←
Vaishnavi13,471 posts
Vaishnavi avatar

Vaishnavi

@_vmlops

Forward Deployed Engineer | Building AI-powered products & turning ideas into reality

◷ Joined December 2024▧ 2,256 Photos and videos
Vaishnavi avatar
Vaishnavi@_vmlops·2h

CLAUDE CAN NOW FIX YOUR SLOW WIFI just open Claude Code and say: "connect to my router and find everything slowing my wifi down" ▫️ finds your router on its own ▫️ catches apps eating your bandwidth ▫️ checks if your neighbour's wifi is crowding your channel ▫️ shows every device on your wifi and blocks the ones you don't know asks before changing anything, then re-tests no more calling your ISP 💀

0:34
Vaishnavi avatar
Vaishnavi@_vmlops·3h

Found this crazy site: HUMAN.md A manifesto about what makes us human in the AI era

Post attachment
Vaishnavi avatar
Vaishnavi@_vmlops·3h

Smart AI engineer: 2023: learned embeddings 2024: learned how RAG actually retrieves 2025: learned evals 2026: gets paid to fix what the average one shipped

Vaishnavi avatar
Vaishnavi@_vmlops·3h

Average AI engineer: 2023: copied prompts from Twitter 2024: copied RAG tutorials from YouTube 2025: copied agent repos from GitHub 2026: copying whatever Claude writes

Vaishnavi avatar
Vaishnavi@_vmlops·6h

GOOGLE DEEPMIND PUBLISHED A PAPER ON HOW AI COULD REWRITE THE FUTURE OF COMPANIES The wildest prediction? Companies could first get smaller as AI makes workers more productive, then suddenly turn into massive AI-run organizations when agents learn to coordinate. It's not just about replacing jobs. It's about reinventing how companies work. 👀

Post attachment
Vaishnavi avatar
Vaishnavi@_vmlops·7h

Why do non-technical HR people get to decide the careers and futures of technical professionals when they don't even understand the work they do.....?

Vaishnavi avatar
Vaishnavi@_vmlops·7h

ANTHROPIC SHOWED HOW TO RUN THE SAME AGENT 13X CHEAPER WITHOUT LOSING A SINGLE TEST it's one notebook in the claude cookbook. 100 recipes, all free what actually matters: → cost: opus high to sonnet medium + cache breakpoints = 13x cheaper, still 10/10 on evals → a datetime.now() in the system prompt kills the cache on every call. moving it to the user turn made the run 44% cheaper → latency: chat agent down to 2.3s p50 by cutting output, not by switching models → advisor: a cheap model calls a stronger one mid-turn only when it's stuck → orchestrator: the frontier model plans, cheap workers search in parallel → dreaming: finished sessions become memory that future runs mount → budgets: hard spend caps per session most teams pick a cheaper model first. the notebook puts that lever last

Post attachment
Vaishnavi avatar
Vaishnavi@_vmlops·8h

Claude is great, but for that human taste, I’ll always prefer ChatGPT

Vaishnavi avatar
Vaishnavi@_vmlops·10h

STANFORD BUILT LLM MEMORY THAT FITS IN 16 TOKENS 👀 Memorilla reads your whole document collection and gives the LLM just 16 memory tokens per question → more accurate than feeding every document on FactKG → 108× fewer tokens Bookmark this 🔖

Post attachment
Vaishnavi avatar
Vaishnavi@_vmlops·13h

I'm at the point where I just want to earn crazy money but don't know how 😭