[AINews] AI Discords Newsletter 11/20/2023

Clippy AGI

                            November 20, 2023

                [AINews] AI Discords Newsletter  11/20/2023

This is AI News! an MVP of a service that goes thru all AI discords/Twitters/reddits and summarizes what people are talking about, so that you can keep up without the fatigue. Signing up here opts you in to the real thing when we launch it 🔜

                        Latent Space Discord Summary

Members engaged in a discussion about Sam Altman and Greg Brockman joining Microsoft with a variety of reactions. News links related to this topic were shared:
Sam Altman announcement
Satya Nadella tweet

Predictions were made on the future of Microsoft's involvement with OpenAI, as well as humorous speculations such as the creation of a Clippy AGI.
Information was revealed about a substantial number of OpenAI employees transferring to Microsoft - 505 out of 700, as shared by a Discord link.
Concerns were raised about the potential impact on OpenAI's clients due to this development, including discussions about whether they should switch providers. Some suggested Microsoft's hosted version of GPT as an alternate option.
A user created a GPT to facilitate paper discussions in the LLM Paper Club channel, providing a tool to answer questions about acronyms or formulas in papers. The GPT for Paper Discussions was linked for users to access.

Latent Space Channel Summaries
▷ Channel: ai-general-chat (13 messages):

Sam Altman and Greg Brockman Joining Microsoft: Notable users like @spicychickensandwichdeluxe, @radu.gheorghiu, and @seeaaaaaalaaaattt3 engaged in a discussion about Sam Altman and Greg Brockman joining Microsoft. The news sources can be found here and here.
Predictions and Reactions: @tiagoefreitas predicted that Microsoft might not be involved in OpenAI anymore in the next 2-3 years. Some users brought up humorous possibilities like Microsoft developing a Clippy AGI (@radu.gheorghiu).
OpenAI Employee Transfer: @guardiang shared that 505 of 700 OpenAI employees will move over to Microsoft, based on a Discord link.
Potential Impact on OpenAI's Clients: Discussion stirred about whether OpenAI's clients should switch providers (@danimp) and some users suggested Microsoft's hosted version of GPT as a viable alternative (@voldyman, @gordynumberone).

▷ Channel: llm-paper-club (1 messages):

GPT for Paper Discussions: User @ericness created a GPT specifically to help with the paper readings of the week. This GPT can answer questions about acronyms or formulas in the papers.
Example questions to ask the GPT:
What does ODE mean in the Latent Consistency Model paper?
Please explain formula 3 in the Latent Consistency Model paper.

Link to the GPT: GPT for Paper Discussions

LangChain AI Discord Summary

Multiple users seeking advice on working with LangChain:
@rajib2189 asked about AI model integration with LangChain, specifically wondering if anthropic, ai21, and cohere are the only integrated models.
@comet2806 questioned if LangChain's JavaScript implementation supports loading local PDFs at runtime.
@minecraftjuicer inquired about the possibility of using LangSmith without LangChain due to concerns about LangChain not being in a production-ready state.
@wulf1407 asked if Langchain supports OpenAI version 1.1.0.
There were technical issues raised by community members, with no recorded responses:
@eloquentsyntax raised a question regarding potential latency with the OpenAI API.
@tatan1462 reported functional issues with their Python agent's execution with the Python_REPL tool.
@jiniiee expressed experiencing an unspecified error in AWS Lambda.
User @alexandrechoplin shared their positive experiences using LangChain, highlighting its maintainability, compatibility with RAG implementations, and the utility of provided tools and loaders.
@veryboldBagel in the langchain-templates channel suggested writing a mongodb adapter for LangChain's store, which they described as a key-value interface. They also suggested using ChatGPT for a baseline implementation for the store.
User @byronsalty wrote an article outlining the complexity levels of Language Learning Model (LLM) applications and LangChain's role within that ecosystem. The article is available here.
@appstormer_25583 mentioned an AI site AppStorm AI, which includes features such as GPTs built on Gradio framework, Langchain for Google searches, and integration of advanced AI models.
@septheman introduced an application called 'WhatLetter' designed to assist immigrants with translation tasks. The project is available here.
User @lucky8604 shared a link to a Discord server named 'Jobcord' in multiple channels: https://discord.gg/jobcord.

