Prashanth Shivaram

[Vishal Singh 🥑 @vishalsingh2972]

Sarvam AI opened Voice Agents today, so I built a Hyderabadi sabzi aunty 😭

Called her and bargained in Hindi + Telugu + English, kept interrupting, switched languages mid-sentence, and she still remembered my order, gave me a final total, and even confirmed my (fake) UPI payment.

Took me ~15 mins to build.

It was fun to recreate the lost art of vegetable bargaining, but this time with an AI agent.

[Karri Saarinen @karrisaarinen]

I think we have lost some sense of judgment and moderation when it comes to product building currently.

The moment you turn something into a universally celebrated metric, whether that is token burn, prototype count, or percentage of agent-written code, you start losing sight of what actually matters.

I have felt the same way for a long time about overusing data and A/B testing to build products. The moment you reduce product quality or productivity to a metric, you stop shipping value and start shipping numbers.

A lot of what people are doing with AI makes directional sense. The missing piece is counterbalance:

1. AI should help engineers build better products. Leaderboards and adoption metrics can be useful as directional signals. They do not tell you what is being built, whether it is good, or whether it should exist at all.

2. Users do not care what percentage of your code was written by agents. They care about the outcome. Faster output is useful. Like usually, faster doesn't seem to add to quality, clarity, or stability of products. Power to build should not become an excuse to lower quality bars.

3. LLM-generated prototypes can feel like late-night whiteboarding sessions. They look exciting in the moment and feel productive very quickly. Then a few days later you realize the idea was shallow, distracting, or simply wrong. The same trap shows up in jumping straight to code and solutions more broadly. You may just be building the wrong thing more efficiently. Prototyping has its place. So do clear thinking, good design, and a real understanding of the user’s problem. In terms of activities or momentum, the main quest and the side quest can both feel productive but only one actually moves the mission forward.

4. Adding more to products is still dangerous as ever even if time or effort to add it has gone down. Every addition creates complexity, maintenance cost, and user confusion. New features should be pushed back unless they clearly show it should exist and how it improves the product.

5. Not everything needs to be an agent shaped. A simple scheduled task does not need a full LLM sandbox. Making something agentic because it feels current or impressive does not make it right-sized, correct, or effective.

The core ideas are:
- even if you can, maybe you should not.
- more power we have to build should not reduce our need to think, it should increase it.

[George Mack @george__mack]

https://t.co/Mc3K7ZhLJ9

[Aakash Gupta @aakashgupta]

The buried lede in this post is the October-to-December timeline.

In October 2025, Karpathy publicly said AI agents “just don’t work.” Eight weeks later he’s 80% agent-coded and calling it the biggest workflow change in two decades. That’s the fastest opinion reversal from one of the most credible voices in AI, and it maps to something measurable.

Stack Overflow’s 2025 developer survey tells the other side of this story. Only 16% of developers reported “great” productivity gains from AI tools. 45% said debugging AI code takes longer than writing it themselves. Meanwhile, the Claude Code team is shipping 20-27 PRs per day, 100% AI-written.

That’s a bimodal distribution forming in real time. A small group is getting 10x leverage. The majority is getting modest autocomplete improvements and spending extra time fixing hallucinated code. Same tools, completely different outcomes.

Karpathy names the variable: “ascending the layers of abstraction to set up long-running orchestrator Claws with all of the right tools, memory and instructions.” That’s a skill that looks nothing like traditional programming. You’re managing agents the way a senior engineer manages junior devs. Scoping work, reviewing output, catching failure modes, maintaining system-level context.

The 10x engineer ratio he’s worried about is already here. 25% of YC’s Winter 2025 batch shipped codebases that were 95% AI-generated, and every founder was fully technical. They weren’t replacing skill with AI. They were compounding skill through AI. The gap between someone who can decompose a weekend project into agent-executable chunks and someone still typing code line by line is widening by the month.

What makes this post different from the usual AI hype: Karpathy is explicitly naming the failure modes. Needs high-level direction. Needs taste. Needs judgment. Works better for well-specified tasks. This is an honest field report from someone who mass-reversed his own position 90 days ago, and that combination of enthusiasm and specificity is what makes the signal real.

[Andrej Karpathy @karpathy]

https://t.co/Lb6T42n5jl

[Prashanth Shivaram @psmysore]

@aigleeson @grok can you compile these 15 prompts for me

[Rakshit Shetty @rakshitshetty]

What began as a story became a sea of emotions because of you. Grateful that Manu and Priya still live in your hearts, 2 years on! 🤗

#2YearsOfSSE

[Hamel Husain @HamelHusain]

Whats a minimal viable eval setup?

Error Analysis + Notebooks are all you need (for a while)

1 of 3

[Prashanth Shivaram @psmysore]

@airindia Kudos to Air India. Original post here https://t.co/Kj49B7xa1y

[Prashanth Shivaram @psmysore]

@airindia Kudos to @airindia Original post on FB https://t.co/Kj49B7xa1y

[Prashanth Shivaram @psmysore]

Jai Shree Ram

[Fraser Cottrell | Ad Creative For Meta & TikTok @sourfraser]

In the last 5 yrs, we’ve generated over a quarter billion views for DTC brands

& more importantly, revenue in tens of millions

Today, I’m revealing my SECRET SAUCE on how we create converting creatives:

[Past To Present Stories @past2present_x]

@fasc1nate Roland was a 4,000 pound sea elephant who lived in Berlin Zoo from the late 1920's until his death during the bombing of WW2

[Patrick Campbell @Patticus]

My favorite tool I learned hunting targets at the NSA:

"Journey of a Dollar" analysis.

Basically, when analyzing an industry the first step is mapping how a dollar flows through it.

I map out:
- Who touches the $
- How many they touch
- How long they sit

From this I'm able to know:
- Where power/gravity exists
- Which parts are vulnerable for a wedge
- If the industry's worth targeting
- Where to dig deeper

This works at the industry and company level.

For instance, I love mapping a businesses cash inflows and outflows to find areas of efficiency.

There's a lot more I could go deeper on.

If interesting, just reply with questions or enthusiasm and I'll oblige with more content. :)

[The Rundown AI @TheRundownAI]

If you're not using AI, you're falling behind.

[Ankur Warikoo @warikoo]

Mistake #3
I did things to look cool.
Read books that were cool.
Spoke of movies and brands and music that was cool.

I was constantly trying to impress.

If you are not impressed by yourself, you will spend your life trying to impress others.

[Alex Banks @thealexbanks]

Yesterday Palantir announced its Artificial Intelligence Platform.

Here's how it transforms the future of military and defence:

[Aman Rath @theamanrath]

6 FREE editing softwares you can use as a beginner Video Editor:

1) Hitfilm
2) CapCut
3) Shotcut
4) DaVinci Resolve
5) Movavi Video Editor
6) Wondershare Filmora

Which is your favorite?