I often get asked which underlying foundation model one should choose for their use cases, and whether an organisation or a country should even train one given the immense compute and talent challenges. It’s a complex space to navigate. 🌐

This question becomes even more perplexing with some recent intriguing discoveries, such as:
1. Picking and merging existing models without significant compute can outperform more resource-intensive approaches. https://lnkd.in/gtCYmAZE
2. Scaling inference-time compute can yield better results than scaling model size. https://lnkd.in/gQJizp8q
3. Simple scaling strategies like “more agents/XYZ-is-all-you-need” can outperform convoluted scaffolding and prompt designs, such as https://lnkd.in/gAqbdbj3

We just released a paper to help you navigate the architecture options for building foundation model-based agents and applications. It covers things from design-time model building, selection, and merging, to runtime model/agent coordination, external capability access, and the usual reflection, planning, and learning considerations. 🛠️

One tantalising takeaway is that an organisation or country might not always need to spend huge resources to “train” a model or settle for adapting at the edge for specific use. There are many options to 𝐬𝐢𝐠𝐧𝐢𝐟𝐢𝐜𝐚𝐧𝐭𝐥𝐲 𝐠𝐫𝐨𝐰 𝐭𝐡𝐞 𝐠𝐞𝐧𝐞𝐫𝐚𝐥 𝐜𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬 𝐨𝐟 𝐚𝐧 𝐀𝐈 𝐬𝐲𝐬𝐭𝐞𝐦—𝐰𝐡𝐞𝐭𝐡𝐞𝐫 𝐬𝐨𝐯𝐞𝐫𝐞𝐢𝐠𝐧 𝐨𝐫 𝐨𝐭𝐡𝐞𝐫𝐰𝐢𝐬𝐞. What we need is more careful research and analysis. 🚀

Paper link https://lnkd.in/gUj_wbmg