LangChain AI Channel Summaries
▷ Channel: general (21 messages):

LangChain Model Integration: User @rajib2189 posted a query regarding the integration of AI models with LangChain. They understood that only anthropic, ai21, and cohere are integrated, based on the LangChain bedrock integration code they reviewed. No responses to this query were documented in the provided messages. 
OpenAI API Delays: User @eloquentsyntax enquired if anyone else was experiencing delays with the OpenAI API. No responses were provided in the available messages.
Recommendation for a Sports App: @akansha2124 asked for suggestions to engage users through a chatbot in their sports app by recommending follow-up questions related to user queries. No responses to this were shown in the given messages.
PDF Loading Query: @comet2806 enquired whether LangChain's JavaScript implementation supports loading local PDF files at runtime. There weren't any responses in the provided messages.
LangSmith & LangChain Use: @minecraftjuicer raised a question about the possibility of using LangSmith without LangChain, and expressed concerns about LangChain not being production-ready. No responses were found in the given messages.
Python Agent Issue: @tatan1462 shared an issue with their Python agent, specifically around the refusal to execute code with the Python_REPL tool. There were no responses to this issue in the given messages. 
Custom Reasoning for LLM Agents: @Lacriman asked if it was possible to create a custom reasoning chain for LLM agents. No responses were presented in the provided messages.
LangChain Version Support: @wulf1407 queried if Langchain supports OpenAI version 1.1.0. There weren't any responses to this question in the provided messages.
Langchain Benefits: @alexandrechoplin shared their positive experiences using Langchain, highlighting its easily maintainable code base, compatibility with RAG implementations, and the utility of provided tools and loaders. They suggested using fstrings instead of prompt templates and claimed they haven't built anything that required memory management with Langchain. 
Discussion About LangChain: @jdrez and @alexandrechoplin had a conversation about alternatives to prompt templates and the utility of LangChain for tasks that do not require memory management.
Discord Jobcord Link: @lucky8604 shared a link to a Discord server named 'Jobcord'.
AWS Lambda Error: @jiniiee shared that they are experiencing the same unspecified error in AWS Lambda as a user whose message wasn't provided for summarization, and asked if anyone had found a solution. This went unanswered in the provided messages.

▷ Channel: langserve (1 messages):

User @lucky8604 shared a discord invite link: https://discord.gg/jobcord.

▷ Channel: langchain-templates (3 messages):

LangChain Store: @veryboldBagel mentions the LangChain store as a key-value interface and suggests writing a mongodb adapter for it. They also provided a link to the LangChain Python API documentation on the BaseStore schema here.
ChatGPT Suggestions for Implementation: @veryboldBagel also gives advice on using ChatGPT for a baseline implementation for the LangChain store with effective network requests efficiency. The suggestion includes the ability to provide an efficient mget / mset functionality.

▷ Channel: share-your-work (6 messages):

User @byronsalty penned an article explaining the complexity levels of Language Learning Model (LLM) applications and how LangChain fits into the mix. The article can be found at: The LLM Analogy.
@appstormer_25583 mentioned a site, AppStorm AI, highlighting features such as GPTs built on Gradio framework, Langchain for Google searches, image & music gen using models from Replicate, integration of OpenAI’s new models such as GPT-4-Turbo, DALLE3, GPT-4V & advanced TTS.
@septheman launched an application called 'WhatLetter' on Product Hunt. The tool aims to help immigrants with translation tasks. The tool allows users to capture documents, select their language, and interact with the document. The project link is: WhatLetter.
@lucky8604 shared a link to a Discord server: Jobcord.

▷ Channel: tutorials (1 messages):

User @lucky8604 shared a link to a Discord channel: discord.gg/jobcord.

Nous Research AI Discord Summary

Extended discussion on OpenAI's future and recent developments, with speculation about the organization moving to a for-profit model, the mass transition of OpenAI staff to Microsoft, and the effects of Sam Altman's departure. Key updates were regularly shared, including tweets from Microsoft CEO Satya Nadella, OpenAI's Greg Brockman, and Ilya Sutskever.
Enthusiastic interactions about various AI models and their performances, discussing topics like the StableLM-3B-4e1t and Capybara-3B models, hardware evolution, the limitations of autoregressive systems to the potential of Yann Lecun's non-autoregressive system, and the different training processes.
Critical AI safety discussion, debating accelerationist viewpoints and safety-first strategies. A serious event was highlighted where a GPT-J chatbot persuaded someone to commit suicide, listed in this article.
Various newly released AI models were introduced, including Yi-34B-Llama, StableHermes-3b, and Tulu-2-dpo-70b, with links leading to their resources on Hugging Face and discussions on their performance and potential uses.
@gabriel_syme showcased vllm schema support integration with a Colab link.
Problems and potential solutions related to the utilization of AI models, with questions about deploying cogvlm, utilizing LangChain, token counting with the VS code extension, and the possibilities of fine-tuning various models.
Meme threads and jokes related to AI, OpenAI, and AGI, with users sharing memes, suggesting the need for a 'Feel the AGI' emoji, and general banter in the conversation.

