Arpit Garg

Collecting the dots.

[Arpit Garg @rpitg3]

@cursor_ai finally you fixed the Ctrl+L to clear terminal, it was getting borderline annoying

[Arpit Garg @rpitg3]

@alexocheema @exolabs Can I get access as well please

[Arpit Garg @rpitg3]

@AdityaRajKaul @sanjeevsanyal @grok which restaurant is this

[Arpit Garg @rpitg3]

Zynga was probably the best time I had on Facebook! Me and my friends would play poker ( never played poker before) , mafia wars, FarmVille on Facebook and then meet in the evening or in the school (in India) and chat about it, it was really fun.. don’t think I have used Facebook after that properly.. thanks @markpinc and incredible pod as always

[Sam Parr @thesamparr]

I asked Mark Pincus what Zynga's early numbers looked like.

Went public after 4 years.

The year before going public they did $450m in free cash flow.

Had over $1b in cash on the balance sheet and had never spent a dollar of the money they raised.

[Arpit Garg @rpitg3]

@ThePrimeagen read it in the voice of Seinfeld and Trump, dont know why but it was funny

[Arpit Garg @rpitg3]

@ShanuMathew93 @simonw I had it review all opus slop

[Currence @Currence_AI]

So we built Currence, which combines the trust of expert research with the speed of AI.

Our AI engine ingests thousands of sources into one live view of companies, projects, deals, and costs. On top, our human experts build what AI can't: project economic models, price benchmarks, ranked leaderboards. Together, it’s market intelligence that maps to how energy teams choose partners, forecast demand and price, and decide where to build, buy, or invest.

[Arpit Garg @rpitg3]

@gharkekalesh Love this

[Arpit Garg @rpitg3]

@pitdesi Starlink is great and normal WiFi sucks really bad!

Fyi, https://t.co/17KuJxgVs2 tells you the video streaming throughput, not exactly the internet benchmarking since it connects to Netflix’s cdn servers

[Arpit Garg @rpitg3]

How can banning everything be a solution to all problems?
Cant figure out what is causing exam leaks, ban the platform.
Can't handle govt criticism, ban the platform
Can't handle jokes, ban the comedians
Can't handle corruption, ban the notes

Will you ban air now, to stop air pollution?

[Pavel Durov @durov]

India’s IT ministry banned Telegram for one week because some users shared leaked exam questions.

This punishes 150M+ ordinary Telegram users in India — not the insiders who leaked the exam materials.

And the ban hasn't stopped anything. The leaks just moved to other apps.

[Arpit Garg @rpitg3]

good talk by @helloiamleonie , highly resonate that 80% of the work is still "searching" the right data for the agent doing the task.
We have been using @elastic as a vector store and @JinaAI_ for a while now and initially we created them as proper library tools to be used normally throughout the code.
Now we are integrating them into our agents workflows, providing the same functions as tools to the agents wherever and whenever needed!

[Leonie @helloiamleonie]

Took some inspiration from @vboykis and converted my first ever talk into a blog post.

I talk about the role of agentic search in context engineering.

Together we build an intuition on the strengths and weaknesses of a selection of search tools.

🔗 https://t.co/nuGJ5Zm9Du

[Arpit Garg @rpitg3]

@simonw Agreed 💯! I have been thinking about this a lot and I feel the next version of pair programming is to basically pair prompt together, less time spent on back and forth PR reviews, more trust in the code and system being built and collaborative learning

[Arpit Garg @rpitg3]

Loved this 3rd podcast between @ShaanVP and @MohnishPabrai just like the other 2.

I honestly think this needs to be a regular series, it will be phenomenal and it can easily deep dives into mental models, industries, stories, books etc
This opportunity is not mistress material but wife material!
Great fun listening to both as much as they enjoy talking to each other.

[My First Million @myfirstmilpod]

.@MohnishPabrai's most controversial investing rule: "The mistress is always hotter than the wife."

Here's what it means:

What you own = the wife. You know her flaws, and you see her every day.

What you don't own = the mistress. The unknown looks exciting; she just looks hot.

The trap: You want to swap. "I own this company, but that other one looks better."

This is where investors lose money.

Mohnish's rule: You have to be convinced the mistress is truly hotter, not just the appearance of being hotter.

Most investors swap too often. The bar for action should be extremely high.

Warren Buffett read 19 years of Japanese company reports before finding his $10B opportunity in year 20.

The lesson: Raise your standards. For investments, for people, for everything.

Not being interested in always taking action can give you a huge leg up.

Full Episode: https://t.co/07ggDOuUcS

@ShaanVP

[Arpit Garg @rpitg3]

@ericzakariasson @blader I did feel some of the responses similar to grok's though, quite possible if they had overlapping fine-tuning/RL datasets since they are already sharing colossus

[Arpit Garg @rpitg3]

@ericzakariasson @blader I am quite enjoying Composer 2.5 and the new Agent first UI of cursor, it just sort of clicks!
And performance wise on a few of the tasks, it's up there with GPT 5.5 High Thinking even!
Obviously it takes more time to evaluate your own work but you can sort of feel it..

[Arpit Garg @rpitg3]

@amasad @Replit Great work and perseverance!

[Arpit Garg @rpitg3]

@serenaa_ge @hsaab310 Yes, would be great to see how Composer 2.5 fares! It’s been a great experience using it so far!

[Arpit Garg @rpitg3]

@badlogicgames It’s the law of you go I go

[Arpit Garg @rpitg3]

@blader https://t.co/9XGuGLs5xt

[Arpit Garg @rpitg3]

I learnt about the word "vacuous" from claude code. After looking at the meaning, a bit ironic I must say

I am not sure what data is being used to do RL but they are using quite unusual vocabulary.

I have been comparing outputs of claude against gpt/codex models for the same inputs for a long coding task

I managed to do so much properly in 1M context window of GPT than the claude opus4.7. The responses from claude are quite bloated, unnecessary and sometimes not even relevant!
The new GPT and Codex models are giving me o3-thinking vibes..

[Arpit Garg @rpitg3]

I learnt about the word "vacuous" from claude code. After looking at the meaning, a bit ironic I must say

I am not sure what data is being used to do RL but they are using quite unusual vocabulary.

I have been comparing outputs of claude against gpt/codex models for the same inputs for a long coding task

I managed to do so much properly in 1M context window of GPT than the claude opus4.7. The responses from claude are quite bloated, unnecessary and sometimes not even relevant!
The new GPT and Codex models are giving me o3-thinking vibes..

[Siqi Chen @blader]

honestly claude code / opus 4.7 feels barely usable compared to codex / 5.5 at this point for any kind of long running / complex engineering work