Mozilla.ai’s Post

Ever wonder what it would take to build a training data set from only legally available materials? There are people working on this! I had the pleasure of facilitating a community of cutting-edge teams working to build training data sets that aren't...horrible. All to support a foundation of work contributing to more open development of emerging technologies. Interested in learning more, check out the blog post below & as always lovely to work with the teams at Mozilla and EleutherAI!

Expert predicts that highly paid coders will be replaced within 10 years by AI after generously supplying code for free to Github used in training machine learning algorithms. I wonder if the terms and conditions of Github have stated this? No doubt you and me can assume everything we posted on social media platforms over the years is also being used to train AI with its mix of knowledge and problematic free speech. I question the lawfulness of this in Europe under Data Protection laws such as the GDPR because posts are mainly opinions and thus personal data but also contain copyright material. Code is protected by laws that depending on the jurisdiction can range from legal issues of copyright in the code, patent rights and contractual licensing terms. Something that might not be considered is that comments in the code by the author of the code and others is likely also to be personal data where it relates to that individual. Given the impact on people's livelihoods this use of code in training AI likely will be tested in court. I see tonnes of infringements of Data Protection and other laws that could return material damages and I am very interested in talking to as many litigation firms as possible whether legal or funders. In my book the Data Protection Implementation Guide, A Legal, Risk and Technology Framework for the GDPR I even detailed some of the more common infringements. Many of these infringements are now being identified in DPAs decisions but this is only a fraction of the infringements. New obligations will also be created under new laws such as the AI Act in Europe. I would be very interested in coders views too. #privacy #gdpr #business #dataprotection #aiact #ai #copyright

every major change in IT is followed by a newcomer taking the top spot, while the old winners are relegated to playing second fiddle, and few notables disappear like digital research. this time around I don't see this scenario will change, and a newcomer will finally emerge from relatively obscure position now. among the top 3, Google is in most vulnerable position right now. just to keep this post short I will look into their crown jewel search. in fact every one of their product space they are facing stiff challenge and failing to meet it. a new Comer perplexity in the search space showed how weak their position is. if I have to bet I will bet on whoever provides the fastest tokens at the cheapest, open source llm, not big but better trained, opensource tools to work with these tools, and interesting user applications, will win over any other combination. perplexity was a wrapper software, and now has its homegrown llm to meet its need. if someone wants try out different open source llm models, and doesn't want to navigate the complex website of huggingface, one try them at labs.perplexity.ai through a chat interface. I started with huggingface.co, a big website for open source llm,m repository. it has steep learning. it's courses are either too basic resemble Hello c, or one needs to know a lot to use this website. they lack learning tools at intermediate levels. one can figure their website out, even use it, but it requires some time commitment. https://lnkd.in/gVz-Nsss https://lnkd.in/gVz-Nsss

🕴️ What is an LLM agent in LlamaIndex? Query engines - built on multiple indexes (indices) and retrievers - takes natural language input and returns “rich” responses. Example pattern (quantitative + qualitative indices): - Natural language query (provided to SQL database)  - SQL query: written automatically from natural language - SQL response: given in natural language - Transformed query given SQL response - Query engine response: Yay or nay about sufficient context Final response: Addresses known and unknown parts of the initial query In code: https://lnkd.in/gCheuRea #agents #llamaindex #llmops

Andrew Katz, our CEO, will speak at FOSDEM 2024 on Sunday 3 February, in the vibrant city of Brussels, Belgium. In a world where technology often outpaces legislation, Andrew’s talk entitled: Using code generated by AI: issues, misconceptions, and solutions will tackle pressing questions, including: ◾What are the copyright issues with ingestion of training data, configuration and transfer of models, and the generation and use of code? ◾Can GPL code inadvertently be incorporated into your non-GPL codebase? ◾Who owns the copyright in generated code? As a legal expert specialising in open source and AI, Andrew will provide insights and consider potential solutions. Read the full blog post here 👇 https://lnkd.in/eM9GvV3W #opensource #technology #AI

My 2¢... generative "AI" is a marketing term that attempts to simplify the commoditization of toolchains that have been built & enhanced over decades. Before DALL·E came around we could muck about with VQCAN+CLIP, where OpenAI's CLIP model guides VQGAN's generative network to produce images based on text prompts. A low-code way to try out these lower-level libraries and mess with different training models is by messing around with Katherine Crowson's Colaboratory where you can easily swap out models, parameters, and even tweak the Python script. If you are interesting in seeing how databases like Flickr and WikiArt contribute to the "Generative AI" output you see today, this is a great way to get into the weeds: https://lnkd.in/d96RYgak Victor Basu's explainer on VQGAN and CLIP is great as well: https://lnkd.in/dYrPcuvN #generativeai #vqgan #clip #ai

The Lazy Predict library! 🚀 ---------------------------------------- This amazing Python tool simplifies the model selection process for data scientists. With just a few lines of code, Lazy Predict quickly builds and evaluates a wide range of machine-learning models, helping you identify the most promising ones for your dataset. It's a game-changer for anyone looking to streamline their model selection and save valuable time. Give it a try and supercharge your data science projects. Docs: https://lnkd.in/gQdHCDeM Colab: https://lnkd.in/g5hY9FSj Shankar P. #DataScience #MachineLearning #LazyPredict #Python #AI #nlp #naturallanguageprocessing #genai #llm

Create synthetic instruction datasets using open source LLM's and bonito🐟! With Bonito, you can generate synthetic datasets with a wide variety of supported tasks. The Bonito model introduces a novel approach for conditional task generation, transforming unannotated text into task-specific training datasets to facilitate zero-shot adaptation of large language models on specialized data. This methodology not only improves the adaptability of LLMs to new domains but also showcases the effectiveness of synthetic instruction tuning datasets in achieving substantial performance gains. AutoBonito🐟: https://lnkd.in/efNm3--J Original Repo: https://lnkd.in/ejVWbSdK Paper: https://lnkd.in/e34JNzrz

In 2022 the OSI started an in-depth global initiative to engage key players, including corporations, academia, the legal community and organizations and nonprofits representing wider civil society, in a collaborative effort to draft a definition of Open Source AI that ensures that society at large can retain agency and control over the technology.

User Interview Summarizer v0.2 A while ago I shared a mini-app for user interview summarization (https://lnkd.in/e_wi-uf4), and now it's time to do an update. Google Colab link: https://lnkd.in/eBiix5td Features added: - Whisper model selection (medium, large-v2, ..) - Language selection (works quite well with Ukrainian) - Structure templates (meeting notes, VPC, generic summary, add your own) - Sources: file, Google Drive, YouTube - Formats: .mp3, .mp4, .txt, youtube url Now it supports records of any length, depending on the size is choosing the summarization method: gpt-4, gpt-3.5-turbo-16k, langchain map_reduce. Bottomline is the same: Basically, this is what you (most likely) will find under the hood of every AI-powered user research assistant app, but customizable as hell 😈 It appeared to be a quite viable and almost free – as long as you are in your trial Colab and OpenAI budget, or run it on your own CUDA hardware. Youtube walkthrough: https://lnkd.in/e3xfneip #python #ai #promtengineering #userresearch #openai #whisper #summarization #langchain