This is AI News! an MVP of a service that goes thru all AI discords and summarizes what people are talking about, so that you can keep up without the fatigue.
@mockapapella, who was seeking more cost-effective and efficient alternatives.@vcarl, albeit without a clear link to existing conversations.@slono.@tiagoefreitas and @swyxio.@youngphlo.@swyxio.Latent Space Channel Summaries
@swyxio discussed exploring model routing and shared some interesting fine-tuning results via a Hacker News link.@swyxio also informed participants about an ongoing situation with the Yi 01 model.@mockapapella expressed difficulties with generating descriptions for a large dataset of part numbers using various models and is seeking suggestions for alternatives while trying to minimize cost and maximize efficiency.@vcarl suggested reading a Tweet by tldraw. The tweet does not have a clear connection to the ongoing discussions.@slono toyed with creating a phrase for the concept of modifying generated code, suggesting "Generacoded" and "Prograsummoned."@tiagoefreitas and @swyxio queried if anyone else had tested the website codegen.com.@swyxio shared an analysis of the costs and market demand for an AI girlfriend product. According to this analysis, it clarified why ElevenLabs has achieved unicorn status.@youngphlo attributed the recent explosion to a paper which can be found at arxiv. @swyxio asked if the code fusion paper has been discussed.Users across different channels expressed difficulties with the 'My Plan' feature on the OpenAI platform, noting it wouldn't load, despite attempts across different browsers and multiple user reports. Both ChatGPT Plus subscriptions and usage limits were under discussion, with users like @katthwren and @dsmagics reporting being logged into the free plan despite having active Plus subscriptions, and users like @gavin5616, @mustard1978, and @18870 discussing usage limit changes from 50 to 40 messages every 3 hours.
Users reported technical and UI issues across the platform. Users @zahmb noted changes in sidebar icons on the OpenAI platform. Others reported slow performance and errors using ChatGPT, possibly related to plugins, with users like @libertadthesecond and @exx1 discussing these issues. Users also reported on difficulties with creating, configuring, and modifying GPTs, addressing problems like URL formatting for actions, changes not saving, and difficulties executing Python code in their custom GPTs.
A significant portion of the discussions involved using the ChatGPT API. Questions were posed about how to pass an API key, connect the ChatGPT API to Discord, and create API actions involving oauth2 and token usage. Users even inquired about building AI-related tools like a scraper for a mobile app.
Differences and comparison between GPT models were explored, with users questioning the performance difference between GPT-3 and GPT-4. In addition, there were in-depth discussions about potential UI enhancements, like embedding custom GPTs on websites, and improving the platform's ability to handle specific tasks, including date formatting, retrieving specific instructions from a website, and improving SEO optimisation.
Community members have shown willingness to share their experience and knowledge. For example, @.kalle97 showcased his own SEO-optimized AI writing GPTs capable of producing long articles - however, he didn't share the precise formulation of the prompt. Other users lent advice on tackling common issues and advised on best practices for using and interacting with AI, like understanding that additional uploaded files to the agent don't continually modify the agent's base knowledge or identifying the potential pitfall with the Configure and Create tabs in the GPT Builder.
OpenAI's credit system and top-up mechanism was also under discussion, with contrasting opinions on the minimum top-up requirement for accessing high-performing GPT models being addressed.
