OpenAI Channel: api-discussions (6 messages🔥):

Summary of OpenAI Discord Chatbot Messages in the api-discussions Channel:

• Humanizing Text in AI Generation: @lodosd asked for prompt ideas to humanize text so it isn't detected by an AI. @lumirix advised using the AI to identify and modify elements that make the text sound non-human.
• Terms of Use: @lumirix reminded participants of OpenAI's terms of use, specifically the clause prohibiting the representation of AI-generated output as human-generated with a link to the OpenAI Terms of Use.
• Prompt Engineering Issue: @flaskie is struggling with a specific issue for prompt engineering where the task requires the GPT model to format solid text into logical order with specific heading levels and blocks of text containing 90-150 words. @eskcanta offered a detailed analysis of potential issues in Flaskie's prompt and suggested revisions.
• Requests for a Command Terminal in ChatGPT: @.pythagoras asked about how to command ChatGPT to open up the command terminal for listing specific functions.
• Assisting with Software Authoring: @taholmes160 shared an issue with using ChatGPT for authoring software, where the output includes unnecessary comment blocks. The possibility of GPT 'moods' affecting the quality of output was raised.
• Issues in Newest GPT Version: Both @.pythagoras and @taholmes160 discussed issues they've encountered with the GPT-4 Turbo, including output length limits and issues with organizing information.
• Considerations for Custom GPTs: @.pythagoras noted that while custom GPTs are interesting to work with, there are significant flaws in their functionality currently.
• Options Table in Markdown: @mnjiman proposed the idea of bringing up an 'option menu' in the form of a markdown table during a conversation with the AI.
• Tweaking AI Parameters: @no.iq raised a concern about the inability to find the tweaking parameters in the new UI for OpenAI, specifically mentioning the parameter 'temperature'.
• Constructing an Argument and Poem: @ertagon posted a complex task for the AI: to construct an argument denying the existence of tomatoes using biological reasoning and a little philosophy, and then turning the argument into a poem.

Guild: LangChain AI

LangChain AI Guild Summary

• Issues, inquiries, and discussions related to LangChain in the general channel:

  • Fishyccy reported certain prompts were causing the LangChain system to time out and was directed to s_papu_25 for assistance.
  • Vudumagic asked about OpenAI Vision API integration with LangChain.
  • Vj19 made a request for tool recommendations for scraping data from websites.
  • Coreyb42 advised data analysts should only learn LangChain if absolutely necessary, and suggested that GPTs might be more suited for non-developer roles.
  • Itscabral asked about creating a RecordManager using local files instead of SQLite.
  • .Cannaboss warned GPT-4 users about failures of direct retrieval with the new OpenAI Assistant API, and recommended using multiple learned "knowledge planes".
  • Jlcases reported changed importing routes in a tutorial and requested for updated ones.
  • Minecraftjuicer proposed a mechanism for automatically generating the perfect prompt for every question.
  • Iloveh8 asked about multimodal RAG.
  • Dennisyurk asked for advanced examples of agents/chains that interact with data warehouse.
  • .Trouble_ offered to pair up on working with LangChain in zoom calls.
  • Ethereon_ queried the possibility of chaining chains together in LCEL, providing a code example.
  • Jinwolf2 asked for assistance on an error with converting an API response.
  • Alimal asked about the best ways to use LangChain and/or APIs for Tabular Data QA.

• Discussion about the limitations of the semi-structured RAG template integration in the langchain-templates channel:

  • Alex_35579 highlighted issues about the integration of the chain.py's docstore and vectorstore with LangChain retriever for a prod app.
  • Alex_35579 also pointed out the limitation of the RAG's current system which does not support persistent data storage and search functionality.

