Arun Manivannan

Muser

[Arun Manivannan @arunma]

Full writeup — failed runs, OOM debugging, eval rubric design, cost breakdown:

https://t.co/Cd1nWgwbTj

Adapter + GGUFs on HF: https://t.co/eIgGLMthGP
Code: https://t.co/IHP6ppeobL

[Arun Manivannan @arunma]

I spent $35 turning a Qwen3-4B model into a foul-mouthed friend named Monty.

The hard parts of fine-tuning were nothing I expected.

Five things I learned 🧵

[Arun Manivannan @arunma]

5/ A system prompt on the vanilla model gets you ~70% of the surface behaviour.

But it still answered like a Stack Overflow reply with a swearing dictionary attached. Fine-tuning is what changes the shape of what the model produces.

[Arun Manivannan @arunma]

Spent some weekends copying out nanochat by hand. Disabled copilot and copied it code-block by code-block. Like a maniac.

https://t.co/m8T1Q8Y4PX

[Arun Manivannan @arunma]

Why your ML/Python container dies after 60 seconds on Cloud Run https://t.co/PpfZxmf8KG via @LinkedIn

[conor | writing @conor_writer]

If you spend 40, 50, maybe 60 hours each week working for someone else, you can find 1 hour each day to build a business to make income online.

[Jason Cohen @asmartbear]

As soon as you have everything figured out, things change.

Competitors, alternatives, employees, customers, the economy, epidemics, and other things whose identities are also unpredictable.

So you can never stop talking to customers or rethinking strategy.

[Arun Manivannan @arunma]

Ordered numbers are so effective for DB indexes. Not sure why time-ordered, high-performance 64-bit Snowflake IDs aren't widely used for non-public identifiers and instead the preference is for 128-bit UUIDs. Convenience?

[Daily Stoic @dailystoic]

Signed and numbered first editions of "Wisdom Takes Work" by author @ryanholiday are available only through the preorder. Order yours before time runs out at https://t.co/TJHOoOp8EG

[Navalism @NavalismHQ]

In most pitch decks, AI stands for “Anonymous Indians”.

@naval

[Avi Chawla @_avichawla]

Build human-like memory for your Agents (open-source)!

Every agentic and RAG system struggles with real-time knowledge updates and fast data retrieval.

Zep solves these issues with its continuously evolving and temporally-aware Knowledge Graph.

Like humans, Zep organizes an Agent's memories into episodes, extracts entities and their relationships from these episodes, and stores them in a knowledge graph:

(refer to the image below as you read)

1) Episode Subgraph: Captures raw data with timestamps, retaining every detail for easy historical lookup.

2) Semantic Entity Subgraph: Extracts entities (e.g., “Alice,” “Google”) and facts (“Alice works at Google”). Everything is versioned, so outdated info gets replaced.

3) Community Subgraph: Groups related entities into clusters, with summaries for faster retrieval.

Zep delivers up to 18.5% higher accuracy with 90% lower latency when compared to tools like MemGPT.

It's fully open-source!

[Navalism @NavalismHQ]

"You get paid for being right first, and to be first, you can’t wait for consensus."

@naval

[Jason Cohen @asmartbear]

It’s true that you can outwork someone else who has more talent, and still win.

But there are competitors with both talent and hard work.

To beat 𝘵𝘩𝘦𝘮, you have to be uniquely yourself, so that it’s no longer about “them” at all.

[Arun Manivannan @arunma]

The browser is the IDE of an average person and every major AI player is racing to become the next Cursor.

[Jason Cohen @asmartbear]

Koan #89

You're never "ready," so start today.
You’re never “finished,” so ship today.
You’re never “perfect,” so enjoy today.

And the student was enlightened.