[AINews] AI Discords Newsletter 11/30/2023
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
- Sharing of AI tools and resources for feedback and learning, including AI Prompts by Guardiang and AI Discords Recap by Swyxio: - guardiang's prompt tool - AI Discords newsletter
- Active discussion on recent industry developments, focusing on entities like Stability AI and Payball: - Information about Stability AI being pressured to change the CEO source, - An issue about Payball bypass source
- Exploration of community members' activities, with a focus on a Reddit user, Emad: - Emad's Reddit profile source - Emad's damage control comment on Reddit source
- Announcement of a new episode for the Latent Space Podcast - Twitter announcement link
- In the academic scene, conversation points included an inquiry about a Context Reversal paper and requests for an updated list of covered papers in the llm-paper-club channel, demonstrating a proactive and studious community.
Latent Space Channel Summaries
▷ Channel: ai-general-chat (11 messages):
- AI Prompts by Guardiang:
@guardiang
shared a prompt tool and invited feedback, linking to the tool at chat.openai.com. - AI Discords Recap by Swyxio:
@swyxio
shared a link to a newsletter collating recent news and discussions from AI-related Discord channels, available at buttondown.email. - Stability AI Developments by Swyxio:
@swyxio
pointed out recent developments surrounding Stability AI, including investor pressure for the CEO to resign. The news is published in a Bloomberg article. - Payball Bypass shared by Jevonm:
@jevonm
shared a link to an archived article concerning the Payball bypass issue. - Chat on Emad's Reddit Activity by Coffeebean6887:
@coffeebean6887
noted Emad's recent Reddit comments, providing links to both his Reddit profile and a specific damage control comment he made.
▷ Channel: ai-event-announcements (1 messages):
- Latent Space Podcast Announcement: User
@swyxio
announced that a new podcast has gone live and shared a link to the announcement made on Twitter.
▷ Channel: llm-paper-club (3 messages):
- Chain of Note paper club notes: User
@swyxio
mentioned the topic of Note paper club notes. - Context Reversal Paper: User
@yikesawjeez
asked for the title of a specific Context Reversal paper. - List of Covered Papers: User
@zf0
inquired about the existence of an updated list of the already covered papers, noting that the current spreadsheet might be outdated.
OpenAI Discord Summary
- Announcement about an update from OpenAI, shared by user
@abdubs
on the OpenAI Twitter account. - In-depth conversation about the use of AI models in facilitating mining exploration reports;
@rjkmelb
shared success stories of having such service adopted by seven companies. - Animated debates about the potential impact of AI and quantum computing on future technology; the conversation showed divergent estimates about when quantum computing will enter commercial space.
- Exploration of the role of AI in education, especially of ChatGPT. Mixed views surfaced concerning whether AI-assisted productivity tools in education should be considered "cheating".
- Engaged discussions around the usefulness of AI in coding, with GitHub Copilot as the central subject. The dialogue presented contrasting views on its pros and cons.
- Conversations on the performance of Local Language Models (LLMs) on different hardware setups. Apple Mac M series chips with a good amount of RAM was identified as the preferred option.
- Commemoration of ChatGPT's one-year anniversary, reflecting on its utility and potential for enhancing learning and problem-solving.
- Concerns expressed about OpenAI's usage policies, suggesting the need for higher-cost, unlimited usage plans for heavy users, and inquiries about the perceived degradation of the core model via the web app.
- Discussions on multi-user chats and visual integration potential, introduced by users
@thetruth3130
and@rewire
. - Anticipation and speculation about future OpenAI releases and management changes, with topics ranging from the release of GPT-5 to potential Quantum Computing integration.
- Clarification about some issues like DNS propagation, issues with GPT-4 providing text instead of solutions, problems with age verification system, GPT usage on desktop, authentication configuration for OpenAI, and several billing issues.
- Introduction of various Custom GPT, prompting techniques, and discussions regarding issues and solutions related to GPT-4.
- Discussions of prompt engineering techniques, experiences with GPT-3, methods for outcome generation, and processes for identifying and resolving conflicts. Users
@madame_architect
,@exh0rt
, and@eskcanta
actively participated in sharing insights and suggestions.
OpenAI Channel Summaries
▷ Channel: annnouncements (1 messages):
- OpenAI Twitter Update:
@abdubs
shared an update from OpenAI.
