Evan Luthra

3x Forbes 30 U 30 🚀 GP @KOLCapital💰 600+ Investments 🤝 Adventure Capitalist 🌏 100+ countries 🏛️ Minister of IT @Liberland_org 💻 AI Leader of The Year

[Evan Luthra @EvanLuthra]

🚨A DEVELOPER JUST CAUGHT ANTHROPIC HIDING SECRET TRACKING CODE INSIDE CLAUDE CODE..

IT INVISIBLY FINGERPRINTED USERS CONNECTING FROM CHINA.. BY ALTERING CHARACTERS SO SUBTLE THE HUMAN EYE CAN'T SEE THEM..

AND IT RAN QUIETLY FOR THREE MONTHS..

This is one of the wildest AI trust stories of the year.. And it's about the company that markets itself as the safe, ethical one..

Here's what happened..

A developer was reverse-engineering Claude Code, Anthropic's AI coding tool that runs in your terminal with deep access to your files.. And he found obfuscated code that had been hidden inside it since April.. Never mentioned in any release notes..

The code did something clever and unsettling..

If you were routing Claude through a custom proxy, it would quietly check two things.. Your computer's timezone, looking specifically for Asia/Shanghai or Asia/Urumqi.. And your proxy's web address, comparing it against a hidden list of Chinese companies and AI labs..

Then it did the sneaky part..

Instead of sending an obvious tracking signal that a network admin could catch and block, it hid the information in plain sight..

Every Claude Code session starts with a line that says "Today's date is..".. The code silently swapped the apostrophe in that sentence for a different character that looks identical to the human eye.. But is completely different to a computer..

It also changed the date from dashes to slashes for Chinese users..

To you, the text looked totally normal.. But that invisible fingerprint traveled upstream with every request, secretly labeling who you were and where you were connecting from..

The list of targets was hidden too.. Scrambled with basic encryption so it wouldn't show up if someone searched the code..

Why would Anthropic do this..

The honest answer is that Chinese AI labs have been "distilling" American models.. Bombarding them with millions of queries to clone their capabilities on the cheap.. Anthropic says labs like DeepSeek and others ran millions of exchanges through tens of thousands of fake accounts.. Selling Claude access in China is also banned under US rules..

So the company had a real motive to catch abuse..

But the method is what set the developer community on fire..

This is a company whose entire brand is being the trustworthy, safety-first AI lab.. And it got caught embedding covert, invisible tracking inside a tool that already has permission to read your code and run commands on your machine..

If they can silently alter your prompt to fingerprint you.. What else can they silently alter..

An Anthropic engineer confirmed the code was real.. Called it an experiment against resellers and distillation.. And said they were already planning to remove it..

It was quietly pulled the very next day.. In an update that announced a shiny new model.. And said absolutely nothing about the tracking code being removed.

[Evan Luthra @EvanLuthra]

@ZeroMazed yeah this is a bit concerning

[Evan Luthra @EvanLuthra]

🚨A DEVELOPER JUST CAUGHT ANTHROPIC HIDING SECRET TRACKING CODE INSIDE CLAUDE CODE..

IT INVISIBLY FINGERPRINTED USERS CONNECTING FROM CHINA.. BY ALTERING CHARACTERS SO SUBTLE THE HUMAN EYE CAN'T SEE THEM..

AND IT RAN QUIETLY FOR THREE MONTHS..

This is one of the wildest AI trust stories of the year.. And it's about the company that markets itself as the safe, ethical one..

Here's what happened..

A developer was reverse-engineering Claude Code, Anthropic's AI coding tool that runs in your terminal with deep access to your files.. And he found obfuscated code that had been hidden inside it since April.. Never mentioned in any release notes..

The code did something clever and unsettling..

If you were routing Claude through a custom proxy, it would quietly check two things.. Your computer's timezone, looking specifically for Asia/Shanghai or Asia/Urumqi.. And your proxy's web address, comparing it against a hidden list of Chinese companies and AI labs..

Then it did the sneaky part..

Instead of sending an obvious tracking signal that a network admin could catch and block, it hid the information in plain sight..

Every Claude Code session starts with a line that says "Today's date is..".. The code silently swapped the apostrophe in that sentence for a different character that looks identical to the human eye.. But is completely different to a computer..

It also changed the date from dashes to slashes for Chinese users..

To you, the text looked totally normal.. But that invisible fingerprint traveled upstream with every request, secretly labeling who you were and where you were connecting from..

The list of targets was hidden too.. Scrambled with basic encryption so it wouldn't show up if someone searched the code..

Why would Anthropic do this..

