Aravind

[Aravind @arangana]

Platform ROAS said 4x. The bank account said flat.

Half the "return" was retargeting people already about to buy. The rest was revenue on a SKU too thin-margin to scale profitably.

ROAS measures the transaction, not whether it was incremental or profitable.

[Aravind @arangana]

The supply side spent years building verification for the data feeding automated decisions. The buy side never did.

Now agents make send and suppress calls off consent records and lead scores nobody's touched in years. Nothing flags it, the data was technically validated once.

[Aravind @arangana]

Marketing AI agents aren't one permission level. There's a ladder: Read, Recommend, Create, Change, Transact.

Most teams are fine with an agent flagging a stalling campaign. Very few have decided whether it's allowed to pull the budget itself, or who signs off when it does.

[Aravind @arangana]

@dupenodi @AsideAI 💯💯💯

[Aravind @arangana]

Aside should've just been called Amaze.

[Aravind @arangana]

@dupenodi @hydra_db How did you make the video

[Aravind @arangana]

Source: Colgate-Palmolive CEO Prabha Narasimhan on the investor call: "It's not a flywheel that we understand as a company." https://t.co/pMvidqi4Dd

[Aravind @arangana]

Colgate told investors this week it doesn't understand the D2C flywheel, so it handed Palmolive's ecommerce ads to Bombay Shaving Company to learn how.

Budget was never the constraint. The decision system connecting spend to outcomes was.

[Aravind @arangana]

Llama-3.1-8B-Instruct is Meta's mid-size open-weight model. Big enough to be useful for chat/coding/agents, small enough to be a realistic stress test for "can consumer hardware handle this." It's become an unofficial benchmark in open source LLM tooling.

[Aravind @arangana]

Soup seems to be the latest idea for local LLM fine-tuning on your laptop. It fine-tunes an 8B model on a 4GB GPU, something that used to need a rented A100. Playing around with it, here's the trick behind it 🧵

[Aravind @arangana]

Macs differ: unified memory (8-24GB+) is shared between CPU and GPU. An M-series Mac with 16GB+ can often handle similar-sized models, just usually slower than a dedicated Nvidia GPU for training.