Nirant (@nirantk) 's Twitter Profile
Nirant

@nirantk

AI engineer that takes companies from RAGs to riches

ID: 1028185370

linkhttps://nirantk.com/ calendar_today22-12-2012 09:58:16

9,9K Tweet

14,14K Takipçi

1,1K Takip Edilen

Nirant (@nirantk) 's Twitter Profile Photo

TIL Facebook promotes it's LLMs on Reddit! This is clearly a LLM generated or template answer. I am not sure why it is signed with human initials. And why are people recommending finetuning models for casual use? This is AI the equivalent of "Want to buy a car? Start by

TIL Facebook promotes it's LLMs on Reddit!

This is clearly a LLM generated or template answer.  I am not sure why it is signed with human initials.

And why are people recommending finetuning models for casual use?

This is AI the equivalent of "Want to buy a car? Start by
Nirant (@nirantk) 's Twitter Profile Photo

AI Tools that I use and how much: Claude Code with API ~2 hrs/day, $400/mo o3 with ChatGPT ~3 hrs/week, $20/mo Littlebird ~1-2 hrs/week Reclaim.ai from Dropbox — critical software for keeping me functional I don't think I can get anything done without this What should I try next?

Nirant (@nirantk) 's Twitter Profile Photo

For the amount of query analysis and exploration work I'd do over a week in 2024 using Langgraph + Cursor -- I am able to do in a long work day using Claude Code + disposable Python scripts

Nirant (@nirantk) 's Twitter Profile Photo

I wake up everyday and run into yet another idea possible now because we can separate storage and compute (ideas like Iceberg, Nixiesearch/Turbopuffer, Neon Database) and realize someone has already done it really well Repeat tomorrow

Nirant (@nirantk) 's Twitter Profile Photo

another day, another company with viral tweets about their amazing AI product and their sign up with Github is broken this influencer marketing thing really makes twitter less trustworthy

Nirant (@nirantk) 's Twitter Profile Photo

Claude Code has officially replaced DSPy for prompt optimizations for me Best part? I can teach Claude Code to use DSPy if I want to — but I've never needed to

Nirant (@nirantk) 's Twitter Profile Photo

Pratilipi continues to be one of the most inspiring, enviable product and company persistence. Every smart person I know in consumer envies their product focus. They don't hire a ton of well paid engineers, doesn't look like they do a ton of ad spend and grow a lot of people

Nirant (@nirantk) 's Twitter Profile Photo

The gap between an engineer with 9 years of professional work experience and 12 years of experience is dramatically smaller than b/w 0 and 3 yoe Indicating the marginal utility of experience in dev careers is very front loaded

Nirant (@nirantk) 's Twitter Profile Photo

most llm apps should not create their own memory! when they do — they should just use the context window and when they need something more — they should just use NER + summary as a strong baseline and finally, try something like zep or mem0 before writing their own mem layer

Nirant (@nirantk) 's Twitter Profile Photo

More and more companies are now openly admitting that most agents are just RPA2.0 There is no need for LLM to figure out a new workflow at every request from scratch. Just use what worked in the past!

More and more companies are now openly admitting that most agents are just RPA2.0

There is no need for LLM to figure out a new workflow at every request from scratch. Just use what worked in the past!
Nirant (@nirantk) 's Twitter Profile Photo

point in time — the local (oss?) llm case is weak: battery, intelligence, speed — everything sucks on local other than some version of corporate puritanism disguised as "compliance" or "security" — the only reason to use local llm is that you are a closet masochist

Nirant (@nirantk) 's Twitter Profile Photo

mark sheets don't capture full understanding or learning of material but definitely point out gaps evals done well should do the same