OpenAI Channel Summaries
@tilanthi expressed concerns about his GPT agent's ability to learn from additional information, beyond its initial training data. @solbus clarified that uploaded files are stored as "knowledge" files for the agent to reference when required, but they do not continually modify the agent's base knowledge.@solbus shared some practical insights on how to work with the Configure and Create tabs in the GPT Builder. Highlighting a potential pitfall, they mentioned that manually filled in information on the Configure tab can be overwritten if the Create tab is subsequently used.@zahmb noted that there have been changes in the sidebar of platform.openai.com. Specifically, they reported that two icons disappeared from the sidebar, one for threads and another for messages.@milestones95 made a query about building a scraper to navigate a mobile app and grab XML using Appium Inspector. In response, @yodababoom clarified that such a task doesn't necessitate the use of AI.@elektronisade informed @ishaka02 that the minimum top-up for OpenAI credit to access GPT-4 or GPT-4-1106 model is one dollar, contradicting the user's belief that it was five dollars. The evidence was presented in the form a link to the OpenAI help article on the topic.@libertadthesecond mentioned possibly being linked to plugins in use. @exx1 commented that Bing uses a modified GPT-4 model which might account for performance differences. @_ciphercode questioned the difference in performance between GPT-3 and GPT-4, stating that he found them to yield similar results.@alfarmer and @busybenss asked for help about DNS validation and domain issues, respectively. They were advised to check their TXT records and the need for a hosting server was mentioned.@mrbr2023, @thesocraticbeard, @sumo_dump, @fujikatsumo, @faikout, @elfonsen, @colwood, @LN, @domvo, @new3ra, @dsmagics, @highsquash, and @tljamesa.@foxabilo and @xh providing tips on how to generate desired outputs.@danielbixo, @ajripon, @playbit, and @gavin5616 narrated facing issues with the app loading, specific functionalities not working, and difficulties in using the chatbot.@vdrizzle_ and @ddantee reported experiencing issues when asking questions on the application, with responses showing "Oops, an error occurred!" message frequently.@gavin5616, @mustard1978, and @18870 discussing the issues around exceeding usage limits and the change from 50 to 40 messages every 3 hours.@xzz3300 and @mjamiv, with questions about how to pass on the API-key, and discussions about how to connect the ChatGPT API to Discord.@juregg discussing whether it would be possible to create a script that fetches and shares an required image (for example, a graph for a SAT question) in a chat.@katthwren and @dsmagics noticed they're logged into the free plan instead of their Plus plan despite active subscriptions. @xh sought help with simplifying LaTeX outputs from GPT to plaintext. Several users such as @solbus and @syndicate47 confirmed the outputs were indeed in LaTeX and suggested asking the AI specifically for plain text responses.@amarnro and @pietman discussed issues with saving and executing actions due to URL formatting. It was suggested to remove trailing slashes from URLs for actions to work correctly.@stealth2077 and @loschess noted difficulties when trying to amend the instructions of their GPTs. The changes were not saving properly, causing the GPTs to revert or ignore the revisions.@.alexandergo inquired about creating API actions involving oauth2 and token usage.@sirkaiwade, @cybector, and @elegante94 were discussing the changes made to usage caps for GPT-4, noting the reduction of limits that have impacted their work.@gordon.freeman.hl2 asked if OpenAI plans to allow embedment of custom GPTs on websites, which was identified as a widely sought-after feature.@no.iq queried about how to customize parameters such as "temperature" in OpenAI's new User Interface (UI), noting that the feature doesn't appear in the playground anymore. @goldmember777 sought advice on making the AI retrieve multiple instructions from a website but share with the user one at a time, to avoid repetitive calls to look up the resource.@.kalle97 shared a list of his AI writing GPTs, emphasizing they are SEO-optimized and can produce long articles. However, he declined @mrposch's request to share some parts of his prompt.@alishank_53783 discussed challenges with prompting for date formatting. @eskcanta provided advice, including explicitly mentioning the desired date format in the instructions and providing good and bad examples.@jungle_jo and @eskcanta discussed how the complexity and clarity of language in prompts can affect the images created by the AI.@cat.hemlock raised an issue with getting their custom GPT to execute Python code. @eskcanta suggested it might be due to the unavailability of the required tiktoken library, advising @cat.hemlock to request OpenAI for support or find an alternative solution.Temperature Parameter Adjustment in New UI: User @no.iq expressed difficulty finding where to tweak parameters like 'temperature' in the new user interface, as it's no longer visible in the playground. No solutions were proposed.
Multiple Instruction Retrieval: User @goldmember777 sought advice on retrieving multiple instructions from a website and sharing them with the user one at a time, with the goal of minimizing the frequency of GPT looking up resources online. No solutions were discussed.
Writing GPTs and SEO Optimisation: @.kalle97 shared links to their SEO-optimized AI writing GPTs capable of producing long articles. They refused to share parts of their prompt when asked by @mrposch, citing significant time investment in building it. Links shared:
Date Formatting in Custom GPTs: User @alishank_53783 reported issues with prompting for date formatting while parsing emails. They had specific need of dates being returned in 'YYYY-MM-DD' format. @eskcanta suggested being more explicit with prompt instructions and gave a few examples.
Executing Python Code in Custom GPTs: @cat.hemlock inquired about how to get their custom GPT to execute python code. They reported that despite having code interpretation selected and internet browsing enabled, the model appeared not to exhibit code execution behavior. @eskcanta suggested potential workarounds but emphasized that certain libraries may not be supported.
