Anand Vemuri

I’ve been thinking about this article from Ajey Gore (former Gojek CTO). His argument (I think): An org is why, what, and how. ‘Why’ defines the strategy. ‘What’ defines the execution of the strategy. ‘How’ is simply translating the ‘what’ into the finished artifacts. AI shrinks the How, so the shape of the organization changes. Fatter at the top, with a wide agent layer in the new basement.

I nodded sagely at the time, but something about this feels off and I’ve been coming back to it. I think I can articulate what now.

I think translation is being used interchangeably with transliteration. It’s not the same. We have a zillion translations of the Tao. There isn’t one right translation, and AI doesn’t seem particularly able to do a better (or even equally good) job.

That’s why it feels off when I try to map these abstract categories to actual work. Most tasks and roles are a mix of why, what, and how. A spec would be transliterated to a feature only if it was a complete definition to one. It has to be translated instead, and that involves inferring micro objectives from larger ones (why), filling the blanks for unspecified implementation detail (what), and then implementing it (how).

I buy that the proportion of types of work will change, that the nature of roles will, that the architecture of the org will follow. Still not sure what that new architecture is though.

Fluentfeeds is inspired by the comprehensible input language learning method/cult. The basic idea: reading something that you 95% understand is the best way to make progress. 5% is the right density of new stuff and the 95% gives you (a) enough context to understand natural usage and (b) hooks into pre-existing knowledge to link the new stuff to.

The knowledge graph for learning a language isn’t as clean as for math or physics, but it is possible to construct a legible one at least locally, around your current state. Naturally, I felt the simplest way for me to learn German was to build a comprehensible input app that tracked my current progress and used a graph to explore new grammar and vocabulary. It started as a claude skill for my own use, then turned into a project when I realized the text interface wasn’t enough. V1 is a simple vocabulary focused graded reader app. Unfortunately, you actually need to know German to correctly eval the fancy knowledge graph stuff, so this lost steam a couple of months ago.

At this point I’m fully focused on a startup idea (more on that later), so I don’t see myself going back to side projects like this. However, I cannot bear the shame of another abandoned project going to the bin, and this v1 is kind of useful (I use it!), so I’ve bought a domain, asked Claude to make a landing page, and put it up. I promise no further development and probably won’t respond to criticism, but hopefully it is of some service to fellow travelers trudging along the road to German mastery.

P.S. I’ll probably open source it at some point if anyone wants to take it over.

Since this blog is mostly read by people I know, a quick personal update:

After 12 years of being founder led, the 91s transition took a little longer than expected. I’m not sure why this surprised me, but it did, and my operational involvement continued through the beginning of 2025. I took a break after that, read a lot of books, started playing with LLMs, and did not learn German.

I’ve been doing some consulting after the break as a way to earn my daily bread and see what’s new in industry in the process (turns out, a lot!). I’m now feeling the itch to jump back into something more substantive and have been coming back to some ideas on using AI to improve how we consume information. Maybe too many ideas, hence the blog.