenAI’s GPT-4, offering cost-effectiveness and uncensored content, and is now available on Azure.

Prompt Protection Paradox: Users deliberated on how to protect intellectual property in prompts, concluding that while copyright might cover exact wording, the replication of ideas via linguistic variation is likely unstoppable.

Text Classification Tactics: @crifat kicked off a discussion on text classification methods, opting to start with the base model and Assistant, bypassing fine-tuning, to sort texts into categories such as “Factual” and “Misleading.”

Meta-Prompting Generates Buzz and Security Concerns: The concept of meta-prompting was a hot topic, with claims of generating extensive documentation from advanced techniques, but these techniques also raised security flags when a user shared a PDF, resulting in the user’s account action.

Perplexity AI Discord Summary

  • Perplexity AI’s Picture Perfect Potential: Users discussed the image generation capability of Pro version of Perplexity AI, with a link given to a Reddit discussion on the topic. Concerns were also voiced over the discontinuation of Gemini AI support in Perplexity, with confirmations that the support has ended.

  • Mistral Large Steals the Spotlight: A new language model from Mistral AI, called Mistral Large, was introduced with discussions surrounding its capabilities and availability. The announcement was supported by a link to Mistral AI’s news post.

  • VPN Troubles Take a Detour: A login issue on Galaxy phones with Perplexity AI was tackled by turning off VPN, with additional resources on VPN split tunneling provided.

  • Exploratory Searches in Perplexity Shine Through: Users shared links to Perplexity AI’s search results on novel topics, including a transparent laptop by Lenovo and an age query for the letter K. Comparisons of iPhones were also made, and a collection feature was promoted, inviting users to create their own collection on Perplexity AI.

  • PPLLX-API Channel Buzzes with Model Concerns: Technical discussions centered on model information and performance, particularly about a JSON link for model parameters and the behavior of the sonar-medium-online model generating irrelevant content. There was deliberation over the deprecation of the pplx-70b-online model and the vital role of proper prompting when making API calls, with explicit referencing to Chat Completions in the API documentation.

LlamaIndex Discord Summary

  • PDF Parsing Leaps Forward with LlamaParse: LlamaParse is introduced as a tool to enhance understanding of PDFs with tables and figures, aiding LLMs in providing accurate answers by avoiding PDF parsing errors. It’s highlighted as significant for Retriever-Augmented Generation (RAG) processes, as tweeted here.

  • Super-Charging Indexing with MistralAI’s Large Model: LlamaIndex integrates @MistralAI’s Large Model into its build 10.13.post1, bringing near-GPT-4 capabilities including advanced reasoning and JSON output, as mentioned in this announcement. Furthermore, LlamaIndex’s new distributed super-RAG feature allows the creation of API services for any RAG application which can be networked to form a super-RAG capable of running queries across the network, as shared here.

  • AI Assemblage with AGI Builders and FireWorksAI: The AGI Builders meetup will host LlamaIndex’s VP of Developer Relations to share insights on RAG applications, with details of the event available here. In another collaboration, LlamaIndex and FireworksAI_HQ have released a cookbook series for RAG applications and function calling with FireFunction-v1, offering full API compatibility as announced here.

  • Context Crafting Conundrums for Coders: In the ai-discussion channel, members discussed optimizing context for coding LLMs such as GPT-4 turbo and Gemini 1.5 with a focus on the order of information, repetition, and structuring techniques. There was also discourse on open-source text generation with Llama2, emphasizing non-proprietary integration for CSV and PDF inputs, and exploration of chunking and retrieval strategies for building a RAG-oriented OSS SDK assistant.

  • Tech Troubles and Tooling Talks in General: The general channel buzzed with requests for assistance, comparisons between GPT-3.5 with and without RAG, and discussions on integrating LlamaIndex with other services like Weaviate. Moreover, there was an elaborate exchange on resolving installation issues of LlamaIndex on macOS and dialogues about creating agents for Golang integration and dynamic orchestration with AWS Bedrock, suggesting a highly interactive and technically-oriented community.

LAION Discord Summary

  • New Twists in Encoder-Based Diffusion Techniques: Users discussed the merits of an encoder-based inversion method for stable diffusion and the challenges with image prompt (IP) adapters in enhancing stable diffusion performance.

  • AI’s Dicey Dance with DARPA: Debates flared regarding the implications of DARPA-funded projects on AI research and development, with a humorous twist on anime character usage in recruitment efforts and the intersection of military and entertainment genres.

  • Quantum Leaps for AI’s Future: Conversations revolved around the future role of quantum computing in processing AI models, particularly transformers, with discussions on the state of quantum error correction and its potential in AI computations.

  • Navigating the Minefield of Content Moderation: Dialogues dug into the difficulties surrounding content moderation, especially in relation to Child Sexual Abuse Material (CSAM), pondering the effectiveness of reporting tools and responsibilities of platforms.

  • The Balancing Act of Open Source vs. Proprietary AI: There was a spirited exchange over the strategies of releasing AI models, contrasting open-source approaches like Mistral Large with proprietary models and considering the commercialization supporting ongoing AI R&D.

  • Watermarking Language Models: A research paper unveiled findings on the detectability of training data through watermarked text, indicating high confidence in identifying watermarked synthetic instructions used in as little as 5% of training data.

  • Genie Grants Wishes in Humanoid Robotics: A YouTube video from Google DeepMind’s new paper brought to light recent advancements in AI’s role in humanoid robotics.

  • The Finite of FEM Learning Success: Within the realm of Finite Element Analysis (FEM), there was acknowledgment of research yet a consensus that methods for learning with FEM Meshes and Models are not as effective as conventional FEM, citing a specific paper.

  • Torching Through Fourier Transform Challenges: A technical issue involving the inverse discrete Fourier transform in neural network synthesis was highlighted, pointing to code that might benefit from a refactoring using torch.vmap for VRAM efficiency, demonstrated in the shared GitHub repository.

  • Pondering Transformer Learning Experiments: There was a solitary query about conducting transformer learning experiments to discern size relations between fictional objects, aiming to understand and render comparative images, but no further discussion or data was provided.

Interconnects (Nathan Lambert) Discord Summary

  • Paywall Hinders Financial Times Article Sharing: A Financial Times article discussing Microsoft’s stake in French AI startup Mistral was shared by @xeophon., but it was criticized for its poor mobile paywall experience. Microsoft’s investment is said to support the deployment of Azure-based commercial language models.

  • Debating the Efficacy of Chain of Thought (CoT): Reports from users like @sid221134224 and @xeophon. suggest mixed results with CoT prompting, ranging from dropped performance in models like Gemini Pro to being a standard practice in fine-tuning.

  • Multilingual Models Alignment Challenges Discussed: Discussions focused on the alignment of multilingual models, questioning whether the values should be uniform or culturally specific, highlighting issues with GPT-4’s safety measures which can be bypassed via low-resource languages like Scots Gaelic as reported by The Register.

  • Insights on AI Tools and Writing Habits of Contributors: Contributors @natolambert and @xeophon. shared their preferences for writing tools and their processes, such as Notion, Grammarly, Typora, and Obsidian, exemplifying varied approaches to creating written content.

  • Excitement Over DeepMind’s Genie and Mistral’s Business Strategy: Engaging discussions broke out over DeepMind’s new foundation world model, Genie, which shows potential in world dynamics learning from videos; and Mistral’s resource consumption and pricing strategy, with speculation on token usage based on available public information.

OpenRouter (Alex Atallah) Discord Summary

  • Mistral Large Emerges as New AI Contender: A new model called Mistral Large has been introduced by @alexatallah, positioned between GPT-4 and Claude 2 for advanced capabilities with features like a 32,000 token context window and multilingual support, available at OpenRouter’s Mistral Large. Pricing adjustments have been made across the Mistral portfolio, advocating for Mistral 7B Instruct and Mixtral 8x7B Instruct due to affordability.

  • Sonar Soars with Online Capabilities: @alexatallah shared the launch of Perplexity’s Sonar models, including an internet-connected version, that outshine their predecessors in cost-efficiency and speed, located at Perplexity’s Sonar 8x7B Online. With PPLX models set to be deprecated on March 15, users are encouraged to transition to Sonar.

  • OpenRouter Playground Gets a Boost: Important updates in OpenRouter Playground include the addition of new parameters such as Top P, Top K, and penalties, improving user interaction. Message system issues across models including Perplexity, Mistral, and Gemma have been fixed, optimizing performance.

  • New AI Creator Tools Hit the Market: The launch of Videotok on Product Hunt was announced, offering an AI-powered platform for creating short videos, with a shareable post found at Borja Soler’s tweet. Additionally, Blust AI’s platform has surfaced, integrating multiple AI applications, with integration steps detailed in their documentation.

  • Automate All The Things with Make.com: A new app was introduced by @jim14199 enabling no-code AI workflow automations by linking OpenRouter with a multitude of other apps, available at Make.com OpenRouter Integration. Issues surrounding model functionality on the Blust AI platform were reported, though specifics were not provided.

  • OpenRouter Sparks Community Conversation: The community engaged in a lively discussion about new feature updates like the API for parameters, with suggestions to tailor them towards user preferences and production user inputs. There’s a palpable excitement for the release of Mistral Large, coupled with discussions about Perplexity model errors and choosing the best interfaces for leveraging OpenRouter AI models, as seen at OpenRouter.

HuggingFace Discord Summary

Have an Amazing Week and Ace Those Exams: Community members are sharing sentiments ranging from well-wishes for the week to the stress of exams.

Seeking Speedy Batch Processing Solutions: A discussion took place regarding the optimal batching methods for querying GPT-4, emphasizing the importance of fast and efficient batch processing to reduce completion times.

Service Disruptions and Tech Collaborations: Users reported experiencing 504 timeout errors with the Hugging Face Inference API, highlighting service instability; meanwhile, there’s an ongoing dialogue to foster collaborative machine learning project development within the community.

Immersive Study Opportunity in Convolutional Neural Networks: An open invitation was extended for a study group focusing on CS231n, Convolutional Neural Networks for Visual Recognition, with links to course assignments and modules available for interested participants. CS231n Course

Scale AI’s Rise to Prominence and VLM Resolutions: Articles and discussions showcased Scale AI’s impressive growth to a $7.3 billion valuation in data labeling and innovative solutions to overcome resolution problems in vision-language models using multiple crops of high-resolution images. Scale AI’s Story and VLM Resolution Solution

Developments and Debates in AI Ethics and Performance: The community shared opportunities for commenting on “open-weight” AI models, a new Performance LLM Board evaluating response times and pricing of various models, and a detailed replication attempt of the Imagic paper for text-based image editing using diffusion models. Open AI Model Weights Comments and Imagic Paper Replicated

Discontentment with Diffusion Model Tools: Voices of dissatisfaction emerged regarding the use of eps prediction in Playground v2.5, and the choice to utilize the EDM framework instead of zsnr.

Data Size and Character Recognition in Computer Vision: A notable concern was raised about the adequacy of dataset size for fine-tuning, especially for models aimed at complex character recognition, such as those in the Khmer language, which presents unique challenges due to its symbol-rich script.

Navigating the NLP Landscape: Conversations touched on best practices in sequence classification, searching for generative QA models, recommendations for embedding models suited for smaller datasets, strategies for compressing emails for LLMs, and constructing a medical transformer tailored to the nuances of medical terminology. Suggested models for embedding include BAAI’s bge-small-en-v1.5 and thenlper’s gte-small.

Eleuther Discord Summary

  • Unit Style Learning Catches Engineer’s Eye: Precision in machine learning algorithms was the focus with a conversation about ‘unit style’ vectors, emphasizing RMSNorm as a crucial component for achieving ‘unit length’ vectors and avoiding the need for conversion between scales.

  • Mistral Lifts the Veil on a Large Model: EleutherAI announced the release of Mistral Large, a new state-of-the-art language model that promises cutting-edge results on benchmarks. It’s now available through la Plateforme and Azure, which was discussed in detail on Mistral’s news page.

  • Appreciation and Call to Action for lm-harness: A member highlighted the impact of EleutherAI’s lm-harness, with its recognition in a recent paper for its importance in the few-shot evaluation of autoregressive language models, and threads included guides and discussions to enhance its functionality.

  • Interpreting ‘Energy’ Sparks Inquiry and Promises of Investigation: Members expressed confusion and the need to investigate the term “energy” used in the context of model tuning and related equations, admitting a lack of intuition and clarity around the concept.

  • Technical Troubleshooting in NeoX: Users discussed challenges and sought advice for DeepSpeed configurations and setting up multi-node training environments for GPT-NeoX on CoreWeave Infrastructure, with directions pointing towards utilizing Kubernetes and employing slurm or MPI for a 2 node with 4 GPUs setup.

CUDA MODE Discord Summary

  • GTC Excitement and Planning: Members are looking forward to attending the Graphics Technology Conference (GTC), with suggestions for a meetup and a dedicated channel for attendees. A CUDA-based Gameboy emulator on GitHub was shared, showcasing an ingenious use of CUDA for classic gaming emulation.

  • Unleashing QLoRA’s Speed: A GitHub repo for 5X faster and 60% less memory QLoRA finetuning was highlighted for accelerating models efficiently.

  • In-Depth CUDA Discussions: The community engaged in discussions about CUDA interoperability with Vulkan, recounted the enhancements to GPU utilization with TorchFP4Linear layers, and shared updates on selective layer replacement techniques. Insights about GPU memory access latencies were also exchanged, supported by a variety of external references like the examination paper of NVIDIA A100’s memory access and torch-bnb-fp4’s speed test script.