LangChain AI Channel Summaries
@jinwolf2 asked for assistance regarding an error they encountered when converting the response from a consumed API in their code, affecting the openAi_function. @alimal inquired about successful methods for using LangChain and/or APIs for quality assurance with tabular data. @brio99 requested clarification on the question. @Siddhi Jain requested help regarding an error with the _aget_relevant_documents() function in ConversationalRetrievalChain, specifically when the 'run_manager' argument needs to be passed. @abhi578 asked about the embedding generation part in LangChain, seeking to create embeddings for prompts in a certain format and find similar embeddings for user-provided context and query. The idea is to match these with user-provided queries for similar queries, and pass the filtered queries with their context and response for response generation. Several strategies were discussed including an initial query reformatting. @abhi578 also asked if there is a way to use vectorDb for retrieval.@0xtogo asked if it was possible to use an existing assistant without calling the create_assistant function in LangChain.@juan_87589 inquired about the future implications of the recent collaboration between Microsoft and LangChain, asking if there will be prioritization of Azure over other cloud providers. Microsoft Collaboration Blogpost@seththunder asked for clarification on the differences between certain syntax structures in LangChain code.@tonyaichamp reported that LangChain was processing unusually slow, potentially indicating server or network issues.@eyueldk inquired if it was possible to get the sources/references used by a LangChain agent in reaching its conclusion.@fishyccy asked if anyone had tried using GPT's Interpreter/Preview API to create live queries to a database, with the aim of having the GPT output an analysis or summary.@endo9001 asked how to save the conversation buffer memory into a json file for external use and for loading into future conversations.@beffy22 reported a problem with vectorStore.asRetriever(), where it always returned documents.@quantumqueenxox asked for guidance on which LangChain web loader to use for loading a certain website and whether the loader would load all the sub URLs as well.@maverick5493 asked for guidance on uploading files into the OpenAI 'retrieval' assistant using LangChain, being specifically interested in corresponding code for creating files and passing the file_ids to OpenAIAssistantRunnable.@rajib2189 asked if there was a way to check if a particular dataset or vocabulary set was used to train commercial models like OpenAI.@andriusem was having difficulties loading the .env file to access API keys. @attila_ibs suggested importing and loading dotenv using the following code:
from dotenv import load_dotenv, find_dotenv
_ = load_dotenv(find_dotenv()) # read local .env file@attila_ibs recommended loading dotenv in server.py.@veryboldbagel suggested @723833811757957142 to upgrade to the latest version of langchain as it uses orjson, which can serialize an ndarray, thus simplifying the process of sending data over the wire and decoding it on the other side.@1121736064751636510, @veryboldbagel suggested the creation of extra endpoints for the ingestion of files. He mentioned that the docstore can be backed by redis or users can implement their own choice of persistence, providing a link to the RedisStore documentation.Introduction of YouTube GPT tool: User @taranjeetio shared a link to a tool that allows users to create YouTube GPTs easily.
Launch of Appstorm.ai: @appstormer_25583 introduced Appstorm.ai, a platform for building custom GPTs for free with different functionalities to meet various needs. The user also shared a series of links to access the different GPTs built on the platform, including ELI5 for STEM topics GPT, Research Assistant GPT, and others.
Sale of a Specialized Text Messaging App: @.broodstar shared the intention to sell a text messaging app developed over two years. This app allows for the saving of long text message conversations within the app and has a GPT chatbot integration. The user shared a link to a demo of the chatbot in the app.
Update on Pantheon: @kingkookri announced updates on the Pantheon app, including improvements in speed and a more intuitive interface. Also, the user shared a link to the app and invited more people to test the app.
Algorithmic Post Creation Platform: @agenda_shaper briefly mentioned a platform equipped with an algorithm for creating posts. However, no additional details or links were provided.
