In this March news episode, our host Ather Gattami sits with Anders Arpteg to unpack a fast-moving period in AI, from NVIDIA GTC and the rise of agentic, always-on systems to the growing demand for inference, physical AI and autonomous driving. They argue that while the industry is moving quickly and companies are racing to build more proactive, capable systems, current models are still far weaker at reasoning and action-taking than the hype often suggests, something they link to the poor results on Arc AGI 3. The conversation also touches on shifting power dynamics between companies like Anthropic, OpenAI and Meta, before ending on a more forward-looking note with Semantic Tube Prediction, a research idea that could offer a more efficient path to reasoning in latent space rather than through token-by-token prediction.