Andrej Karpathy shipped 'Wiki'. We made the same bet in February.

Andrej Karpathy shipped “Wiki”.

An AI-native knowledge base that continuously updates from your reading stream. Built with Pinecone. The architectural bet: feed the system everything you read, and it becomes smarter than you at remembering patterns.

We made the same bet in February. Here’s what 90 days of running it nightly produced.

The system:

AIRA runs at 2am IST every day. Ingests LinkedIn feeds, YouTube transcripts, WhatsApp research, web signals. Routes everything to a structured knowledge graph. By 8am there’s a briefing that used to take analyst teams a week.

As of now: 107,633+ entities. 4,980+ investor files. 297 research pattern insights. 2-person operation.

What Karpathy’s system will hit — and what we had to solve first:

Jean-Michel Lemieux (former Shopify CTO) published something that names the real bottleneck precisely. It’s not model capability. It’s agent context initialization — what does the agent know on Day 1?

The agents that perform well know where they are, what they’re optimising for, and what’s already been learned. The knowledge graph is the answer to that. 107,633 entities isn’t just data. It’s 90 days of accumulated context that every agent wakes up knowing.

Without that, you have a powerful tool solving a different problem every morning.

The McKinsey number that should focus every AI team:

94% of enterprises are running AI pilots. 7% have scaled to production. (McKinsey Global Survey, May 2026)

That gap is almost entirely a context problem, not a model problem. Pilots work in sandboxes. They fail when they hit real institutional memory — the kind that lives in analyst notes, deal memos, historical patterns, and three years of market observation.

What Karpathy shipped is what Manthan Intelligence discovered by building it: you can’t bolt AI onto knowledge work. You have to rebuild the knowledge layer underneath it first.

Three months in, the compounding is real. Day 90 is qualitatively different from Day 1. The pipeline is more important than the model. Continuous beats comprehensive — we don’t try to analyse everything, we analyse what’s relevant nightly and let the graph grow.

If you’re building AI infrastructure for financial services, legal, research, or any knowledge-intensive workflow — Karpathy’s architectural validation of this pattern should move it to the top of your roadmap.

The question isn’t whether your team can use AI.

It’s whether your AI knows what your team knows.

Full write-up of how we built the architecture at getmanthan.com.

---

Read more at getmanthan.com