@teknium happily announced that OpenHermes 2.5 is finally on the HF Leaderboard, bagging the 2nd place in 7B models..wooser.Nous Research AI Channel Summaries
@yorth_night started a conversation about finding a service that can hold terabytes of data for free because of the storage limitations they are facing with their project. Several users participated in the thread, discussing potential solutions.@.benxh suggested using Hugging Face and stream the dataset so it wouldn't need to be held on disk. Following the Hugging Face suggestion, @crainmaker shared a link to an Hugging Face discussion where it was clarified that a user can upload as much data as they want, provided each file is less than 50GB.@tsunemoto and @crainmaker proposed a method of hashing the image data using MD5 and then matching the hashes with the image set in order to efficiently process and store the dataset.@tsunemoto expressed intention to adopt the suggested strategy of chunking the files and then pushing to a Hugging Face repository using the company's Python implementation, as advised by @.benxh.@crainmaker pointed out that relying solely on Hugging Face could be a potential issue of centralization, suggesting the need for distributed backups to prevent a single point of failure.Discussion on the Skeleton of Thought Paper: User @georgejrjrjr mentioned an academic paper titled "Skeleton of Thought" paper here claiming that it has done something similar to the ongoing discussion.
Link share and Discussion on Decontaminator: User @bloc97 shared a link to a blog post on "Training on the rephrased test set is all you need" blog post here.
Tree of Thought Dataset Generation Technique: User @yorth_night shared a link to a tweet discussing a novel technique called "Tree of thought" for dataset generation tweet here.
DeepMind's AI in Music Creation: User @yorth_night shared a link to a post by DeepMind discussing how AI is transforming music creation post here. The discussion shifted to the prospects of open-source vs closed-source models in this space. User @.wooser opined that, despite being behind in understanding the state of the art AI tech, if Google can develop such technologies, others can too eventually.
Meta's New Text-to-Video model: User @tsunemoto shared a link to a Text-to-Video model developed by Meta here. The model was described as "animate-diff on steroids". The project's PDF was also linked here. There was a positive response from user @qasb towards it.
@teknium announced that OpenHermes 2.5 is finally on the HF Leaderboard, achieving the 2nd place in 7B models.@f3l1p3_lv asked a series of questions on probability. The questions were related to translation and calculation of probability based on given scenarios, such as chance of a student being a woman given that she doesn't wear glasses and the probability of the sum of two dice roll being 8 given that the numbers are odd.@gpt4 and @compbot responded to these questions, providing detailed and comprehensive answers.@f3l1p3_lv expressed that Claude v2 was the best.@f3l1p3_lv asked an AI bot if it was alive. To which, @gpt4 responded by stating it's not alive in a biological sense, merely a program designed to simulate conversation and assist users.@f3l1p3_lv posed. These went from a straightforward counting scenario to one that includes conditional probability. Questions included the chances of a student being a woman given she doesn't wear glasses, the probability of the sum of two dice rolls being 8 as well as the probability of getting a 3 on the first roll of a dice given that the total sum is 7. All these queries coloured an engaging discussion in the bots channel.@yorth_night discusses Suno.AI's AI music generation capabilities, particularly praising the instrumental aspects. He included a link to Suno.AI's service (https://app.suno.ai/).@cue, @.wooser, @teknium, @.wooser, @euclaise, among others. @cue shared a link to a blog post on this topic (https://huggingface.co/blog/ram-efficient-pytorch-fsdp).@alexatallah announces the launch of Capybara 34B API and playground available at https://openrouter.ai/models/nousresearch/nous-capybara-34b.@yorth_night, @.wooser, @euclaise and @teknium weighing in. @.wooser provides a link to a study on this topic (https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2).@00brad, @teknium @alpindale, @giftedgummybee discuss training models and managing 'slop', undesirable input/output behaviors during training. @alpindale shares a link to a GitHub project for contributing to this effort (https://github.com/AlpinDale/gptslop).@teknium, @.wooser discussed how languages required a large dataset for full fine-tuning. Particularly for non-roman languages, developing a new tokenizer might be necessary. They also shared a HuggingFace tutorial on how to train a tokenizer.@.wooser about the difference between continued pretraining and a full finetune, to which @teknium replied that the difference lies mainly in the size of the dataset.@.wooser discussed his intent to utilize a large Japanese fiction text dataset to create a language model akin to NovelAI or AI Dungeon. They were seeking advice on whether their approach —having an instruction to "Finish the following section of text" followed by an incomplete sentence— was a viable method of structuring their dataset.@teknium and @.wooser shared personal observations about public sentiments towards AI, noting some levels of apprehension due to risks associated with jobs and deepfakes.