▷ Channel: ai-discussions (262 messages🔥):
- Use of AI in Mining Exploration Reports: User
@rjkmelb
shared their successful implementation of using a AI model to read old geotechnical mining exploration reports and simplify them for non-technical researchers. This solution is employed by seven companies that pay around $1200 AUD per month to use it. There's a plan to scale the service to potential prospectors too. - Discussions on AI & Quantum Computing: Users debated the future of AI, predicting a significant acceleration in technological advancement, particularly in relation to quantum computing. While some argue that quantum computing will enter the commercial space in the next 10 years, others believe its commercial viability is still decades away.
- The Role of AI in Education: Users discussed the role of AI in education, especially the use of ChatGPT as a tool for students. Some argue that it should be treated like any other tool that enhances productivity and should not be considered "cheating", others voice concerns about students potentially using AI to bypass learning and understanding critical course content.
- Value of AI in Coding: Participants debated the usefulness of AI in coding, particularly the application of GitHub copilot. Some users find caution in its application due to its drawbacks while proponents argue about its value as a productivity multiplier, but only in the hands of those with domain knowledge.
- Discussion on Local Language Models (LLMs):
@thewizzard___
and@rjkmelb
discussed using local LLMs on different hardware setups. They noted that Apple Mac M series chips with a good amount of RAM performs very well for this task as opposed to a PC setup.
▷ Channel: openai-chatter (672 messages🔥):
- ChatGPT's Birthday and Relevance in Personal Development: Users celebrated ChatGPT's one-year anniversary, expressing their appreciation of the tool's utility and potential for enhancing learning and solving problems. User
@mysticmarks1
referred to using the tool for prompting complex neural network models, while@dyhr
used it for code optimization (source: https://chat.openai.com). - Discussion on Use of AI in Education: Users debated the role of AI, particularly ChatGPT, in completing schoolwork. Potential misuse of the tool for cheating was discussed, with concerns raised about if (and how) rules against such usage could realistically be enforced (source: https://chat.openai.com).
- Concerns about OpenAI's Policy and ChatGPT Limitations: Some users expressed frustration about policies on limiting model usage and code generation. Discussions revolved around whether it would make sense to have higher-cost, unlimited usage plans for heavy users. There were also concerns about the perceived degradation of the core model via the web app and a request for public clarification (source: https://chat.openai.com).
- Potential of Multi-user Chats and Visual Integration: Users
@thetruth3130
and@rewire
inquired about the possibility of having multiple active users in a chat and the potential integration of image upload while using a plugin, respectively (source: https://chat.openai.com). - Anticipation of Future OpenAI Announcements: Speculation and anticipation were high among users for future OpenAI releases and changes in company management. The topics ranged from the release of GPT-5 to changes in company leadership and the potential for Quantum Computing integration (source: https://chat.openai.com).
▷ Channel: openai-questions (120 messages):
- DNS Propagation Issue:
@zero1.ai
asked if anyone faced issues with DNS propagation after adding TXT record to their domain. - Issues with GPT-4 Providing Text Instead of Solutions:
@selekcjoner.
expressed frustration with GPT-4's inability to solve problems directly and its tendency to generate extensive text.@elektronisade
responded that this was a known bug which OpenAI is working on.@selekcjoner.
suggested allowing usage of older GPT versions as a quick fix. - Age Verification System:
@memenorio
discussed receiving an email notification about an age verification system from OpenAI. The notification was confirmed as legitimate following a link provided in the email. - Issues with GPT Usage on Desktop:
@apelambo
reported having problems running the GPT on his desktop, both on Chrome and Brave browsers.@solbus
suggested trying Firefox and inquired about the use of a VPN. Disabling the VPN appeared to fix the issue. - Discussion on Propagation of Code Errors:
@woodenrobot
shared issues of a circular dependency in a yaml code causing errors despite commenting out every reference. The issue persists and further discussion was suggested in another channel. - Discussion on GPT-4 Waitlist and Subscription: There were numerous discussions about the wait time for GPT-4, subscription issues, payment methods, and usage limitations. Users expressed concern regarding these issues but no definitive answers were provided.
