Last week in AI | 23 March

Welcome to Last Week in AI. Danilo was away for last week’s livestream so I was very lucky to be joined by Ricardo Teixeira, a Principal Data Architect at Thoughtworks. 

On this week’s show we spent a lot of time discussing NVIDIA's GTC conference, which we think showcased a massive leap in both hardware and software integration. The Vera Rubin platform news — which looks like an expansion beyond a chip to something more akin to an “AI factory in a box — was particularly notable, suggesting that NVIDIA is continuing to redefine its products and how it distributes them to its ever-growing customer base.

Also at GTC was the launch of NemoClaw, which could represent the next phase in the rapidly evolving OpenClaw story. NemoClaw essentially runs on top of OpenClaw, and is designed to give users greater control over what OpenClaw can and cannot do by providing the necessary guardrails, sandboxes and security required to make agentic capabilities safe.

Finally, we discussed research that looks to have found a novel way to expand LLMs’ ability to perform complex multi-step tasks. While tool invocation and agent scheduling are two techniques that can help here, they both have some limitations — namely, being very intensive in terms of tokens and computation. However, a team experimented by essentially running a computer inside a transformer model.