Welcome to the November edition of the Token! Halloween 🎃 has come and gone and the good news is that AI has not decided to take over yet!
October was a bit more quiet on AI news as the community was digesting OpenAI’s GPT-4o1. Google keeps pushing on AI overviews, adding ads to the mix. At the same time, OpenAI released a real time API and Meta introduced their first multimodal Llama models.
In this issue, we also cover a new case study with a large NGO client, helping them to build an AI system that extracts medical information from technical documents. We also share our thoughts on why sometimes AI projects can go wrong ❌
📰 Bites from AI news
🖱️Google introduces ads to AI overviews. Now you can buy the glue to make your pizza 😅 https://techcrunch.com/2024/10/03/google-brings-ads-to-ai-overviews-and-rolls-out-ai-organized-pages/
🧪 Jina embeddings v3 were released with near state of the art performance, using a tiny LLM of 570M params coupled with task specific adapters 🚀 https://arxiv.org/html/2409.10173v3
🖱️OpenAI announced a couple of new features last week, including a real-time API with speech input and output 😮, vision tuning, caching to reduce costs of repeated calls, and a canvas to interact with the output of ChatGPT more effectively https://openai.com/index/introducing-the-realtime-api/
🧪 GliClass is a small zero-shot text classification model inspired by GliNER. There is no accompanying paper or comparison yet, but if it works as well as GliNER, this will be our suggested go-to in the case of few or no examples 🔥 https://github.com/Knowledgator/GLiClass
🧠 Meta releases Llama 3.2, which is the new state of the art for small models of 1B and 3B, outperforming phi and Gemma 🔥 It also brings vision to the Llama series with the 11B and 90B variants. The latter is on par to GPT4o-mini. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/
💼 Extracting medical information using AI
We worked with a large NGO to extract medical characteristics of drugs that are missing from the market using LLMs. Automating this process can enable funders to direct their funding more efficient towards areas of need.
You can read the entire case study here.
🎛️ Customising AI
While we are all used to using AI out of the box with ChatGPT and similar tools, in order for AI to be useful to your organisation, you need to customise it to a specific use case. We wrote a guide on different ways to do that aimed at executives.
Read it here.