famous last words but it's a quiet day and everyone is heading out to neurips (so are we). Andrej called out sources of alpha [here](https://twitter.com/karpathy/status/1733968385472704548) and we're considering adding them, please shoutout what reddits/discords/anime pfp anons we should add.

swyx

[TOC]

Nous Research AI Discord Summary

  • Members expressed interest in attending NeurIPS and meeting up, with suggestions for future AI events in Australia. @richardblythman urged those interested in an open-source, decentralized AI project to reach out to them. Users shared their projects, like @cyborg_1552’s photo GPT AI tool and @pradeep1148’s introduction of StableLM Zephyr 3B.
  • User @gabriel_syme triggered interest around Mixtral by sharing a GitHub link. Performance comparisons between Mixtral and GPT-3.5 heated discussions. @mihai4256 unveiled their fine-tuned model, Pallas-0.2, available on Hugging Face. A [Youtube video](https://youtu.be/y9k-U9AuDeM? si=2X5j64_cdsdKwWEw) discussing open-source LLMs usage sparked brief reactions.
  • Both OpenHermes-2.5-neural-chat-v3-3-Slerp and Mixtral were the topic of hype for their performances, with debating dictates on the latter’s GPU requirement. Tools such as Tensorboard, Wandb, evalplus, llamahub were stated beneficial for fine-tuning and evaluating models. User experiences on model hosting platforms like Ollama and LM Studio were exchanged with contrasting opinions favoring both.
  • A robust conversation on MoE led by @gabriel_syme clarified why Mixtral, a model based on Mistral, is set apart from previous implementations. Discussions on fine-tuning LLMs suggested limited data requirements. The potential of Mixtral being competitive with GPT-4 after finetuning was proposed. @wlrd explained how open-source LLMs could be implemented, leading to the OpenHermes 2.5 - Mistral 7B model. Speculations on GPT-3.5 suggested it’s a 20B model and forecasted its soon open-source release. Inference optimization possibilities for ChatGPT touched on strategic batching, potential caching, and user base size.
  • The memes channel saw an array of emojis and memes shared by members for fun and communication. Specific interests in speakers like Yann and Karpathy were expressed. User @teknium amusingly delineated a character as being heavily concerned about x risk.

Nous Research AI Channel Summaries

▷ #off-topic (19 messagesđŸ”„):

  • NeurIPS Meetup:

    • @blue_matcha asked if anyone was attending NeurIPS in the hopes of meeting up, and @teknium said they might be available on a Thursday and Friday.
    • @gabriel_syme expressed disappointment in NeurIPS consistently being in the US, later revealing they are based in Australia. @gabriel_syme also proposed hosting an event in Australia the following year.

  • Open Source and Decentralized AI Co-founder Search:

    • @richardblythman is in search of a co-founder for a project in the open-source and decentralized AI space and asked anyone interested to DM them.

  • Interest in Australian AI Conferences:

    • @deki04 pointed out that there would be considerable interest in Australian-based AI conferences, recounting a well-attended in-person fastAI course held in Brisbane led by Jeremy Howard.

  • Photo GPT AI Development:

    • @cyborg_1552 mentioned the development of a tool using Stable Diffusion, offering to write a blog post if people are interested. They also provided a link to their github for those wishing to explore further.

  • Introduction of StableLM Zephyr 3B:

    • @pradeep1148 shared a YouTube video introducing StableLM Zephyr 3B, a large language model.

▷ #benchmarks-log (1 messages):

nonameusr: i think he used markdown

▷ #interesting-links (24 messagesđŸ”„):

  • Discussion about Mixtral and its Architecture: @gabriel_syme shared a GitHub link to MixtralKit – a toolkit for the mixtral-8x7b-32kseqlen model. @cyborgdream posted a twitter link, sharing that Mixtral outperforms GPT-3.5 in benchmarks even before fine-tuning. The subsequent discussion involved @nonameusr, @euclaise, and @chhillee debating the benefits and uniqueness of Mixtral’s Transformer-based architecture.

  • Release of New Fine-Tuned Model: @mihai4256 announced the release of their fine-tuned model, Pallas-0.2, hosted on Hugging Face. This model, a fine-tune of Tess-34B-v1.4, is designed for reasoning tasks and performs well with long system prompts.

