[AINews] AI Discords Newsletter 11/16/2023
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 🔜
Guild: Latent Space
Latent Space Guild Summary
- Discussion on the potential issue of copyright violation and the consequential lawsuits concerning training data for machine learning models. Notable mention of the suggestion for different types of licenses similar to coding licenses: Twitter post by Edward Newton-Rex.
- Conversation about the Yi 01 model, including drama related to it: A discussion on Hugging Face.
- Inquiry on the practical implementation of AI for item description generation, with focus on trial and error on a few model implementations, such as Mistral, Zephyr, Llama 2 C++, and ChatGPT-3.5-1106 Turbo.
- Question on the preferred model routing companies, among Martian, Open Router, and Pulze.
- Sharing of insights on fine-tuning results and an interesting area on 'AI girlfriend product', with a focus on cost and market demands: YCombinator News item and Blog on the AI girlfriend product.
- Call for coding terminology recommendations for describing generated and tweaked code.
- Query on experiences with the platform Codegen.com.
- In a separate channel, questions on the Code Fusion paper and speculation of a recent paper's influence on latest developments: Link to the paper.
Latent Space Channel Summaries
Latent Space Channel: ai-general-chat (2 messages):
AI-General-Chat Discussion Summary:
- Concerns about Copyright and Training Data: @swyxio highlighted a potential issue regarding a copyright disagreement that could significantly affect lawsuits. @mitch3x3 hopes for some nuance to be introduced into these lawsuits, bringing up the suggestion of different types of licenses similar to coding licenses.Twitter post made by Edward Newton-Rex
- Drama around Yi 01 Model: There was a discussion about the drama related to the Yi 01 model. A Hugging Face post discusses this issue.
- Usage of AI for Item Description Generation and Performance Concerns: @mockapapella raised a query regarding AI usage for item description generation based upon part numbers. He detailed his attempts with Mistral, Zephyr, Llama 2 C++, and ChatGPT-3.5-1106 Turbo, with varying results and challenges.
- Inquiry on Model Routing Companies: @coffeebean6887 queried for recommendations related to model routing companies like Martian, Open Router, and Pulze.
- Interest in Fine-Tuning Results and AI Girlfriend Product: @swyxio shared a link on interesting fine-tuning results and provided another on a breakdown of costs and market demands of 'AI girlfriend product'. Link to YCombinator News and Link to Mazzzystar's blog.
- Suggestions for Coding Terminology: @slono asked for suggestions on terms to use for the description of generated and tweaked code.
- Query on Codegen.com: @tiagoefreitas asked if anyone had tried Codegen.com, a query also echoed by @swyxio.
Latent Space Channel: llm-paper-club (2 messages):
Discussion on the Code Fusion Paper:
- Question about Code Fusion Paper Discussion: @slono asked if the group had discussed the code fusion paper during their meeting.
- Source of Recent Development: @youngphlo suggested that a recent paper could be the explanation behind the latest developments, mentioning that it likely builds on previous works.
- Links:
Guild: OpenAI
OpenAI Guild Summary
- Conversation surrounding the functionality and practicality of OpenAI's Chatbot agents, especially with regard to their ability to store and access knowledge, as well as their capacity for app navigation and web scraping.
- Discussions on various technical challenges including disappearing platform icons, difficulty with creating an assistant instance, and restrictions on random responses in the Assistant Playground.
- Numerous instances of problems with GPT-4 and ChatGPT, including issues with accessibility, missing features, and server instability. Additionally, conversations surrounding getting the desired output from GPT-4, including generating graffiti tag images and handling large data sets.
- Reports of specific problems users are experiencing, such as issues with uploading a picture in chat, and difficulties in formatting date prompts.
- In-depth forum on API functionality and usage, including discussions on the difference between GPTs created through the web interface and via the API, and the possibility of embedding custom GPTs on websites.
- Diverse range of prompt engineering topics including brainstorming for task assistance, efficient ways to extract step-to-step instructions from web sources, and advice on prompt engineering techniques.
- Discussions also highlighted concerns related to OpenAI credit top-ups, access to GPT-4, restrictions on random responses, and ongoing site issues.
- The OpenAI Community also shared useful resources and discussions on training models, and model performance.
