Shiv Deepak

My purpose in life is to inspire people to create technology with joy and wonder, and accelerate innovation!

Principal Eng @youdotcom; Prev Eng #1 @hackerrank

[Shiv Deepak @shivdeepak]

Although they are slower, I have completely switched over to reasoning models!

[Shiv Deepak @shivdeepak]

Claude’s Plan! ❤️❤️❤️

https://t.co/CwYl0QgvfX

[Shiv Deepak @shivdeepak]

Mind blown 🤯 with how Fast Grok + Cursor is!

@SpaceXAI has become a very capable Frontier AI company that is also publicly traded!

Although Anthropic and OpenAI are winning for now. There is some serious competition!

And this is good for us, the humans! 🙂

[Shiv Deepak @shivdeepak]

It feels like such a handicap when a software project is not set up for agentic engineering.

It’s the new tech debt! Pay it up, people!

[Shiv Deepak @shivdeepak]

Tread with Caution! ⚠️😅

[Prompter @PromptLLM]

Insane Fable 5 Prompt

[rvivek @rvivek]

The more capable AI becomes in software engineering, the more we will realize it’s an art more than science.

[Shiv Deepak @shivdeepak]

@ishandeveloper That’s Claude sandboxes if you are using Claude code or Claude desktop heavily. Or it could be just docker containers cache. I have to deal with this every few months.

[Shiv Deepak @shivdeepak]

AI can be a great brainstorming tool. But think of it as an extension to you. You still are the top level Harness, using your brain as the underlying model! Don’t outsource your understanding, and decision making to a sub agent. 😅

[Shiv Deepak @shivdeepak]

I can't believe I'm saying this, but the best way to think and have original thoughts right now is to grab a pen and paper and stay away from your phone and laptop!

[Shiv Deepak @shivdeepak]

These days, I don’t look at a screen for at least the first 30 minutes of the day. This has improved my clarity of thought and general happiness.

[Shiv Deepak @shivdeepak]

Capping on tokens burnt is the bare minimum one could to manage AI spend.

A better metric would be ROI, but it’s quite impractical to instrument metrics around calculating ROI for AI spend at this point.

[Chamath Palihapitiya @chamath]

Tesla is one of the smartest, cracked and most advanced engineering companies in the world.

If they actually did this, then it is likely verifiably true that a dollar above $200/week is waste.

[Shiv Deepak @shivdeepak]

Now imagine sharing memory with the rest of the team. Insane!

Do we now need granular access control on memory? How will we even do it?

[María Benavente @merybenavente]

… and they say romance is dead.

[Shiv Deepak @shivdeepak]

It also understands and evolves specs better than most developers.

So, what shall we do as engineers?

We should get really good at reading and understanding huge walls of text and writing thoughtful prompts.

[Can Vardar @icanvardar]

one of the best things about llms is that they understand git better than most developers

[Shiv Deepak @shivdeepak]

That said, it’s only a matter of time that tokens will get cheaper and we will hopefully think of spending tokens like running EC2 instances.

[Shiv Deepak @shivdeepak]

Loop engineering does seem like the obvious next step for software engineering. But tokens cost $$$ and burning tokens in a loop doesn’t feel like a very responsible thing to do.

One of the luxuries @bcherny has that rest of the world doesn’t is virtually unlimited tokens! 🙂

[Shiv Deepak @shivdeepak]

Or we could say Mental Model changes over Architecture Changes! 😅

[Shiv Deepak @shivdeepak]

A good software engineering practice would be to update the memory along with the PR, and engineer reviewing the PR would focus more on reviewing the memory update and lean less on reviewing the code.

[Shiv Deepak @shivdeepak]

Building agent memory is easier and more useful than writing documentation for software!

[Shiv Deepak @shivdeepak]

So Code Review could become obsolete in favor of Memory Reviews. And these reviews will focus more on architecture and behavior changes!