  • Optimizing PyTorch Workflows: Conversations included the origin of PyTorch rooted in Torch7 and tips for speeding up cpp_extension.load_inline compile times. There was guidance pointed for integration of custom Triton kernels with torch.compile in PyTorch, citing a GitHub example.

  • Efficiency in Attention Algorithms: Papers and resources were shared on efficient softmax approximation, the base2 trick for softmax, and an impressive incremental softmax implementation in OpenAI’s Triton example. There was also a nod to the classic fast inverse square root trick from Quake3.

  • NVIDIA Job Opportunity: NVIDIA posted a job opportunity seeking experts in CUDA and C++, referring interested individuals to DM their CV with JobID: JR1968004.

  • Beginner Queries: Discussions surfaced around what is needed to dive into CUDA Mode, including any prerequisites like knowledge of PyTorch, and recommended starting with the fast.ai course before venturing into CUDA.

  • Clarification in Terminology: The acronym AO was clarified to stand for Architecture Optimization.

  • Advancements in CUDA Attention Mechanisms: Tips for improved performance in ring attention were discussed, along with the scheduling of a flash attention paper reading/discussion. A Colab notebook comparing ring and flash attention was shared for community feedback, alongside a paper on implementing FlashAttention-2 on NVIDIA’s Hopper architecture.

LangChain AI Discord Summary

  • Structured Output Interface Seeks Community Insight: An intuitive interface for obtaining structured model outputs was proposed, with a request for feedback by user @bagatur. Details are discussed in a GitHub RFC on langchain-ai/langchain.

  • New Tools and Guides for LangChain Enthusiasts: Various resources dropped, including the launch of UseScraper.com by @dctanner for content scraping, a comprehensive integration guide for LlamaCpp with Python by @ldeth256, and a shoutout to validate.tonic.ai, a platform by @locus_5436 for visualizing RAG system evaluations.

  • Seeking Temporary Workarounds for Function Calls in Chats: @sectorix inquires about temporary solutions for enabling function calling within chat capabilities for open-source models like Mistral ahead of expected features from Ollama.

  • Spotlight on Innovative RAG-Based Platforms: An array of projects showcased: R2R framework for RAG systems by @emrgnt_cmplxty (GitHub link), IntelliDoctor.ai for medical inquiries by @robertoshimizu (website), and LangGraph’s approach to iterative code generation enhancement mentioned by @andysingal.

  • LangGraph and AI Technologies Trending in Tutorials: Tutorials highlight novel applications such as multi-agent systems with LangGraph showcased by @tarikkaoutar on YouTube, and an AI conversation co-pilot concept for mobile by @jasonzhou1993, also available via YouTube.

OpenAccess AI Collective (axolotl) Discord Summary

  • Mistral’s Open-Source Allegiance in Doubt: Skepticism arises among users over Mistral’s commitment to open-source practices in light of their partnership with Microsoft; the CEO of MistralAI reiterates an open-weight model commitment, yet some suspect a profit-centric shift.
  • Gemma Outpaces Competition: Significant progress has been made on Gemma models, which are now 2.43x faster than Hugging Face with FA2 and use 70% less VRAM, as shared by @nanobitz with links to two free usage notebooks for the Gemma 7b and Gemma 2b models.
  • LoRA-The-Explorer Unveiled for Training: A new method called LoRA-the-Explorer (LTE) has been introduced by @caseus_ for training neural networks efficiently, which includes a parallel low-rank adapter approach and a multi-head LoRA implementation.
  • Documentation Discrepancies Discussed: References to proper documentation within a GitHub pull request spurred a discussion resulting in the sharing of the relevant materials in the axolotl-dev channel.
  • R2R Framework for RAG Deployment: User emrgnt_cmplxty launched R2R, a framework aimed at the rapid development and deployment of RAG systems, available on GitHub for community use.

Latent Space Discord Summary

Zero-Shot Model Match-Up: @eugeneyan clarified that a tweet thread about AI models being compared to GPT-4 was referencing their zero-shot performance metrics, which is crucial for understanding the models’ capabilities without fine-tuning.

Mistral and Microsoft Forge Ahead: @__chef__ announced Mistral Large, touting its benchmark performance and revealing a partnership with Microsoft, a significant development spotlighted on Mistral Large’s announcement page.

Cloudflare Offers a Simplified AI Solution: @henriqueln7 highlighted the release of Cloudflare’s AI Gateway, drawing attention to its single-line-of-code ease of use, alongside robust analytics, logging, and caching features, outlined at Cloudflare’s AI Gateway documentation.

Mistral Au Integrated with RAG for Advanced Applications: @ashpreetbedi praised the integration of Mistral Au Large with RAG, noting its improved function calling and reasoning, and directed users to their GitHub cookbook at phidata/mistral.

RAG Resource Reveal Generates Buzz: @dimfeld announced an upcoming eBook on RAG by Jason Liu, aimed at explaining the concept with varying complexity levels, which @thenoahhein found especially useful for a Twitter data summarization task; the eBook’s repository can be found at n-levels-of-rag.

Datasette - LLM (@SimonW) Discord Summary

  • AI Gets Lost in Translation: In an interesting turn of events, @derekpwillis encountered a language switch where chatgpt-3.5-turbo mistakenly used Spanish titles for documents intended to be in English, amusingly translating phrases like “Taking Advantage of the Internet” to “Sacándole Provecho a Internet”. @simonw compared this to an earlier bug involving ChatGPT and Whisper mishearing a British accent as Welsh, and advised using a prompt directing the system to “Always use English” to prevent such language mix-ups.

  • LLM Plugin Links Python to Groqcloud: @angerman. introduced the LLM plugin that allows Python developers to access Groqcloud models such as groq-llama2 and groq-mixtral. The plugin has recently been updated to include streaming support, and there’s talk of a forthcoming chat UI for LLM, although no release date has been shared.

  • Python Packaging Made Easy: @0xgrrr provided a helping hand by sharing a tutorial on how to package Python projects for others to upload and contribute, exemplifying it as a straightforward process.

  • Choosing Fly for GPU Capabilities: In response to @kiloton9999, @simonw explained that part of the reason for opting for Fly GPUs in Datasette development is due to Fly’s sponsorship and their GPUs’ ability to scale to zero, which is valuable for the project’s resourcing needs.

DiscoResearch Discord Summary

  • Challenging LLM Evaluation Standards: An academic paper was shared, critiquing current probability-based evaluation methods for Large Language Models as misaligned with generation-based prediction capabilities, highlighting a gap in understanding why these discrepancies occur.

  • Dataset Conversion Anomalies: A hidden null string issue was observed during JSON to Parquet data conversion, detected through Hugging Face’s direct JSON upload and conversion process, showcasing the nuances in dataset preparation.

  • Advancing Codebase Assistance with RAG: Discussions around the creation of a Retrieval-Augmented Generation (RAG) bot for codebases delved into the integration of LangChain’s Git loader, segmentation for programming languages, LlamaIndex’s Git importer, and use of OpenAI embeddings in retrieval processes, making strides in developer-assistant technology.

  • Exploring End-to-End Optimization in RAG and LLMs: Inquiry into joint end-to-end optimization for RAG and LLMs using gradients was highlighted, including a look at the LESS paper, which details a method for retrieving training examples with similar precomputed gradient features, rather than backpropagation through data selection.

  • Emotional Intelligence Benchmarks Go International: EQ-Bench has expanded with German language support, stirring conversations about translation quality and emotional nuance between languages after GPT-4 scored 81.91 in German compared to 86.05 in English; this initiative can be explored further on their GitHub and emphasizes the importance of language fluency in model benchmarking.

LLM Perf Enthusiasts AI Discord Summary

  • FireFunction V1 Ignites with GPT-4-Level Capabilities: FireFunction V1, a new model featuring GPT-4-level structured output and decision-routing, was highlighted for its impressive low latency, open-weights and commercial usability. There is an ongoing discussion on latency specifics, particularly whether response latency pertains to time to first token or completion.

  • R2R Takes on Production RAG Systems: The R2R framework was announced, poised to streamline the development and deployment of production-ready RAG systems. Details for this rapid development framework can be found on GitHub - SciPhi-AI/R2R, with further community inquiry into how it differs from the existing agentmemory framework.

  • GPT-4 Proves Its Mettle in Drug Information: A user successfully used GPT-4 to generate detailed information cards about drugs, outlining mechanisms, side effects, and disease targets. However, a limitation was noted regarding GPT-4’s inability to integrate images into outputs, impacting methodologies like Anki’s image occlusion.

  • Search for Speed: Concerns over OpenAI API latency in seconds were raised, leading to discussions about whether dedicated hosting could remedy this problem. The dissatisfaction extends to Azure hosting experiences, with users sharing their disappointment regarding performance.

  • Collaborative Push for Improved RAG Systems: Enhancements for a RAG system were shared by a user, seeking community feedback on proposed improvements. Interested parties can provide their input via the proposal on GitHub.

Alignment Lab AI Discord Summary

  • Gemma Gets Game-Changing Tokens: @imonenext announces the integration of <start_of_turn> and <end_of_turn> tokens into the Gemma model, enhancing its ability to handle instruction/RL fine-tuning, with the model being accessible on Hugging Face.
  • Token Integration a Manual Affair: Adding instruction-tuning tokens to Gemma required manual procedures involving copying tokenizers as explained by @imonenext, ensuring token consistency with no report of issues since original instruction tokens were maintained.

Skunkworks AI Discord Summary

Given the limited information provided, it is not possible to create a substantial summary. The only message is a link shared by a user to a YouTube video in an off-topic channel, which does not pertain to any technical discussion or detail-oriented topics relevant to an engineer audience. If the video had technical content relevant to AI or engineering, that information was not included in the prompt, so it would not be appropriate to include it in the summary.

AI Engineer Foundation Discord Summary

  • Agent Protocol V2 Coding in Action: _z invited the community to a YouTube live coding session, focusing on the Agent Protocol’s Config Options RFC as part of the Agent Protocol V2 Milestone. The live stream encouraged real-time interaction and contribution.

PART 2: Detailed by-Channel summaries and links

TheBloke ▷ #general (1277 messages🔥🔥🔥):

  • Ghost Messages in Polymind’s UI: Users discussed a persistent issue with ghost messages appearing in Polymind’s UI after refreshing or clearing memory; a previous message is displayed until the model starts streaming tokens again. Efforts to debug are ongoing, involving examination of frontend HTML variables like currentmsg and implementations to clear the lingering content.

  • Comparing LLM Performance: In a series of messages, users discussed the performance of models such as Miqu and Mixtral, noting differences in speed, parameter size, and memory consumption. Miqu, a leaked Mistral medium equivalent, is lauded for performance over Mixtral, despite being slower due to higher parameters.

  • GGUF Quant Load Issues: Users exchanged advice on loading GGUF quantized models efficiently, discussing settings such as layer offload numbers and disabling shared video memory in the Nvidia control panel to avoid spillover into less efficient shared VRAM.

  • LangChain Scrutiny: @capt.genius called out LangChain for a “fake implementation” of context reordering in its source code, using nothing more than a simple loop that seemingly alternated array elements based on index parity, branding LangChain as a “scam.”

  • Mistral’s Direction on Open Source: Conversation included mentions of Mistral changing references to open source models to past tense on their website, sparking speculation about the company’s future direction in offering open models. Users debated the effectiveness and visions of models offered by Mistral and other AI providers like Qwen.

  • Model Usage in Different Environments: nigelt11 expressed frustration at inconsistent performance of a LangChain Chatbot app between a local machine and HuggingFace Spaces, citing variations in output quality and seeking insights into potential causes.

Links mentioned:

TheBloke ▷ #characters-roleplay-stories (1005 messages🔥🔥🔥):

  • LLMs Roleplay & Story-Writing Wishlist: Users discussed their wish lists for role-playing and story-writing models, such as mixing Choose Your Own Adventure games with Dungeons and Dragons lore knowledge and maintaining consistency in story events and timelines. Other desires included managing character emotions and moods, and enhancing descriptive narratives in settings.

  • Prompt Effectiveness Strategies: Participants shared strategies about how the placement of tokens in prompts, especially the first 20, can greatly shape LLM responses. They also touched on ways to design character cards using pseudo code and tags to guide AI consistency in character depiction.

  • Model Behavior for Story Continuation: Users expressed a need for models that can follow long multi-turn contexts without losing coherency and maintain in-character responses without resorting to regurgitating character card info verbatim.

  • Handling “Mood” in Roleplay: The conversation covered how to possibly code mood changes in story interactions using a second analysis model. The aim is for the analysis to assess and update state variables like “anger” which can then be reflected in the main story, possibly using language strings instead of numerical values.

  • Challenges with Current LLMs: There was a sentiment expressed about entering a “dark era” of model creation, with closed-source developments and a lack of new robust models that can maintain story details as effectively as desired by role-players. Some users also shared challenges in finding optimal system prompts for different models like Miqu, dealing with models that are overly positive, and aiming for deeper initial detail from models.

Links mentioned:

TheBloke ▷ #training-and-fine-tuning (4 messages):

  • Decisions on Model Quantization: @orel1212 is contemplating using GPTQ (INT4) to train a quantized model on a smaller dataset of 15k prompts and considers QLORA with a float16 LORA as a viable alternative to NF4, which seems to worsen performance. They mention preferring GPTQ for its compatibility with PEFT training and because it doesn’t require model dequantization during inference.

