This is the September issue, hopefully you had a nice break over the summer 😎
August was a bit quiet with the most notable development being Google releasing new versions of its Gemini models that momentarily overtook GPT4o in terms of performance. Of course, OpenAI rapidly responded with their own update to reclaim top spot 👑
In this issue, we also cover a real case study with a health tech client, on building an AI assistant for pharma reps. We also share our thoughts on GenAI costs and future abilities in a new blog post ✨
📰 Bites from AI news
🖱️ Google released a new version of Gemini 1.5 Pro that momentarily ranked first 🥇 before taking the second place🥈 when OpenAI also released its latest version GPT4o https://venturebeat.com/ai/googles-gemini-1-5-pro-leaps-ahead-in-ai-race-challenging-gpt-4o/
🖱️Further cuts in LLM costs were announced across the board with frontier models now sitting at around $3 per million tokens, and efficient models like GPT4o mini and Gemini Flash costing around $0.1 per million tokens 💸 https://developers.googleblog.com/en/gemini-15-flash-updates-google-ai-studio-gemini-api/
🧪 Sakana introduces a system that can conduct AI research autonomously 😮: from developing new research ideas, testing them, and writing up the work in a research paper 🤖https://venturebeat.com/ai/sakana-ai-scientist-conducts-research-autonomously-challenging-scientific-norms/
🖱️ Google announced its response to GPT4o called Gemini Live, which seems to be taking over from Google assistant. Given the market share of Android and Google Chrome, this has the potential to impact billions of users 🔥 https://www.wired.com/story/what-is-gemini-live/
🧪 Microsoft released phi 3.5 which includes a mini models with 3.8 Billion parameters, a mixture of experts 16 x 3.8 B, as well as a multi-modal version. 🚀 Performance-wise the mini model seems to be similar to the small llama (8B) and Google’s Gemma (9B) while the mixture of models seems on par with Gemini Flash and GPT4o mini https://huggingface.co/collections/microsoft/phi-3-6626e15e9585a200d2d761e3
✍️ AI medical assistant for pharma representatives
We recently worked with a client that develops software for pharma companies in order to help them deal with questions from patients, doctors or sales representatives. We worked with them to develop a RAG based AI assistant on top of verified medical information that can act as a better means to retrieve related material to a question. The material and answers is always double checked by experts to ensure correctness.
You can read the entire case study here.
🤔 How we are thinking about genAI: costs and abilities
AI models are becoming more cost efficient. This is partly because we overtrain them to be inference optimal, essentially lowering the cost at inference time but also because we tune smaller models to do the specialised tasks we want, a good example being Apple Intelligence. Their abilities are also increasing with the new generation of models being able to speak, hear, see and write. Lastly, AI is integrating with more and more tasks that humans do and we believe this trend will continue to enhance human performance without replacing them.
Read our entire piece here.