  • Video about Open Source LLMs usage: @teknium shared a Youtube video answering the question “Should You Use Open Source Large Language Models?” @n8programs and @nonameusr gave one-word responses to the question, with conflicting opinions.

▷ #general (639 messagesđŸ”„đŸ”„đŸ”„):

  • Fine-tuning and Performance Discussions: Users discuss the fine-tuning and performance of several models, including Hermes 2.5, Mistral, and GPTs Agent. For instance, @nonameusr suggests that OpenHermes-2.5-neural-chat-v3-3-Slerp (also nicknamed “Slurpy”) outperforms the original Hermes in some regards but notes inconsistencies. Several users also discuss the performance of Mixtral (or Mixtral MoE), discussing topics like its GPU requirements and its behavior when quantized.

  • Model Hosting and Management Platforms: Multiple users compare their experiences using Ollama and LM Studio for hosting and managing AI models. While some users express a preference for Ollama, others point out that LM Studio may be more customizable and better support a wider range of models.

  • Compute and Training Resources: Users like @vatsadev and @gabriel_syme discuss their computing resources, with the discussion also touching on the potential of university resources.

  • Useful Tools: Discussion also touched on various tools like Tensorboard, Wandb, evalplus, and llamahub, which can be useful for fine-tuning, testing, and evaluating models.

  • New Models and Techniques: The channel saw mentions of new models and techniques, like ‘slerp’ (in the context of OpenHermes-2.5-neural-chat-v3-3-Slerp). Some users also speculate about the Mixtral and StripedHyena models and the potential for further improvements to them via fine-tuning or merging strategies. Finally, @ldj suggests that Mixtral’s method of choosing “experts” during its computation could influence its performance.

▷ #ask-about-llms (123 messagesđŸ”„đŸ”„):

  • Mixture of Experts (MoE) Discussion: Users @akhxl, @cyborgdream, and @gabriel_syme engaged in a conversation about MoE, with @akhxl initially expressing confusion about the sudden hype over a technique that’s been around for some time. @gabriel_syme provided an explanation, stating that previous implementations didn’t yield useful models and that Mixtral, based on Mistral, has shown practical utility.
  • Finetuning Large Language Models (LLMs): In a dialogue involving @akhxl and @gabriel_syme, clarifications about the amount of data needed for finetuning were offered. @gabriel_syme noted that recent advancements didn’t require substantial data to finetune a good model due to the quality of base models and expansive pretraining data availability. A discourse on the potential of Mixtral to perform comparably to GPT-4 after finetuning ensued with @cyborgdream predicting such an outcome.
  • Open Source LLMs Usage:@.plot and @wlrd held a conversation regarding the acquisition and implementation of open-source LLMs. @wlrd pointed out that the models’ weights are open-sourced and can be fetched from Hugging Face and gave an example link to the OpenHermes 2.5 - Mistral 7B model.
  • GPT-3.5 Turbo Discussion: A nuanced discussion over the GPT-3.5 Turbo specifications occurred, primarily involving @cyborgdream,@agcobra1, and @n8programs. The discourse ranged from its performance compared to both smaller and larger models, with @cyborgdream suggesting the model is possibly a 20B model, basing on the leaked G3PO information and predicting its open-source release soon.
  • Inference Optimization for ChatGPT: User @zohad_sikder initiated a conversation regarding potential optimizations for faster inference in ChatGPT. Speculations from @teknium, @bjoernp, @eas2535 and @skadeskoten ranged from the unlikely use of quantization to strategic batching and potential caching for frequently asked questions. The fast response time of ChatGPT was discussed, with @zohad_sikder hypothesizing a robust caching mechanism due to the substantial user base.

▷ #memes (10 messagesđŸ”„):

  • Meme Sharing and Reactions: Users in this channel, namely @teknium and @Error.PDF, frequently share emojis and meme reactions. Notable mentions include the “Y not both” and <:pepeshy:1151280286345207819> emojis.
  • Desire for Certain Speakers: @teknium expressed a desire for individuals such as Yann and Karpathy to speak, leading to responses and discussions among the users.
  • Character Evaluation: @teknium expressed their opinion on an unidentified individual, characterizing them as “crazy psycho about x risk”.

