Krishan Gupta

AI Evals @GoogleDeepMind. Obsessed with Intelligence (real and artificial).

[Krishan Gupta @NowImKrishan]

Gemini is 🔥

[Mattia @mattiapomelli]

GPT-5.2 vs Gemini 3.0 vs Opus 4.5

Same prompt.

Which one do you prefer?

[Krishan Gupta @NowImKrishan]

I remember hearing a theory about how AIs can't be funny...

[Nilesh Trivedi @nileshtrivedi]

Someone asked Gemini to imagine HackerNews frontpage 10 years in the future from now:

[Krishan Gupta @NowImKrishan]

Interesting that music is expanding in two directions:

How many people can listen to a song?
Live → records → radio → streaming
One person can make music for billions.

How many people can make a song?
Bands → sampling packs → Large audio model
One person can listen to music made by billions.

[Krishan Gupta @NowImKrishan]

@nrmehta @Ganeshuor I knew that sobriety eval I added would pay off!

[Krishan Gupta @NowImKrishan]

@nrmehta @Yuvnesh Please read the X terms of service you two!

[Krishan Gupta @NowImKrishan]

Counterpoint: https://t.co/wFwXO2mN4l

[Our World in Data @OurWorldInData]

In the last decades, the world has made fantastic progress against extreme poverty. In 1990, 2.3 billion people lived in extreme poverty. Since then, the number of extremely poor people has declined by 1.5 *billion* people.

This means on any average day in the last 35 years, about 115,000 people left extreme poverty behind.

Leaving the very worst poverty behind doesn’t mean a life free of want, but it does mean a big change. Additional income matters most for those who have the least.

It means having the chance to leave hunger behind, to gain access to clean water, to access better healthcare, and to have at least some electricity — for light at night and perhaps even to cook and heat.

Can we expect this rapid progress to continue? Unfortunately, we cannot. Based on current trends, progress against extreme poverty will come to a halt.

As you can see in the chart, the number of people in extreme poverty is projected to decline, from 831 million people in 2025 to 793 million people in 2030. After 2030, the number of extremely poor people is expected to increase.

To understand why the rapid progress against deep poverty will not continue into the future, we need to know why the world made progress in the past.

Extreme poverty declined in the last three decades because, back in the 1990s, the majority of the poorest people on the planet lived in countries that subsequently achieved very fast economic growth.

In Indonesia and China, more than two-thirds of the population lived in extreme poverty. But these economies then grew rapidly, so that by today, the share has declined to less than 10%.

Other large Asian countries — including India, Pakistan, Bangladesh, and the Philippines — also achieved strong growth, and as a consequence, the share living in extreme poverty declined rapidly.

Much of the progress happened in Asia, but conditions in other regions improved too: the share living in extreme poverty also declined in Ghana, Cape Verde, Cameroon, Panama, Bolivia, Mexico, Brazil, and many other countries.

What is different today is that the majority of the world’s poorest people are stuck in economies that have been stagnating for a long time. Unless the poorest economies start growing, this period of progress against the worst form of poverty is over.

[Krishan Gupta @NowImKrishan]

@Austen It's not about speed, it's about parallelism.

Quantum computing leverages superposition to explore multiple computational paths simultaneously (each in a separate universe, if you believe Everett).

[Krishan Gupta @NowImKrishan]

I'm not even caught up on email and that presentation was due last Friday 😭

[Matt Turck @mattturck]

Us: “Super productive week, I’m fully caught up on email and almost done with my big presentation!”

Elon:

[Delip Rao e/σ @deliprao]

Is Boston Legal the first documented LLM jailbreak attempt? Watch with sound on.

[Krishan Gupta @NowImKrishan]

@Yuvnesh It's the example most used to test if an LLM is functional (and included in lots of unit tests, documentation for the whole ecosystem).

[Krishan Gupta @NowImKrishan]

If the capital of France ever changes from Paris we're going to have a real y2k moment in LLMs

[Krishan Gupta @NowImKrishan]

Resetting my p(doom) back a few years…

[Krishan Gupta @NowImKrishan]

@nrmehta I would 100% read " How to Succeed at Success" 😂

[Krishan Gupta @NowImKrishan]

I always thought my coding skill would atrophy. Turns out coding just gets easier and easier so it's OK.

[Google for Developers @googledevs]

🚀 Build from anywhere with #ProjectIDX!

Level up your dev game with AI-powered coding, one-click deployment, and tons of starter templates —all from your browser.

Celebrate its one-year anniversary 🎉 → https://t.co/VqMPbWnhSt

[Krishan Gupta @NowImKrishan]

I laughed, and laughed, and laughed… cried a little if I’m going to be honest.

[Krishan Gupta @NowImKrishan]

@Yuvnesh 😂😂😅

[Krishan Gupta @NowImKrishan]

It's a good thing no one ever forced Picasso to tell them what metrics his paintings were going to drive

[Krishan Gupta @NowImKrishan]

@Yuvnesh Not everything that matters can be measured!

[Greg Brockman @gdb]

evals are surprisingly often all you need

[Krishan Gupta @NowImKrishan]

The Quiet Quitting guy turns out to be super thoughtful.

[Quiet quitting] was followed by an existential dread of: “What am I actually doing with my life?”

If you can afford to quiet quit, you can afford to quit. Be all in or all out.

https://t.co/JHunT2zibH