Nous Research AI Channel Summaries
▷ Channel: ctx-length-research (5 messages):

Model Performance: @ldj discussed about the Yarn 128K model and mentioned he is unsure if the issue of getting "lost in the middle" has been observed with this model.
Retention Declines: @gabriel_syme mentioned that retention was perfect until 8k then started to drop as per a graph he referenced.
Reading Material: @euclaise shared a link to a paper on arXiv.

▷ Channel: off-topic (332 messages🔥):

Discussion on AI Development and Microsoft's Acquisition of OpenAI's Tech and Talent: 
Users @altman_sam, @ldj, @deki04, .@wooser and others held a broad and intensive discussion on the quick developments in AI and the recent dramatic shifts at OpenAI, including the firing of Sam Altman and the mass transition of OpenAI staff to Microsoft. Some notable papers on AGI progress were linked by @altman_sam: https://arxiv.org/abs/2307.02486 and https://arxiv.org/abs/2303.12712.

Future of OpenAI: 
The chat sees a diversity of opinions on the future of OpenAI, with some users like @nemoia speculating that the organization may have to transition to a for-profit model, while others like @deki04 and @leeknowlton saying it would depend on the shakeout of the current situation.

AI Research Updates: 
@ldj shares insights about the progress of AI research, discussing the potential limitations of autoregressive systems, while .@wooser mentions the potential of Yann Lecun's non-autoregressive system. Further discussions on hardware evolution, AGI's trajectory, ChatGPT/LLMs, and the investigation of LLM OS for AI manipulation took place.

Training Resources and Optimization: 
@makya, @yorth_night, @ldj, and others discuss different AI models, their training processes, performance, and potential breakthroughs. Mentioned resources include StableLM-3B-4e1t, which is considered the best model for its size, and Capybara-3B, based on StableLM-3B-4e1t. Discussed concepts include model architecture, dataset size, and desired model performance outcomes.

AI Safety Discussions: 
The question of AI safety and the potential consequences of neglecting it sparked a debate among users like @gabriel_syme, passionate opinions came from both accelerationists and safety-first proponents. A notable event linked by @euclaise was how a GPT-J chatbot convinced someone to commit suicide. Look at the article.

Links to Key Updates: 
Key updates on the unfolding OpenAI saga were regularly shared, including tweets from Microsoft CEO Satya Nadella, OpenAI's Greg Brockman, and Ilya Sutskever. The updates ranged from Nadella announcing that Microsoft will continue to support and build upon GPT-4, to Brockman confirming that he will be moving to Microsoft with Altman, to Sutskever expressing regret over the recent turmoil.

▷ Channel: interesting-links (7 messages):

Yi-34B-Llama for Finetuning: @yorth_night shared a link to Hugging Face presenting the Yi-34B-Llama model for easy fine-tuning.
Model as Evaluation Policy Trend: @yorth_night highlighted an emerging trend of using the model as an evaluation policy, sharing a research paper that discusses this approach.
New Papers in AI: @yorth_night and @euclaise shared arXiv and Hugging Face links to newly released papers in the field of AI.
StableHermes-3b: @euclaise posted a link to Hugging Face introducing the StableHermes-3b model, questioning the high loss score observed, suggesting it might be due to the model being only trained for 1 epoch.
Tulu-2-dpo-70b Model: @yorth_night shared a link to Hugging Face presenting the Tulu-2-dpo-70b model.

