Harsh (@devloperhs) 's Twitter Profile
Harsh

@devloperhs

Learning AI Agents | AI Engineer + Technical Writer @composiohq | Sharing what I learn

ID: 1763427814735265792

linkhttps://linktr.ee/devloper_hs calendar_today01-03-2024 04:55:39

921 Tweet

80 Takipçi

36 Takip Edilen

Harsh (@devloperhs) 's Twitter Profile Photo

Languages are just a medium to express intent , real leverage lies in understanding the nuances of language & use it to build powerful systems that rewrite and evolve itself over time. (self healing & improvement loops) On light note , go might be a good alternative.

alphaXiv (@askalphaxiv) 's Twitter Profile Photo

If you're using Claude Code for research: stop making it read directly from PDFs We've introduced a SKILL.md that fetches structured, AI-friendly paper overviews from alphaXiv 👀

Harsh (@devloperhs) 's Twitter Profile Photo

A few days back , I was inquiring about an open claw docker setup. I believed that the gemini didn't captures the memory until unless explicitly told to. Today I was proved wrong , when I prompted it to generate this image. Guess it's time to switch to oss models : gemini 😂

A few days back , I was inquiring about an open claw docker setup.

I believed that the gemini didn't captures the memory until unless explicitly told to. 

Today I was proved wrong , when I prompted it to generate this image. 

Guess it's time to switch to oss models : gemini 😂
Harsh (@devloperhs) 's Twitter Profile Photo

Morning, Slack message: "We need to ship updates in 3 days, get this done" Opened GitHub, 300+ issues listed 😵 Me? > Opened Claude Code. > Created a todo.md. > Spun up subagents,assigned each an issue. Evening, raised a PR with 100 commits. Rest tommorow

Morning, Slack message: 

"We need to ship updates in 3 days, get this done" 

Opened GitHub, 300+ issues listed 😵

Me?

> Opened Claude Code.
> Created a todo.md.
> Spun up subagents,assigned each an issue.

Evening, raised a PR with 100 commits.

Rest tommorow
Harsh (@devloperhs) 's Twitter Profile Photo

Top 1% people thriving in the blue region are those doing 10 tasks at once (synthesis). Single ones are left behind / switch to physical jobs. Bottom line: companies now prioritize outcomes over inputs - that's where we are heading! Link to graph sources in comments 👇

Top 1% people thriving in the blue region are those doing 10 tasks at once (synthesis).

Single ones are left behind / switch to physical jobs.

Bottom line: companies now prioritize outcomes over inputs - that's where we are heading!

Link to graph sources in comments 👇
Dhruv Jaradi (@jdhruv14) 's Twitter Profile Photo

You can now run Claude Code for FREE. → Use any LLM via API (Gemini, OpenAI, etc.) → Locally and fully private → Full agent + tool workflows Here's the FULL step-by-step tutorial to get it all set up in less than 10 minutes:

You can now run Claude Code for FREE.  

→ Use any LLM via API (Gemini, OpenAI, etc.)
→ Locally and fully private 
→ Full agent + tool workflows

Here's the FULL step-by-step tutorial to get it all set up in less than 10 minutes:
Harsh (@devloperhs) 's Twitter Profile Photo

We spent years abstracting math away from developers. Now the only people who can fix broken agents are the ones who never stopped doing maths.

Harsh (@devloperhs) 's Twitter Profile Photo

Stuck with neural text degeneration / repetition collapse! This happens when a llm keep predicting the same high probability character (whitespace's or block) as most likely next token. Basically LLM get's stuck due to an endless self-reinforcing pattern.

Stuck with neural text degeneration / repetition collapse!

This happens when a llm keep predicting the same high probability character (whitespace's or block) as most likely next token.  

Basically LLM get's stuck due to an endless self-reinforcing pattern.
Harsh (@devloperhs) 's Twitter Profile Photo

Most people building ‘AI’ today are just prompt engineers around black boxes ! They’ve never trained a model from scratch, yet they’re the ones who get to decide how intelligence should behave.

Sayujya Gupta (@guptasayujya) 's Twitter Profile Photo

Every day, 1000s of Job Posts are posted on X. Yet you can't find them because X search SUCK! So, I built an app that finds all those leads for you. - Scrapes all job posts - Qualifies with AI - Export in Excel I'll be launching it for FREE on my 22nd bday (9th April) This

Harsh (@devloperhs) 's Twitter Profile Photo

Built Contextio - turns every group chat into a multi-tenant memory system that understands your conversations before you even ask. Almost pulled an all-nighter… HydraDB , Harnoor Singh , Nishkarsh demo (2:13) + how it works (0:43) 👇

Harsh (@devloperhs) 's Twitter Profile Photo

Most people make static portfolios. I made them editable. You can click on each block and edit inline in editor mode ( left tab ) Part 1 of rewamping kcron.com done. Few fixes and then moving to add a new template. See ya 🙌

Harsh (@devloperhs) 's Twitter Profile Photo

After a very tiring day , I learned one thing! In the world of AI , anybody can write code & make an app but not everybody can upscale it. Writing code is the easy part , the real moat is how you engineer features. One wrong approach and you need redo , so use git Night 🙌

After a very tiring day , I learned one thing!

In the world of AI , anybody can write code & make an app but not everybody can upscale it. 

Writing code is the easy part , the real moat is how you engineer features.

One wrong approach and you need redo , so use git

Night 🙌
Sayujya Gupta (@guptasayujya) 's Twitter Profile Photo

Everyday 1000s of job posts are posted on X, Yet you can't find any, cause X search sucks. So, I built an app that finds out all the recent jobs for you. I could easily charge $40/month for this But today I'm giving it away for FREE (cause it's my birthday) How to get it? -

HackerRank (@hackerrank) 's Twitter Profile Photo

LeetCode is dead. Developers don't write code line-by-line anymore. They orchestrate AI agents working in parallel, review AI-generated code, and make architectural decisions. That's the job now. But most interview processes haven't caught up. They still test algorithm

Harsh (@devloperhs) 's Twitter Profile Photo

Automated my financial analysis. Hermes now sends me financial analysis reports based on the market sentiments, fear index and war repercussions every morning & on a mazor movement. Sent one prompt & it got it done Teaching how to build this for yourself in the next blog :)

Automated my financial analysis.

Hermes now sends me financial analysis reports based on the market sentiments, fear index and war repercussions every morning & on a mazor movement.

Sent one prompt & it got it done

Teaching how to build this for yourself in the next blog :)
Harsh (@devloperhs) 's Twitter Profile Photo

I met a guy yesterday, and asked what he was doing ? His answer shocked me 🤯 I was getting bored so I am creating an app to put my digital version online using avatar and voice. I asked him what tech stack he is using for the same? - I don't know AI making us dumb 😖

I met a guy yesterday, and asked what he was doing ? 

His answer shocked me 🤯

I was getting bored so I am creating an app to put my digital version online using avatar and voice.

I asked him what tech stack he is using for the same? - I don't know

AI making us dumb 😖