OpenAI Discord Summary

  • An ongoing discussion centered around the topic of AI bias, morality, and fair use in the context of copyrighted content and AI. Conversations delved into issues such as biases in large language models (LLMs) and the philosophy of truth, alongside speculations surrounding Google’s new AI, Gemini, and alternative AI technology options like Mistral Instruct and gpt4all.
  • Members engaged in various technical discussions regarding GPT-4, touching upon ‘Dynamic Limits’, waitlist duration, prefix prompt exploration, ChatGPT’s performance and access issues, and differences in features across various devices. Speculations were made about the development of GPT-5 and the opening of GPT Store in the new year.
  • Issues with and improvements for GPT usage have been a pressing topic, with dissatisfaction expressed over the dialogue summarization by GPT, missing features in GPT Builder, and the absence of a feature allowing Inline editing or trim for AI responses. A parallel conversation took place regarding the acquisition of developer access for ChatGPT plugins, clarification of OpenAI’s Terms of Service, and the need for comprehensive guides on custom GPTs.
  • Conversations about game development using GPT and chatbot performance indicated a healthy interest in the potential applications of AI technology. Issues with captcha during API key generation, searching specific conversations, and perceived changes in GPT output fueled the debate on current limitations and areas for improvement in the AI system.
  • A notable topic in the guild was prompt engineering, digging deep into the usage of emotional language and the implementation of personalities in PPM. The community also dived into issues concerning text chunking, embeddings, and creation of detailed prompts. The sharing of a series of detailed prompt guidelines and command protocols for GPT-4, dalle, and browser tools reflected collaborative efforts to enhance utilisation of the AI model.

OpenAI Channel Summaries

▷ #ai-discussions (123 messagesđŸ”„đŸ”„):

  • Discussion on AI bias and morality: Users @whynot66k20ni, @light.grey.labs, @solbus, @lhc1921 engaged in a deep conversation regarding the inherent nature of biases in large language models (LLMs), the philosophy of truth, and the potential self-awareness of AIs.
  • ChatGPT’s AI ethics and ‘fair use’: @.dooz, @lhc1921, @light.grey.labs discussed the ‘fair use’ in the context of copyrighted content and AI. .dooz suggested that transformative use of copyrighted content could be constituted as fair use.
  • Discussion about OpenAI’s GPT Store release: @lumirix shared an excerpt from an email received by GPT creators promising the release of the GPT Store early next year and promising other great updates to ChatGPT.
  • Alternatives to OpenAI ChatGPT: @mysticmarks1 recommended Mistral Instruct and gpt4all as alternatives or additions to OpenAI’s ChatGPT for @sneakobrah who was seeking alternative chat AIs.
  • Discussion on Google’s AI Gemini: @prajwal_345 shared a link about Google’s Gemini AI suggesting that it was announced under pressure, and it outperformed OpenAI’s GPT-4 on various benchmarks.

▷ #openai-chatter (112 messagesđŸ”„đŸ”„):

  • GPT-4 Dynamic Limits and Waitlist Discussion: @dr.youvi.avant asked about the new GPT-4 ‘Dynamic-Limits’. @stefatorus mentioned that unlocking older GPT versions is possible but can be expensive, with his usage amounting to approximately 200 EUR per month. @killer.5643 inquired about the GPT-4 waitlist duration, with @7877 mentioning the upcoming GPT Store, and @jonathan_91672 sharing that he waited about a month for his invitation.

  • GPT-4 Prefix Prompt Exploration: @israel_a4 shared a YouTube tip from Wes Roth which allows users to see GPT-4’s Prefix or Secret Prompt by using a certain code. When asked about a potential patch to prevent this, @elektronisade stated that no such plans were in place due to the inherent functioning of the models.

  • ChatGPT Performance and Access Issues: Several users reported issues with ChatGPT, with @mrcrack_ mentioning consistent network errors and ADA’s ineffective image reading. @zz99mz mentioned the issue of the domain not loading at all. @pruo indicated trouble with their custom instructions, and @mrcrack_ also voiced dissatisfaction with the dynamic limits.