• Projects shared on the share-your-work channel:

  • Arcypojeb shared a project about a hierarchical cooperative multi-agent framework that uses websockets for LLM communication.
    • This integration involves multiple AI platforms/frameworks/interfaces including Gradio, Chainlit, Tkinkter, and PySimpleGUI.
    • The project's purpose is to create a multi-purpose AI assistance platform that joins existing AI-driven applications/tools.
    • Shared GitHub, HuggingFace, and LangChain Multiagent cookbook links related to the project.

LangChain AI Channel Summaries

LangChain AI Channel: general (4 messages):

Summary of LangChain AI Discord Chatbot Messages:

• Issue with Certain Prompts Timing Out: @fishyccy reported that certain prompts were causing the system to time out and was directed to DM @s_papu_25 for assistance.
• Inquiry about OpenAI Vision API Integration: @vudumagic asked if the new openai vision api is part of langchain yet.
• Request for Tools or Integrations to Scrape Data from Website: @vj19 asked for recommendations about tools or integrations for scraping data from websites where he has to enter a date range and keywords in the searchbar and scrape the generated articles.
• Discussion about the Use of LangChain for Data Analysts not Developers: @coreyb42 advised that data analysts should only learn a heavy application framework like LangChain if absolutely necessary, suggesting that GPTs might be a better fit for non-developer roles.
• Query about `RecordManager` using Local Files: @itscabral asked if successful, it possible to create a `RecordManager` using local files instead of SQLite.
• Issue with Direct Retrieval Failures in the GPT-4 OpenAI Assistant API: @.cannaboss warned about failures of direct retrieval with GPT-4 using the new OpenAI Assistant API, recommending the use of multiple learned "knowledge planes" with trained weight tensors for specific tasks as the solution.
• Request for Updated Routes for Importing: @jlcases reported that routes for importing in a notebook he was following from some videos have changed and inquired for the updated ones.
• Discussion on Generating Perfect Prompts: @minecraftjuicer proposed a mechanism for automatically generating the perfect prompt for every question, providing a templated senior data scientist example. He expressed concern on feasibility and response time.
• Question about Multimodal RAG: @iloveh8 asked if anyone has ever done multimodal RAG.
• Inquiry about Advanced Examples of interacting with Data Warehouses: @dennisyurk inquired about more advanced examples of agents/chains that interact with data warehouses.
• Offer to Pair Up on LangChain: @.trouble_ offered to pair up on working with LangChain in zoom calls.
• Question about Chaining Chains in LCEL: @ethereon_ asked for confirmation on the possibility of chaining chains together in LCEL, providing a code example.
• Issue with Converting API Response: @jinwolf2 reported an error with converting an API response and asked for some assistance.
• Best way to use LangChain and/or APIs for Tabular Data QA: @alimal questioned what the best way to use langchain and/or APIs for tabular data QA.

LangChain AI Channel: langchain-templates (4 messages):

Semi-Structured RAG Template Integrations:

• @alex_35579 highlighted issues regarding the semi-structured RAG template, specifically about integrating the chain.py's docstore and vectorstore with Langchain retriever for a prod app. The user noted that the current setup loads the data on the fly and utilizes similarity search. However, the user pointed out that integrating real RAG requires persistent data storage which the current system doesn't support.
• Furthermore, @alex_35579 noted the limitation of the RAG template which loads the entire file content into its context without supporting search functionality. The user asserts that this isn't feasible for real-world applications and the issue also hinders any search that isn't dependent on vector/semantic search.

LangChain AI Channel: share-your-work (4 messages):

Hierarchical Cooperative Multi-agent Framework Project by arcypojeb:

