Congrats @chheplo / it’s awesome to see this effort going into such a critical problem. 30% of incoming questions from small shareholder farmers on ai tools like farmer.chat contain photos. Bravo.

Deepseek shows that the global investment and competition in AI continues to drive down marginal inference costs dramatically. This is particularly important to those of us working in the AI for impact space (e.g. with workers that don’t make $200/ hour and have thinner margins to show productivity gains). For everyone trying to leverage AI to help technicians, community health workers, small farmers, teachers, etc (eg. gooey.ai/impact), it means that at the pilot stage that we should be using the latest, most expensive intelligence models - like GPT4o-audio- with the assumption that per user costs will fall by the time we iterate & discover which AI use-cases are actually valuable and ready for scale.

The Deepseek model is smart and was cheap to make BUT biased in that it follows the Internet restrictions of the CCP (in the same way that every model is biased based on its training dataset). My hope and expectation is the open-source community (read Europe, India, Africa, SEA, South A) takes the engineering lessons of deepseek and applies to them to make better Public AI models that have its intelligence but a broader view of the world, so folks can ask questions like “What happened on the 3rd of June 1989?”

Given #NJDRONES, 3 films seem worth re-watching (and why): The Abyss - undersea aliens that mimic humans, then threaten our sovereignty over earth unless we stop climate change and nuclear escalation Close Encounters of the 3rd kind - most accurate depiction of UFOs says @LueElizondo Annihilation - unknowable mimicry

It’s absolutely reasonable that all GenAI companies (even small ones like ours⁩) should pay .1% of sales and 1% of profits to cover the massive IP theft that made this wave of AI possible. Hence ⁦@GooeyAI⁩ commits to pay into these funds. scanthehorizon.org/p/un-puts…

This fall, 1 billion people with hearing-loss are going to get access to a non-stigmatizing, best in class hearing aids at 15% of current prices (see Monday’s @Apple AirPod news) because of bi-partisan legislation by @SenWarren and Trump. And he didn’t mention it last night. #fool

This is an incredible video on the flagella motors of bacteria. It shows that modern molecular biology is often a series of AI workflows (doing image classification, mapping images to 3D shapes, etc). There’s a large opportunity here for shared AI Recipes in the sciences on @GooeyAI. DM me to collaborate youtu.be/VPSm9gJkPxU?si=VoRN…

This is a good argument for data trusts and the value of high quality, human created data. In short, models collapse when trained on recursive synthetic data nature.com/articles/s41586-0…

The AI sum is greater than the parts. Our belief at @GooeyAI is model innovation will come from big private and open sources (like @AIatMeta, @Google, @MistralAI, @OpenAI) + great specialized ones (eg @chheplo’s agriculture and @GhanaNLP’s Africa models) and all that innovation should be composable & instantly available via shared AI recipes.

We've been working our butts off over the last year to offer a chatbot builder - gooey.ai/copilot - that really is a super-set of all the best LLMs, vectorDBs, speech reco and lipsync models. It's my sincere hope that by aggregating the best of AI into a simple tools, we can aid the productivity of everyone. Here's our latest: blog.gooey.ai/from-bland-to-… The TL;DR: 1. All the best LLMs! LLaMA3, Claude3, Mistral, Gemini Pro 1.5 & Gemma support - Use any of the best LLM (& 3 vision models!) in Copilot including GPT4 Turbo + GPT Vision. 2. WhatsApp Self Serve - Get yourself a spare SIM and connect your chatbot to WA (or Facebook, Slack or your own app). 3. Fast Speech Recognition in 1000+ languages- We now run Meta's MMS-Large speech recognition model "hot" meaning you can transcribe virtually any language in near real-time. 4. OpenAI + Azure Voices - great for realistic and African accents! 5. Transparent Pricing - 3 credit (~$.03) / message + your LLM, TTS and speech recognition token costs. 6. <1s RAG with 1000s of docs, pdfs and video - Add 1000s of PDFs, files or youtube videos to your knowledge base and we'll query them in milliseconds (after the first run). 7. Built-Analytics and Conversation Analysis - Dashboard and analytics for monitoring AI Copilot interactions, retention, 👍🏾 👎🏽 feedback and messages categorization. 8. Bring Your Own Copilot Evaluation - Get yourself a few golden questions and answers and then test how changes to your copilot's LLM, instructions or knowledge docs improve its answer quality. 9. Auto-embed Documents + Compare Vectors Feature - Just add a PDF or doc to copilot, and we'll visually understand its contents and vectorize it. Works for YouTube links too. 10. Built-in Synthetic Data Maker - Every video or PDF really needs a FAQ to help your copilot better answer questions. Our synthetic data maker has you covered.