- There are also ongoing links being shared for different resources and issues, like AI writing GPTs, adversarial code collaborative agent, and various OpenAI Guilds and Channels. For instance: link1, link2, link3, Adversarial Code Collaborative Agent.
- At last, discussions encapsulated a range of topics and user's concerns, serving as a significant snapshot of ongoing community interests and areas for OpenAI's potential improvements.
OpenAI Channel Summaries
OpenAI Channel: ai-discussions (6 messages🔥):
OpenAI Chatbot Discussions Summary:
- Discussion on Training ChatGPT Agents: @tilanthi expressed concerns about the ability to train chatbot agents beyond the April 2023 training cutoff date. @solbus clarified that the expert knowledge provided via files are stored as "knowledge" files different from stored instructions, and accessed by the GPT on specific queries.
- Discussion on App Navigation and XML Scraper: @milestones95 asked about AI's necessity in building a scraper for navigating mobile apps to grab XMLs. This was countered by yodababoom, who suggested that it requires a good old scraper, not AI.
- Censored AI: @jaimd joked about restricted AI features related to restrictions in asking about certain topics.
- ChatGPT Assistant Usage: @magical_beagle_14362 inquired if one has to create an assistant instance every time for usage.
- Thread and Message Icons on OpenAI Platform: @zahmb discussed the disappearance of Thread and Message icons on platform.openai.com interface with @foxabilo.
- ChatGPT Notice Misinterpretation: @trumantaurus asked about the notice from ChatGPT regarding errors; @lugui, @solbus provided some context.
- Restriction on Random Responses in Assistant Playground: @crash.tech sought options to restrict random responses from the Assistant Playground, leading to a discussion involving @elektronisade and @lugui.
- OpenAI Credit Top-Up and Access: @ishaka02 asked about the minimum amount for top-up to gain access to the GPT-4 model. @elektronisade clarified that the minimum is one dollar.
- API Functionality Inclusion in GPT Model: @kyper discussed implementing functionality in their company's API enabling GPT models to make some API calls. @lugui mentioned that such requirements might require a fine-tuned model.
- Issues with OpenAI Threads Endpoint: @criptobroh reported issues with the Threads Endpoint of OpenAI after POST RUN THREAD, sparking a discussion involving @elektronisade and others.
OpenAI Channel: openai-chatter (6 messages🔥):
OpenAI Discord Server Summaries:
- Issues with ChatGPT Performance and Models: Numerous users reported technical issues with ChatGPT, including @becausereasons who expressed frustration with the lack of basic functionality, @dabonemm who questioned if the chat's history has an impact on performance, and @satanhashtag who suggested trying a new chat without CI. There was a discussion about Hermes 2.5 vs Hermes 2 performance and the limitations of extending Mistral beyond 8k.
- Accessibility and Site Navigation Concerns: @zahmb noticed a change in the platform's interface, wherein previously visible threads and messages were missing. @foxabilo offered assistance on the topic and suggested moving to a different section for troubleshooting.
- Errors and Missing Features: Several mentions were made about various system failures, ranging from the inability to access the "My Plan" page (@antony0000 and @solbus) to disappearing chat history (@stunspot and @windrider30). There were also reports about ongoing site issues and server instability (@lugui).
- Access to ChatGPT Plus and GPT4: The sudden pause on new ChatGPT Plus subscriptions prompted discussions. Users shared experiences about the waitlist (@mrjazzhop and @milou4dev) and others voiced concerns regarding access to the upgraded model GPT4 (@shiv0613 and @Philipsman).
OpenAI Channel: openai-questions (6 messages🔥):
OpenAI Discord channel Issues & discussions:
- Adding website verification txt file to Godaddy: User @soltzu asked for help to solve a problem adding website verification txt file to Godaddy.
- Generating a Few Graffiti Tag Images: @_johanc pondered about generating few graffiti images in a specific style. @foxabilo suggested finding a popular artist who creates graffiti in the preferred style and command the AI to create graffiti in that style. @solbus recommended trying to provide a visual description of the desired image to GPT-4 and then bringing the description back to DALL·E.
- Issues with 'My Plan' in ChatGPT: Several users including @mrbr2023, @thesocraticbeard, @sumo_dump, and @fujikatsumo reported issues accessing the 'My Plan' page in their ChatGPT account settings, across browsers such as Firefox and Chrome, and also on different platforms including Edge and Safari.