  • Features in Different Devices: @gd2x inquired about the speech feature absence in the Android version of ChatGPT, which @elektronisade attributed to the use of an adblocker. A discrepancy between features available in Android and iOS versions was also discussed.

  • GPT-3 Extensions and GPT Store Speculations: @youraveragedev speculated about GPT-5’s development, but @clockrelativity2003 denied its current training. A discussion about GPT Store’s opening in the new year was held by @lugui.

▷ #openai-questions (158 messagesđŸ”„đŸ”„):

  • Issues and Improvements in GPT: User @stealth2077 expressed concerns about GPT ending dialogues with a concluding summary paragraph, even after providing explicit instructions not to do so. @stealth2077 has also proposed a feature for inline editing or trim for AI responses for easier control over generated conversations, a topic joined by @ath0rus. @stealth2077 voiced dissatisfaction over the reduction of GPT usage from 50 to 40 and the removal of additional 10 usages reserved for custom GPT testing.
  • GPT Builder Limitations: @amanshrestha experienced issues in GPT Builder, which seemed to stem from the Python environment. @stealth2077 also expressed frustration over the restrictions in changing custom instructions mid-chat, and he highlighted the need for a better functionality to edit a chat’s context.
  • ChatGPT Plugins: @keebs1995 inquired about gaining developer access for ChatGPT plugins for building a calculator app for their industry. @elektronisade informed that plugins are being phased out and suggested using custom GPTs instead.
  • Terms of Service (ToS) Clarifications: User @eric.turnr sought elaboration on the OpenAI ToS section mentioning “Automatically or Programmatically extract data or Output (defined below).” @lumirix clarified that “Output” is defined in the Content section of the ToS.
  • Performance Issues & Enhancements: A few users, including @Shunrai and @lucianah, reported lagging and network error issues with GPT. @Rock asked for comprehensive guides about the workings of custom GPTs, and @strange073 sought clarification on how to access the GPT-4 API with a single dollar donation.

▷ #gpt-4-discussions (25 messagesđŸ”„):

  • Use of GPT for Game Development: @cerebrocortex shared their experience working on a Civilization-like game, expressing surprise at how well ChatGPT manages tasks like inventory management. They requested people’s feedback on their game.
  • ChatGPT Plus Invites: @pietman and @mlgpro0225 mentioned people receiving invites to join ChatGPT Plus, indicating that the waitlist might be moving forward.
  • Debugging GPT builder: @cerebrocortex asked about updating instructions for a custom GPT and @Capcon suggested saving changes to the draft and using the “update” button to publish changes.
  • Searching Specific Conversations in ChatGPT: @q16.kr asked if it is possible to search a specific conversation made with ChatGPT and @pietman replied it’s not currently available.
  • ChatGPT API Key Generation Issue: @realspacekangaroo reported an issue with captcha while trying to generate a new API key, deeming it excessively difficult and leading to them being locked out from generating new API keys.
  • Change in GPT Output: @victronwolfson noticed a change in the outputs of gpt-4-1106-preview over the last week.

▷ #prompt-engineering (36 messagesđŸ”„):

  • Using emotion in prompts: @eskcanta discusses the use of emotional language in prompts and its impact on the ChatGPT during a conversation about a paper named “ai emotional prompt”. They not that they could not find a specific prompt used in the paper for testing and cannot thereby reproduce the results.
  • Introducing personalities in PPM: @eligump and @mysticmarks1 engaged in a dialogue regarding the development of a PPM (persistent personality mode) with two personalities. @mysticmarks1 shares a link to illustrate how to implement behaviors like stutters and airheadedness in dialogues.
  • Creating detailed prompts: @cybector shares a draft of a detailed prompt for the python programming language and invites other users for feedback and suggestions to improve it.
  • Issues with text chunking and embeddings: @merpnderp requests for resources or discussions about strategies for text chunking and embeddings due to costs for density experiments. @eskcanta suggests experimenting with web interface ChatGPT to find potential cost-saving solutions. @m0bsta expresses difficulties in this approach due to the limit in messages.
  • Prompt and Guidelines for GPT-4: @cat.hemlock shares a series of detailed prompt guidelines and command protocols for GPT-4, dalle, and browser tools in markdown form. This consisted of the base information, tools used, and various policies to guide the use of the AI model. She also goes on to show the JSON format of what a typical detailed prompt would look like.

