Substack Show Notes
Sanjay Rakshit joins the podcast for what feels like his fourth appearance now, and this one was sparked by a single quote. Speaking at DSE, the digital signage industry event run by Invidis, Sanjay told a room that LLMs are the most overrated part of AI. Bhairav wanted to know what he actually meant by that, and the conversation that follows is a proper deep dive into agentic AI, from what separates an agent from a chatbot through to why context engineering, not the model itself, is where most of the real work happens.
What You Will Learn From Listening
The story behind Sanjay’s “LLMs are the most overrated part of AI” quote and the DSE event where he said it
A clear definition of what makes something an agent, versus a chatbot that simply responds
Why an LLM is just one tool inside an agentic system, not the system itself
Why Sanjay estimates 70% of agentic work is context generation and systems engineering, not calling the model
The difference between data quality and data fidelity, and why fidelity matters more
How understanding the underlying maths of AI helps cut through hype and confirmation bias
Where quantum computing might genuinely move the needle for AI going forward
Sanjay’s “power tools” analogy for AI and jobs, and why his belief in it has strengthened
Why he remains sceptical about AGI, and how he separates AGI from robotics
His work with the Bayes Institute in Scotland and what he’s seeing in robotics right now
Memorable Quotes
On agents versus chatbots: a chatbot just talks, an agent thinks, plans, and acts until it actually achieves the outcome you asked for.
On where the real work is: calling an LLM is the smallest part of building an agentic system, most of the effort goes into getting the context right.
On data: high quality data and relevant data aren’t the same thing, and confusing the two is a common mistake.
On AGI: claiming a machine now has general intelligence is like claiming we’ve fully modelled the human brain, and I’d like to meet the person who believes that.
On AI and jobs: Sanjay’s power tools analogy has only gotten stronger since he’s seen the metrics from his own team’s AI-assisted development.
Summary
The episode opens with Sanjay explaining the moment his now-famous quote came from, delivered at a talk at DSE where he was representing Poppulo. From there, Bhairav pushes him to properly define agentic AI for listeners who are newer to the term, and Sanjay lays out the core distinction that runs through the whole conversation: an agent has agency, it thinks, plans, and acts toward an outcome, while a chatbot is fundamentally reactive. The mistake most people make, in his view, is treating a chatbot with a nice interface as if it were an agent.
From there the conversation moves into the mechanics of what actually makes agentic systems work, and this is where Sanjay makes his strongest case. He argues that context engineering, not the language model itself, is the unglamorous discipline doing most of the heavy lifting, something he puts at around 70% of the total effort. He draws a sharp line between data quality and data fidelity too, making the point that having good data isn’t the same as having the right data for your specific use case.
The back half covers ground that will resonate with anyone who’s been in tech long enough to be sceptical of buzzwords. Sanjay talks about why he goes back to first principles and the underlying mathematics whenever something new lands, rather than trusting whoever has the biggest following that week. He and Bhairav also get into AI’s impact on jobs, where Sanjay’s power tools analogy holds up better than ever, and close with his measured scepticism on AGI and a genuinely interesting look at where robotics is heading through his work with the Bayes Institute.
If you’ve sat through enough AI hype to be tired of it, this is the corrective. Sanjay isn’t dismissing AI, he’s making the case for taking it more seriously by understanding what’s actually happening underneath it.