- Trouble logging in with an active subscription: @katthwren and @dsmagics mentioned having troubles logging in to their active ChatGPT+ subscriptions on the app while the webpages were still accessible.
- Usage limit on ChatGPT: A discussion led by @mustard1978, @gavin5616, and @18870 shed light on a possible reduction in the usage limit for the ChatGPT service from 50 to 40 messages every 3 hours.
- Error faced while accessing Chat: @ajripon shared concerns about an error received when asking a question in the chat that has been persistent for the last 24 hours. Other users, including @vdrizzle_, also articulated facing similar challenges.
- Working with a large dataset for GPT-4: @playbit raised a question about handling large data sets in GPT-4. @thesocraticbeard suggested using a different manner of reading the CSV like using Modin or converting the CSV to JSON as the GPT performed better with JSON.
- Accessing ChatGPT service with a VPN: @danielbixo brought up a problem accessing ChatGPT on his personal computer while using OpenVPN GUI, even though the service worked well on Android. @xh and @solbus suggested trying a different OpenVPN server.
OpenAI Channel: gpt-4-discussions (6 messages🔥):
"GPT-4-Discussions":
- Performance and Cap Limit Messages: Users like @amarcerro, @elegante94, @cybector, and @mustard1978 discussed about message cap and performance of GPTs. Questions about API usage and cost-modelling for exceeding cap limit were raised by @satanhashtag and @elegante94.
- GPTs - Creation, Issues, and Conversational Behavior: Users including @stealth2077, @fran9000, @gubok, and @Derpgore shared their experiences creating, training and fine-tuning GPTs, mentioning the issues related to instructions and GPTs reverting to regular chat.
- GPTs and API - Discussion and Queries: @gubok asked about the difference between GPTs created through the web interface and via the API and if the GPTs created through the web could be invoked via API. @hhf0363 sought a simple explanation for what API is for a non-developer person.
- Technical Issues and Questions: @cbuh reported an issue with uploading a picture in chatgpt while @fran9000 shared how they struggled to get their doc-searching GPT to work. @gordon.freeman.hl2 inquired about the possibility of embedding custom GPTs on websites.
- Links:
- Adversarial Code Collaborative Agent shared by @Derpgore
OpenAI Channel: prompt-engineering (6 messages🔥):
Prompt Engineering Discussions:
- Chat Log Brainstorming and Context Simulation: User @mnjiman discussed the idea of using context simulation to facilitate tasks in the conversation. They also suggested the use of a markdown table to provide an options menu.
- Parameter Tweaking in New UI: User @no.iq sought for advice on where to adjust parameters like 'temperature' within the new UI, as it was no longer present on the right side of the playground.
- Instructions Extraction from Online Sources: User @goldmember777 asked about the most efficient way to extract multiple instructions from a website and share them one at a time with the user using GPT.
- Usage of AI Writing GPTs: User @.kalle97 recommended their SEO-optimized AI writing GPTs and shared links to their product for others to utilize: link1, link2, link3.
- Formatting Date Prompts: User @alishank_53783 raised a question about formatting date prompts in 'YYYY-MM-DD' type where they were experiencing difficulties parsing emails. @eskcanta advised on making instructions for the model extremely clear and unambiguous.
- Prompt Engineering Tips: @eskcanta also shared advice in general about prompt engineering, emphasizing the importance of clear language usage, accurate expectations, and careful fact-checking.
OpenAI Channel: api-discussions (6 messages🔥):
"OpenAI Discussion API Topics":
- Parameter Tweaking in New UI: @no.iq asked about modifying parameters such as 'temperature' in the new UI, stating they couldn't find the option anymore.
- Using ChatGPT for Task Facilitation: @mnjiman discussed various ideas about utilizing ChatGPT for task assistance, including presenting options in a markdown table.
- Handling Multiple Instructions from a Website: @goldmember777 raised a concern about sharing step-by-step information with users one step at a time without repeatedly searching for online resources.
- Prompt Engineering with Kalle97's GPT and Soliciting Usage: @.kalle97 shared links to their AI writing GPTs optimized for SEO and solicited users to try and upvote them if they found them useful. @mrposch asked for sample prompts, but @.kalle97 declined to share specifics due to the time and effort put into building it. @syndicate47 later shared their GPT and asked users for their feedback.