- Issues with Rough Drafts and GPT Upgrades:
@a_chepurnyi
expressed disappointment after losing a month-long chat while trying to write a novel using GPT. It switched from version 4 to version 3.5 after exceeding the message limit.@solbus
confirmed that there is no option to revert back to version 4 once switched to 3.5, and suggested starting a new chat with copied content certain points from the older chat. - Auth Configuration for OpenAI:
@rosarioemmi
was having troubles configuring Auth for OpenAI, and asked for help. No specific solution was offered in the channel. - OpenAI API Questions: Discussion on consuming the OpenAI API, encountered errors, and rate limits was orchestrated by users including
@pandaterminator
,@rudrajadaun_09
, and@bombenstein
. Various solutions and tips are offered by the community. - Billing Help:
@coddent
expressed frustration over unresolved billing issues with OpenAI's API access and the customer support's inability to assist appropriately. No specific answers were offered.
▷ Channel: gpt-4-discussions (66 messages):
- GPT-4 Discussion Highlights:
-
@ajlbs
introduced a Custom GPT that provides insights inspired by the Bhagavad Gita. -@perfectpixels
presented an idea that uses program monitoring chat to decide which custom GPTs the message should go to,@rjkmelb
suggested this relates to Mixture of Expert (MoE) and provided a link for more information. -@niyohn.
asked for tools to create an embeddable chat app featuring a Custom GPT.@rjkmelb
suggested no current tools but it can be achieved via building out with Assistants API. -@leoalvarenga
sought advice for providing instructions to GPTs.@.pythagoras
and@adx2061
suggested a format mimicking programming languages using colons and brackets for better accuracy. -@8020
reported that their Custom GPT, which extracts text from a file and summarizes it, faced issues with the GPT-4 API, suggesting the issue lies within custom GPTs, not the API itself.
▷ Channel: prompt-engineering (83 messages):
- Prompt Engineering Techniques:
@madame_architect
shared recently published prompting techniques such as Chain of Note, Contrastive Chain of Thoughts, System 2 Attention, Thread of Thoughts, Take a Deep Breath, Emotional Prompts, Simulation Theory of Mind, and Step Back Prompting. Each technique was supplemented with a corresponding ArXiv reference number for further reading. - Experience with GPT-3:
@exh0rt
discussed their experience of working on a prompt designed like a game with specific rules and outlined challenges in maneuvering GPT's interpretations of the rules, having it maintain a desired tone, and dealing with instructional conflicts. - Generating Outcomes:
@eskcanta
suggested that@exh0rt
provide clear instructions that detail what the AI is supposed to do rather than stating what it should not do, to avoid confusion and foster better outcome generation. - Identifying Conflicts:
@eskcanta
highlighted the importance of identifying instructional conflicts within the prompts given to the AI. They suggested asking the AI to evaluate instructions for ambiguity, potential conflict, and confusion, and to list points of conflict. - Discussion on Tone and Language:
@exh0rt
and@eskcanta
had an extensive discussion on setting a tone for the AI, dealing with instructional conflicts concerning language use (jargon vs technical terms), and communicating expectations to the AI effectively. They further detailed processes for conflict resolution, understanding the AI's interpretation of instructions, and setting up efficient communication rules.
▷ Channel: api-discussions (83 messages):
- API-disucssions on OpenAI Discord Server: Members
@exh0rt
and@eskcanta
discuss issues with the former's instructions to a language model. Key suggestions from@eskcanta
included: - Directly telling the AI what to do instead of using negative language (i.e., what not to do). - Identifying conflicting instructions and rectifying them to avoid confusion. - Clearly defining contexts and expectations. - Avoiding repetition as the AI has a good memory and can comprehend from the first instance. - Making sure the instructions don't have unnecessary aggression as the AI is programmed to be compliant and not argue. @exh0rt
contemplated incorporating GPT's suggestions for clarification and resolving conflicts within his script-generating game prompt setup. He planned to run the modified game and see if the newer set of rules were easier for the AI to handle.- Later,
@madame_architect
shared new techniques for prompt engineering identified in various research papers published in the last 30 days. Some ideas included "Chain of Note", "Contrastive Chain of Thoughts", "System 2 Attention", "Thread of Thoughts", "Take a Deep Breath", "Emotional Prompts", "Simulation Theory of Mind", and "Step Back Prompting". @leoalvarenga
was seeking a template for creating custom GPT setups, but no responses were given to his query within the captured discussion.@exh0rt
summarized his understanding and main takeaways from the detailed discussions into a succinct list of rules for future reference, which@eskcanta
further clarified.