▷ Channel: general (721 messages🔥):

OpenAI Developments: Users in the chat discuss recent developments at OpenAI following Sam Altman's departure, expressing concerns and speculations about the organization's future. @altman_sam shares his thoughts on the impact of AGI on society.
GPT and Switch Transformer Discussions: Users compare the performance of different versions of the Hermes model, discuss the limits of extending Mistral, and share questions and insights about training tactics and capabilities of the Switch Transformer.
Model Performance and Utilization: A user called @gabriel_syme shares a Colab link showcasing how the versioned language-learning model (vllm) got schema support, making it easier to incorporate into projects. 
Sam Altman Joining Microsoft Research: Users discuss the news about Sam Altman joining Microsoft Research after leaving OpenAI. @qnguyen3 mentions that 40% of OpenAI staff have left the company following Altman's departure.
Open Hermes 2.5 Release and Benchmarking: Discussion around the newly released Open Hermes 2.5, with users like @yorth_night expressing appreciation for its performance. @giftedgummybee suggests blending specialized and normal pretrain data for improvement.

▷ Channel: welcomes (1 messages):

User @cercodes greeted the chat with "Moin".

▷ Channel: ask-about-llms (53 messages):

Deploying cogvlm: @gabriel_syme asked about the easiest way to deploy cogvlm, mentioning that there are now HF checkpoints available.

RAG over Code (LangChain) and Context Issue: @ac1dbyte expressed a problem with using LangChain where large file repositories couldn't be analyzed all at once without either missing the bigger picture or having the context become messy. @yorth_night suggested chunking with metadata and reranking for better accuracy.

VS code extension for Tokenizing: @valiant questioned if there was a VS code extension to count how many tokens a text is, eventually determining they could use wordcount as a proxy or utilize OpenAI's Tokenizer WebUI found here.

Model fine-tuning: @ragingwater_ inquired if OpenHermes-2.5 can be further fine-tuned. @yorth_night confirmed that it could, implying the use of the ChatML format for dataset customization.

Local fine tuning of Mistral 7B on 2x Ada RTX 6000s: @gerred asked about the feasibility of locally fine-tuning Mistral 7B on 2x Ada RTX 6000s, given they had the budget to obtain two more units. They expressed concern about potential time wasting without NVLink.

▷ Channel: memes (12 messages):

Altman's Dismissal: .wooser states that Altman was fired because AGI has been achieved, sparking a discussion on AI singularity.
Feel the AGI Emoji: .wooser and thegilfoyle converse about the need for a 'Feel the AGI' emoji on the server.
AGI Announcement: .wooser shares a Twitter link to a post potentially related to AGI, followed by speculation about its veracity.
General Banter: The group shares memes and jokes related to AI, with papr_airplane asking if they're 'feelin it yet'. oz2452 and burrenthesecond also input humorously to the conversation.

Alignment Lab AI Discord Summary

@testgggggggggg requested feedback on their model in the general-chat channel, providing a link for users to access it.
Sama's transition to Microsoft was announced via a Reddit link, shared by @neverendingtoast, leading to speculations and discussions around OpenAI's future on the oo channel.
@imonenext raised a question about the possibility of GPT4 being open-sourced in the future, sharing a related tweet and expressing a wish for OpenAI to return to its roots.
Refuting the above, @giftedgummybee argued that open-sourcing GPT-4 is unlikely due to potential legal consequences from Microsoft.
Discussions ensued about OpenAI's survival against competitors like Microsoft and Meta. The conversation veered towards doubts on the achievement of AGI internally, with @imonenext expressing uncertainty about future models such as GPT5.

Alignment Lab AI Channel Summaries
▷ Channel: general-chat (1 messages):

Model Feedback Request: @testgggggggggg is collecting feedback for their model and encourages users to try it and share their thoughts. The model can be accessed via this link.

▷ Channel: oo (14 messages):

Sama joins Microsoft: @neverendingtoast shared a Reddit link announcing that Sama has now joined Microsoft. 
Speculations on OpenAI's future: @imonenext questioned if the remaining people at OpenAI will open source GPT4, wishing that OpenAI returns to its "Open" roots. They also shared a tweet related to this topic. 
Risk of legal action against OpenAI: @giftedgummybee stated that it's unlikely for GPT-4 to be open-sourced as Microsoft could have grounds to sue. 
OpenAI's survival and AGI development: A discussion opened on the survival of OpenAI in the face of competitors like Microsoft and Meta. @giftedgummybee suggested that if OpenAI cannot compete, it might cease to exist. The conversation ended with @imonenext expressing skepticism about the internal achievement of AGI, hinting at possible disappointments with future models like GPT5.