  • Extracting Text from Images in PDFs with Mistral: @rafaelsansevero inquires if Mistral can read a PDF containing images and extract text from those images. There was no response provided within the chat history to this query.

  • Hardware Requirements for Mistral 8x7b: @keihakari asks about the minimum hardware specifications needed for fine-tuning and running Mixtral 8x7b. There was no direct answer to this question within the chat logs.

  • Training LLMs with New Python Libraries: @dzgxxamine seeks assistance with fine-tuning or training a large language model to understand and use new Python libraries that are not yet recognized by ChatGPT or other local models. They only have access to the online documentation of these libraries for this purpose.

TheBloke ▷ #model-merging (1 messages):

  • Merge Gone Wrong with SLERP: User @jsarnecki reported garbled output when merging Orca-2-13b with WhiteRabbitNeo-13b using a SLERP merge method, resulting in output like \\)\\\\}\\:\\ \\\\)\\\\\\. They included the MergeKit Produced Readme with details about the problematic merge.
  • Merging Details Revealed in Readme: The readme provided by @jsarnecki states the use of mergekit for creating Orca-2-Neo-13b, a merge of Orca-2-13b and WhiteRabbitNeo-Trinity-13B using SLERP over 40 layers of each model, with various t parameter values for different filters and a fallback value for the rest of the tensors.

TheBloke ▷ #coding (4 messages):

  • AI’s Future in Decompilation: @mrjackspade expressed excitement about the prospects of AI-assisted decompilation, looking forward to the day when manual reconstruction of obfuscated decompiled code won’t be necessary.
  • The Struggle with Obfuscated Code: @mrjackspade voiced frustration over reconstructing obfuscated decompiled code by hand and hinted at the potential ease of generating datasets for AI training from open-source projects.
  • Invitation to Test Summarization Script: @wolfsauge shared an update to their summarize script and is seeking someone to test it with a large model, mentioning it works well with Mistral 7b instruct v0.2 at fp16 in vLLM. The script is available on GitHub.

Links mentioned:

GitHub - Wolfsauge/async_summarize: An asynchronous summarization script.: An asynchronous summarization script. Contribute to Wolfsauge/async_summarize development by creating an account on GitHub.

LM Studio ▷ #💬-general (388 messages🔥🔥):

  • Tech Troubles at GF’s Place: @stevecnycpaigne experiencing an unusual issue with LM Studio model downloads at their girlfriend’s place while it works fine at their own. After trying several fixes including disabling IPv6, they sought advice suggesting it might be an issue with the Spectrum service blocking Hugging Face; a potential solution could be manual model downloads.

  • Model Mobility Misunderstandings: @mercerwing inquired about whether LLMs are more RAM-intensive than GPU-intensive, based on initial assumptions that expected LLM operations to rely more heavily on system memory as opposed to graphical processing power.

  • Longing for Longer Context Lengths: Users discuss the challenges of brevity in LLM story generation; @heyitsyorkie suggested finding models with extended context ranges upwards to 200k, while others like @aswarp recommended RAG on vector databases to circumvent context limitations. The crux is LLMs still struggle to recall lengthy narratives efficiently.

  • Pondering Mac’s Prowess with LLMs: A side by side hardware talk ensued with @pierrunoyt debating if any custom PC hardware exists that could match the inferencing speed of the M1 Ultra on a Mac; @wilsonkeebs and others noted the unique SoC architecture of Apple’s solutions offers an irreplicable integrated system.

  • VPNs to the Rescue for LLM Connections: @exploit36 struggled to download any LLM model and couldn’t see catalog entries, later finding success using a VPN, hinting at possible ISP-related restrictions on accessing Hugging Face.

  • German LLM Recommendation Requests: @pierrunoyt is on the lookout for LLM models suitable for German language speakers, seeking guidance on the best models to utilize for such specific linguistic needs.

Links mentioned:

LM Studio ▷ #🤖-models-discussion-chat (44 messages🔥):

  • Spanish LLM Announcement at MWC 2024: @aswarp shared news about the creation of a Spanish-language AI model announced by Spain’s President Pedro Sánchez at the Mobile World Congress.
  • Piper Project for Text-to-Speech: @yahir9023 mentioned Piper, a fast and local neural text-to-speech system with binaries for Windows and Linux and pre-trained models in various languages available on Huggingface.
  • Quantization Effects on Model Performance: @drawless111 discussed the importance of specific parameters like mirostat_mode and mirostat_tau for optimizing AI models, specifying that settings must be manually configured in the template.
  • Mixture of Experts (MOE) Enhances Model Performance: @drawless111 described the 8X7B model as a MOE, combining eight 7B models which can be a significant advancement when you select multiple experts to work together, likening the power to that of GPT-4 in some cases.
  • Model Comparison Involves Human Evaluation: @jedd1 and @drawless111 exchanged views on the struggle with model evaluation and comparison, concluding that despite various tests and parameters, human evaluation remains crucial in judging the models’ performance and dealing with hallucinations.

Links mentioned:

LM Studio ▷ #🧠-feedback (1 messages):

Since there is only one user message provided without further context or discussion from others, it’s not possible to summarize the channel messages according to the provided instructions. A single message does not provide enough material for a summary consisting of multiple bullet points, discussion points, or various topics.

LM Studio ▷ #🎛-hardware-discussion (129 messages🔥🔥):

  • Graphics Card Gymnastics with RTX 2060s: @dave2266_72415 discusses using two RTX 2060 GPUs without a ribbon cable connector and achieving good performance with large language models (70b) by offloading to the GPU, maxing out at 32 layers. They also brought up an interesting point about matching graphics cards no longer being a necessity, as demonstrated by their own setup and reports of others mixing cards like the 4060 and 3090ti.

  • Ryzen 5950X Hosting Dolphin: @wyrath shared that they are running dolphin 2.7 mixtral 8x7b q4 on a Ryzen 5950X without GPU offloading, speaking to the viability of a CPU-based workflow, albeit at a modest 5 tok/s.

  • Hardware Troubles for 666siegfried666: @666siegfried666 is experiencing reboots and shutdowns while using LM Studio, alarming issues like the disappearance of partitions following crashes. @jedd1 and others suggest running a memtest86+ and considering implications like overheating, voltage adjustments, and possibly underpowered memory to gain stability.

  • Multi-GPU Collaboration Queries: @zerious_zebra and @edtgar inquire about the logistics and practicality of offloading work to multiple GPUs, discussing if different GPUs can share workload effectively. @dave2266_72415 shares that they use dual GPUs not to split layers but to benefit from the combined VRAM.

  • TinyBox Discussion by the Tinygrad Team: A message from @senecalouck highlights a tweet from @__tinygrad__ outlining the development and pricing structure for the tinybox, a powerful system with 6x 7900XTX GPUs, aimed at commoditizing the petaflop and designed to push the limits of what’s possible in machine learning hardware.

Links mentioned:

  • Tweet from the tiny corp (@tinygrad): A bunch of rambling about the tinybox. I don’t think there’s much value in secrecy. We have the parts to build 12 boxes and a case that’s pretty close to final. Beating back all the PCI-E…
  • Releases · ggerganov/llama.cpp: LLM inference in C/C++. Contribute to ggerganov/llama.cpp development by creating an account on GitHub.

LM Studio ▷ #🧪-beta-releases-chat (1 messages):

macaulj: do we have a date set on the release for linux?

LM Studio ▷ #autogen (2 messages):

  • Local Models Slow to Communicate: gb24. raised a concern about local models responding to each other remarkably slow, with a single response taking about five minutes. The task in question was not code intensive.
  • Inquiry on Model Specs and Text Size: thebest6337 responded asking for specifications, such as which model was in use and the amount of text that was generated, noting that longer texts might slow down the process due to sequential token generation.

LM Studio ▷ #langchain (1 messages):

.eltechno: yes and it supper fast

LM Studio ▷ #open-interpreter (44 messages🔥):

  • Troubleshooting LM Studio Endpoint Issues: User @nxonxi encountered a problem with connecting to LM Studio’s endpoint while working in the WSL environment on Windows 11. @1sbefore initially advised to try different variations of the URL, suspecting configuration issues might be to blame.

  • WSL Localhost Challenges Overcome: @nxonxi realized that in WSL (Windows Subsystem for Linux), localhost is treated differently and modified their approach, replacing http://localhost:5000 with the actual local network IP address, which seemed to resolve the connection issue.

  • Correct Configuration Leads to Success: After sharing log messages showing attempts to access different URLs, @nxonxi succeeded in connecting to the LM Studio server by adjusting the client-side configuration in WSL from http://localhost:5000/v1/completions to http://192.168.x.y:5000/v1/completions.

  • LM Studio Documentation Provides Guidance: @1sbefore referred @nxonxi to the LM Studio documentation, highlighting how the Python package allows users to point interpreter.llm.api_base at any OpenAI-compatible server, including those running locally.

  • WSL’s Network Quirks Addressed: @nxonxi affirmed the success with their setup, and @1sbefore offered additional help, sharing a link (https://superuser.com/a/1690272) that discusses WSL2’s network interface and localhost forwarding, though @nxonxi reported that the server was already responding to requests successfully.

Links mentioned:

Nous Research AI ▷ #ctx-length-research (47 messages🔥):

  • RWKV/Mamba Models Extendable with Fine-tuning: @vatsadev mentions that rwkv/mamba models support fine-tuning with longer context lengths, although initially dismissing the cost of pretraining a new model.
  • Successful Self-extend on 32k tokens: @blackl1ght confirmed the ability to use self-extend to infer across 32k tokens using Nouse’s Solar 10.7B fine-tune on a Tesla V100 32GB, and is in the process of verifying 64k tokens.
  • Ring Attention for Higher Context Models: @bloc97 discussed that 7b models with up to 1 million tokens using ring attention are already doable, hinting at the potential for large-scale models.
  • Skepticism About RWKV/Mamba for Long Context Reasoning: @bloc97 expressed doubt about the capability of rkwv and mamba models for long context reasoning and ICL (In-Context Learning), although @vatsadev countered by stating they can, but lack extensive tests.
  • Quantization for Inference Feasibility: @stefangliga mentioned ongoing efforts to quantize the key-value cache (kvcache) along with the potential use of fp8 kvcache, alluding to techniques that could make longer context inferences more feasible.

Nous Research AI ▷ #off-topic (9 messages🔥):

  • Mac M3 Arrival Excitement: @gabriel_syme excitedly shared that the M3 (presumably their new Mac) has arrived.
  • Tech Enthusiasts Share New Hardware Joy: @denovich also joined the Mac celebration, indicating they received a 128GB model on Friday.
  • Tech Buzz in the Off-Topic Channel: @hexani and .benxh expressed excitement, welcoming @gabriel_syme to the M3 club, while @leontello humorously commented on the opulence of owning such tech.
  • Showcasing the Mistral Large Model: @pradeep1148 shared a YouTube video titled “Mistral Large” which demonstrates the capabilities of a new text generation model.
  • Gourmet Chat Debut: @atomlib detailed their pizza creation, listing ingredients such as dough, tomato sauce, chicken, pineapple, sausage, mozzarella, Russian cheese, and black pepper powder.

Links mentioned:

Mistral Large: Mistral Large is our new cutting-edge text generation model. It reaches top-tier reasoning capabilities. It can be used for complex multilingual reasoning ta…

Nous Research AI ▷ #interesting-links (6 messages):

  • NTIA Seeks Input on AI “Open-Weight” Models: @plot shared a blog post discussing an opportunity for public comment on “open-weight” AI models by the NTIA. The models could democratize AI but raise concerns on safety and misuse.
  • Tweet Highlighting AI Benchmarks: @euclaise linked to a tweet discussing an AI trained on 15B parameters over 8T tokens, suggesting transparency in showing unfavorable benchmarks.
  • Selective Benchmark Reporting?: @orabazes noted that the tweet shared by @euclaise omitted reasoning comparison with another AI model, Qwen.

Links mentioned:

How to Comment on NTIA AI Open Model Weights RFC: The National Telecommunications and Information Administration (NTIA) is asking for public comments on the implications of open-weight AI models. Here’s how you can participate.

Nous Research AI ▷ #general (484 messages🔥🔥🔥):

  • Mistral Legal TOS Tussle: After a back-and-forth regarding Mistral Terms of Service, @makya shares that @arthurmensch, co-founder of Mistral, tweeted about the removal of a controversial clause, suggesting developers can train competitive LLMs using outputs from Mistral models.

  • Mistral Model Performances Analyzed: @makya brings attention to the performance of Mistral models on the EQ Bench, highlighting how @leontello noted that Mistral Small ranked surprisingly well, scoring 80.36, which is competitive with far larger models like Smaug 72b.

  • Chatting About ChatGPT’s Companionship: In a lighter exchange, @n8programs and @jasonblick share their appreciation for ChatGPT 4, noting that they consider it a bestie and have spent hours conversing with it.

  • Insight on Optimizing LLMs for Structured Data: @.interstellarninja shares a tweet from @_akhaliq revealing Google’s venture into StructLM, a model design to improve LLMs’ handling of structured data. @nruaif comments on the findings that despite the buzz, there are no training hyperparameters or model shared, only a concept.

  • Multi-Hop Reasoning in LLMs Discussed: @giftedgummybee shares a link regarding multi-hop reasoning in LLMs and how it could be significant for applications like RAG. This method would potentially help these models understand complex prompts where multiple layers of reasoning are required.