LangChain AI Discord Summary
- Extensive exploration into LangChain's capabilities, with members discussing intricate functionality, like setting similarity threshold within RAG Retriever to prevent hallucination, implementing interleaved streaming in React and ascertaining specific file locations in LangChain documentation.
- Streaming responses to FastAPI endpoints was a topic brought up as a direct query concerning OPENAI_FUNCTIONS_AGENT and AsyncCallbackHandler function.
- A bounty was offered for the integration of LangChain into E2B, an open-source sandbox for AI agents with details found on GitHub.
- Debugging Python code in VS Code whilst using LangServe was addressed, with successful application of the "Debug currently file and add path" configuration and further alternative debugging approach suggestions.
- The discovery that
qaTemplate
is now deprecated in JavaScript led to the suggestion of the usage of QChainoption{} instead. - Several helpful resources and potential collaborative opportunities were shared like the
analyticsvidhya.com
article on model quantization for large-scale deployment, and the invitation for contributions in the novelspace.tech project. - Unresolved requests for guidance, including a LangChain JavaScript starter boilerplate and assistance in implementing streaming with langchain + GPTCache.
- Member
@jasonzhou1993
showcased a YouTube video of a project where they created a research agent with autogen and sought community feedback.
LangChain AI Channel Summaries
▷ Channel: general (18 messages):
- Interleaved Streaming in React Framework:
@liminalstvte
asked about implementing interleaved streaming in the React framework, specifically he wants to detect when a function in Langchain LLM is called and then provide an answer and continue the conversation. - Location of Documented Files in LangChain: Users
@daii3696
and@baytaew
sought guidance on how to find the exact file locations that are frequently referenced in LangChain documentation, such as 'Defined in docs/api_refs/langchain/src/chains/conversational_retrieval_chain.ts:33'. - Streaming Responses to FastAPI Endpoints:
@sid.pocketmail
requested assistance in streaming responses to a FastAPI endpoint for OPENAI_FUNCTIONS_AGENT, asking for specifics on the AsyncCallbackHandler function. - Preventing Hallucination in RAG Retriever:
@legendary_pony_33278
inquired about setting a similarity threshold for the RAG Retriever (Faiss or Chroma) such that if no document in the vector database exceeds the threshold, no source is returned. This, they believe, would help prevent hallucination. - Bounty for LangChain Integration into E2B: A financial reward was offered by
@unicorn1997
for integrating LangChain into E2B, an open-source sandbox for AI agents. They provided a link to the bounty details on GitHub.
▷ Channel: langserve (3 messages):
- Debugging Python code in VS Code while using langserve:
@virtualmasterpieces
initially struggled with setting up debugging for his Python code in VS Code while using LangServe. His initial attempt using a specific configuration inlaunch.json
didn't yield the desired result. - Found Solution for Debugging: Later,
@virtualmasterpieces
found a solution to his debugging issue by using the "Debug currently file and add path" configuration. - Alternative Debugging Approach:
@veryboldbagel
suggested an alternative approach to debugging, advising to first try debugging the chain itself instead of debugging the chain through the server. This tactic could also be useful for writing unit tests/integration tests for the chain exposed by the server.
▷ Channel: langchain-templates (5 messages):
- Setting Input Keys in the Prompt: User
@liminalstvte
advised to set input keys in the prompt. - QA Template Deprecation in JS: User
@menny9762
discovered that theqaTemplate
is now deprecated in JavaScript, suggesting that this could be why the prompt was ignored. He stated adapting his usage to QChainoption{}.
▷ Channel: share-your-work (3 messages):
- Model Quantization for Large-Scale Deployment:
@soumyadarshani
shared an article on analyticsvidhya.com discussing model quantization intended for large-scale deployment.
- The Next Frontier of Email Efficiency with LLMS:
@soumyadarshani
also introduced an article on The Next Frontier of Email Efficiency with LLMS.
- Novelspace Project Collaboration:
@benji8214
is seeking collaboration in building out infrastructure at novelspace.tech and elysium.novelspace.tech, web apps using Langchain for chatbots. The project focuses on creating images for each chat response using the LCM SDXL model on Replicate.
- AI Smart Career Start:
@shving90
posted a link to AI Smart Career, a platform dedicated to helping users define a career path in AI.
