AI-podden

In this very special episode, our host Ather Gattami welcomes Robby Stein, VP of Product at Google Search, to reveal some truly exciting Google news: the global rollout of AI Mode. This new feature introduces conversational, multimodal search in nearly 50 countries and 36 languages, allowing users to ask complex questions through text, voice, or images - and receive richer, link-backed responses powered by Google's real-time information systems. Robby explains how AI mode builds on tools like Lens and AI Overviews, while carefully balancing innovation with reliability, addressing challenges such as compute costs and hallucinations through Google's long-established quality systems. From planning trips with multiple constraints to troubleshooting appliances with a photo, he describes these breakthrough experiences as "AI magic moments."

In this week's update, host Ather Gattami is joined by regular guest Anders Arpteg, Head of AI and Data at GlobalConnect, to unpack the summer's biggest AI developments. They review the release of GPT-5, which introduced cheaper APIs, longer context windows, and fewer hallucinations but ultimately fell short of lofty expectations, particularly when compared with Claude and Gemini. The discussion explores possible reasons, including talent losses, outdated training data, and speculation of withheld advances, while also spotlighting AI's Olympiad successes, DeepMind's pursuit of hard mathematical problems, and video models like Veo 3 that demonstrate an emerging grasp of physical laws. They conclude that although omni-models may appear, specialized systems are more likely to dominate, creating significant opportunities for startups and enterprises despite the concentration of power in a few frontier labs.

In this episode, Ather Gattami and one of our favourite guests; Anders Arpteg, Head of AI and Data at GlobalConnect, explore Meta's shifting AI strategy, including its investment in Scale AI amid doubts about LeCun's JEPA model and LLaMA 4 performance. They compare Tesla's vision-based Robotaxi rollout to Waymo's sensor-heavy approach, highlighting Tesla's edge in scalability and real-world data. They also dive into Apple's "Illusion of Thinking" paper, questioning whether today's LLMs genuinely reason or simply excel at recall, and discuss the need for memory, planning, and architectural breakthroughs to push AI beyond current limits.

In this episode, Stephen Jenkins, VP of Technology Strategy at Magna Electronics, shares practical insights into how AI is reshaping the automotive industry - from cutting development timelines with synthetic data to building smarter, more efficient autonomous parking systems. He explores the challenges of trust, safety, and explainability in AI, the current state of self-driving technology, and why full autonomy may be closer than we think.

In this week's episode, Ather chats with Magnus Hambleton, investor at byFounders and former data lead at Natural Cycles. Magnus has successfully applied transformers to improve fertility predictions in a medical setting. As an investor, he focuses on agentic AI and backs startups slightly ahead of what's currently possible, betting on near-future model capabilities. He remains cautious about LLMs' reliability and generalization, believing new architectures may be needed for AGI. While big tech is dominant, he sees room for startups in complex, workflow-driven applications.

In this week's episode we host John Lamb, Solutions Lead AI & Analytics at TietoEvry. John emphasises that while tools like LLMs are useful for automating repetitive tasks such as code generation or data prep, they fall short when it comes to reasoning and interpreting structured data. He explains that data cleaning still dominates much of a data scientist's time and that current AI tools, though helpful, cannot yet be trusted for end-to-end analysis. John also discusses why he believes AGI is still years away, requiring a fundamental shift in technology, and shares practical advice for using AI responsibly, underscoring the importance of human judgment in any AI-driven workflow.