Anjaneya Tripathi

learning one token at a time

[Anjaneya Tripathi @spamjaneya]

social engineering at it’s best.

[rosie @therosieum]

https://t.co/F2G1usjT2E

[Anjaneya Tripathi @spamjaneya]

@abeshek452000 @pjparties vehicle numbers match the app unlike in india

[Anjaneya Tripathi @spamjaneya]

@eshamanideep reminds me of a bold, but accurate statement made by @Virender_VA a few years back — conversation is the next ui.

[Anjaneya Tripathi @spamjaneya]

wish i could function under pressure like argentina… 🥵

[Anjaneya Tripathi @spamjaneya]

@janvikalra_ been using the 5.6 beta for over a week and it’s a beast!

[Anjaneya Tripathi @spamjaneya]

Just from experience:
An apple a day keeps the doctor away,
but a mango a day brings the doctor your way. 💀

[Anjaneya Tripathi @spamjaneya]

cheaper models + canonical logs >>> frontier models + raw log dump

[Anjaneya Tripathi @spamjaneya]

Great read.

[Rachitt @rachittshah]

It's been over a year since we have seen agents, harnesses and AI go fully mainstream.

I used to run an AI consulting firm called Transfrm Labs, so this is mainly my 2 cents on things I found interesting.

I spent the last few months building through it, mostly in the open: cheap inference, local models, verifiable output. A dozen repos, most of them papers I didn't quite believe and rebuilt to check.

 TLDR of what I learned:

 - The model is a commodity now. Same one your competitor rents, same price, on a schedule neither of you controls. The system around it is what you own.
 - Local is real. KV-cache compression buys 3-5x more context on a laptop. A 4B model you control beats a frontier API for most production work.
 - Long context is a programming problem. Stop stuffing the window. Recursive setups go from 0% to 91% on million-token tasks.
 - Stop hand-writing prompts. With a good eval, the system optimizes them itself.
 - Never trust one model call. Cross-check across model families. Verification is the product.
 - Agents are still amnesiacs. Memory and a real harness matter more than the model.

Link in the comments. Views personal.

[Anjaneya Tripathi @spamjaneya]

@abeshek452000 2 AM debugging helped in the end. 😜