▷ Channel: tutorials (3 messages):
- Langchain JavaScript Starter Boilerplate: User
@Faizul Ahemed
requested if anyone has a langchain JavaScript starter boilerplate repository. - Streaming Implementation with Langchian + GPTCache:
@sako_70104
expressed need for help on how to implement streaming with langchian + GPTCache as they were unable to find any relevant information. - Research Agent with Autogen:
@jasonzhou1993
shared a YouTube link to a project where they created a research agent with autogen, where they verify each other's work, and asked for feedback from the community.
Nous Research AI Discord Summary
- Discussion regarding a new concept coined as "knowledge-guided attention" for AI models introduced by
@maxwellandrews
, aiming at efficient handling of long documents by creating a temporary entity graph. This approach is designed to resemble human recall processes by focusing on entity relationships rather than verbatim text. A GitHub repository for relevant research on a knowledge graph attention network was shared. - Conversations about AI/ML career paths, with
@jaisel
asking about how others knew they wanted to focus on ML/AI.@teknium
shared his personal experience that was sparked with the advent of stable diffusion and the release of chatgpt. Also, during this discussion, two YouTube videos were posted, however, no discussion around these videos took place. - Detailed metric scores of the Qwen/Qwen-72B model were shared in the benchmarks log. An issue was raised concerning a potential problem with the
bigbench
task by main.ai, while@gabriel_syme
engaged the group to explain the meaning behind the benchmarking data. - Sharing of posts highlighting new techniques, including one that enables a 70b LLM inference on a single 4GB GPU, and another introducing Adept Experiments as a possible technique for fine-tuning agents. The size of StableLM was discussed with HuggingFace listing it as 2.8B. Various links were shared without context.
- Discussions on various language models, including 1.8B Qwen, 72B Qwen, and 70B LLaMA-2. Models were compared considering the performance and mentioned a need for independent evaluations. Community interest was expressed in Open Source Contributions, and multilingual data was discussed. The potential for future models like Hermes 2.6, Qwen 3b, Qwen 14b, Yi 34B, a 70B model, Mistral 70b or Yi 100b were also a point of discussion.
- A pizza emoji (🍕) was shared in the welcoming channel by user
@rgbkrk
without any provided context. - Discussion on the best stack for pretraining, conversations around hardware requirements for GGUF models, and instability of Quant models with function calling. User
@b_mc2
posted a link and asked for details about vision Hermes. Discussion on SERP API pricing and alternatives took place. - Queries about the Collective Cognition Models on Huggingface Leaderboard and clarification on project ownership concerning Open Access AI Collective.
- A comic relief message mentioning LIGHT MODE, a common display option in applications, was shared suggesting an ongoing discussion or inside joke within the community about preferred display modes.
Nous Research AI Channel Summaries
▷ Channel: ctx-length-research (1 messages):
- Knowledge-guided Attention:
@maxwellandrews
proposed a concept of focusing the attention of AI models more efficiently when dealing with long documents. His concept, termed "knowledge-guided attention," involves creating a temporal entity graph in a "scratchpad" while sliding over the context with a fixed-sized window (e.g., 2048 tokens). The model would then attend between the user's query and the knowledge graph and potentially the original passages that were used to create the relevant nodes in the knowledge graph. He emphasized that this approach is more reflective of how the human mind works - instead of recalling verbatim text, we remember a compressed representation of entity relationships (essentially an entity embedding). He shared a potentially relevant GitHub repository of a knowledge graph attention network.
▷ Channel: off-topic (7 messages):
- AI/ML Career Choices: User
@jaisel
instigated a discussion about career paths in AI/ML, asking how others knew they wanted to dedicate their time to Machine Learning (ML)/Artificial Intelligence (AI)—and how they found their work focus.@teknium
shared his personal experience, stating that he became obsessed with AI since stable diffusion and the release of chatgpt. - Shared Videos: Two YouTube videos were shared during the discussion.
@pradeep1148
shared a YouTube video. Later,@jaisel
shared another YouTube video, stating it made him reflective after listening to Sam. The topics of these videos were not discussed in the conversation.
▷ Channel: benchmarks-log (4 messages):
- Benchmark Results for Qwen/Qwen-72B model: main.ai shared detailed metric scores of Qwen/Qwen-72B model on tasks such as
truthfulqa_mc
,arc_challenge
,arc_easy
,boolq
,hellaswag
,openbookqa
,piqa
,winogrande
,agieval_aqua_rat
,agieval_logiqa_en
,agieval_lsat_ar
,agieval_lsat_lr
,agieval_lsat_rc
,agieval_sat_en
,agieval_sat_en_without_passage
andagieval_sat_math
. Both the accuracy (acc
) and normalized accuracy (acc_norm
) were reported for every given task. - Potential Issue with BigBench Task: main.ai mentioned that there may be a problem with the
bigbench
task due to a EOS/EOD issue. - Question about Benchmarks:
@gabriel_syme
asked for explanations of what the benchmark data means.