• @arcypojeb is working on a project of a hierarchical cooperative multi-agent framework which uses websockets for Low Level Manager (LLM) communication. The aim of the project is to create a multi-purpose AI assistance platform that integrates existing AI-powered applications/tools.
• The user has managed to establish websocket connectivity between several AI platforms/frameworks/interfaces including Gradio, Chainlit, Tkinkter, and PySimpleGUI. The user has also created multiple interfaces for both clients and servers, and has established a basic mechanism of question-answering AI-driven logic.
• @arcypojeb shared links to the project on GitHub and HuggingFace for others to try out:
• Links:
  • https://github.com/CognitiveCodes/NeuralGPT/blob/main/Chat-center/ServerMain.py
  • https://github.com/CognitiveCodes/NeuralGPT/blob/main/Chat-center/ChainlitCli.py
  • https://huggingface.co/spaces/Arcypojeb/ServerNeural
  • https://huggingface.co/spaces/Arcypojeb/QA-Docs-Chainlit-Langchain
  • https://github.com/langchain-ai/langchain/blob/master/cookbook/multiagent_authoritarian.ipynb

LangChain AI Channel: tutorials (4 messages):

As this message does not provide any information related to specific topics or discussions, no summary can be derived from it.

Guild: Nous Research AI

Nous Research AI Guild Summary

• Discussions surrounding handling and processing of Parquet Dataset, by retaining the dataset in parquet and storing images as blobs, and improvements on dataset handling process through batch processing and script development.
• Debates on how the dataset was originally constructed, with speculations on the use of unofficial workarounds or possible data leaks from Midjourney.
• Prospective discussions on the use of Language Model (LLM) to rephase the dataset's prompts into more instructional language.
• Acknowledgement of collaborative efforts in handling and processing the dataset, with division of workload and resource sharing between @tsunemoto and @yorth_night.
• Encounters on technical issues of RAM crashes during resource-intensive processing, and proposed solutions like writing processed dataframes instead of end-task concatenation.
  • GitHub Midjourney Messages Dataset
  • Shared Solution (OpenAI chat link)
• Discussions on AI performance on specific tasks, the use cases for AI especially in audio classification, and queries on model licensing, closures and research-based restrictions.
  • Twitter link on AI performance
  • Twitter link on AI Audio Classification
• Communications on model merging strategies, technical updates on guidance-ai/guidance, price drops for Nous Hermes and OpenHermes APIs via OpenRouter, performance reporting for OpenHermes 2.5 Beta, and trouble shooting for the issues faced when inferring the Yi Capybara model.
• Basic interactions with new channel members.
• Informations and solutions on multiple issues like merging Qlora with PEFT, fast inference with OpenHermes, performance speed of AWQ model, sequence lengths for chatbots, structured output from models through fine-tuning and the feasibility of function calling for output structuring, and possibilities of fine-tuning a model for low-resourced languages.
  • GitHub link for merging Qlora
  • JSON grammar builder
  • Reddit link on fine-tuning for low-resourced languages
• Recommendations on the blog post "Do Things that Don't Scale" by Paul Graham, and a related post discussion on Hacker News shared by @_automagic.
  • Do Things that Don't Scale blog post
  • Hacker News post

Nous Research AI Channel Summaries

Nous Research AI Channel: off-topic (6 messages🔥):

Discussion & Progress on Parquet Dataset Handling:

• Improvements on Dataset Handling Process: @tsunemoto revealed an improved method for handling the dataset. He mentioned that retaining the dataset in parquet and storing all images as blobs, as well as integrating batch processing, significantly increased the process speed. He also developed a new script that allows exploring the blob'd parquet, including searching through it using regex and automatically converting the blobs back to images upon query.
• Discussion Around Dataset Origin: @tsunemoto and @yorth_night mused about how the dataset was initially constructed, as Midjourney doesn't provide an API. They speculated it could have been created through an unofficial workaround or possibly a data leak.
• Exploration of Future Possibilities: @yorth_night suggested that an LLM (Language Model) could potentially rephrase the prompts in the dataset to become more instructional, allowing for generation with natural language instead of merely tags. This idea was well-received by @tsunemoto, who noted it might aid in multimodal benchmarking.
• Collaborative Work and Shared Resources: Both @tsunemoto and @yorth_night decided to share the workload, each taking a different section of the dataset to process. They also agreed to share scripts and resources to accomplish the task more efficiently. @yorth_night selected parquet file `00001` and @tsunemoto worked on `00002`.
• Challenges and Proposed Solutions: @yorth_night reported that the resource-intense processing resulted in RAM crashes. @tsunemoto suggested possible solution such as writing the processed dataframes as they processed, rather than trying to concatenate them at the end. However, concerns were raised about potential data corruption if the process was interrupted mid-way.
• Links:
  • Original Discord Message
  • Midjourney Messages Dataset on Hugging Face
  • OpenAI Chat with a shared solution.
  • Archive.org