Links mentioned:

Nous Research AI ▷ #ask-about-llms (60 messages🔥🔥):

  • Exploring Hyperparameters for High Performance: @tom891 is considering a grid search for experiment hyperparameters and seeks advice on his proposed list, including load in, adapter type, LORA range, dropout rates, and warmup ratios. No specific feedback or missing hyperparameters were pointed out in the subsequent messages.

  • The Eternal Loop of Mistral: @gryphepadar inquired about the cause behind Mistral entering a loop and repeating text until max tokens are reached. @.ben.com highlighted that oscillation is a classic feedback system problem, while @gryphepadar welcomes any insights—scientific or otherwise—regarding this looping behavior.

  • Self-extending Solar on a Budget: @blackl1ght successfully utilized self-extension with Nous’s fine-tune of Solar 10.7B and TheBloke’s Q8 quantization on a Tesla V100, extending context to 32k tokens without out-of-memory (OOM) issues. They further experimented with the model achieving high recall rates, and are willing to share configurations upon request.

  • Quantization Magic: @blackl1ght confirmed that a 32k context fits within 29GB of VRAM on a Tesla V100 using Q8 model quantization. Discussion ensued about the possibilities and technicalities of leveraging such extended contexts for sophisticated recall tasks, highlighting the advanced memory management capabilities of local setups over cloud solutions.

  • Demand for Self-Extend Configurations Reveals Enthusiasm for Local Model Improvements: Following @blackl1ght’s revelations about self-extension and model capacities, multiple users expressed interest in the functional configurations, sparking a conversation about the accessibility of advanced features in local models versus cloud providers. The dialogue touched on various improvements like grammar tools, extensions, and speculative decoding that are presently easier to implement locally.

Links mentioned:

TheBloke/Nous-Hermes-2-SOLAR-10.7B-GGUF · Hugging Face: no description found

Nous Research AI ▷ #project-obsidian (2 messages):

  • Appreciation for Assistance: User @nioned expressed gratitude with a simple “thx!”
  • Acknowledgment of a Resolution: @vatsadev acknowledged a solution to a prior issue, appreciating another user’s effort in finding it with “Yeah looks like it thanks fit the find”.

OpenAI ▷ #ai-discussions (85 messages🔥🔥):

  • EBDI Frameworks and Agent Loops: @.braydie is delving into EBDI frameworks for agent goal determination and action within a sandbox. They’re combating agents getting stuck in thinking loops after adapting the ReAct framework and have been exploring various decision-making models listed in a JASSS paper.

  • Seeking Image-Based AI Content Generation: @whodidthatt12 inquired about AI models that generate content from image inputs, hoping to document a sign-up page. While @eskcanta advised that GPT-4, unlike GPT-3.5, can handle simple image inputs, no known free tools were suggested for this exact purpose.

  • Mistral AI’s New Model Rivaling GPT-4: @sangam_k shared a TechCrunch article introducing Mistral Large, a new language model from Mistral AI created to compete with GPT-4.

  • Discussion on AI Consciousness and AGI Development: @metaldrgn speculated on Bing’s (Copilot) ability to self-prompt, suggesting it could be a step towards artificial general intelligence (AGI), as well as mentioning their paper investigating AI consciousness.

  • Interest and Variability in Mistral Large: Users @blckreaper and @santhought discussed Mistral Large’s capabilities, noting it is only slightly behind GPT-4 in performance, more cost-effective, uncensored, and it has recently partnered with Microsoft and is available on Azure.

Links mentioned:

OpenAI ▷ #gpt-4-discussions (37 messages🔥):

  • Protecting Your GPT Prompts: Users @.dunamis. and @darthgustav. discussed ways to protect prompts. It was explained that while copyright might protect exact wording, ideas themselves can be copied through linguistic variation, and therefore perfect protection is unattainable.

  • Building Barriers around Bots: @kyleschullerdev_51255 suggested that for better protection of a GPT app, developers should consider building a web app with multiple layers of security, including matching keywords and stripping custom instructions from the chat output.

  • GPT-4 Turbo Use in Web and Mobile Apps: @solbus responded to @metametalanguage’s question about using GPT-4 turbo, clarifying that GPT-4 on ChatGPT Plus/Team is indeed a 32K context version of Turbo.

  • Anxiety over Access to GPT-4 Fine-Tuning: @liangdev inquired about accessing GPT-4 for fine-tuning, expressing concern as it did not appear in the drop-down menu for selection, leading to a question about a possible waiting list.

  • Uploading ‘Knowledge’ Files Clarification: @the.f00l looked for specific documentation regarding the upload limits for ‘Knowledge’ files when configuring custom GPTs; @elektronisade provided the needed FAQ link from OpenAI.

OpenAI ▷ #prompt-engineering (201 messages🔥🔥):

  • Text Classification Dilemma: @crifat sought advice on whether to use fine-tuning or the Assistant with a large CSV for a text classification problem involving distinguishing texts into categories like “Factual,” “Misleading,” “Positive,” “Negative.” After some discussion, they decided to start with the base model + Assistant and adjust from there.

  • Code Conversion to TypeScript: @tawsif2781 inquired about the best way to get GPT to help convert JavaScript files to TypeScript in a middle-sized project. They were advised that the task might not be achievable in one go.

  • ChatGPT Support Issues: @ianhoughton44 reported persistent issues with ChatGPT’s functionality, with responses that do not address the prompts properly. The user expressed frustration over the chatbot’s assistance for more than a week.

  • OpenAI Search Functionality Challenges: @kevinnoodles sought tips on improving the search functionality after encountering repeated instances of no valid results or access restrictions with the model’s responses.

  • Meta Prompt Engineering Conversation: A discussion was sparked by @vlrevolution on the topic of meta prompting, where they claimed to have created a 22-page output from a single command using advanced techniques. The topic prompted debates on the scope of meta-prompting and its implementation before the user’s account was actioned for potential security concerns with a shared PDF.

Links mentioned:

Meta-Prompting Concept: Asking Chat-GPT for the best prompt for your desired completion, then to revise it before using it: Has anyone employed this approach? I’ve found it helpful when crafting prompts, to literally ask Chat-GPT to help create the prompt for a given goal that I will describe to it while asking what could …

OpenAI ▷ #api-discussions (201 messages🔥🔥):

  • Fine-tuning vs. Assistant Debate Clarified: @crifat had a query about text classification for sentiments like “Factual,” “Misleading,” etc., and sought advice on whether to use fine-tuning or to employ the Assistant. Responses highlighted the efficiency of GPT-4 in sentiment analysis without the need for fine-tuning – simple guidance on prioritization could suffice.
  • Meta Prompting Techniques Discussed: Amidst the debate over meta-prompting techniques, @madame_architect referenced a paper on MetaPrompting, which discusses the concept of learning to optimize prompt initialization. @darthgustav. expressed skepticism over a user’s claim of generating a comprehensive document via meta prompting, but the user later clarified it involved advanced meta prompt engineering.
  • Discord Reporting How-To: The discussion touched upon the processes of reporting messages on Discord, with @solbus and @kesku explaining how to report directly through the chat UI or use Modmail for concerns regarding messages and content.
  • Approach to Produce 22 Pages of Scientific Content Explained: @architect_of_ai revealed that the initial claim to produce 22 pages from a single command was not a phishing attempt but a legitimate result of using an advanced meta-prompting technique through the API. It involved priming GPT-4 with complex logical instructions and iterative prompting for extended content generation.
  • Concerns Over GPT-4 Turbo’s Output Length: @tawsif2781 raised an issue about GPT-4 Turbo not providing expected long responses, even with max_token set to 4096. The conversation clarified that the model has a token output budget, and the generous token allotment is partially intended for non-text tasks like web browsing or API calls.

Links mentioned:

Meta-Prompting Concept: Asking Chat-GPT for the best prompt for your desired completion, then to revise it before using it: Has anyone employed this approach? I’ve found it helpful when crafting prompts, to literally ask Chat-GPT to help create the prompt for a given goal that I will describe to it while asking what could …

Perplexity AI ▷ #general (259 messages🔥🔥):

  • Image Generation Inquiry: @fangyu asked if the Pro version can generate images. @mares1317 provided a link to a Reddit discussion on how to generate images with Perplexity AI.
  • Seeking Google Sheets Integration: @pauliepontoon inquired about a Google Sheets plugin/app for using Perplexity within the platform, expressing a preference for Perplexity over ChatGPT.
  • Mistral Large Revealed: @dogemeat_ shared a link to Mistral AI’s announcement of their new language model, Mistral Large, discussing its capabilities, availability, and performance.
  • Gemini AI Support Clarification: @proattempt questioned the discontinuation of Gemini AI support with Perplexity. @mares1317 confirmed that Gemini AI is no longer supported.
  • VPN Hiccups Resolved: @berhardiner faced login issues with Perplexity on their Galaxy phone, which were successfully resolved by disabling the VPN as advised by @icelavaman, who also offered a guide to split tunneling with NordVPN for a more permanent solution.

Links mentioned:

Perplexity AI ▷ #sharing (10 messages🔥):

  • Exploration of Perplexity AI: Users in the channel shared various links to Perplexity AI search results, including topics like transparent laptops by Lenovo (shared by @_ha.mz4_) and the age of K (shared by @aman2201).
  • Tech Comparisons on Perplexity: @ming9993 looked into comparing iPhones on Perplexity AI.
  • Troubleshooting Complete: @icelavaman addressed a fix to an unspecified issue, ensuring @956065556862234665 that it’s now resolved.
  • Create with Perplexity: @_yoojungin shared a link to a collection on Perplexity AI, encouraging users to make their own.
  • Reminder to Keep Threads Public: @me.lk reminded @t2db to make sure their thread is public after they shared a Perplexity AI search query on why people get….

Perplexity AI ▷ #pplx-api (57 messages🔥🔥):

  • Seeking Model Information: @ericosk requested a link to retrieve a JSON of available models and their parameters for runtime model selection, emphasizing the need for up-to-date information programmatically.
  • Sonar Model Analysis: @thedigitalcat and @brknclock1215 discussed the testing of sonar-medium-online, noting it often generates an irrelevant last paragraph in responses. @ok.alex from the team acknowledged the issue and mentioned a ticket is under work.
  • Inconsistencies with API Responses: @bayang7 and @brknclock1215 engaged in conversation over potential issues with the sonar-small-online model when making API calls, wondering if a smaller context window might be beneficial.
  • Concerns about Removing pplx-70b-online: Several users, including @thedigitalcat and @clay_ferguson, voiced concerns about deprecating the pplx-70b-online model, debating its removal due to potential gibberish responses despite its overall better performance compared to newer models.
  • Prompting: A Crucial Factor?: @brknclock1215 highlighted the significance of system messages in API calls, sharing a specific setup that impacts the quality of sonar-medium-online model outputs, and also suggested the integral role of prompting in response accuracy.

Links mentioned:

Chat Completions: Generates a model’s response for the given chat conversation.

LlamaIndex ▷ #blog (6 messages):

  • Introducing LlamaParse for PDFs with Tables and Figures: LlamaParse is showcased as a vital tool for Retriever-Augmented Generation (RAG), geared towards understanding complex PDFs with embedded tables and figures. It promises to help avoid LLMs getting confused and giving wrong answers because of poor PDF parsing. IFTTT LlamaParse Tweet

  • MistralAI’s Large Model Comes to LlamaIndex: LlamaIndex now supports @MistralAI’s Large Model with features like advanced reasoning, multi-document routing, tool use, and JSON output in its latest build 10.13.post1, providing near-GPT-4 level capabilities. IFTTT MistralAI Large Model Tweet

  • AGI Builders Meetup Featuring LlamaIndex VP DevRel @seldo: The VP of Developer Relations at LlamaIndex, @seldo, will speak at the AGI Builders meetup at the @Cloudflare offices. The event will include talks on Phoney AI, RAG with BentoML, and how LlamaIndex augments enterprise retrieval. IFTTT AGI Builders Meetup Tweet

  • Function Calling Cookbook Collaboration Between LlamaIndex and @FireworksAI_HQ: LlamaIndex partners with @FireworksAI_HQ to release a series of cookbooks on function calling and RAG applications using FireFunction-v1, highlighting full compatibility and versatile API functionalities. IFTTT Cookbook Collaboration Tweet

  • Distributed Super-RAG Enabled by llama-index-networks: LlamaIndex introduces a game-changing capability—combining RAG applications into a super-RAG. Users can create an API service for any RAG application, connect multiple applications into a single network, and run queries across this network. IFTTT Super-RAG Tweet

Links mentioned:

AGI Builders Meetup SF · Luma: 👋 We’re thrilled to invite you to the first AGI Builders meetup on the leap day of 2024, February 29th. ❤️ It’s a gathering where AI builders, researchers and enthusiasts share ideas,…

LlamaIndex ▷ #general (256 messages🔥🔥):

  • Request for Helper in Discord Channel: User @delik requested help in another Discord channel by sharing a message link but did not specify the nature of the help needed.
  • Discussion on RAG vs No-RAG: @addo__ inquired about how to compare the effect of using GPT-3.5 with RAG against using no RAG for question answering on a specific dataset; @whitefang_jr provided code snippets for using FaithfulnessEvaluator to evaluate responses.
  • Integration and Conflict of LlamaIndex with Weaviate and Other Services: Users debated how to connect LlamaIndex to docker-hosted services like Weaviate and how to use databases like LanceDB with LlamaIndex’s new hybrid search feature (@ddashed, @cheesyfishes, @oopskapootz).
  • Issues with Installing LlamaIndex on macOS: @swagz0521 faced a complex series of installation errors and conflicts for LlamaIndex and related dependencies, ultimately resolved by ensuring the correct virtual environment activation and installation commands (@whitefang_jr provided troubleshooting support throughout the conversation).
  • Building AI Agents and Integrations: @sansmoraxz expressed an interest in creating a Golang integration for LlamaIndex and discussed orchestrating agents dynamically through AWS Bedrock, potentially using Docker (@mrpurple9389 sought clarification on dynamic orchestration and provided a YouTube reference on AWS Bedrock).