▷ Channel: interesting-links (16 messages):
- Performance of New_Models:
@metaldragon01
shared a blog regarding a new technique that enables a 70b LLM inference on a single 4GB GPU. In a later discussion,@euclaise
and@teknium
had a discussion about the impressive results of a 1.8B GSM8K model, although some suspicions were raised. - Agent Fine-tuning:
@metaldragon01
queried whether there should be more focus on fine-tuning agents after sharing a blog introducing Adept Experiments. - Introduction of Adept Experiments:
@coffeebean6887
announced the introduction of Adept Experiments which might be related to fine-tuning agents. - Mistral and Bugatti Collaboration:
@atgctg
shared a video about an interesting partnership between Mistral and Bugatti. - Model Comparisons:
@teknium
raised a question about the size of stableLM. In response,@euclaise
clarified that Hugging Face lists it as 2.8B. - Shared Links: Various links were shared by the users:
- vxtwitter.com by
@tsunemoto
. - fxtwitter.com by@sanketpatrikar
. - twitter.com by@nods
. - Google Docs link by@atgctg
. - Hugging Face models and Hugging Face models by@euclaise
. -
▷ Channel: general (240 messages🔥):
- New Models & Evaluation: Users discussed various language models, including 1.8B Qwen, 72B Qwen, and 70B LLaMA-2. Comparisons were made between these models and Mistral, StableLM, GPT-3.5, and GPT-4. Qwen models reportedly outperformed the baselines on various tasks, but independent evaluations were deemed necessary to gain a comprehensive understanding of their performance (
@teknium
,@teknium
). - WeAreOnlyHuman Community Livestreams:
-
@altryne
announced a series of livestreams to learn how to use platforms such as WandB for models and open-source Language Learning Models (LLMs). - Open Source Contributions: User
@xyzyrz
showed interest in contributing to open-source ML efforts and sought advice for beginners looking to become involved. - Models and Datasets for Non-English Languages: User
@zhil_arf
suggested translating pre-existing datasets to non-English languages to produce multilingual LLMs. - Potential Future Model Updates:
- ML model makers discussed strategies for creating future versions of Hermes, such as Hermes 2.6, with training possibly planned over models like Qwen 3b, Qwen 14b, Yi 34B, and a 70B model (
@teknium
). - Talks about future models like a potential Mistral 70b or Yi 100b were also mentioned (@metaldragon01
). - Training for Diverse Task: User
@mihai4256
mentioned the fascinating possibilities of using an audio language model to fine-tune for various tasks.
▷ Channel: welcomes (1 messages):
- Participant
@rgbkrk
offered a pizza emoji (🍕) in the discussion. However, the context or purpose of this contribution is not clear from the provided message.
▷ Channel: ask-about-llms (36 messages):
- Discussion on best stack for pretraining: User
@mtybadger
inquired about the best stack for pretraining, expressing dissatisfaction with the mosaicml composer/streaming dataset. - Information about Vision Hermes: User
@b_mc2
posted a link about a multimodal vision Hermes and asked for notable details such as image size for training, prompt format etc. [link]. - Discussion on SERP API pricing and alternatives: Users
@teknium
and@markopolojarvi
had a discussion concerning the high cost of the SERP API and the potential alternatives, including building their own SERP scraping infrastructures and utilizing Google Custom Search Engine (CSE) or serper API. - Hardware requirements for running GGUF models: User
@umarigan
made inquiries about the hardware needed to run GGUF models swiftly. Subsequent discussions with other users revealed that they were attempting to run these models without a GPU, and were also using different bit versions (q5_k_m and q2_k). - Instability of Quant Models with Function Calling: User
@raddka
noted that quant models, even at Q6/Q8 versions, are unstable at executing system instructions for function calling when compared to original models like OpenHermes-2.5.