- Date Formatting Challenge: @alishank_53783 mentioned a struggle with date formatting in email parsing and @eskcanta suggested additional examples and gave a solution that seems to have worked. He also mentioned the limitations of AI when it comes to interpreting date formats.
- ChatGPT's Summarizing Tendency and Style Marks: @mnjiman and @eskcanta had a discussion about the tendency of ChatGPT to summarize its responses in the last sentence/paragraph. @eskcanta provided examples showing variability depending on the context and noted that it acts like a style marker, comparing it to a person's unique mannerism.
Links shared:
- https://chat.openai.com/g/g-H8gTAgiLm
- https://chat.openai.com/g/g-zH8HltElu
- https://chat.openai.com/g/g-f8pALzMrj
- https://discord.com/channels/974519864045756446/1171911165861970001
- https://chat.openai.com/share/9acbbacb-13be-41b8-9f61-2c9fd50b10e5
- https://chat.openai.com/share/1fcb9df9-f3d8-41c0-a7f9-a4fa4c14c870
Guild: LangChain AI
LangChain AI Guild Summary
- Discussion on chaining chains in LCEL, with a query presented by @ethereon_ on whether chains could be linked in LCEL.
- An inquiry about converting the API response error in OpenAi function, proposed by @jinwolf2.
- @alimal led the conversation regarding the effective use of langchain and/or APIs for tabular data QA.
- @Siddhi Jain expressed an issue with a missing 'run_manager' argument in _aget_relevant_documents().
- Discussion initiated by @abhi578 on langchain embedding generation and its potential applicability for a specific case.
- @0xtogo proposed a query on the potential to use an existing assistant with langchain without the need to call "create_assistant".
- @juan_87589 ignited a dialogue on the implications of the recent Microsoft collaboration announcement and its possible effects on LangChain's focus and development. Link of interest: LangChain Expands Collaboration with Microsoft.
- A query from @seththunder concerning the difference in run commands such as "llm_response = conversation.run/apply/invoke/batch({"query": question})".
- @tonyaichamp highlighted a performance issue in LangChain, stimulating discussion on the subject.
- Inquiry on the possibility of obtaining the sources/references used by a langchain agent to conclude a point, proposed by @eyueldk.
- @beffy22 shared an issue with vectorStore.asRetriever() always returning documents.
- Queries about LangChain web loader initiated by @quantumqueenxox, raising questions about the loading mechanism for sub URLs.
- @maverick5493 sought tips on uploading files into the OpenAI 'retrieval' assistant with langchain.
-
@rajib2189 ignited a conversation on checking whether a specific dataset or vocabulary was used to train commercial models like OpenAI.
-
@andriusem sought advice concerning loading .env files for an API project in LangChain. @attila_ibs recommended adding the .env file in the server.py.
-
Discussions and sharing of solutions surrounding LangChain upgrades and user file ingestion; an update to the latest LangChain version was suggested for resolving ndarray serialization issues, and creating extra endpoints was proposed for file ingestion. Link: RedisStore
-
Various AI projects and tools were introduced, with examples like YouTube GPT by taranjeetio, Appstorm.ai by appstormer_25583, Sophists App by .broodstar, Pantheon AI by kingkookri, and AI Tool by agenda_shaper. These discussions included links to the platforms and specific examples of what they are offering.
LangChain AI Channel Summaries
LangChain AI Channel: general (4 messages):
Discussions and Questions:
- Chaining Chains in LCEL: @ethereon_ sought clarification on whether chains could be chained together in LCEL and shared an example.
- Converting Response Error in OpenAi Function: @jinwolf2 inquired about converting the API response error in OpenAi function.
- Using langchain and/or APIs for Tabular Data QA: @alimal sought suggestions on the effective usage of langchain and/or APIs for handling tabular data QA.
- Error on run_manager Argument in _aget_relevant_documents(): @Siddhi Jain expressed an issue with missing the 'run_manager' argument in _aget_relevant_documents().
- Langchain Embedding Generation: @abhi578 speculated on langchain embedding generation and how it could be employed for a particular case.
- Using an Existing Assistant with Langchain: @0xtogo questioned whether an existing assistant could be utilized without calling "create_assistant" with langchain.
