Anthropic made a nice article about how they have implemented web search in Claude using a multi-agent system: anthropic.com/engineering/bu… I do recommend this article if you are building an agentic application because it gives you some ideas about how your system could be architected. It mentions things like: - Having a central large LLM act as an orchestrator and many smaller LLMs act as workers - Parallelized tasks vs sequential tasks - Memorizing key information - Dealing with contexts - Interacting with MCP servers - Controlling costs - Evaluating accuracy of agentic pipelines Multi-agent systems are clearly still in their infancy, and everyone is learning on the go. It's a very interesting topic that will require strong system design skills. An additional take: RAG pipelines are going to be replaced with multi-agent search because it's more flexible and more accurate. Do you agree with that?

There has been a backlash against Cursor over the last couple of days. It seems that the Cursor support system is 100% based on AI, and it clearly gave very bad answers to users who could not log into Cursor because of a bug, leading to many customers cancelling their subscription. As an AI developer myself, I think I understand the limitations of generative AI pretty well, and I find it shocking that some companies implement support chatbots to give such definitive answers without a human in the loop. This is even more shocking when done by a team of people - the Cursor team - who know AI very well... Unfortunately this is tempting to cut support costs like this, but this is a big mistake in my opinion. AI is not mature enough for that.

Google TPUs have been around for quite some time now, and I've rarely seen any company seriously use them in production... At NLP Cloud we used TPUs at some point behind our training and fine-tuning platform. But they were tricky to set up and not necessarily faster than NVIDIA GPUs. We also worked on a POC for TPU-based inference, but it was a failure because GCP lacked many must-have features on their TPU platform: no fixed IP address, no serious observability tools, slow TPU instance provisioning process, XLA being sometimes hard to debug... Researchers may be interested in TPUs but is it because of TPUs themselves or because of the generous Google TRC program that gives access to a bunch of free TPUs? Also, the fact that Google TPUs cannot be purchased but only rented through the GCP platform might scare many organizations trying to avoid vendor lock-in. Maybe this new generation of TPUs is different and GCP has matured the TPU ecosystem on GCP? If some of you have experience using TPUs in production, I'd love to hear your story 🙂

Wow this is powerful. Grok is able to accurately give me the last closing price of a stock option, and correctly explain the reason for this price. Congrats @grok team, your RAG capabilities are very impressive and useful, Grok will be my exclusive personal assistant from now on 👍🏻