Podcast Ruined by a Software Engineer

When the conversation reaches the technical substance (~halfway through), there are legitimately useful explanations of chunking mechanics, the spatial awareness problem in RAG, and traversal-path chunk sets. However, the first half of a 119-minute episode is almost entirely childhood reminiscence, travel anecdotes, and programming language nostalgia that delivers zero actionable insight for a B2B operator.

“90% of systems out there, what they do is okay, character window, full cut. So that means like text messages in the 90s”

“PDF is a format for printers. PDF is humanity's nemesis... everyone thinks PDF is machine readable. No, it's printer readable. And we literally printed out in our system to then take a photo and then reverse engineer what was on it”

The tree-based chunking approach and traversal-path chunk sets as retrieval units are a genuinely non-obvious framing of a real infrastructure problem, and the vision of replacing PDF with a Puma file format is a fresh, first-principles idea. The personal sections, however, trade in entirely conventional entrepreneurial wisdom (first principles, Musk/Jobs biographies, ego is expensive).

“We let it regurgitate the same text to us verbatim, like, literally word by word. But let's say make, make, make, take. Take pauses when you speak, uh, AKA formatted like code”

“RAG is so dead. RAG is so last Friday. Now it's a context engine. That's cool. I'm cool with that. You know, I can now say rag is dead. The long live Rag”

Kihm is a genuine multi-exit practitioner - legal tech at 18, PhD in applied econometrics, fintech sold to a unicorn in 2019 - with real technical depth and a patent on the core chunking method. The deduction is that Puma AI appears to be very early stage (1,000-page free tier just launching), so the scale of demonstrated success is moderate rather than exceptional.

“We actually processed half a million Volkswagen Diesel scandal claims and it's still MySQL PHP”

“in 2019, so six years after founding, we sold the company to them and integrated as their pension division”

There are useful concrete details - 57 file formats, 8,192 token embedding limit, 500k Volkswagen claims processed, a 2019 exit, token costs exceeding $1 per question during the birthday demo - but the headline performance claim of "50 to 90% less material" is a wide, unsourced range, and no customer names, benchmark datasets, or accuracy comparisons are offered for Puma AI itself.

“Every question burned more than a dollar. I mean, yeah, it was a bit more expensive than now”

“we take in 57 file formats so we don't need pure text anymore”

The host consistently validates rather than challenges, regularly derails into his own tangents (DMV systems, childhood technology, his own coding history) that consume substantial airtime, and never pushes on key claims such as the 50 - 90% efficiency figure, competitive differentiation from unstructured.io, or actual customer outcomes. Questions are broad and invitation-style rather than probing.

“I absolutely love it. Like a lot of people. Okay, this is the last comment before we move on to the actual tech part. Because, like, there's a text show”

“That is a quote that I, I cannot reuse more”