Revenue Engine Masters

The episode delivers a handful of genuine insights - eval strategy depth, model commoditisation, data as the last differentiator, and the IP-leakage risk of cloud-hosted models - but these are interspersed with lengthy tangents about Italy, the Renaissance, and the host's own company experiences, significantly diluting the signal-to-noise ratio.

“Demoing an agent is extremely easy. Shipping it to production in a reliable, trusted way, evaluated the way you. In a way that actually matters is actually a lot harder.”

“there's a true like privacy beyond just personal privacy...it's so easy for AI systems to learn from the prompts, from the data you send them, from the patterns of workflow...you're going to transfer your proprietary differentiation to a tech company very quickly”

The IP-leakage framing - that cloud-hosted frontier models are vacuuming up your proprietary workflows and eroding your competitive moat - is a usefully sharp take, and the Tapestry consortium-training concept is genuinely novel; however, most other arguments (model commoditisation, open source wins long term, Linux/Kubernetes analogies) are well-worn in AI discourse.

“the key reason Anthropic has been so successful is because they've built this really good product experience right around the model which has a lot of sophistication”

“we are going to build a globally sourced uh, foundation model...do it in a, in a privacy and ownership preserving way and train a model that could be better than anyone else could train”

Annunziata is a genuine practitioner who has built new business lines inside IBM (quantum computing, AI for science, the AI Alliance), holds patents, and was in an early Y Combinator cohort - he has done real things at scale, not just advised on them; however, he is not an operator who has built and scaled an independent commercial AI product, limiting the practitioner sharpness.

“I had several opportunities inside IBM to build a new quantum computing business, to build a new AI for science platform which we launched during the pandemic”

“We have 205 uh, partner companies in 29 countries doing that.”

There are a handful of concrete anchors - 205 AI Alliance partners across 29 countries, named models (Granite, Gemma, Mistral), named projects (Tapestry, Kubernetes, Linux) - but the episode is largely absent of customer case studies, revenue figures, measured outcomes, or timelines that would let a listener benchmark against their own situation.

“We have 205 uh, partner companies in 29 countries doing that.”

“There's a new one I'll tell you more about in just a moment called Tapestry, which is a very ambitious project led by Yann McCun”

The host asks a few substantive questions (eval strategy, data preparedness, on-prem) but routinely hijacks the conversation with multi-paragraph personal anecdotes about Italy, Amazon, and ScaleStack, rarely pushes back on under-supported claims, and lets important threads drop rather than drilling for specifics.

“Anthony, this was super interesting. I recognize that I may not have followed strictly the guy that I said to, but we went into many interesting areas.”

“So I was trying to ask myself, like, what is that? Like, you know, is different. In other words, like, is it. You become successful at what you do now because you're more entrepreneurial”