▷ Channel: collective-cognition (5 messages):
- Collective Cognition Models on Huggingface Leaderboard: User
@yobibyte
inquired about a collective cognition model that seemingly disappeared from the Huggingface leaderboard.@teknium
clarified that the model was not deleted and provided the link to the model on Huggingface. - Project Ownership Clarification: Upon further clarification,
@yobibyte
mentioned the Open Access AI Collective project associated with the Jackalope model.@teknium
directed this to<@257999024458563585>
, stating that these were their projects.
▷ Channel: memes (1 messages):
- User
@nruaif
shared a comic relief message, referencing the LIGHT MODE function, typically a display setting in applications. This may suggest an ongoing discussion or inside joke within the community about preferred display modes. No further context or follow-up discussion was provided.
Alignment Lab AI Discord Summary
- Deep learning application in materials discovery:
@entropi
shared a DeepMind article detailing how AI has been used to discover millions of new materials. - Invite to a WandB live stream:
@altryne
invited users to a live stream discussing how to use WandB and encouraged those using WandB to connect for future collaborative opportunities. The stream link was provided. - Questions regarding Synthia 1.3's generation:
@imonenext
inquired whether Synthia 1.3 was the output of GPT4 and sought clarification about the 'system' and 'instruction' fields in its HuggingFace dataset page. - Inquiry about the Alignment Labs launch:
@zolandinho
asked about the expected launch date for Alignment Labs. - The user
@puffy310
indicated intent to investigate an unspecified issue or topic. - Work and skills overview:
@frankyan.ai
detailed professional experience and skills in full-stack development, software engineering, DevOps practices, and AI project integrations, citing specific experience with ChatGPT. - Upcoming plans and author of Synthia:
@giftedgummybee
and@imonenext
discussed upcoming work on system prompts.@imonenext
and@nanobitz
conversed about the author of Synthia, with@nanobitz
providing the author's name (Miggel Tissera) and a link to his Discord profile.
Alignment Lab AI Channel Summaries
▷ Channel: looking-for-collabs (1 messages):
As the task requirement, this prompt is designed to test how an assistant can follow the instruction to summarize a chatbot message history by applying a specific format that matches the given instruction. But it appears to be a misunderstanding as no history messages are provided here, and thus the assistant can't provide a meaningful response. It suggests reviewing the task setup or providing a proper dataset for the assistants to generate the expected output.
▷ Channel: general-chat (7 messages):
- DeepMind's Discovery of Millions of New Materials:
@entropi
shared an article from DeepMind, detailing how deep learning has been used to discover millions of new materials. - Live Stream on Using WandB with LDJ:
@altryne
extended an invitation to join a live stream withLDJ
where they will be discussing how to use WandB. The event can be accessed via this twitter post and Altryne also encouraged others who use WandB to reach out for future learning experiences. - Query about Synthia 1.3's Generation:
@imonenext
asked whether Synthia 1.3 was generated by GPT4 and provided a link to its huggingface dataset. The user also inquired if the 'system' field is the system prompt to GPT4 and the 'instruction' as the user's message. - Alignment Labs Launch Inquiry:
@zolandinho
inquired about the launch date for Alignment Labs. The user also sought the assistance of admins to answer their questions.
▷ Channel: oo (1 messages):
@puffy310
indicated they will look into a particular issue or topic, though the specific topic or issue was not mentioned within the provided context.
▷ Channel: looking-for-work (1 messages):
- Work Experience and Skills: User
@frankyan.ai
brings more than 25 years of full stack development and software engineering experience, with a focus on financial sector projects. His skillset includes creating websites and single page applications, designing and implementing microservices, using open-source components and automation tools, and building CI/CD pipelines.
- DevOps Experience:
@frankyan.ai
practices DevOps skills, working with Terraform and Ansible scripts to manage cloud resources and set up CI/CD pipelines, logging, monitoring, and tracing infrastructures.
- AI Projects:
@frankyan.ai
has worked with ChatGPT and integrated it into various projects. These include an AI-enhanced resume matching system, making ChatGPT accessible to friends in China, and researching FastChat to support LangChain integration.
▷ Channel: oo2 (9 messages):
- Upcoming Plans:
@giftedgummybee
asked about any upcoming plans.@imonenext
replied that they are working on system prompts. - Synthia's Author:
@imonenext
inquired about the author of Synthia.@nanobitz
provided the author's name as Miggel Tissera and shared that they had seen him on Axolotl discord. When@imonenext
couldn't find him on OpenAccess AI Collective,@nanobitz
provided a direct Discord link to his profile.