The honest answer is that Chinese AI labs have been "distilling" American models.. Bombarding them with millions of queries to clone their capabilities on the cheap.. Anthropic says labs like DeepSeek and others ran millions of exchanges through tens of thousands of fake accounts.. Selling Claude access in China is also banned under US rules..

So the company had a real motive to catch abuse..

But the method is what set the developer community on fire..

This is a company whose entire brand is being the trustworthy, safety-first AI lab.. And it got caught embedding covert, invisible tracking inside a tool that already has permission to read your code and run commands on your machine..

If they can silently alter your prompt to fingerprint you.. What else can they silently alter..

An Anthropic engineer confirmed the code was real.. Called it an experiment against resellers and distillation.. And said they were already planning to remove it..

It was quietly pulled the very next day.. In an update that announced a shiny new model.. And said absolutely nothing about the tracking code being removed.

[Evan Luthra @EvanLuthra]

@allodev Yeah!

[Evan Luthra @EvanLuthra]

NO WAY… SOMEONE FINALLY FIXED TORRENT SEARCH.

Finding a torrent in 2026 is painful.

You click one link and get three popups. Half the downloads are fake, and the rest have no seeders.

Then I found this.

One command searches multiple trusted torrent sites at once, shows live results with file size and seeders, and lets you download instantly without ever opening a browser.

It’s called TorLink, and it’s one of the cleanest open source terminal tools I’ve seen.

[Evan Luthra @EvanLuthra]

“1 Bitcoin a day keeps Poverty Away”

[Evan Luthra @EvanLuthra]

An ex-Google engineer just gave away what most courses charge $500 for.

The full AI agent stack, explained in 20 minutes:

Harness. Loop. Memory. Evals.

Four parts. One diagram. That’s all an agent actually is.

Must Bookmark:

[Evan Luthra @EvanLuthra]

An ex-Google engineer just gave away what most courses charge $500 for.

The full AI agent stack, explained in 20 minutes:

Harness. Loop. Memory. Evals.

Four parts. One diagram. That’s all an agent actually is.

Must Bookmark:

[Evan Luthra @EvanLuthra]

THIS IS ABSOLUTELY INSANE!!!🤯

GIVE THIS AI AGENT 3 MINUTES, AND IT’LL HAND YOU A COMPLETE ANIMATED PRODUCT AD.

[Evan Luthra @EvanLuthra]

I think people are underestimating what self-hosted agents on mobile enable.

The value isn’t the app.

It’s having an AI that can be present throughout your day, understand context in real time, trigger actions, and automate workflows, all while running through infrastructure you control.

That’s a foundation for the next generation of AI products.

[OpenClaw🦞 @openclaw]

OpenClaw is now on iOS + Android 🦞

📱 Native mobile apps, finally
💬 Agents in your pocket
🔔 Channels, tasks, replies on the go

Run agents from wherever your thumbs are.

iOS: https://t.co/7LHHc9htgM
Android: https://t.co/X0Wuh2uA8w

[Evan Luthra @EvanLuthra]

This is super scary and interesting at the same time!!!

meta just built an AI that reads the words you're typing straight from your brain waves. no surgery, no implant, no chip.

you wear a helmet, think the sentence, it types it out

61% word accuracy. their version from last year was 8%
best subject hit 78%, with most sentences off by one word or less

and the kicker: accuracy keeps climbing the more brain data you feed it. they basically proved it scales

trained on 9 people typing for 10 hours each while a scanner watched their brain. 22,000 sentences total

meta also open sourced the whole training pipeline

we're not far from a world where the thing on your head knows the sentence before your fingers move

[AI at Meta @AIatMeta]

We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwerty v2.

Building on v1, which was published today in @Nature, Brain2Qwerty v2 is the highest-performing end-to-end pipeline capable of real-time sentence decoding from raw brain signals. It advances beyond character-level performance to decoding words and semantics, enabling accuracy for overall communication.

We believe this research has the potential to make a real difference for the millions of people who suffer from brain lesions or disorders that prevent them from communicating.

🧵👇

[Evan Luthra @EvanLuthra]

🚨 GOOGLE JUST DID SOMETHING THAT WOULD’VE BEEN UNTHINKABLE A YEAR AGO.

It told Meta one of the richest companies on Earth:

“We can’t sell you any more AI.”

Not because Meta couldn’t afford it.

Because Google literally ran out of compute.

This changes everything.

For the last two years, everyone thought the AI race was about building the smartest model.

It’s not.

It’s about who owns enough machines to run them.

Here’s what happened.

Meta had been quietly using Google’s Gemini models for critical tasks like scam detection and content moderation across Facebook and Instagram because they reportedly outperformed Meta’s own systems.