▷ #api-discussions (36 messagesđŸ”„):

  • eskcanta discussed EmotionPrompt’s use in language models, questioning its implementation and effectiveness due to the lack of clear prompt examples in the referenced paper.
  • madame_architect highlighted part of EmotionPrompt’s implementation from the available documentation. They provided examples of emotional stimuli and mentioned that the base prompts & template to which these stimuli were added were also present in the companion documents.
  • In a series of messages, eligump and mysticmarks1 discussed the creation and manipulation of Private Playground Models (PPMs), particularly how to incorporate roleplay and specific language styles.
  • A user named mattiacastioni asked for help in a linked conversation thread. The nature of this request was not further discussed.
  • cybector shared a template for engaging with ChatGPT surrounding Python programming language discussions, specifically instructing the model to source information from the official Python documentation.
  • merpnderp asked for recommendations of resources related to strategies for text chunking and embeddings, aiming to decrease costs in production. eskcanta suggested discussing cost-saving strategies with ChatGPT.
  • Lastly, cat.hemlock shared guidelines for using the markdown, dalle, python, and browser tools in OpenAI’s ChatGPT, as well as an example of how to construct a “default prompt”.

OpenAccess AI Collective (axolotl) Discord Summary

  • Active discussion and developments around the Mixtral integration prompted by @caseus_, with a focus on sample packing, sharding, and addressing various technical issues. The creation of the mixtral-multipack branch highlighted alongside relevant GitHub Links.
  • Release of a new dataset Verified-Camel-zh on Hugging Face by @noobmaster29 with direct access to the dataset.
  • A conversation identifying common issues in model error reporting and proposed solutions, such as changing model_type and disabling is_mistral_derived_model.
  • Sharing and exploration of various scientific paper processing libraries, such as the allenai/papermage, axa-group/Parsr, and the Unstructured-IO/unstructured libraries, for transforming PDFs, documents, and images into structured data.
  • Dialogues on the RLHF channel about the upcoming Data Programming Override (DPO) strategy for data set creation; specifically, the need for two distinct DPO datasets to handle “unalignment” and provision “quality answers”.
  • Miscellaneous conversations including a podcast with an axolotl representative, AI projects, tokens in coding, and a YouTube video titled The Insane Biology of: The Axolotl.

OpenAccess AI Collective (axolotl) Channel Summaries

▷ #general (25 messagesđŸ”„):

  • Mixtral Integration and Development:

    • @caseus_ shared updates on Mixtral integration with axolotl, including the addition of a mixtral-multipack branch and the merge of Mixtral MoE finetuning w multipack.
    • To use the updated features, users must install the latest version of transformers from git main.
    • For further development, @caseus_ shared a link to a work-in-progress branch by @214834317774422028 (GitHub link).

  • New Dataset Release:

    • @noobmaster29 announced a new dataset on Hugging Face called Verified-Camel-zh (link to dataset).

  • Miscellaneous Discussions:

    • @swyxio highlighted a podcast featuring an axolotl representative, and shared several AI-related resource and project links.
    • A conversation took place on the use and naming of tokens in coding, notably the use of the start and stop tokens.
    • @noobmaster29 shared a YouTube video titled The Insane Biology of: The Axolotl (link to video).

▷ #axolotl-dev (170 messagesđŸ”„đŸ”„):

  • Mixtral Sample Packing: @caseus_ has been working on implementing sample packing for Mixtral and has created a mixtral-multipack branch. There were reports of initial high loss that decreases, indicating the potential effectiveness of this approach. @faldore has been using the mixtral-multipack branch and reported stable operation and decreasing loss rates.

  • Fixes and Workarounds: Certain errors were encountered by users, for which workarounds and fixes were suggested. Specifically, disabling is_mistral_derived_model: true and changing model_type: AutoTokenizerForCausalLM seemed to resolve some issues. There was also a suggestion from @casper_ai to remove deepspeed if using a single GPU.