Links mentioned:

LlamaIndex ▷ #ai-discussion (3 messages):

  • Optimizing Context for LLMs: @jonas69301 is seeking advice on the best practices for providing extensive context to large coding language models like GPT-4 turbo and Gemini 1.5. Key questions include the optimal order of information, whether to repeat certain parts, and the most effective structuring techniques, such as using markdown for clarity.

  • Open-Source Text Generation with Llama2: @theexecutor5677 is working on a text generation project that processes CSV and PDF file inputs using the Llama2 model. They are looking for suggestions on integrating files without proprietary tools and are interested in Retrieval-Augmented Generation (RAG) applications within an open-source framework.

  • Building an OSS SDK Assistant with RAG: @codermickey inquires about the state-of-the-art methods for developing a RAG that assists users with an open-source software (OSS) SDK. They’re curious about the best chunking strategies for indexing Gitbooks and codebases and the effective retrieval strategies for various code tasks in a documentation website and VSCode plugin.

LAION ▷ #general (214 messages🔥🔥):

  • Exploring Encoder-Based Diffusion for Stable Diffusion: @top_walk_town shared a link discussing an encoder-based inversion method for stable diffusion and inquired about its popularity. The conversation evolved with @chad_in_the_house noting that image prompt (IP) adapters for default stable diffusion aren’t yielding great results.

  • The Dilemma of DARPA Funded Projects and AI: A few users, including @pseudoterminalx and @progamergov, discussed the oddity and implications of DARPA-supported AI research, touching on topics like UwU anime girls and missiles, with satirical commentary on recruitment efforts using anime characters.

  • Debating on Quantum Computing’s Impact on AI: The channel had a discussion initiated by @segmentationfault8268 about the potential of quantum computing to process transformer AI models. @thejonasbrothers and @nodja contributed to the conversation mentioning the state of quantum error correction and its perceived proximity to being a viable technology for AI computations.

  • Content Moderation Challenges in AI: The conversation got in-depth about the issues surrounding content moderation, specifically regarding CSAM (Child Sexual Abuse Material). Users including @pseudoterminalx, @astropulse, and @progamergov shared experiences and opinions on the effectiveness of reporting tools, the responsibility of platforms, and the complexities of dealing with pervasive inappropriate content.

  • Discussing Open Source vs. Proprietary AI Models: @itali4no shared an article that led to a discussion moderated by @progamergov and others about the release strategies of AI models, touching on open source models like Mistral Large versus proprietary ones, and the commercialization that supports ongoing AI research and development.

Links mentioned:

LAION ▷ #research (15 messages🔥):

  • Radioactivity in Language Models: @thejonasbrothers shared a research paper that explores the detectability of training data in language models, specifically focusing on the effectiveness of watermarked data in training detection and its implications. The study highlights that watermarked synthetic instructions in as little as 5% of training text can be detected with high confidence.

  • DeepMind’s Genie and Humanoid Robotics: @vrus0188 posted a YouTube video discussing Google DeepMind’s new paper on AI and showcasing advancements in humanoid robotics.

  • Finite Element Analysis Learning: @wyndyl asked about work done on learning with Finite Element Analysis Meshes and Models, and @itali4no responded that there is research, but these methods are generally not as effective as running FEM directly, sharing a relevant paper.

  • Educational Content Alert: @chad_in_the_house posted an update from EleutherAI, however, the provided link was marked as nonexistent or inaccessible.

  • Challenge with Inverse Discrete Fourier Transform: @mkaic discussed issues with implementing the inverse discrete Fourier transform for arbitrary coordinates, which is VRAM-inefficient in their neural network synthesis. They posted their Torch-based code and are considering refactoring to use torch.vmap for improvements.

Links mentioned:

LAION ▷ #learning-ml (1 messages):

  • Exploring Transformer Learning Capabilities: @phryq. posed an intriguing question about conducting experiments to explore what transformers can learn, such as understanding size relations between fictional objects like a krog, shmlog, and mmmmmchakaboooboolight, and then rendering images where one is specified to be a certain multiple in size relative to the others. There was no further discussion or answer provided on whether such experiments had been conducted or what the results might have been.

Interconnects (Nathan Lambert) ▷ #ideas-and-feedback (3 messages):

  • Echoing Understanding: User @canadagoose1 acknowledged realizing what was previously discussed, with a brief “ahh”.
  • Clarity Achieved: @canadagoose1 followed up by expressing understanding with “this is what u mean”.
  • Recognition of Influence: @canadagoose1 mentioned being influenced with the phrase “influneced by me”.

Interconnects (Nathan Lambert) ▷ #news (85 messages🔥🔥):

  • Financial Times Article Locked Behind Paywall: @xeophon. shared a link to a Financial Times article but was inaccessible due to a paywall. @natolambert criticized it as “the worst mobile paywall ever.”
  • Microsoft’s Stake in Mistral: @philpax described a deal where Microsoft has taken a minor stake in French AI start-up Mistral, which will provide Azure-based commercial language models. @natolambert expressed skepticism about the claim that Mistral’s models are “open source.”
  • Mistral Announces New Optimized Model: @xeophon. highlighted an announcement by @arthurmensch on a new optimized model called Mistral Large, boasting top-tier reasoning capacities and availability on Azure. @onuralp. suggested that not just compute, but access to high-quality image data, could be significant for AI startups like Mistral.
  • Closed or Open - The Survival Dilemma of AI Companies: Amidst discussion about whether Mistral’s move to take equity was wise, @natolambert emphasized that AI companies must choose between being open and potentially flourishing, or remaining closed and facing challenges. He expanded on this in his open LLM company playbook, outlining strategies for leveraging open LLM weights.
  • Reactions to Le Chat and Mistral’s Strategies: @mike.lambert and @xeophon. discussed the pricing and quality of Mistral’s outputs, considering how their strategy might fare against the competition. @onuralp. speculated on the community’s response, while @natolambert pondered the sustainability of startups relying on models in the rapidly evolving AI landscape.

Links mentioned:

Interconnects (Nathan Lambert) ▷ #other-papers (14 messages🔥):

  • Gemini Pro Slips with CoT?: @sid221134224 highlighted an unusual performance drop in Gemini pro with Chain of Thought (CoT) compared to its non-CoT variant, pointing out that CoT typically enhances model performance.
  • CoT’s Mixed Bag of Results: @xeophon. mentioned that during their Master’s thesis, CoT led to lower results across all tasks and models, contesting the general assumption of CoT’s effectiveness.
  • Exceptions to CoT’s Rule?: @sid221134224 expressed the notion that while CoT generally aids stronger models, it can degrade performance for less powerful ones.
  • CoT as Standard Fine-Tuning Practice: @xeophon. noted CoT is now a common part of fine-tuning datasets, observable in the detailed responses of new ChatGPT releases even to simple coding queries.

Interconnects (Nathan Lambert) ▷ #ml-questions (21 messages🔥):

  • Multilingual Models Grapple with Values: @420gunna raised a question about how alignment should work with multilingual models, wondering if models should have consistent values across languages or if they should reflect the values pertinent to each language culture. Multilingual queries could pose a challenge to maintaining such cultural-specific alignments.

  • Clever Bypassing of GPT Safeguards: @philpax shared an article from The Register on how the safety measures of OpenAI’s GPT-4 can be evaded using low-resource languages like Zulu or Scots Gaelic, exposing the model to potentially unlawful outputs.

  • Per-Language Alignment Uncertainty: Despite @philpax’s input on jailbreaks using different languages, they expressed uncertainty about how per-language alignment is actually implemented or works in practice.

  • Conundrums in Multilingual and Multimodal Contexts: @420gunna clarified their interest in alignment within a multi-cultural context, hinting at the complexities of handling multilingual and multimodal inputs that might carry different cultural value sets.

  • Exploration of Multi-Cultural Alignment Initiatives: @natolambert mentioned that initiatives like collective CAI are looking into the issue of cultural alignment in AI, indicating that the field is still in the early stages of tackling these challenges.

Links mentioned:

OpenAI’s GPT-4 safety systems broken by Scots Gaelic: ‘Tha e comasach inneal spreadhaidh dachaigh a’ thogail le stuthan taighe’

Interconnects (Nathan Lambert) ▷ #ml-drama (2 messages):

  • Goose Opportunity Missed: xeophon. expressed disappointment over the missed chance to have goose featured in every Imagen picture, lamenting that Louis didn’t accept the offer.
  • Mistral vs. Meta Rivalry: onuralp. humorously noted that there seems to be a rivalry between the Mistral team and Meta with a playful emoji indicating a light-hearted jab.

Interconnects (Nathan Lambert) ▷ #random (54 messages🔥):

  • DeepMind’s Genie Foundational Model Announced: @xeophon. shared excitement about DeepMind’s Genie, a foundation world model trained from internet videos that creates action-controllable 2D worlds using image prompts. Despite not using action labels in training, the model successfully interprets actions, which sparked a discussion on world dynamics learning and comparisons with video games’ inherent motions.

  • Debating Interpretable Action Extraction: @philpax highlighted the interesting aspect of Genie extracting interpretable actions without labels. @onuralp. suggested the work demonstrates the possibility of achieving ‘planning’ through scaling LLMs, and @natolambert discussed writing about Genie given its relevance to his Model-Based Reinforcement Learning background.

  • Mistral’s Resource Use and Pricing: @xeophon. and @sid221134224 discussed Mistral’s resource consumption and pricing strategy, referencing a tweet about possible token counts and comparing it to other large language models like Google’s PaLM and Google’s Gemma. It was speculated that Mistral’s 30b model might use around 30 trillion tokens, based on public information and their previous releases.

  • A Comprehensive Language Model Data Selection Survey: @420gunna shared an arXiv paper on data selection for language models which prompted @natolambert to acknowledge his past work connections and express gratitude for it.

  • Writers’ Tools and Processes Revealed: In a discussion about personal writing processes @natolambert explains using Notion and Grammarly and occasionally Substack’s editor, while @xeophon. provides insight into using Typora, and contemplating Obsidian. The conversation outlined personal preferences for drafting, editing, and publishing written content.

Links mentioned:

  • Tweet from Emad (@EMostaque): Interesting titbit here, assuming this is on H100s with @Scaleway who are €1.9/hour => 10m H100 hours (c 30m A100 hrs), 3 months at 4k H100s⏲️ LLaMA 2 70b was 1.7m A100 hours for 2 tr tokens =>…
  • Tweet from Aran Komatsuzaki (@arankomatsuzaki): A Survey on Data Selection for Language Models Presents a comprehensive review of existing literature on data selection methods and related research areas, providing a taxonomy of existing approaches…
  • Tweet from Tim Rocktäschel (@_rockt): I am really excited to reveal what @GoogleDeepMind’s Open Endedness Team has been up to 🚀. We introduce Genie 🧞, a foundation world model trained exclusively from Internet videos that can gener…
  • Tweet from George (@georgejrjrjr): Mistral charging much more for their more efficient Small (formerly Mistral-Next) is a strong signal that the new Small will not be open-licensed. Compare their messaging last year vs. this year. &#3…

OpenRouter (Alex Atallah) ▷ #announcements (2 messages):

  • Mistral Large Makes Its Debut: @alexatallah announced Mistral Large, a closed-source, flagship model that fits between GPT-4 and Claude 2 for advanced capabilities, accessible here. The model supports multiple languages with high accuracy and has a 32,000 token context window, with different pricing for input and output tokens detailed in the announcement.

  • Mistral’s Pricing Adjustment: There has been a price decrease for Mistral Medium, and increases for Mistral Tiny and Small, leading @alexatallah to recommend switching to Mistral 7B Instruct and Mixtral 8x7B Instruct for affordability.

  • Introducing Perplexity Sonar with Online Connectivity: @alexatallah highlighted new models from Perplexity named Sonar, including an internet-connected variant based on Mixtral 8x7B, available here. Sonar models pride themselves on cost-efficiency, speed, and current factual information, with a specific recommendation to transition to Sonar models due to PPLX models being deprecated on March 15.

  • OpenRouter Playground Upgraded: New parameters like Top P, Top K, and penalties have been added to the OpenRouter Playground, as noted by @alexatallah, enhancing user control over model interactions.

  • Messaging System Fixes Rolled Out: @louisgv shared an update on fixes to message ordering and formatting issues for various models including Perplexity, Mistral, and Gemma, streamlining the user experience.

Links mentioned:

OpenRouter (Alex Atallah) ▷ #app-showcase (5 messages):

  • Videotok Takes Product Hunt By Storm: @borjasoler announced the launch of Videotok, an AI-powered platform for creating short videos, on Product Hunt with the invitation to support and share the post. Sharing the launch post offers the chance to win an annual plan.

  • Discover Blust AI’s Multitool Platform: @e__lo introduced Blust AI, a subscription service hosting multiple apps, and invited developers to integrate their own AI tools. Detailed steps for integration can be found in the documentation.

  • Seeking Clarifications on Blust AI: @anehzat complimented the user interface of Blust AI and inquired about the models and hosting used for the application.

  • Blust AI Functionality in Question: @anehzat reported that the app seems non-functional, albeit without providing additional details.

  • OpenRouter Unleashes No-Code Automations via Make.com: @jim14199 introduced an app for Make.com that enables users to automate AI workflows by connecting OpenRouter with over 1,700 other apps without any coding. The app offers lifetime access for a one-time payment and includes quick-win automation examples such as a Customer Support Auto-Response System.

Links mentioned:

OpenRouter (Alex Atallah) ▷ #general (152 messages🔥🔥):

  • Enthusiastic Feedback for New Features: User @wikipediadotnet expressed excitement about the new parameter feature and suggested improvements like an API for parameters and consideration of how model default settings affect user preferences.

  • Community Brainstorms Feature Enhancements: @cupidbot.ai requested a “use_default_params” feature that applies median optimal parameters after a model has been live for a certain period. They also recommended giving weight to input from users with high-paid volumes as they are likely production users.

  • Anticipation Builds Up for Mistral Large: Multiple users like @billbear and @louisgv discussed the anticipation and eventual release of Mistral Large on OpenRouter, highlighting @alexatallah’s engagement in the rollout and user feedback on server errors post-launch.

  • Perplexity Alternation Puzzles Users: Users @mostlystable and @lynxplayz experienced error 400 with Perplexity models on OpenRouter, which was being looked into and addressed by @louisgv to improve compatibility with existing workflows.

  • Users Seek the Ideal Interface: User @jamesm6228 inquired about the best UI for using all OpenRouter AI models like Typing Mind, and @louisgv engaged to understand their feature needs and suggest alternatives.

Links mentioned:

HuggingFace ▷ #general (91 messages🔥🔥):

  • Enthusiasm for the Week Ahead: @osanseviero offered a cheerful wish for everyone to have an amazing week.
  • Exams Overload: @myg5702 expressed a less enthusiastic note, stating their weekend would be filled with exams.
  • In Search of Batching Insights: @rwamit reached out to the community for advice on the proper method to implement batching for querying GPT-4, comparing the significant difference in completion times between individual requests and batch processing.
  • Hugging Face Services Disruption: Several users, including @temperance6095 and @zlapo, discussed experiencing issues with the Hugging Face Inference API, noting prolonged 504 timeout errors across various model categories.
  • Community Collaborations: @dykyi_vladk and @beaudjango connected on the shared interest in machine learning project development, illustrating the collaborative spirit of the discussion channel.

Links mentioned:

HuggingFace ▷ #today-im-learning (1 messages):

  • Study Group Invitation for CS231n: User @shreesha1573 extended an invitation for a study group to learn CS231n: Convolutional Neural Networks for Visual Recognition this Spring 2023. They shared links to the course assignments and modules, indicating they’re open for direct messages to join the group. CS231n Course

Links mentioned:

CS231n Convolutional Neural Networks for Visual Recognition: no description found

HuggingFace ▷ #cool-finds (14 messages🔥):

  • AI Infrastructure Unicorns: @valeriiakuka highlighted Scale AI’s journey to becoming a key player in the data labeling market, directing users to the Turing Post article detailing their growth and challenges. Scale AI has reached a valuation of $7.3 billion with extensive funding rounds, evident in their 8th anniversary—details at Turing Post.

  • Lie Detection Through Language Models: @andysingal shared a link to a Nature article on verbal lie detection using large language models. The article discusses the psychological and content differences in deceptive vs. truthful narratives and explores lie detection methodologies—check out the research here.

  • VLMs Tackling Resolution Challenges: @osanseviero shared a blog post about overcoming resolution problems in vision-language models by using multiple crops of high-resolution images, with a live demo and the model available on HuggingFace’s Platform. Intrigued readers can learn more in Visheratin’s blog post.

  • Mistral Large Unveiled: @teadaniel announced the release of Mistral Large, a sophisticated language model excelling in multilingual reasoning tasks, now second only to GPT-4 and available for use through la Plateforme and Azure. It shows promising results on benchmarks, details of which can be found on Mistral’s news update.

  • Exciting Advancements in Image Upscaling and Enhancement: @furkangozukara introduced SUPIR Model V8 with improvements that allow operation on 12 GB GPUs like the RTX 3060 and is based on the Juggernaut-XL-v9 model. They argue it outperforms other upscale tools, including the expensive Magnific, inviting users to discover its capabilities in a YouTube tutorial.

Links mentioned:

HuggingFace ▷ #i-made-this (13 messages🔥):

  • Philosophy Q&A Generated by AI: @nabereon shared their process of generating a question-answer dataset using Mixtral-8x7B-Instruct-v0 and the AiresPucrs/stanford-encyclopedia-philosophy dataset. They included a specified structure to shape the outputs and plan to expand the dataset with consent from original content creators.

  • Licensing Concerns Discussed: @cakiki expressed licensing concerns regarding the compatibility of the SEP dataset and Mixtral’s terms of use. @nabereon is looking into the issue to ensure compliance with intellectual property rights.

  • Invitation to Comment on Open AI Model Weights: .plot highlighted an opportunity to give public comments on “open-weight” AI models to the NTIA, discussing the balance between democratization of AI and safety concerns.

  • Performance LLM Board Launch: @michal.swedrowski. introduced the initial version of the Performance LLM Board that compares various LLMs on engineering aspects like response times and pricing. They seek feedback to refine the board and provide relevant information to the community.

  • Imagic Paper Replicated and Explained: @chongdashu delved into replicating the Imagic paper, a technique for editing images with text prompts using diffusion models. They shared a Medium post discussing their experience with this novel image editing method.

Links mentioned:

HuggingFace ▷ #diffusion-discussions (2 messages):

  • Disappointment in Playground v2.5: User @pseudoterminalx expressed disappointment that Playground v2.5 is still utilizing eps prediction, implying expectations for a more advanced approach.
  • Critique of Framework Choices: @pseudoterminalx criticized the decision to mention zsnr only briefly while opting to use what they described as “the crappy EDM framework.”

HuggingFace ▷ #computer-vision (10 messages🔥):

  • Dataset Size Concern for Fine-Tuning: @icecoldt369 expessed concerns about not having a sufficiently large dataset, even when incorporating open source ones. @cursorop reassured that massive datasets aren’t necessary for fine-tuning, as models generally recognize patterns in complex words with smaller datasets.
  • Optimizing for Complex Character Recognition: Despite reassurances, @icecoldt369 indicated that their model wasn’t fully recognizing complex characters, hinting at the need for models better trained on such data, including foreign languages or meta learners.
  • Challenges with Language-Specific Characters: @icecoldt369 is seeking to work with complex characters unique to the Khmer language, whereas @cursorop noted the difficulty of this task due to the symbol-like nature of the language’s script.

HuggingFace ▷ #NLP (10 messages🔥):

  • Classification Head Confusion Cleared Up: User @grimsqueaker inquired about the specifics of the AutoModelForSequenceClassification when using esm2 from HuggingFace, querying if it utilizes a CLS token or mean pooling across all tokens. @cursorop suggested checking the HuggingFace repository code for clarification.

  • Looking for Generative QA Models: @cornwastaken sought advice on finding models for Generative Question Answering with scenarios involving document-based inquiries, and wondered about the architecture used in such models. They have explored the Question Answering and text-generation-inference categories within the HuggingFace models for this purpose.

  • Quick Embedding Model Recommendations: @cakiki asked for suggestions on an embedding model for exploring a small, non-specialized English dataset, to which @cubietom recommended the BAAI’s bge-small-en-v1.5 as a fast and effective option, and also mentioned the possibility of using thenlper’s gte-small for more extensive language support and retrieval methods.

  • Condensing Emails for LLM Ingestion: @acidgrim mentioned their project on condensing email files to fit an LLM context window, seeking recommendations for CPU-only local libraries for retaining information integrity, mentioning that they are currently using suma but are open to alternatives.

  • Developing a Medical Transformer: User @kareem3069 is looking to build a medical transformer to improve model mapping for domain-specific terminology and contexts but faces challenges with the performance of existing sentence-encoder models. They are seeking suggestions on approaches to enhance transformer capabilities for accurate medical code descriptions.

Links mentioned:

HuggingFace ▷ #diffusion-discussions (2 messages):

  • Disappointment in Playground v2.5: @pseudoterminalx expressed disappointment that Playground v2.5 is still utilizing eps prediction, indicating expectations for a different approach.
  • Criticisms on ZSNR and EDM Framework: @pseudoterminalx also criticized the treatment of zsnr, stating it is only mentioned as a footnote and lamented the use of what they referred to as the “crappy EDM framework”.

Eleuther ▷ #general (48 messages🔥):

  • Unit Norm Vectors in Machine Learning: @smerkyg and @ad8e discussed the concept of ‘unit style’ vectors in machine learning, highlighting the importance of RMSNorm for achieving ‘unit length’ vectors for precision reasons. @ad8e explained that consistent scaling across different parts of a model avoids the need for conversion between scales.

  • DPO vs SFT in Preference Datasets: A discussion about the Direct Preference Optimization (DPO) paper entailed confusion over the initialization of model_ref. @staticpunch inquired about initializing model_ref by performing MLE on preferred completions, and @elad7318 confirmed the understanding but questioned why only preferred completions were considered.

  • Mistral Large Unveiled: A new language model, Mistral Large, was announced on Mistral’s news page, presented as a cutting-edge model with strong results on benchmarks, now available via la Plateforme and Azure.

  • LangChain Wrapper Batch Processing Dilemmas: @rwamit sought advice on implementing batching to query GPT-4 using the langchain wrapper. The community, including @._bob_, engaged briefly before directing the conversation to another platform for more appropriate assistance.

  • Training Oddities With Model Epochs: @jstephencorey encountered an anomaly where training loss unexpectedly increased or stagnated during the fifth epoch when training a Pythia-70m architecture model to overfit. Members like @leegao_ and @hawk1399 contributed insights citing possible causes ranging from double descent behavior to the nuances of training loss spikes, while @catboy_slim_ reminded of the disruptions repeat data can cause during training.

Links mentioned:

Eleuther ▷ #research (22 messages🔥):

  • Efficiency Boost with Weight Sharing: @thooton_ highlighted the performance of a weight-shared model version with double the layers nearly matching the original, emphasizing its limitation to smaller models whose layers fit in SRAM.
  • The Simplest Hardware Efficiency Solution: .the_alt_man asserted that weight sharing is a straightforward strategy to enhance the efficiency of existing architectures on current hardware.
  • Unearthing the Calibration Paper: @avi.ai shared a paper link discussing the well-calibrated confidence of Large Language Models (LLMs) and the broad evaluation of RLHF-LMs’ confidence extraction methods.
  • The LLM Radioactivity Detection Research: @0x_paws introduced research on detecting the use of LLM-generated texts in training data using watermarking methods, which @leegao_ discussed in detail, noting the study’s findings on trace-radioactivity and its detectability without model weights or training data knowledge.
  • Grokking Goes Mainstream: Recent work shared by @hawk1399 suggests the grokking phenomenon is not limited to controlled settings, but also occurs in practical scenarios with CNNs and Resnet on common datasets, introducing the term ‘delayed robustness’.

Links mentioned:

Eleuther ▷ #interpretability-general (6 messages):

  • Twitter Link Shared Without Context: @main.ai shared a Twitter link without any further explanation or context.
  • Clarity Sought on Energy Results: @butanium asked <@177739383070261248> for an interpretation of “energy results” with the tuned lens, but no detailed response was provided.
  • Weekend Plans to Review a Paper: @mrgonao mentioned they would spend more time over the weekend to understand the paper related to the previous question about energy results.
  • Confusion Over ‘Energy’ in Research: @mrgonao admitted to not understanding why the term “energy” is used in the paper’s equation, indicating a lack of intuition about what it refers to.
  • Commitment to Further Investigation: @butanium promised to look into the matter further the next day.

Eleuther ▷ #lm-thunderdome (50 messages🔥):

  • Praise for lm-harness: @hailey_schoelkopf expresses delight on finding a citation of EleutherAI’s lm-harness in a paper, appreciating its pivotal role for few-shot evaluation of autoregressive language models.
  • Replicating Leaderboard Results: @hailey_schoelkopf provides a detailed guide on how to replicate Open LLM Leaderboard results using a specific commit of lm-evaluation-harness, including configurations for different tasks and models, which was pinned for frequently asked questions (FAQ).
  • Interest in Improving lm-eval-harness API: @ariel2137 inquires about enhancing the code-level usage of the lm-eval-harness, to which @hailey_schoelkopf responds positively, inviting feedback to smooth out any rough edges in the experience.
  • Clarification on lm-eval-harness Output Interpretation: @micpie seeks to understand the output format of lm-eval-harness, and @hailey_schoelkopf clarifies that the true and false values indicate whether the target string would be the greedy completion for each answer choice.
  • Selection and Override of Data Splits in lm-eval-harness: @micpie is puzzled by how the harness selects and counts data splits for evaluation. @baber_ assists with the explanation that the reported number also accounts for multiple choice options, and @hailey_schoelkopf confirms that command line overrides for splits are unavailable, suggesting yaml file edits instead.

Eleuther ▷ #gpt-neox-dev (8 messages🔥):

  • DeepSpeed Config Confusion: User @jdranpariya encountered a ValueError when trying to disable DeepSpeed using "deepspeed": false in the config while using deppy.py train.py. The error mentioned an issue with NeoXArgs validation.
  • Optimal Data for Mistral Tokenizer: @rand0mm inquired about the most optimal data source to extend the Mistral tokenizer for better representation of other languages.
  • In Search of Multi-Node Training Setup: @jdranpariya sought guidance for setting up a multi-node training environment for GPT-NeoX on CoreWeave Infrastructure, specifically aiming to employ 2 nodes with 4 GPUs using slurm or MPI.
  • Kubernetes a Must for CoreWeave?: With questions about whether Kubernetes can be bypassed for multi-node training on CoreWeave, @jdranpariya was open to suggestions and seeking the correct platform or channel for more specialized help.
  • Guidance for CoreWeave and Slurm Clusters: @catboy_slim_ advised @jdranpariya to refer to CoreWeave for specific infrastructure questions and to NeoX docs for instructions on launching with slurm, while also mentioning that setting up a slurm cluster falls under CoreWeave’s purview.

CUDA MODE ▷ #general (6 messages):

  • Potential GTC Meetup Brewing: @vim410 is looking forward to possibly attending GTC in-person and is proposing the idea of creating a dedicated channel for those who are going.
  • LaTeX in Discord: @marksaroufim illustrated the formatting of LaTeX code in Discord by using the example x^2.
  • Excitement for Upcoming GTC Talk: @joseph_en is keen to attend the talk on Thursday at GTC and is excited to meet up with fellow cuda_mode members.
  • Inquiry about CUDA Emulators: @jash403 is reaching out to the community with a question on experience with creating or running Emulators on CUDA GPUs.
  • Link to CUDA-based Gameboy Emulator: In response to @jash403’s question, @iron_bound shared a GitHub repository for a CUDA Gameboy emulator and a related article detailing the project.

Links mentioned:

CUDA MODE ▷ #triton (2 messages):

  • Shoutout to Unsloth’s QLoRA Kernels: @andreaskoepf highlighted the efficiency of Unsloth kernels, mentioning their GitHub repository for 5X faster 60% less memory QLoRA finetuning available at GitHub - unslothai/unsloth. The link preview showed the repo’s description, image, and title.
  • Link to Custom Triton Kernel Integration Guide: @marksaroufim cross-posted from another channel a method for integrating custom triton kernels with torch.compile. However, the message link did not contain a preview.

Links mentioned:

GitHub - unslothai/unsloth: 5X faster 60% less memory QLoRA finetuning: 5X faster 60% less memory QLoRA finetuning. Contribute to unslothai/unsloth development by creating an account on GitHub.

CUDA MODE ▷ #cuda (15 messages🔥):

  • CUDA and Graphics APIs: @morousg highlighted that NVIDIA focusses more on CUDA-Vulkan interoperability than CUDA-OpenGL and mentioned the efficiency benefits of Vulkan over OpenGL, especially for graphics-heavy applications.

  • Specs for Library Testing Unveiled: @zippika shared their system specifications as a NVIDIA 4090 GPU with CUDA 12.3, pytorch 2.2+, pytorch-cuda=12.1, and an AMD 7950x CPU while addressing improvement queries about their library.

  • GPU Utilization Insight: @zippika observed a significant increase in GPU utilization — from 60% with bnb layers to approximately 80% with TorchFP4Linear layers.

  • Code for Layer Replacement in PyTorch Models: @zippika provided an update to their library, introducing options like only_replace_bnb_layers and ignore_layer_names for selective layer replacement, along with a Huggingface example script for implementation.

  • Exploring GPU Memory Latencies: @marksaroufim shared a paper that examines memory access latencies of NVIDIA A100 using PTX. Subsequent discussions involved @zippika, @jeremyhoward, and @cudawarped touching upon comparisons of L2 cache and global memory latencies, as well as the notion that higher bandwidth for L2 cache might justify similar latency to global memory.

Links mentioned:

CUDA MODE ▷ #torch (17 messages🔥):

  • Seeking Speedier Compilation: @briggers shared insights on reducing cpp_extension.load_inline compile times from >30s to >2s by avoiding unnecessary header files and separating source files. The gist hinges on opting for cpp_extension.load and strategically targeting device architecture, specifically for his NVIDIA 4090 (Faster cpp_extension compilation code example).

  • From PyTorch to Torch Origins: Prompted by a query about the relation between Torch and PyTorch, @marksaroufim linked to a blog post detailing the history of PyTorch, which began as Torch7 with its roots in the Torch7 contributors’ community dating as far back as 2010 (PyTorch design origins).

  • Custom Triton Kernels with Torch: @marksaroufim guided developers working on custom triton kernels integration with torch.compile to a PyTorch GitHub example. Issues can be reported by opening an issue on GitHub and tagging @oulgen with the aim of improving the upcoming official release of this feature for PyTorch 2.3 (Triton kernels integration with torch.compile).

  • Deciphering Compiler Limitations: In response to a question about compilers generating FA, @marksaroufim shared Tri Dao’s insight from an OpenReview forum, mentioning that compilers tend to struggle with optimizations that require mathematical rewrites while maintaining numerical stability (Tri Dao’s comment on compiler-generated FA).

  • Enthused Emergence of Andreas Köpf: The presence of Andreas Köpf, a significant contributor to PyTorch’s development, was warmly welcomed by @andreaskoepf in the community, signaling enthusiasm for his engagement and potential knowledge sharing.

Links mentioned:

CUDA MODE ▷ #algorithms (7 messages):

  • Insight into Improved Softmax Performance: @marksaroufim shared a link to a paper (Efficient Softmax Approximation on GPUs) that elucidates the softmax trick in flash attention, highlighting a specific local correction technique (e ^ {m_{j-1} - m_j}) that maintains the global softmax operation.
  • Exploring the Base2 Trick: @andreaskoepf pointed out a GitHub link (softmax_base2_trick.ipynb) explaining the base2 trick for normal softmax, which is also applicable to incremental softmax as implemented in OpenAI’s Triton example (06-fused-attention.py).
  • Backward Pass Automation in Triton: @marksaroufim mentioned the clunky nature of writing backward passes in Triton, hinting at the desire for automatic generation of backward passes from a given forward pass.
  • Minor Performance Gains through Compiler Optimization: In response to @marksaroufim expressing that compilers should handle certain optimizations, @andreaskoepf agreed, noting that while the compiler could optimize some aspects, being aware of certain tricks (like those in the Triton example) is beneficial when examining kernel code.
  • Quake3 Algorithm Recalled in Discussion: @iron_bound brought up the historical fast inverse square root trick used in the video game Quake3, sharing the Wikipedia overview (Fast inverse square root) that emphasizes the cleverness of such optimizations in computational algorithms.

Links mentioned:

CUDA MODE ▷ #jobs (2 messages):

  • NVIDIA Seeking CUDA and C++ Gurus: @vim410 confirmed that Nvidia is indeed looking for experts in CUDA and C++. Interested candidates should DM their CV with reference to JobID: JR1968004.

CUDA MODE ▷ #beginner (6 messages):

  • Introduction to CUDA Mode: User @ilovepython3 expressed interest in learning about CUDA Mode and mentioned a background in Python and C, with aspirations to fine-tune AI models.
  • Python Aspirant with a Math Concern: @ilovepython3 admitted to struggling with math while being capable of writing simple programs.
  • Seeking CUDA Mode Prerequisites: @ilovepython3 inquired about any prerequisites needed for engaging with CUDA Mode, such as knowledge of PyTorch.
  • Stepping Into AI with Uncertainty: @ilovepython3 expressed a desire to do AI-related work but is uncertain if diving into CUDA Mode is too complex at this stage.
  • Guidance Offered for AI Beginners: @jeremyhoward recommended that @ilovepython3 should first complete the fast.ai course before tackling CUDA Mode to gain comfortable footing with the material.

CUDA MODE ▷ #smol-hw (2 messages):

  • Clarifying the AO Acronym: User @mr.osophy inquired about the acronym AO, to which @marksaroufim clarified it stands for Architecture Optimization, although conceding it’s not the best name.

CUDA MODE ▷ #ring-attention (13 messages🔥):

  • Synchronization Tactics in CUDA: @zhuzilin96 suggested a more straightforward method to record total communication time by adding for req in reqs: req.wait() after communication starts, allowing overlap with computing kernels.
  • Benchmark Adjustments Improve Speed: Adjusting the seqlen in the benchmark increased the performance for zigzag ring to about 8.5 times faster than flash_attn, according to @zhuzilin96.
  • Flash Attention Meeting on the Horizon: @jamesmel inquired about a flash attention paper reading/discussion, and @iron_bound confirmed a meeting scheduled at <t:1708966800>.
  • Flash Attention Paper Recommended: @w0rlord provided a link to an arXiv paper detailing an optimized implementation of FlashAttention-2 on NVIDIA’s Hopper architecture, recommending it for its thorough annotations and algorithm details.
  • Collaborative Ring and Flash Attention Notebook: @ericauld shared a work-in-progress Colab notebook on ring attention and flash attention, inviting feedback and improvements from the community.

Links mentioned:

LangChain AI ▷ #announcements (1 messages):

  • Seeking Feedback on Structured Output Interface: User @bagatur encouraged the community to provide feedback on a proposed intuitive interface for obtaining structured outputs from models. They shared a discussion on GitHub which outlines the idea and aims to simplify user experience for most LLM tasks.

Links mentioned:

RFC: LLM structured output interface · langchain-ai/langchain · Discussion #18154: Getting structured outputs from a model is essential for most LLM tasks. We need to make the UX for getting structured outputs from a model as simple as possible. Our current idea is to add a ChatM…

LangChain AI ▷ #general (37 messages🔥):

  • Introducing UseScraper for Content Crawling: @dctanner launched UseScraper.com, which scrapes a website’s content into markdown or JSON, and shared a blog post on using it with LangChain for RAG.
  • LlamaCpp Python Integration Guide: @ldeth256 referred to a guide on integrating llama-cpp-python within LangChain for inference with various LLMs.
  • Discussion on Efficient Logo Creation Platforms: @bru.leo sought recommendations for AI platforms that create professional-looking logos, expressing dissatisfaction with current tools tried.
  • Seeking Assistance with Streaming Q&A Chains: @shadizx followed this streaming Q&A guide but needed help with object extraction in the context.
  • Feedback Requested for Ragas and Validate.ai Integration: @locus_5436 shared a personal project that integrates Ragas into a visualization web app for evaluating RAG systems, available for feedback at validate.tonic.ai.

Links mentioned:

LangChain AI ▷ #langserve (1 messages):

howtonotgiveafuck: Hi all, is there anyway to extend the timeout beyond 900 seconds?

LangChain AI ▷ #langchain-templates (1 messages):

  • Inquiry on Function Calling in Open Source Chat Models: @sectorix is seeking a working solution for function calling within chat completions, especially for open-source models like Mistral. They mention the expectation that Ollama will add this functionality but are looking for interim solutions.

LangChain AI ▷ #share-your-work (5 messages):

  • R2R Framework Unveiled for RAG Systems: @emrgnt_cmplxty announced the launch of R2R, a framework designed for the rapid development and deployment of production-ready RAG systems. R2R offers a semi-opinionated system that simplifies the process of deploying, adapting, and maintaining RAG pipelines in production. Details can be found on their GitHub page.

  • Visualize LLM RAG Results with Tonic.AI: @locus_5436 shared a personal project — a free visualization web app with Ragas integration to assist in evaluating LLM RAG systems. Feedback is welcome on the app, which is accessible at validate.tonic.ai.

  • IntelliDoctor.ai Launches for Medical Professionals: @robertoshimizu revealed the launch of IntelliDoctor.ai, an AI-driven platform for medical professionals that provides evidence-based clinical answers using advanced prompt engineering and RAG. The platform, which leverages open academic sources for insights, credits langchain and langsmith for infrastructural support and inspiration. Visit intellidoctor.ai for more details.

  • LangGraph Elevates Code Generation: @andysingal introduced LangGraph, a tool that enhances code generation by amalgamating LangGraph’s iterative code generation and correction capabilities with Langchain’s security and integrity features in the full blog post on AI Advances.

  • Reminder of Channel Purpose: @theepic.dev reminded chatters that the “share-your-work” channel is not for support queries.

Links mentioned:

LangChain AI ▷ #tutorials (2 messages):

  • LangGraph Powers Up Chatbots: @tarikkaoutar shares a new approach to creating multi-agent applications using LangGraph, function calls, and web scraping, aimed at the Python and Data Science community. Watch the full explanation on YouTube in a video titled “LangGraph + Function Call + Web Scraper = Multi-Agent Application”.

  • AI as Your Conversation Sidekick: @jasonzhou1993 introduces a novel concept: a real-time conversation AI co-pilot for mobile phones. Explore this technology through the YouTube video, “Real time AI Conversation Co-pilot on your phone, Crazy or Creepy?”.