- Implications of the Microsoft Collaboration Announcement: @juan_87589 questioned the implications of the latest Microsoft collaboration announcement on the future focus and development strategies of LangChain.
- Difference in Run Commands: @seththunder questioned about the difference in commands "llm_response = conversation.run/apply/invoke/batch({"query": question})".
- Performance Issue in LangChain: @tonyaichamp raised a concern regarding performance issues in LangChain.
- Presence of Source/References: @eyueldk questioned the possibility of getting the sources/references used by a langchain agent to reach a conclusion.
- Issues with VectorStore.asRetriever(): @beffy22 expressed having issues with vectorStore.asRetriever() always returning documents.
- Loading a Website for LangChain Webloader: @quantumqueenxox expressed confusion on loading a website using langchain web loader, and whether it would load all the sub URLs as well.
- Uploading Files into OpenAI ‘retrieval’ Assistant using Langchain: @maverick5493 sought guidance on how to upload files into the OpenAI 'retrieval' assistant using langchain.
- Dataset/Vocabulary Check in Commercial Models: @rajib2189 inquired about checking whether a specific dataset or set of vocabulary was used to train commercial models like OpenAI.
Link of Interest:
- LangChain Expands Collaboration with Microsoft: Shared by @juan_87589 announcing the collaboration between LangChain and Microsoft.
LangChain AI Channel: langserve (4 messages):
Discussion on Loading .env Files in LangChain:
- @andriusem asked for help regarding how to load .env files for their project. They specifically wanted to include API keys in their LangChain application and attempted to do so by adding the .env file in the root directory.
- @attila_ibs suggested a solution which involved importing and loading dotenv using the following python code snippet:
from dotenv import load_dotenv, find_dotenv load_dotenv(find_dotenv()) - When questioned where to load the .env file, between server.py or chain.py, @attila_ibs recommended doing so in server.py.
LangChain AI Channel: langchain-templates (4 messages):
LangChain Upgrades and User File Ingestion:
- @veryboldbagel suggested to <@723833811757957142> that an upgrade to the latest version of LangChain could resolve issues with ndarray serialization, as the new version uses orjson.
- For setting up RAG over user-uploaded files, @veryboldbagel advised <@1121736064751636510> to create extra endpoints for file ingestion. They also mentioned that the docstore could be backed by Redis, or a persistence of their choice could be implemented.
- Links:
LangChain AI Channel: share-your-work (4 messages):
AI Projects and Tools Introduced in LangChain AI Discord:
- YouTube GPT by taranjeetio: An AI that keeps you updated about the latest YouTube video. Shared link: https://twitter.com/taranjeetio/status/1725023491387486451
- Appstorm.ai by appstormer_25583: Introduced a platform for building custom GPTs for free with various examples shared. Claimed their GPTs offer dynamic Google searches, integration with OpenAI models & Assistant API, collaboration with models from replicate.com, and support for diverse tasks. Links:
- Sophists App by .broodstar: A specialized text messaging app with GPT chatbot integration. Offered for sale by the developer. They also shared a link to a demo with the chatbot in the app: https://www.sophists.app/c/1a5f1c0f
- Pantheon AI by kingkookri: An upgrade on the Pantheon AI, now offering faster services and improved UI. Invited more users, especially in the STEM fields. Shared links: https://www.pantheon-ai.com/ and Discord for feedback
- AI Tool by agenda_shaper: Shared a brief mention of a platform with an algorithm creating posts.
Guild: Nous Research AI
Nous Research AI Guild Summary
-
Discussions around large-scale data storage and management challenges, focusing on potential solutions like Hugging Face, issues around centralized storage, and the exploration of concepts like MD5 hashing and data compression for efficient storage: HuggingFace size limit, YouTube Video.
-
Exchange of interesting links and thoughts around datasets, AI, and music including discussions about "Skeleton of Thoughts" paper, the "LLM Decontaminator", tweets on "Tree of thought for dataset generation", DeepMind's music revolution project, Python-based projects, and Meta's New Text-to-Video Model: Skeleton of Thought, LLM Decontaminator, Tree of Thought Tweet, DeepMind Music Creation, Related YouTube Link, Python Project YT, Meta's Text-to-Video Model, Meta's Text-to-Video Platform, Synthia-v1.3 Dataset.