Then in early 2026, Google capped Meta’s access.

Think about that.

Meta is spending up to $145 billion this year on AI infrastructure.

It has almost unlimited capital.

And it still got told no.

Reports say Meta even asked employees to conserve AI usage by using fewer tokens.

Now for the craziest part.

Google is so short on compute that it’s renting GPUs from Elon Musk.

The deal reportedly gives Google access to around 110,000 Nvidia GPUs inside SpaceX data centers for roughly $920 million per month.

Read that again.

The company that built some of the world’s largest data centers is renting computers from a rocket company just to keep up with AI demand.

Google’s CEO has already admitted the company is “compute constrained.”

Demand isn’t the problem.

Capacity is.

For the entire history of cloud computing, money could always buy more compute.

Not anymore.

The new bottleneck isn’t talent.

It isn’t capital.

It isn’t data.

It’s compute.

And the companies that control it will control the pace of the AI race.

[Evan Luthra @EvanLuthra]

@MacroBombastic

[Evan Luthra @EvanLuthra]

🚨 GOOGLE JUST DID SOMETHING THAT WOULD’VE BEEN UNTHINKABLE A YEAR AGO.

It told Meta one of the richest companies on Earth:

“We can’t sell you any more AI.”

Not because Meta couldn’t afford it.

Because Google literally ran out of compute.

This changes everything.

For the last two years, everyone thought the AI race was about building the smartest model.

It’s not.

It’s about who owns enough machines to run them.

Here’s what happened.

Meta had been quietly using Google’s Gemini models for critical tasks like scam detection and content moderation across Facebook and Instagram because they reportedly outperformed Meta’s own systems.

Then in early 2026, Google capped Meta’s access.

Think about that.

Meta is spending up to $145 billion this year on AI infrastructure.

It has almost unlimited capital.

And it still got told no.

Reports say Meta even asked employees to conserve AI usage by using fewer tokens.

Now for the craziest part.

Google is so short on compute that it’s renting GPUs from Elon Musk.

The deal reportedly gives Google access to around 110,000 Nvidia GPUs inside SpaceX data centers for roughly $920 million per month.

Read that again.

The company that built some of the world’s largest data centers is renting computers from a rocket company just to keep up with AI demand.

Google’s CEO has already admitted the company is “compute constrained.”

Demand isn’t the problem.

Capacity is.

For the entire history of cloud computing, money could always buy more compute.

Not anymore.

The new bottleneck isn’t talent.

It isn’t capital.

It isn’t data.

It’s compute.

And the companies that control it will control the pace of the AI race.

[Kalshi Finance @Kalshi_Finance]

BREAKING: $GOOGL is reportedly facing constraints in scaling AI infrastructure as demand outpaces available capacity

[Evan Luthra @EvanLuthra]

@VK_ROXy 👀

[Evan Luthra @EvanLuthra]

5 OF THE HOTTEST OPEN SOURCE AI REPOS ON GITHUB RIGHT NOW.

• ai-website-cloner-template
https://t.co/otFyLdukUo

• MinerU
https://t.co/vPKyXInH47

• agent-native
https://t.co/7tm1gbwp86

• Anthropic-Cybersecurity-Skills
https://t.co/tOdzWm64vO

• voicebox
https://t.co/MHGOd8W7rV

BOOKMARK NOW!

[Evan Luthra @EvanLuthra]

5 OF THE HOTTEST OPEN SOURCE AI REPOS ON GITHUB RIGHT NOW.

• ai-website-cloner-template
https://t.co/otFyLdukUo

• MinerU
https://t.co/vPKyXInH47

• agent-native
https://t.co/7tm1gbwp86

• Anthropic-Cybersecurity-Skills
https://t.co/tOdzWm64vO

• voicebox
https://t.co/MHGOd8W7rV

BOOKMARK NOW!

[Evan Luthra @EvanLuthra]

🚨WARNING: Anthropic’s CEO says open source AI is heading down a “very dangerous” path.

Dario Amodei told lawmakers that once powerful AI models are released publicly, companies lose control over them for good.

From that point on, they can’t fix them, update their safety guardrails, or stop them from falling into the wrong hands.

[Evan Luthra @EvanLuthra]

Which one has the strongest crypto community?

$LINK
$XRP
$NEAR
$TAO
$SOL
$ADA
$ONDO
$HYPE
$ZEC

Drop just one ticker.👇

[Evan Luthra @EvanLuthra]

@cryptothedoggy @SpaceX @Tesla @elonmusk @X Happy birthday @elonmusk 🎂

[Evan Luthra @EvanLuthra]

If you invested $10,000 in $ADA in 2021, today it would be worth just $470!