@ac1dbyte sought advice on how to utilize the AI, Hermes, for analyzing Rust code for vulnerabilities and consolidating the report in a semi-normalized JSON format for post-processing and insertion into MongoDB. Their existing approach involved looping through individual files with a base prompt and produced satisfactory reports._automagic referring to a News Ycombinator post..wooser expressed enthusiasm for an unspecified field, stating it's "The perfect field for me, then!".@rusch, discussing the challenges of preventing information leakage during model updates, highlighting the potential of H100's TEE (Trusted Execution Environment) Support in secure enclaves for private learning.@nanobitz in the context of looking for potential collaborations.@imonenext for hand-grading the OpenChat 3.5 Hungarian exam, prepared using the few-shot template by <@748528982034612226>, which had scored 100% when graded by GPT4.@imonenext, explaining that in the original repository, fine-tunes were evaluated in a zero-shot manner, and base models were evaluated in a five-shot manner.@alpindale and an invitation to contribute. Relevant repository link.@teknium, using a playful emoji. The content may not hold topical relevance.Alignment Lab AI Channel Summaries
@rusch discussed the challenge of preventing the leakage of private information in model updates, indicating this as a potential setback for private learning.@rusch showed a preference for secure enclaves, especially H100's TEE (Trusted Execution Environment) support, citing its usefulness for private learning.@rusch mentioned that Federated Learning (FL) and homomorphic encryption possess certain performance issues that limit their application to only niche use cases.@rusch, TEEs have approximately 10% overhead, requiring the same code and approach. However, they are more vulnerable to side channels and, therefore, cautioned against their use with Mossad or the NSA.@nanobitz brought up the topic of Non-English Language Models (LLM).@imonenext put forward an opportunity for other users to hand-grade the OpenChat 3.5 Hungarian exam. This exam was generated using a "few-shot" template by <@748528982034612226> and scored 100% when graded by GPT4.@imonenext explained that in the original repo, fine-tunes were evaluated in a 0-shot manner while base models were evaluated 5-shot, which might have caused fine-tunes to seem overfitted.@imonenext brought to attention some frequent hallucinations experienced with the GPT4 of <@748528982034612226>, stating they were worse than those in chat mode.@alpindale introduced a new repository created with the objective of gathering all the commonly generated phrases by GPT and Claude models. The aim of this is to easily filter these terms from training datasets. He invited other users to contribute. The link to the project is here.@teknium made jovial inputs in the conversation such as using playful emoji "@oleegg sending good morning wishes in the 'skunkworks' channel. @pradeep1148 in the off-topic channel.@occupying_mars asking questions about the model's process, as well as its ability to maintain logical coherence between two images. User @far_el proposed that GPT4v might be creating embeddings from each image for processing.Skunkworks AI Channel Summaries
@oleegg posted a general morning greeting to the 'skunkworks' channel.@pradeep1148 shared a YouTube link.@occupying_mars asked about how GPT4v supports multiple images and whether the projections are treated as vectors. @far_el suggested that the model likely captures embeddings from each image and then passes them to the model.@occupying_mars also raised a query about the apparent coherent logic between two images that GPT4v seems to maintain. However, they expressed uncertainty, noting the unpredictable nature of Transformer models. ## [MLOps @Chipro](https://discord.com/channels/814557108065534033) Discord Summary
Only 1 channel had activity, so no need to summarize...
**MLOps @Chipro Channel Summaries**
### ▷ Channel: [events](https://discord.com/channels/814557108065534033/869270934773727272) (2 messages):
@jovana0450 announced an online event on November 16, 2023, regarding real-time data infrastructure in crypto fintech. Speakers like Yaroslav Tkachenko from Goldsky and James Corbett from Superchain Network are part of the lineup. According to jovana0450, this event is geared towards data engineers, data scientists, data analysts, crypto, and blockchain enthusiasts. Interested participants can register for the event here. @jonnyhsu shared details about two TechBio mixer events. Hosted by Valence Labs, the first event is scheduled for November 22nd at Oxford, co-hosted by Michael Bronstein here. The second event will be held at Cambridge on November 23rd here. The events focus on accelerating drug discovery with AI and feature dinner, drinks, and discussions.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.
@stevekamman sharing a comprehensive Databricks blog post that defines and explores the concept.@zorkian, who shared an informative blog post from Discord detailing the incident.YAIG (a16z Infra) Channel Summaries
@stevekamman shared a link to a Databricks blog post discussing what a Data Intelligence Platform is.@zorkian shared a blog post discussing Discord's recent 1-hour outage. You can read more about it here.