  • VRAM requirements: Concerns regarding VRAM usage were discussed, with @caseus_ suggesting strategies to reduce VRAM usage, such as freezing early layers of the model. Running Mixtral on 2xA6000 and 4xA100 GPUs was mentioned, with ambitions to achieve full finetuning on 4 to 8xA6000s. @casper_ai created a branch with parts of sharding to optimize VRAM usage, but it is still a work in progress.

  • Model Error Reporting: @ludis___ reported a RuntimeError when running Mixtral which read “output tensor must have the same type as input tensor”. This was resolved by the removal of certain configuration parameters.

  • LoRA and qLoRA usage: There were successful runs of Mixtral using qLoRA on GPU configurations such as 4xA100 and A40. However, attempts to run with LoRA resulted in errors related to the bnb package.

Links:

▷ #other-llms (3 messages):

  • Potential Hiring Discussion: @faldore expressed a sentiment that a certain situation could have been improved if they were hired.
  • Elon Musk Employment Opinion: In response, @nruaif suggested that working under Elon Musk might not be desirable.

▷ #general-help (5 messages):

  • Merging Qlora Chat Mixtral Issue: @matts9903 reported an error received while attempting to merge the mixtral model with the Axolotl tool. The issue is with a validation error for repo id: huggingface_hub.utils._validators.HFValidationError: Repo id must use alphanumeric chars or '-', '_', '.', '--' and '..' are forbidden, '-' and '.' cannot start or end the name, max length is 96: './qlora-out'.

  • @caseus_ suggested using an absolute path to the qlora-out directory but the suggestion didn’t resolve the issue.

  • @caseus_ then shared a recent change to model merging GitHub link and requested a stack trace for further troubleshooting.

▷ #datasets (4 messages):

  • PaperMage Library: @noobmaster29 shared a link to GitHub for the allenai/papermage library, suggesting that it might be worth testing. This library supports NLP and CV research on scientific papers.
  • Parsr Library: @visuallyadequate is currently experimenting with the axa-group/Parsr library, which transforms PDFs, documents, and images into enriched structured data.
  • Tika Library: @visuallyadequate mentions having used the Tika library, describing it as having provided the best solution so far, but they have not yet tested PaperMage.
  • Unstructured Library: @joshuasundance shared a link to the Unstructured-IO/unstructured GitHub library, which provides open-source libraries and APIs for building custom preprocessing pipelines.

▷ #rlhf (5 messages):

  • DPO Completion: @caseus_ mentioned needing to finish the DPO (Data Programming Override), after having been sidetracked by work on Mixtral.
  • Unalignment and Quality Answers DPO Dataset: @faldore discussed the idea of needing two DPO datasets, one for “unalignment” and another for providing “quality answers”.
  • Rejected Field Inquiry and Comparison: @nruaif suggested asking Llama 2 7B chat for the rejected field and additionally compared it with GPT 4, suggesting that in 90% of cases, the Llama 2 7B chat would yield worse answers.

LangChain AI Discord Summary

  • Extensive discussion on using local models with chat LLMs in LangChain, featuring insights from @_egeres on the potential use of environment variables and subclassing LLM and ideas from @lhc1921 surrounding the use of a backend like llama.cpp for handling constrained grammar.
  • Queries raised by various members but remained unanswered, including:
    • @analyticsrepo’s question on Gemini integration from Google into LangChain.
    • @_ashisharya’s request for comprehensive resources on agent coding and deployment.
    • @xstepz’s guidance request on limiting the usability of pandas functions in Kork package.
    • @yasuke007’s seeking advice on learning pathway for AI development with a specific focus on the necessity of Python knowledge when using langchain with React.js.
    • @rajib2189’s inquiry about the potential use cases for running language models locally.
  • Announcement by user @reletreby regarding the Askly December Release, now integrating OpenAI ChatGPT 3.5 and HuggingFaceH4/zephyr-7b-beta from HuggingFace. New features include multi-file reasoning, summarization, web search, necessitating users to delete and re-upload old files to enable the new functionalities. Full details shared via Askly’s blog.