-
Announcement of Hermes 2.5's ranking second place in the 7B models category on the HF Leaderboard.
-
Deep dives into probability calculations using given classroom scenarios focused on varied instances like probability of a French speaker being male, probability of sum being 8 when two dice are thrown and have odd numbers, etc.
-
Accord on Music Generation with AI, training large models, and Nous Research updates including the feasibility of using AI for music creation, difficulties of finetuning large models, and the availability of a new Capybara 34B API and playground: Suno AI, Blog on Ram-Efficient PyTorch FSDP, Capybara 34B API and playground.
-
Conversations on LLMs, including the difference between pretraining and finetuning methods, using LLMs for code review, challenges of training models for non-roman languages like Greek, and considerations on training new tokenizers: HuggingFace Link.
-
Meme shared by @_automagic connected to the tech field, with comments from @.wooser: Link.
Nous Research AI Channel Summaries
Nous Research AI Channel: off-topic (7 messages🔥):
Discussion about storing and managing large datasets:
- Data storage challenges: @yorth_night discussed the challenges of storing potentially dozens of terabytes of data, including cost considerations and the need for an easily accessible, open source solution.
- Hugging Face as a possible solution: @.benxh, @crainmaker, and @tsunemoto discussed the possibility of using Hugging Face's dataset hosting, which offers up to a petabyte per repo with a 50GB limit per file. They also discussed potentially using md5 hashing to manage images and compressing the data for more efficient storage. The idea of streaming the dataset to avoid disk storage issues was also proposed.
- Concerns about centralized storage: @crainmaker raised the concern that relying on a single platform for hosting the data creates a single point of failure and suggested the need for distributed backups.
- Links:
Nous Research AI Channel: interesting-links (7 messages🔥):
Interesting discussions and links shared in Nous Research AI Discord Chat:
- Skeleton of Thought Discussion: @georgejrjrjr posited that the idea presented in the "Skeleton of Thought" paper was similar to a different concept. He shared the link to the paper: https://arxiv.org/abs/2307.15337
- LLM Decontaminator: @bloc97 shared a blog post about the "LLM Decontaminator": https://lmsys.org/blog/2023-11-14-llm-decontaminator/
- Tree of thought for dataset generation: @yorth_night shared a tweet regarding this topic: https://vxtwitter.com/migtissera/status/1725028677124235288
- DeepMind's Music Creation: Various members discussed DeepMind's project to revolutionize music creation, sharing the project link: https://deepmind.google/discover/blog/transforming-the-future-of-music-creation/. @ldj shared a related YouTube link: https://youtu.be/rrk1t_h2iSQ?si=njkk-ajNonaiTum4
- Python-based Project: @f3l1p3_lv shared an older Python project of his: https://www.youtube.com/watch?v=eQC1JGMIxU0
- New Text-to-Video Model from Meta: @tsunemoto shared a link about this topic: https://emu-video.metademolab.com/assets/emu_video_v1.pdf and https://emu-video.metademolab.com/
- Synthia-v1.3: @teknium shared the dataset Synthia-v1.3: https://huggingface.co/datasets/migtissera/Synthia-v1.3
Nous Research AI Channel: announcements (7 messages🔥):
Announcement Summary:
- @teknium announced that **Hermes 2.5** has made it to the HF Leaderboard, currently ranking second place in the 7B models category.
Nous Research AI Channel: bots (7 messages🔥):
Probability Calculations Based on Given Conditions:
- Probability of being a woman among non-glasses wearers: @f3l1p3_lv presented a classroom scenario and asked for the probability of a randomly selected student being a female knowing the student does not wear glasses. @gpt4 provided the calculative answer indicating 60% probability based on conditional probability method.
- Probability of sum being 8 when two dice are thrown: @f3l1p3_lv asked for the probability of sum being 8 when two dice are thrown and have odd numbers on them.
- Probability of a French speaker being male: @f3l1p3_lv asked for the probability of a randomly selected French speaker from a group being male using the given data distribution of the group.
- Probability of getting a 3 on first throw of a dice with sum 7: @f3l1p3_lv ended with a question asking for the probability of getting 3 on first throw when two dice are thrown and the sum of the numbers on the dice is 7.