Skunkworks AI Discord Summary
Only 1 channel had activity, so no need to summarize...
Skunkworks AI Channel Summaries
▷ Channel: off-topic (1 messages):
@pradeep1148
shared a YouTube link. The content of the video was not discussed.
LLM Perf Enthusiasts AI Discord Summary
- OpenAI Leadership Update: User
@joshcho_
informed about Sam Altman returning as CEO of OpenAI and the formation of a new initial board as posted on OpenAI’s blog. - Discussion on Performance Issues: Users reported slowness with Azure OpenAI calls and slower than usual operation speeds on both unnamed main platform and Azure. On contrast, user
@jet0266_71389
highlighted that Neuralmagic tends to demonstrate good CPU performance. - Open Source Models and Deployment: Conversation on potential performance impact of larger context window, with the concern raised by
@thebaghdaddy
about the possible reduction in performance when using larger context window upto GPT-4. The community also discussed plans to test the Starling-lm-7b model and@thisisnotawill
expressed the idea of creating a Hugging Face endpoint. User@robotums
queried about the exposure of log-probabilities in pplx-API to examine perplexities in cross-encoder models.@thisisnotawill
also looked for advice on running open source models locally and deploying them to Azure, stating a rate of $200/hr for guidance. A helpful resource was shared by@pantsforbirds
, pointing towards Mozilla-Ocho’s Llamafile as a potential guide. - Event and Project Notification:
@res6969
announced investor interest for an event in New York City and briefly shared plans to create a Luma either in the current day or the next. - Appreciation for New Channel Creation: User
@joshcho_
acknowledged<@757392677280022549>
for creating the new prompting channel.
LLM Perf Enthusiasts AI Channel Summaries
▷ Channel: general (1 messages):
- OpenAI's New CEO and Initial Board: User
@joshcho_
shared a link to OpenAI's blog post announcing that Sam Altman is returning as CEO and unveiling the company's new initial board.
▷ Channel: gpt4 (1 messages):
- Azzure OpenAI Calls Performance Issues:
@psychickoala
reported experiencing slowness with Azzure OpenAI calls.
▷ Channel: opensource (17 messages):
- Performance Impact of Large Context Window:
@dongdong0755
suggested prompting an entire dataset, but@thebaghdaddy
cautioned that a larger context window often decreases performance, a phenomenon observed till GPT4.
- Testing Starling-lm-7b:
@thisisnotawill
announced plans to test Starling-lm-7b and even create a Hugging Face endpoint.
- Query About log-probs Exposure in pplx-API:
@robotums
asked if the pplx-API exposes log probabilities for examination of perplexities in cross-encoder models.
- Request for Guidance with Open Source Models:
@thisisnotawill
requested assistance with running open source models locally and deploying them to Azure, offering $200/hr for 1-2 hours of guidance.
- Potential Useful Resource Shared:
@pantsforbirds
shared a potentially useful resource, the Llamafile by Mozilla-Ocho, hoping it could be of some help.
▷ Channel: speed (1 messages):
- Neuralmagic's CPU Performance:
@jet0266_71389
indicated that Neuralmagic seems to have a good CPU performance.
▷ Channel: irl (2 messages):
- Upcoming NYC Event: User
@res6969
mentioned that they have investors interested in attending an upcoming event in NYC. - Creating a Luma:
@res6969
also mentioned planning to create a Luma either today or the next day.
▷ Channel: openai (3 messages):
- Performance Issues: Users
@kev.o
and.psychicKoala
reported experiencing slower than usual operation speeds, both on the unspecified main platform and on Azure.
▷ Channel: prompting (1 messages):
@joshcho_
expressed gratitude for the creation of the channel, specifically thanking<@757392677280022549>
.
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.
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.
AI Engineer Foundation Discord Summary
Only 1 channel had activity, so no need to summarize...
AI Engineer Foundation Channel Summaries
▷ Channel: general (4 messages):
- AWS re:Invent Sessions: User
@juanreds
initiated a discussion about the AWS re:Invent sessions. - Challenges with AI Adoption:
@juanreds
expressed concerns on how organizations can start using AI, including decisions around AI applications and processes. - Framework of AI Best Practices:
@juanreds
proposed the idea for an AI Engineer Foundation to create a standard, a framework of best practices, to guide organizations wanting to adopt AI.
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.