LangChain AI Channel Summaries

▷ #general (72 messagesđŸ”„đŸ”„):

  • Gemini from Google integration: A user @analyticsrepo asked about the status of integrating Gemini from Google to LangChain, but no answer was provided.
  • LangChain with Local Models: @_egeres and @lhc1921 discussed extensively the possibility of using local models with chat LLMs in LangChain. @_egeres mentioned the possibility of tweaking API endpoints via environment variables and sub-classing LLM. @lhc1921 suggested the use of a backend like llama.cpp that is capable of taking constrained grammar.
  • Resources for Agent Coding and Deployment: @_ashisharya asked for comprehensive resources on Agent coding and deployment, but didn’t receive any response.
  • Kork Package with Pandas: @xstepz sought guidance on how to limit the pandas functions accessible to their agent using the Kork package, but didn’t receive any response.
  • Learning Pathway for AI Development: @yasuke007, a new AI developer, asked for advice on whether Python would be necessary in their AI development journey using langchain with React.js, but received no response.
  • Use Cases for Running Language Models Locally: @rajib2189 asked about the possible use cases for running language models locally, like personal assistant or edge type of analytics, but received no response.

▷ #share-your-work (1 messages):

  • Askly December Release: User @reletreby announced the latest version of Askly which is significantly upgraded with the integration of OpenAI ChatGPT 3.5 and the open-source model HuggingFaceH4/zephyr-7b-beta from HuggingFace. The new features include multi-file reasoning, summarization, web search, and more. However, to access these features, users who had uploaded files on or before December 1st, 2023, need to delete their old files and reupload them. This is critical to activate the new functionalities. The complete details were shared on the Askly’s blog.

Alignment Lab AI Discord Summary

  • Interaction between @astra1337 and others after demo presentations, highlighting the interest shown by the audience for additional explanation. Additionally, @astra1337 raised a query about the awareness of Pygmalion AI with respect to a video game demo.
  • Query by @mister_poodle about the fine-tuning process of Mistral-OpenOrca for specific tasks, with a particular focus on enhancing performance for a Named Entity Recognition (NER) task with JSON outputs.
  • Dialogue around diagramming tools, with Whimsical and Excalidraw being highlighted.
    • Whimsical was introduced by @teknium and tested by @gabriel_syme, noting its tendency for collaborative features.
    • Excalidraw was suggested by @lightningralf who provided the link Excalidraw and noted the existence of an Obsidian plugin.

Alignment Lab AI Channel Summaries

▷ #oo (3 messages):

  • Astra1337 Interaction with Others regarding Demos: User @astra1337 mentioned that people approached them for additional information after some demo presentations.
  • Discussion on Pygmalion AI: @astra1337 asked someone from a video game demo if they were aware of Pygmalion AI, a research group known for creating video game characters with memory.

▷ #open-orca-community-chat (1 messages):

  • Fine-tuning Mistral-OpenOrca: @mister_poodle inquired about fine-tuning Mistral-OpenOrca for specific tasks using personal datasets, showcasing an intention to improve the model’s performance on a Named Entity Recognition (NER) task with JSON outputs. No link or additional information was provided by @mister_poodle in this context.

▷ #oo2 (8 messagesđŸ”„):

  • Discussion on diagramming tools: @teknium introduced Whimsical as a diagramming website. Upon trying it, @gabriel_syme thought that it had collaborative features since it prompted for creating a workspace.
  • Excalidraw recommendation: @lightningralf recommended Excalidraw as another option, linking to the website, and additionally, mentioned a plugin for Obsidian. Here is his recommended link: Excalidraw.

Latent Space Discord Summary

Only 1 channel had activity, so no need to summarize


  • Using qlora with small batches and context window: In a response to a query, @eugeneyan shared that a 24gb GPU should work for using qlora with a small batch size and decent context window (batch of 2, context window 512 - 1024).
  • Features query about HumanLoop: @jozexotic expressed concerns about the slow development of new features in HumanLoop, specifically the lack of access to models outside of OpenAI and asked if anyone knew about these additions being on the near term agenda for the platform.
  • Frustrations with chatgpt+: @slono expressed a considering to cancel their chatgpt+ subscription due to the slow progress and recurrent stream errors.

Skunkworks AI Discord Summary

Only 1 channel had activity, so no need to summarize


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

LLM Perf Enthusiasts AI Discord Summary

Only 1 channel had activity, so no need to summarize


.psychickoala: any of you seen best practices to force parallel function calling