HUGE Immigration News! The birthright citizenship executive order Trump signed today means children of non-citizens or non-permanent residents will not be US citizens.

Children of Indians who are in the 100+yr green card wait WILL NOT be US citizens.

Applies from Feb 20, 2025. Source: whitehouse.gov/presidential-a…

HUGE: DeepSeek R1 from China reaches o1 level performance! Most important paper of 2025.

They show that using pure RL using GRPO can teach models to reason better and better with more tokens: the first at this scale.

Here are the 4 top research learnings from the paper:

1/5 1. Distilling more powerful models like R1 on smaller base models (Qwen / Llama) outperforms large scale RL training on those base models themselves (like QwQ-32B).

2/5

MEGATHREAD: EB-1s for Indians in the US 2025!

— It's been 2yrs since the priority date has moved past Feb '22
— Indian EB-1 petitions stayed flat this year, at ~12k
— The backlog moved forward by ~200/mo.
The EB-1 backlog is basically stuck.

Let's break it down..

1/12 What is the EB-1?

Indians under the standard EB-2/3 green card category have a 100+yr long wait. This is because of the 7% country cap by category.

This was the initial misconstrued old remark by Sriram that was taken out of context and set off this whole H-1B drama.

2/12

The Superhuman AI for Poker is a timeless research paper.

Poker, unlike Go/Chess, is an imperfect info game. Pluribus, made by now OpenAI engineer, beat 5 pros decisively in 10,000 poker hands, winning 48 milli big blinds/game.

Made a mini demo on rock-paper-scissors too:

1/5 Like humans, Pluribus confirms that limping, calling the “big blind” rather than folding or raising, is suboptimal.

However it disagrees with the wisdom that “donk betting” (starting a round by betting when one ended the previous betting round with a call) is a mistake.

2/5

BOMBSHELL: New Hindenburg report calls Carvana "a father-son accounting grift for the ages."

The online car dealer's stock went 60x in the last 2yrs after dropping 99%, hitting a $44B valuation, while the CEO's father sold $5B in stock.

Here's the wild story of $CVNA...

1/10 From Aug 2021 to Dec 2022, Carvana stock plunged 99%. CEO's father Ernest Garcia II sold $3.6B in stock before that.

After the 63x rebound, he just sold another $1.4B.

Funny timing?

2/10

H-1B DATA! SALARY VS INDIAN % show the 2 sides of the H-1B

~380K accepted H-1Bs (not LCAs) from 2021-24. We plot the top 200 employers by volume: avg base salary vs %age of H-1Bs that were Indian.

The entire H-1B is NOT low wage Indian. ~30-50% is.

Let's unpack the data🧵

1/7 The cluster of large employers that pay <$100k and 95%+ of their applicants are Indian are the IT staffing firms

Infosys
TCS
Cognizant
Wipro
IBM
Capgemini
Accenture
L&T
HCL
Tech Mahindra

Including the long tail of smaller firms shows ~50% of H-1Bs are in this bucket.

2/7

We can't argue about the H-1B without data, so here it is:
Top 200 H-1B Employers by Salary from 2021-2024!

The red bars are companies with 95%+ Indians — the IT staffing firms, all in the bottom 50%. Bottom 70 are <$100k.

~30-50% H-1B has a low-wage problem. Source: github.com/BloombergGraph…

Starlink is one of the seminal feats of engineering in history.

It will enable internet that's
— fast 100-300mbps
— uncensored
— cheap $1500/yr
in:
— the most remote areas
— ships in the ocean
— airplanes in the sky
— poles

But few even know what this picture is..

1/10 Traditional satellite internet uses geostationary orbit (GEO) - satellites at 36,000km altitude. The physics is simple but the latency is brutal: 600ms+ for signals to make the round trip.

Online gaming? Video calls? Forget it.

2/10

OpenAI o3 scored 91.5% on public ARC-AGI test despite taking 57M tokens and ~$5000/task. The problems are fairly simple!

I made an app with Claude to visualize them. Strangely, Sonnet can even identify the pattern sometimes.

Let's look at the 34 problems o3 got wrong...

1/13 The spiral with a margin of one.

2/13

HUGE immigration news on the H-1B!

— Cap-exempt H-1B (no lottery) applies to more companies
— Founders can self-petition and own >50%
— Easier hiring for startups
— F-1 students get cap-gap until April 1
— Crack down on IT consulting fraud

Let's get into it 🧵

1/7 More orgs are cap-exempt employers.

Beyond the 85k H-1B limit, certain employers could bypass the lottery. Now the definition changes from "primarily engaged" in research to it being a "fundamental activity", unlocking more companies. Part-time work is allowed.

2/7

o1-preview is far superior to doctors on reasoning tasks and it's not even close, according to OpenAI's latest paper.