Nous Research AI Channel: general (7 messages🔥):
Discussion on AI in Music Generation, Training Large Models, and Nous Research:
- AI in Music Generation: @yorth_night shared a story about his musician friend who couldn't detect that a song was generated by an AI, specifically using https://app.suno.ai/.
- Training Large Models: Various members discussed the challenges and techniques related to finetuning large models. @cue shared their attempts to finetune a 70B model using deepspeed/FSDP but encountered Out-of-Memory (OOM) errors, referencing a Hugging Face blogpost on the subject: https://huggingface.co/blog/ram-efficient-pytorch-fsdp. @teknium suggested that finetuning a 70B model might require multiple nodes with 16x 80GB GPUs. @euclaise noted that LoRA can be equivalent to full finetuning if a large enough rank is picked and the embedding layers are also finetuned.
- Nous Research Updates: @alexatallah announced the availability of a Capybara 34B API and playground: https://openrouter.ai/models/nousresearch/nous-capybara-34b. @teknium clarified that Nous has four founders, including @alexatallah and himself. @f3l1p3_lv asked about the organizational structure of Nous.
Nous Research AI Channel: ask-about-llms (7 messages🔥):
Topics Discussed in "Ask about LLMS" Channel:
- Difference Between Continued Pretraining and Finetuning: The users discussed the distinctions between continued pretraining and finetuning. @teknium defined the difference as the size of the dataset, interpreting all finetunes as continued pretraining. @.wooser questioned whether 200k tokens used for fine-tuning and 20B tokens for continued pretraining could be seen as criteria to differentiate the two, with @teknium affirming this presumption.
- Usage of LLM for Code Review and Analysis: @ac1dbyte discussed their method of conducting Rust codebase analysis using Hermes for code review, and expressed interest in leveraging it for generating in-depth reports with specific elements such as vulnerability name, severity, and vulnerable functions for further normalization into JSON and insertion into MongoDB.
- Challenges in Training Non-roman Language Models: @.wooser shared insights about the complexity of training AI models for non-roman languages like Greek. Emphasizing on the need for a large dataset and probably a new tokenizer, .wooser mentioned about Japanese companies succeeding in training AI models using at least 40b tokens worth dataset.
- Training a New Tokenizer: @.wooser discussed the topic of training a new tokenizer mentioning although it seemed a complicated process, it appeared possible after going through a Huggingface link shared in the conversation.
- Dataset Creation for Japanese Fiction Text Data: @.wooser initiated a thread about creating a dataset from a large collection of Japanese fiction text data. Aimed at transforming LLMs into interactive storytelling AI, .wooser raised questions about the repetitive nature of the instructional prompts to the AI during this process.
- Public Perception of AI: A general discussion about the public fear and misunderstanding of AI ensued, emphasizing the need for broader acceptance of AI as a tool, as @.wooser and @teknium shared perspectives on people's fear of job threats and misuse of AI for deepfakes.
- LLM model types: @f3l1p3_lv enquired whether other models, apart from the Transformer neural entanglement model, can be used with LLMs.
- Links:
Nous Research AI Channel: memes (7 messages🔥):
(No specific topic title given in messages):
- @_automagic shared a link: https://news.ycombinator.com/item?id=36225723
- @.wooser commented on the link stating, "The perfect field for me, then!"
Guild: Alignment Lab AI
Alignment Lab AI Guild Summary
- Open discussion on federated learning; it was introduced by @erogol in relation to recent adapter methods and the potential utility of federated learning was acknowledged. A discussion on privacy concerns followed, in light of potential information leaks through model updates. @rusch suggested secure enclaves such as H100's TEE support as an alternative considering its 10% overhead and with attending cautions related to possible side-channel vulnerability. This was substantiated with links to adapter methods with federated learning and a GitHub repository.
- @nanobitz expressed eagerness towards undertaking projects revolving around language localization, specifically with regards to language learning models for non-English languages in the #looking-for-collabs channel.
- @far_el touched upon AI model operations and the immutability of base models in the #general-chat channel.
- Substantial discussion in the #oo channel around the OpenChat 3.5 Hungarian exam, which included grading invitations by @imonenext. The topic shifted towards defining 'fine-tunes' with @imonenext and @teknium. Snape remarks about GPT4's propensity to hallucinate. @alpindale's GitHub Repository, which is designed to filter common generated phrases from GPT and Claude models during training, was mentioned.