Nous Research AI Channel: interesting-links (6 messages🔥):

Discussion on AI Research and Benchmark Performance:

• AI Performance on Specific Tasks: @max_paperclips shared a tweet about AI performance on a specific task, saying that he suspects Transformers would be good at it but he also posited that its performance can likely be beaten with something smaller. @ldj expressed skepticism about the benchmark being a test of general logical reasoning, suggesting it's might be specifically tailored towards proofwriting.
• Use Cases for AI: @teknium challenged proposed use cases for AI in speech editing, stating that he found it to have even fewer use cases than vision. @tsunemoto pointed out potential in audio classification, using it to sort a large list of audio tracks into predefined categories.
• Models and Licensing: @_automagic and @teknium discussed the possible release of a model on HuggingFace, with @teknium noting that it is research only. @_automagic expressed disappointment if the model retains its closed license. @tokenbender joined the discussion suggesting that fine-tuned versions of the phi 1.5 model reason well, but expressed concern for those wanting to use it commercially.
• Links:
  • AI Performance on Task - Twitter by @max_paperclips
  • AI Audio Classification - Twitter by @tsunemoto
  • Research Video - Youtube by @joey00072
  • Big Brain AI - Twitter by @yorth_night
  • Skeleton of Thought Paper - arXiv by @georgejrjrjr

Nous Research AI Channel: general (6 messages🔥):

Summary of Hermes 2.5 vs Hermes 2 Performance Discussions:

• Model Merging Strategies: @alpindale explained the process of merging models like the Goliath-120B. This involves taking slices from different layer ranges from each model and stacking them on top of each other. Alpindale also mentioned a plan to finetune 120B so the layer stitches "heal" so to speak. More details here.
• Guidance-ai/guidance Updates : @mihai4256 reported a major redesign of guidance-ai/guidance, noting it now accepts GGUFs (llama.cpp). Visit GitHub Page.
• OpenRouter Price Drops: @alexatallah announced price drops for Nous Hermes and OpenHermes APIs over at OpenRouter. See Discord Invitation.
• OpenHermes 2.5 Beta Performance: @fullstack6209 reported 1.5-2.5 tokens per second performance on a $60 chromebook with OpenHermes 2.5 Beta using TheBloke's gguf.
• Issues with Inferring Capybara 34B: A discussion led by @yorth_night occurred about difficulties inferring the Yi Capybara model. @teknium provided advise about loading the model with LlamaModelForCausal and also provided a sample code for inference with HF Transformers.

Nous Research AI Channel: welcomes (6 messages🔥):

Welcome Messages Summary:

• SpaceCraft (DJ/VJ) greeted the channel with: "Blessings".
• teknium responded with a greeting of his own using a wave emoji: "Hi "

Nous Research AI Channel: ask-about-llms (6 messages🔥):

Discussion Summary of AI Models and Fine-tuning:

• Merging Qlora with PEFT: @yorth_night suggested a method for merging Qlora using PEFT, and shared a Github link here. @lightninglemons expressed interest in trying out this method.
• OpenHermes and Fast Inference: @ac1dbyte queried if OpenHermes allows fast inference with vllm, to which @teknium confirmed it should.
• Speed Issue with AWQ Model: @gabriel_syme reported an issue about AWQ model in vllm, which became twice as slow compared to the original model. He speculated it might be due to something with max_seq_len.
• Sequence Lengths for Chatbots: In continuation to AWQ model's performance issue, @teknium remarked that chatbots usually would not be intelligible past a 4-8k context, or around a max of 32k if Mistral has been trained on it.
• Fine-tuning for Structured Output: @ac1dbyte raised a requirement for his fine-tuned OpenHermes to produce outputs as a (semi-)structured JSON object for post-processing purposes. @teknium suggested starting with prompting before venturing into fine-tuning, and asked @ac1dbyte for specific input-output examples.
• Function Calling for Structured Output: @tsunemoto provided a different perspective, suggesting that function calling might also achieve @ac1dbyte's requirements if the model supports it. A JSON grammar builder that's llama.cpp compatible was shared here.
• Fine-tuning for Low-Resourced Languages: @4biddden inquired if fine-tuning a model on a low-resourced language dataset would enable the model to speak that language. @crainmaker and @giftedgummybee noted that continued pretraining might be needed, and it would also depend on the complexity of the language and the information in the dataset. @4biddden later shared a Reddit link here to support the discussion.

Nous Research AI Channel: memes (6 messages🔥):

Discussion on Scaling:

• Doing Things that Don't Scale: @teknium shared a link to a blog post by Paul Graham titled "Do Things that Won't Scale". You can read it here.
• Related Post on YCombinator: @_automagic shared a related discussion on Hacker News, specifically linked to a post with the id 36225723. The link to that discussion is here.

Guild: Alignment Lab AI

Alignment Lab AI Guild Summary

• Creating Budget Datasets for ML Models: A discussion initiated by @igoforth on the cost-effectiveness of producing datasets for finetuned machine learning models, comparing the costs of GPU time and using a GPT-4 generated dataset.
• Dialogue on Federated Learning with Adapter Methods: @erogol shared insights and resources about the benefits and capabilities of federated learning with adapter methods, especially for scaling model training and offering conveniences for parties who can update models but cannot share data.
  • NVIDIA Developer Blog on Federated Learning with Adapters
  • Big Science Workshop Petals GitHub Repo
• A request from @5811g, an MLE, seeking collaboration for a Techstars Startup Weekend event in San Francisco. The project idea focuses on using Large Language Models (LLMs) to automate operations for e-commerce businesses.
  • Techstars Startup Weekend San Francisco AI Tickets
• Discussion on Model Merging and Fine-tuning: @tural.sadik highlighted the variability in responses from separately fine-tuned models; @teknium shared insights on model merging, suggesting the 'original model' stays frozen until merged; @far_el emphasized that the base model remains unchanged.
• Conversation on CRLFT and Weighted Cross-Entropy in OpenChat: @turgutluk sought clarification on CRLFT's implementation in OpenChat, confirming its entails prompt conditioning and weighted CE, which was affirmed by @imonenext.
• Experiments with weights for GPT 3.5 and GPT 4 were proposed by @turgutluk and @imonenext. The recurring need to remind the model condition in multi-turn tasks was also pointed out.
  • MergeLM GitHub Repo

Alignment Lab AI Channel Summaries

Alignment Lab AI Channel: ai-and-ml-discussion (4 messages):

AI and ML Discussion:

• Creating Budget Datasets for ML Models: @igoforth initiated a discussion about the cost-effectiveness of producing datasets in relation to building finetuned machine learning models. They raised the comparison between the costs of GPU time and using a GPT-4 generated dataset.
• Federated Learning with Adapter Methods: @erogol shared two links about federated learning with adapter methods and its potential benefits, mentioning its capabilities for scaling model training and conveniences for parties unable to share data but can update models.
• Links:
  • https://developer.nvidia.com/blog/adapting-llms-to-downstream-tasks-using-federated-learning-on-distributed-datasets/
  • https://github.com/bigscience-workshop/petals

Alignment Lab AI Channel: looking-for-collabs (4 messages):

Request for Team Members for a Techstars Startup Event:

• @5811g is an MLE seeking individuals to join their team for the Techstars Startup Weekend event in San Francisco. The team's project idea involves using Large Language Models (LLMs) to automate operations for e-commerce businesses.
• Links:
  • Techstars Startup Weekend San Francisco AI Tickets

Alignment Lab AI Channel: general-chat (4 messages):

Discussion on Model Merging and Fine-tuning:

• Fine-tuned models giving different responses: @tural.sadik brought to attention that separately fine-tuned models (A and B) are likely to give different responses to the same input. They demonstrated this with the following code: model.generate(evalPrompt1), peftA.generate(evalPrompt1), peftB.generate(evalPrompt1).
• On merging models: @teknium clarified that they always merge models, indicating the 'original model' stays frozen until merged or attached. They also pointed out that they've never inferenced a Lora that they attach, suggesting other users, including @563068096747798529, @1117586410774470818, and @317006433797537792, for better advice.
• Unchanging base model: @far_el emphasized that the base model remains intact, echoing @teknium's point on the 'original model' staying frozen.

Alignment Lab AI Channel: oo (4 messages):

Discussion on CRLFT and Weighted Cross-Entropy in OpenChat:

• Clarification on CRLFT in OpenChat: @turgutluk asked about CRLFT's implementation in the OpenChat codebase and paper, confirming it entails prompt conditioning and weighted CE. @imonenext affirmed this understanding.
• Experimenting with Weights: @imonenext suggested beginning with weights of 0.1 for GPT 3.5 and 1 for GPT 4 while @turgutluk is planning to test the approach on a single task data to text generation SFT dataset using GPT 3.5 (85%) and GPT 4 (15%).
• Conditioning and Templates: @imonenext advised that in multi-turn tasks, the model tends to forget the condition unless reminded each turn, stating that their template establishes this recurring reminder.

• Links:
  • https://github.com/yule-BUAA/MergeLM

Only 1 channel had activity, so no need to summarize...

Only 1 channel had activity, so no need to summarize...

The Ontocord (MDEL discord) Discord has no new messages. If this guild has been quiet for too long, let us know and we will remove it.

The AI Engineer Foundation Discord has no new messages. If this guild has been quiet for too long, let us know and we will remove it.

The Perplexity AI Discord has no new messages. If this guild has been quiet for too long, let us know and we will remove it.

Guild: YAIG (a16z Infra)

YAIG (a16z Infra) Guild Summary

• Discussion on Cloudflare Workers AI Integration with Mistral and Stable Diffusion by @stevekamman, highlighting the integration's upcoming debut.

  • Discord link to the announcement

• Dialogue about constructing an Open Source Infrastructure for large scale data ingestion designed for semantic search by @danme. The team expressed the need to connect with individuals dealing with knowledge graphs on a large scale, and are interested in exploring the issues faced when frequently embedding or dealing with massive quantities of data.

  • GitHub Link to the project

• Analysis of a recent Discord Outage by @zorkian offering insights into a blogging post, that discusses an outage which lasted approximately one hour.

  • Link to the Blog post

YAIG (a16z Infra) Channel Summaries

YAIG (a16z Infra) Channel: ai-ml (2 messages):

Discussions and Announcements:

• Cloudflare Workers AI Integration with Mistral and Stable Diffusion: @stevekamman mentioned that Mistral and Stable Diffusion will be coming on Cloudflare Workers AI.
  Links:
  • https://discord.com/channels/595317990191398933/1154819662161395742/1173496868811059200
• Open Source Infrastructure for Large Data Ingestion for Semantic Search: @danme conveyed that they are building an open source infrastructure to ingest large volumes of data for semantic search and are looking to talk to individuals working with knowledge graphs at scale. They also expressed interest in understanding the challenges faced by those who embed data frequently or in large amounts.
  Links:
  • https://github.com/dgarnitz/vectorflow