AI does ~80% vs ~30% on the 143 hard NEJM CPC diagnoses.

It's dangerous now to trust your doctor and NOT consult an AI model.

Here are some actual tasks:

1/5 Here's an example case looking at phosphate wasting and elevated FGF23, then proceeded to imaging to localize a potential tumor.

o1-preview suggested testing plan takes a broader, more methodical approach, systematically ruling out other causes of hypophosphatemia.

2/5

Want to design better AI agents? Take notes from code writing systems.

Techniques include
— Multi-agent
— Tool choice
— Underlying model
— Diff format
— Innovative Signals
— Code retrieval + knowledge graphs
— LSP
— Fault localization

Let's dive deeper with real examples:

1/10 Multi-agent

Use agents with different roles / prompts that have access to different tools and can hand off to another agent.

Some roles used in coding: searcher, planner, reproducer, coder, tester, editor

2/10

As cool as the new Sora is, gymnastics is still very much the Turing test for AI video.

1/4 Attempt 2.

2/4

HUGE Immigration News!

We have the first EVER look at H-1B lottery data. Did you also suspect the lottery wasn't truly random? They're not.

Certain companies like Tiktok and Bytedance have 50% higher odds than average.

I broke it down by nationality, company and age...

1/5
By Age.

In 2024, there were 350,084 applications and 85,304 were selected, a 24.4% acceptance rate.

The process seems to have rampant ageism, with only ages 26-32 having above average acceptance rates.

A 26yo has 50% higher odds than a 36yo+!

2/5

A small company in the <1M city of Niigata, Japan has a monopoly on the equipment that makes every single modern iPhone and TV display on the planet.

Here's the story of Tokki, the most important company you've never heard of...

1/8 Tokki makes just ~10 ELVESS machines a year.

Each one is a clean room within a clean room, stretching longer than an Olympic swimming pool, and can costs $ 100M+. They're the only ones who can do it.

2/8

The middle manager is the biggest culprit of the "quiet quitting" SWE epidemic.

They have 0 incentive to fire. The entire job is bargaining for more headcount so they can get promoted.

They'll say "we are understaffed, we need more people" no matter how little they do.

1/4 If you have multiple levels of this, its often dysfunctional down the entire chain.

The non technical middle manager are the worst culprits. ICs can (and will) swindle them endlessly into their infinite timelines. I've never worked with a good one.

2/4

The youth of India spends money they don't have to play status games.

Multiple people making <₹50k ($600)/mo in their 20s are buying
— iPhone & other Apple products (on interest)
— Coldplay / Dua Lipa concerts
— trips to Thailand / Vietnam / Goa

They save little to no money. I’ve been in India and multiple people have told me they’re essentially broke after multiple years of working.

Income asymmetry amongst similar social groups is high which can force this “suddenly my college buddy makes 90LPA but I’m at 12”

The peer pressure to “fit in” is high

A viral thread about a growing number of "quiet quitting" SWEs was seen 15M+ times (and led to a crack down at Box)

My DMs have been full of ppl telling me that people badge in for 30mins/day and go on 5+ intl vacations/yr at these companies.

Here's what went down..

1/4


— Stanford researcher concluded ~10% of SWEs are ghosts
— Box Founder / CEO said "fml", his mom texted him, and then he sent a company-wide email
— Karpathy backed it up saying he's heard it happen

2/4

Software engineers aren't lazy.

They just have no incentive to be productive at bigCos.

Here's why:
— No reward for simplicity / deleting things
— No prizes for doing it quickly (everyone assumes it was easy)
— Little interesting work. Business needs are often boring.

1/5 — No dearth of alternate career options. Get fired? Just change jobs
— No accelerated career growth.
— No innovation. Middle managers only get fired for rocking the boat.
— No $$ reward. Stock goes up for reasons beyond your control.

2/5

Using light as a neural network, as this viral video depicts, is actually closer than you think. In 5-10yrs, we could have matrix multiplications in constant time O(1) with 95% less energy. This is the next era of Moore's Law.

Let's talk about Silicon Photonics...

1/9 The core concept: Replace electrical signals with photons.

While current processors push electrons through metal pathways, photonic systems use light beams, operating at fundamentally higher speeds (electronic signals in copper are 3x slower) with minimal heat generation.

2/9

A single fairly unknown Dutch company makes maybe the most expensive and complex non-military device ($200M) that builds on 40 years of Physics and has a monopoly responsible for all AI advancement today.

Here's the story of ASML, the company powering Moore's Law..

1/9 ASML's extreme ultraviolet (EUV) machines are engineering marvels.

They shoot molten tin droplets 50,000x/s with a 25kW laser turning it into plasma as hot as the sun's surface to create 13.5nm UV light —so energetic it's absorbed by air itself.

2/9