Alignment Lab AI Channel Summaries
Alignment Lab AI Channel: ai-and-ml-discussion (4 messages):
Federated Learning and Privacy Admission:
- Discussion on Federated Learning: @erogol initiated a discussion on federated learning with reference to recent adapter methods. The potential of this technology to be useful when data cannot be shared but model updates can, was highlighted.
- Concerns over Privacy Leakage: @rusch expressed concerns around the potential leakage of private information through model updates. Simultaneously, the usefulness of such updates was recognized.
- Preference for Secure Enclaves: @rusch indicated a preference for secure enclaves like H100's TEE support, which can perform private learning with just a 10% overhead. It was also warned that potential vulnerability to side-channels should be considered.
- Links:
Alignment Lab AI Channel: looking-for-collabs (4 messages):
Non-English LLM Project Proposals:
- @nanobitz expressed interest in projects related to the localization and usability of language learning models for Non-English languages. No specific details were given regarding the approach or challenges involved in these potential projects.
Alignment Lab AI Channel: general-chat (4 messages):
Discussion on AI Model Operations (Base Model):
- @far_el, while speaking of AI model operations, mentioned that the base model remains intact. However, without further information on the context, the exact topic or issue under discussion could not be determined.
Alignment Lab AI Channel: oo (4 messages):
OpenChat 3.5 Hungarian Exam Discussion:
- Hand-Grading Invitation: @imonenext invited anyone interested to help hand-grade the OpenChat 3.5 Hungarian exam. He noted that the exam was generated using a template and resulted in a 100% score with GPT4 grading.
- Disscussion on Fine-tunes: @imonenext and @teknium discussed what was considered to be fine-tunes, with Imonenext confirming that the 'Codellama' was considered a base.
- Report on GPT4 Hallucinations: @imonenext noted that GPT4 is hallucinating a lot, even worse than in chat mode.
- Links:
- @alpindale's GitHub Repository: @alpindale created a repository to gather all common phrases generated by GPT and Claude models in order to easily filter them out from training datasets. Contributors are invited to help.
Guild: Skunkworks AI
Skunkworks AI Guild Summary
- @pradeep1148 shared a YouTube video link in the off-topic channel, with no further conversation or context provided.
- In the bakklava-1 channel, there was a discussion regarding how GPT4v model handles multiple images:
- @occupying_mars posed a query on whether GPT4v makes comparisons akin to vector projections when presented with multiple images.
- @far_el hypothesized that the model could be capturing embeddings from each image and passing them in. This sparked a debate on the capacity of the model to create coherent logic between two images, understanding state changes, and the generality of using transformers.
Skunkworks AI Channel Summaries
Skunkworks AI Channel: general (3 messages):
Since the provided text is just a greeting and does not provide any substantial discussion points, URLs, or quotes, no summary can be generated based on the provided text.
Skunkworks AI Channel: off-topic (3 messages):
Shared Video Link:
- @pradeep1148 shared a YouTube video link.
Skunkworks AI Channel: bakklava-1 (3 messages):
Discussion on GPT4v's Support for Multiple Images:
- Query on GPT4v's Support for Multiple Images: @occupying_mars asked if anyone knows how GPT4v supports multiple images and further queried if they just compare the projections like vectors.
- Explanation and Speculation: @far_el suggested the process is likely capturing embeddings from each image and passing them to the model. @occupying_mars expressed confusion about the model's capacity to create a coherent logic between two images as if understanding state change but noted that such generality is possible with transformers.
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
- Data Intelligence Platform discussions surfaced, with @stevekamman sharing a Databricks blog post that explains what a data intelligence platform is.
- Conversation regarding Discord's Outage where @zorkian shed light on a recently published blog post detailing a near one-hour long Discord outage.
YAIG (a16z Infra) Channel Summaries
YAIG (a16z Infra) Channel: ai-ml (2 messages):
Data Intelligence Platform Discussion:
- @stevekamman shares a link to a Databricks blog post that explains what a data intelligence platform is.
- Links:
YAIG (a16z Infra) Channel: tech-discussion (2 messages):
Discussion on Discord's Outage:
- Discord's Outage: @zorkian shared about a recently published blog post regarding discord's nearly one hour long outage.
- Links:
Don't miss what's next. Subscribe to AI News (MOVED TO news.smol.ai!):