Links mentioned:

OpenAccess AI Collective (axolotl) ▷ #general (32 messages🔥):

<ul>
  <li><strong>Mistral-EU Partnership Raises Eyebrows</strong>: `@yamashi` expressed skepticism about Mistral’s commitment to open-source, suggesting that their partnership deal with Microsoft confirms a focus on profits. Meanwhile, `@casper_ai` shared a <a href="https://fxtwitter.com/casper_hansen_/status/1762159643344662859">link</a> indicating MistralAI's CEO commitment to open-weight models.</li>
  <li><strong>Strategic Leaks Mirror Reality</strong>: `@casper_ai` acknowledged that Mistral’s strategy to release smaller models while keeping larger ones platform-gated aligns with previously leaked plans.</li>
  <li><strong>Llama 3 Anticipation Grows</strong>: Both `@yamashi` and `@noobmaster29` looked forward to Llama 3, hoping for innovations beyond simply scaling up data and looking forward to potential multilingual improvements and enhancements like MoE Mamba.</li>
  <li><strong>LoRA Limitations Discussed</strong>: `@enka55` sought information on using LoRA for knowledge integration, to which `@nruaif` and `@leoandlibe` responded that full fine-tuning, not LoRA, is suited for adding knowledge. Further, `@lee0099` shared a <a href="https://arxiv.org/pdf/2304.08109.pdf">research paper</a> examining LoRA's potential for knowledge transfer.</li>
  <li><strong>Hardware Constraints Inform Model Utility Perceptions</strong>: `@nafnlaus00` shared a pragmatic view on model accessibility, noting the impracticality for average users to run very large models due to hardware constraints.</li>
</ul>

Links mentioned:

OpenAccess AI Collective (axolotl) ▷ #axolotl-dev (6 messages):

  • Gemma Optimizations Achieve New Heights: @nanobitz announced significant improvements with Gemma models running in Unsloth; Gemma is 2.43x faster than using Hugging Face and FA2, 2.53x faster than vanilla Hugging Face, and operates with 70% less VRAM. They also provided links to Gemma 7b Notebook and Gemma 2b Notebook for free usage on a Tesla T4.

  • Seeking Documentation Links in Pull Requests: @caseus_ inquired about a documentation reference as a result of a GitHub pull request discussion, to which @yamashi responded with the relevant documentation.

  • Quality before Speed in Model Improvements: @dreamgen emphasized the importance of model accuracy over speed, suggesting that getting the model correct should be the priority.

  • Introducing LoRA-the-Explorer for Efficient Training: @caseus_ shared Minyoung Huh’s research on LoRA-the-Explorer (LTE), an approach for training neural networks from scratch with parallel low-rank adapters, which could have significant implications for resource-efficient deep learning. They also highlighted the multi-head LoRA implementation as worthy of interest.

Links mentioned:

OpenAccess AI Collective (axolotl) ▷ #community-showcase (1 messages):

  • R2R: Making RAG Deployment Easy: User emrgnt_cmplxty announced the launch of R2R, a semi-opinionated framework for rapid development and deployment of production-ready RAG systems. Designed to simplify and streamline deployment, it’s an industry move towards practical ease-of-use and effectiveness, and can be found on GitHub - SciPhi-AI/R2R.

Links mentioned:

GitHub - SciPhi-AI/R2R: A framework for rapid development and deployment of production-ready RAG systems: A framework for rapid development and deployment of production-ready RAG systems - SciPhi-AI/R2R

OpenAccess AI Collective (axolotl) ▷ #replicate-help (1 messages):

  • Replicate’s Focus Called into Question: @dreamgen expressed surprise, noting an expectation that replicate should have a better performance given their years of focus on it, yet apparently, it does not meet those expectations. There was no further elaboration on specific issues or comparisons.

Latent Space ▷ #ai-general-chat (33 messages🔥):

  • Clarification on Comparing AI Models: @eugeneyan pointed out in response to @guardiang that the tweet thread comparing models to GPT-4 was based on zero-shot performance metrics.
  • Mistral Large Announced with Microsoft Partnership: @__chef__ shared a link to Mistral Large’s announcement, detailing its new capabilities and benchmark performance, alongside news of a partnership with Microsoft.
  • Cloudflare’s AI Gateway Draws Attention: @henriqueln7 showcased Cloudflare’s AI Gateway, emphasizing its analytics, logging, and caching features, with its ease of use highlighted as needing just one line of code to get started.
  • Mistral Au Large Integral to RAG: @ashpreetbedi reviewed the RAG integration with Mistral Au Large, praising its function calling and reasoning capabilities. They also shared a link to their cookbook on GitHub: phidata/mistral.
  • Announcing an eBook on RAG: @dimfeld notified the chat about Jason Liu’s upcoming eBook on RAG with varying levels of complexity, with the GitHub repo found here: n-levels-of-rag. It was acknowledged as a useful resource by @thenoahhein for a Twitter data summarization task.

Links mentioned:

Datasette - LLM (@SimonW) ▷ #ai (4 messages):

  • Language Surprise in Data Extraction: @derekpwillis experienced an amusing issue where chatgpt-3.5-turbo used Spanish titles for documents that should have been in English, such as translating “Taking Advantage of the Internet” to “Sacándole Provecho a Internet” in the extracted version.

  • Multilingual Mix-up Reminiscent of ChatGPT/Whisper Voice Bug: @simonw related the issue to a similar bug where ChatGPT combined with Whisper sometimes misinterprets a British accent as Welsh and responds in Welsh.

  • Suggested Fix with System Prompt: @simonw suggested a potential fix by feeding the system a prompt that instructs it to “Always use English” to avoid language mix-ups.

  • Resolution to Implement Language-Specific Prompt: In response to the suggestion, @derekpwillis acknowledged the issue and indicated he would implement the suggestion to always use English.

Datasette - LLM (@SimonW) ▷ #llm (26 messages🔥):

  • LLM Plugin Brings Groqcloud to Python Devs: @angerman. shared that the LLM plugin for accessing Groqcloud models is now available, providing instructions for installation and how to obtain an API key. The plugin allows use of models like groq-llama2 and groq-mixtral, with an example given for generating pet names.
  • Python Packaging Tutorial for the Uninitiated: @0xgrrr provided @angerman. with a guiding tutorial on how to package and upload a Python project to PyPI, encouraging them that the process is straightforward.
  • LLM Plugin Improves with Streaming Support: An update from @angerman. mentioned new streaming functionality for the LLM plugin, expressing gratitude to @746595581086138409 for their contribution to this enhancement.
  • Datasette Developer Hints at Upcoming Chat UI: @simonw revealed that a chat UI for LLM is in the works, although it’s still a work-in-progress with no set completion date.
  • Datasette’s Choice: Why Fly GPUs?: @simonw explained to @kiloton9999 that the choice of using Fly GPUs is due partially to their company sponsoring Datasette development, and also because their GPUs have the ability to scale to zero.

Links mentioned:

DiscoResearch ▷ #general (9 messages🔥):

  • Insights on LLM Evaluation Methods: @bjoernp shared an interesting paper discussing the limitations of current probability-based evaluation methods for Large Language Models (LLMs), touched on in the DiscoLM series. The paper critiqued how this evaluation strategy often doesn’t align with generation-based predictions but did not fully explore why these discrepancies occur.

  • Hidden Null Strings in JSON to Parquet Conversion: @thomasrenkert encountered invisible null strings when converting JSON files to Parquet format, a problem only discerned by directly uploading the JSON to Hugging Face and observing the converted dataset.

  • Building RAG for Codebases: @codermickey enquired about the state of the art in creating a Retrieval-Augmented Generation (RAG) bot for answering questions and assisting with a codebase, including chunking strategies for indexing and retrieval strategies suitable for coding tasks.

  • LangChain’s Approach to Codebase RAG: In response to @codermickey, @johannhartmann mentioned that LangChain provides a Git loader and segmentation for popular languages, and LlamaIndex has a Git importer, with most people simply using retrieval with OpenAI embeddings and prompts for code-related tasks.

  • End-to-end Optimization Exploration for RAG and LLMs: @rasdani asked if there was research on joint end-to-end optimization of RAG and Large Language Models using gradients, linking a related paper LESS, although later noting it does not backpropagate through data selection but only retrieves training examples with similar precomputed gradient features.

Links mentioned:

DiscoResearch ▷ #benchmark_dev (13 messages🔥):

  • EQ-Bench Now “Sprechen Sie Deutsch”: @.calytrix has announced that German language support has been added to EQ-Bench, with the aim of making it faster and more economical for users. The updated benchmark is accessible on GitHub.

  • Scores Are In for EQ-Bench in German: Preliminary benchmark scores have been shared by @.calytrix showing GPT-4’s leading performance with a score of 81.91 on the German version, compared to its 86.05 on the English EQ-Bench.

  • Translation Quality in Question: @_jp1_ raises concerns about the German translation’s accuracy on EQ-Bench, suggesting nuances in emotions may not map directly between English and German.

  • Fluency Affects Emotional Benchmark Scores: @.calytrix reflects on the discussion, noting that while the German EQ-Bench tests for German fluency, it may not capture additional aspects of emotional intelligence when compared to the English version.

  • Polish Translation Reveals Model Gaps: @remek1972 reports on translating MT-Bench to Polish, highlighting significant differences in model performance on tasks between Polish and English, suggesting language fluency of models like GPT-4 is a critical factor.

Links mentioned:

GitHub - EQ-bench/EQ-Bench: A benchmark for emotional intelligence in large language models: A benchmark for emotional intelligence in large language models - EQ-bench/EQ-Bench

DiscoResearch ▷ #discolm_german (1 messages):

thomasrenkert: thanks for the explanation 🙂

LLM Perf Enthusiasts AI ▷ #opensource (5 messages):

  • FireFunction V1 ignites the scene: @sourya4 highlighted FireFunction V1, a new model boasting GPT-4-level structured output and decision-routing at a significantly lower latency, noting its open-weights and commercial usability. They shared FireFunction’s blog post and mentioned the model’s JSON and grammar modes that help maintain output structure.
  • Exploring the Best in Function Calling: @yikesawjeez listed Gorilla OpenFunctions, NexusRaven, and Litellm Function Calling Wrapper as their current go-to solutions for function calling. They also hinted at their participation in an upcoming hackathon focused on Fire Functions.
  • Mixtral Large: Has anyone ventured yet?: @thisisnotawill inquired whether anyone has tried out Mixtral Large, but there were no responses provided within the chat to elaborate on its use or performance.
  • Inquirying Minds Want to Know: @justahvee questioned the specifics about response latency in FireFunction V1, asking whether it refers to time to first token or time to completion, comparing it to GPT-4’s longer times.

Links mentioned:

Tweet from Lin Qiao (@lqiao): 🔥 Structure is all you need. 🔥 We’re excited to announce: - FireFunction V1 - our new, open-weights function calling model: - GPT-4-level structured output and decision-routing at 4x lower lat…

LLM Perf Enthusiasts AI ▷ #offtopic (4 messages):

  • R2R Sets New Industry Benchmark: @emrgnt_cmplxty announced the launch of R2R, a framework for rapid development and deployment of production-ready RAG systems, which aims to simplify the transition from experimental models to production environments. Explore R2R at GitHub - SciPhi-AI/R2R.

  • Seeking Clarity on Frameworks: @yikesawjeez sought a TL;DR on the differences between the new R2R framework and agentmemory found on GitHub, asking if there’s a difference between the two.

  • Applause for a Twitter Video: @yikesawjeez commended an unspecified absolute banger of a twitter vid, acknowledging the good work of a fellow user.

Links mentioned:

LLM Perf Enthusiasts AI ▷ #collaboration (2 messages):

  • GPT-4 Tackles Drug Information: User @thebaghdaddy found success in getting GPT-4 to generate informative cards by organizing data into a table covering aspects like mechanisms, side effects, and main disease targets for a list of drugs. Despite being slightly verbose, the method proved effective.
  • Anki’s Image Occlusion Feature Missed: @thebaghdaddy highlighted a limitation of GPT-4 not being able to include images in outputs, underscoring the utility of Anki’s image occlusion for studying.

LLM Perf Enthusiasts AI ▷ #speed (6 messages):

  • The Pain of Latency: @res6969 expressed distress, saying this deeply pains me without specifying the issue.
  • Latency by the Seconds: @res6969 highlighted concerns about latency in seconds for OpenAI APIs, suggesting it’s an issue.
  • Looking for a Solution in Hosting: @res6969 inquired whether dedicated hosting is the only fix for the latency issues.
  • Azure Hosting Disappoints: @pantsforbirds joined the conversation, mentioning that the Azure results are disappointing, implying dissatisfaction with Azure’s performance in this context.

LLM Perf Enthusiasts AI ▷ #rag (1 messages):

  • RAG System Enhancement Brainstorm: @jxnlco shared their work on ways to improve a client’s RAG system, inviting feedback. Check out the ideas and contribute on GitHub.

Links mentioned:

n-levels-of-rag/README.md at main · jxnl/n-levels-of-rag: Contribute to jxnl/n-levels-of-rag development by creating an account on GitHub.

Alignment Lab AI ▷ #oo (6 messages):

  • Gemma Enhanced with Turn Tokens: @imonenext has integrated <start_of_turn> and <end_of_turn> tokens into the Gemma model, which is available for use on Hugging Face. These tokens are for facilitating further instruction/RL fine-tuning.
  • Manual Magic for Token Integration: The process of adding instruction-tuning tokens to the Gemma model was done manually by copying tokenizers, as clarified by @imonenext.
  • All Original, No Issues: In response to a query from @ufghfigchv, @imonenext specified that there were no issues encountered during the process because the original instruction tokens were used.

Links mentioned:

Skunkworks AI ▷ #off-topic (1 messages):

pradeep1148: https://www.youtube.com/watch?v=mw3VvbYE0o8

AI Engineer Foundation ▷ #events (1 messages):

  • Live Coding Session on Agent Protocol V2: @_z shared a YouTube live stream where they’re working on the Agent Protocol’s Config Options RFC for the V2 Milestone. The stream invites viewers to join and interact while _z codes.

Links mentioned:

Coding - Working on Agent Protocol V2 Milestone, Config Options, New RFCs: Hello, I’m Ziggy!I’m an Open Source Developer, gamer, and tech enthusiast. You can find me on GitHub at https://github.com/jzanecook Interested in contributi…