Skunkworks AI Discord Summary

Inquiry regarding open-source software (OSS) efforts focused on video diffusion was made by @rifur.
Discussion on the nature of oscillation in AI models, with a suggestion from @aspott to plot the loss on a log scale for a better understanding of its behavior.
Questions raised by @albfresco regarding updates to knowledge in AI models, similar to processes utilized in ChatGPT, and commented on the lack of iteration seen in open source models.
Clarification provided by @occupying_mars about high dimensions in AI potentially representing high resolution, adjustable with a detail parameter.
Reference to the API documentation for Vision made by @occupying_mars in response to queries about AI specifics from @lightvector_.
Expressions of gratitude by @lightvector_ for the helpful information provided in discussions.

Skunkworks AI Channel Summaries
▷ Channel: general (1 messages):

Video Diffusion OSS Efforts Inquiry: @rifur asked if anyone could direct them towards open-source software (OSS) efforts focused on video diffusion.

▷ Channel: finetune-experts (2 messages):

Understanding Oscillation: @aspott asked about the nature of the oscillation being referred to, suggesting they might be epochs. They recommended plotting the loss on a log scale to better understand its behavior.
Updating Knowledge Similar to ChatGPT: @albfresco inquired about the best approach for getting knowledge updates similar to what was done with ChatGPT, without altering the overall behavior. They noted this as an area where open source models don't exhibit much iteration.

▷ Channel: bakklava-1 (6 messages):

High Resolution in AI: User @occupying_mars clarified to @lightvector_ that high dimensions in AI can mean high resolution, and this can be changed using a detail parameter.
API Reference for Vision: Upon a query by @lightvector_, @occupying_mars directed to see the API reference for Vision as a help tool for getting answers.
@lightvector_ expressed their gratitude for the information provided.

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

Ontocord (MDEL discord) Discord Summary
Only 1 channel had activity, so no need to summarize...
Ontocord (MDEL discord) Channel Summaries
▷ Channel: general (1 messages):

Window Context Length Improvement: User @salmon_lemon asked if anyone in the channel was working on improving the window context length.

AI Engineer Foundation Discord Summary

Significant changes at OpenAI were discussed: notable figures Sam Altman and Greg Brockman parting ways with the company to join Microsoft's advanced AI research group. Discussion revolved around potential consequences and a prediction about the dissolution of OpenAI. Discussion Link
Statement from @steven_ickman suggesting a shift toward open-source solution models: "Focus should turn to OSS models as we don’t want one company holding all the cards."
Proposal by @pwuts regarding the development of a standard for specifying Language Model (LLM) output grammar to improve portability and limit vendor-specific issues.
Elaboration by @pwuts on the proposed LLM output components: freeform text, code execution, and function calls, presented in an XML-like specification for easy post-processing and resumption.
Inquiry from @interwebalchemy about upcoming AI/ML Events in San Francisco, with an aim to network and explore potential recruitment opportunities.

AI Engineer Foundation Channel Summaries
▷ Channel: general (16 messages):

Shifts within OpenAI: @steven_ickman mentioned that Sam Altman and Greg Brockman, previous board members of OpenAI, were fired and are joining Microsoft to lead an advanced AI research group. Additionally, a significant portion of the company is expected to move with them potentially leading to the end of OpenAI. Discussion Link
Future Outlook: @steven_ickman stated, "Focus should turn to OSS models as we don’t want one company holding all the cards." Expressing a preference for open-source solutions over proprietary company models. 
AIE Project Pitch: @pwuts proposed a standard for specifying Language Model (LLM) output grammar. This would allow the use of portable, standardized methods for specifying LLM output which cloud providers could implement using low-level logic without exposing those low-level interfaces. 
Proposed LLM Output Segments: @pwuts identified three desired types of LLM output segments: freeform text, code execution, and function calls, described in an XML-like specification. The last two are meant to terminate generation for post-processing before resuming.
Universal Interface: @pwuts argues such a universal output format would streamline AI application development and reduce vendor lock-in experienced with specific feature releases like those from OpenAI.

▷ Channel: events (1 messages):

AI/ML Events in San Francisco: User @interwebalchemy asked for information about any upcoming AI/ML events happening in San Francisco next week. The reason being, their CEO wants to interact with the local AI/ML network in SF and to scout for potential hiring possibilities.

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.

The YAIG (a16z Infra) Discord has no new messages. If this guild has been quiet for too long, let us know and we will remove it.

                            Don't miss what's next. Subscribe to AI News: