8 Google engineers wrote the paper that every AI company now uses as their bible. OpenAI built GPT on it, Anthropic built Claude on it, and Meta built LLaMA on it.

Every LLM worth billions uses this paper's transformer architecture as the foundation... Before 2017, teaching computers human language was torture.
AI would read text like humans reading through a keyhole - one word at a time.

They were slow, forgot context, and choked on long passages.
Then 8 researchers decided to flip things up...

A Swedish non-profit operates the DNS billions of internet users depend on from Cold War bunkers.

They've had 100% uptime for 20+ years & they developed the invisible signatures that protect domains from hackers.

How they do it is interesting... Without these signatures, hackers can hijack your DNS requests.

Which means that when you visit a website, you can be redirected to a fake site that looks same.

It's known as DNS poisoning & hackers just need to intercept your DNS request & send back a fake IP address to do it.

Three German brothers emailed eBay in 1999: "Let us run Germany for you."

eBay ignored them. So they cloned eBay, called it Alando, and made it so big that 100 days later eBay had to buy it for $43 million.

But what happened next was even more interesting... The brothers - Marc, Oliver, and Alexander Samwer - turned this into a formula:

> Find successful US startups that hadn't expanded to Europe.
> Copy them exactly.
> Scale faster than the originals could expand.
> Sell it back to them or dominate.

They did this 100+ times.

Everyone thinks Apple is losing the AI race.

But Apple made their Neural Engine 60x more powerful.
Its M4 chip processes AI inputs 2X faster than rivals.

And they're quietly using the picks and shovel strategy used by Levi's during the California Gold Rush.

Thread
Let's first go back to 1849.

A news headline about California having a lot of gold broke out.

Hundreds and thousands of people rushed to California digging for gold.

But most of them died or went completely broke.

However, there was a guy named Levi Strauss...

Everyone's freaking out about Microsoft's deal with Nebius for $19.4 billion.

Two years ago, the same company was sanctioned and delisted from Nasdaq.

The founder fled from Russia with 1,300 engineers after condemning Putin's war.

Here's the wild story: Microsoft's deal sent Nebius from $64 to $90 in hours.

$19.4 billion through 2031. That's 13x what Nebius made in all of 2024.

Microsoft had no choice though. They'd just lost their main GPU supplier to OpenAI...

Pentagon can't operate without it.
Netflix can't stream without it.
And banks can't trade without it.

Yet most people have never heard of Akamai.

How a $11 billion company operating on a 25-year-old mathematical equation secures 2 trillion of your interactions 🧵 In 2024 alone, Akamai blocked 311 billion web attacks (that's 850 million attacks per day)

But the irony is that the Israeli commando who co-founded Akamai was the first victim to be stabbed on the 9/11 flight.

While Danny Lewin was dying, his algorithm was being tested...

We can now read AI's personality like a brain scan - and change it with basic arithmetic.

Anthropic proved traits like evil and hallucination are just mathematical patterns in neural networks. You can literally add or subtract it.

Here's how you do it🧵 When an AI lies, specific neurons fire in a pattern. Same when it's helpful or deceptive.

Like finding what makes someone angry by comparing their brain when calm vs furious.

Take the difference between "lying AI" and "honest AI" brain patterns. That's the lying vector.

Masayoshi Son had lost $59B in 24 months.

The worst venture capital failure in history. Softbank's fund became a laughingstock.

He had broken down: "I'm 65 and haven't done anything."

3 years later, Softbank is up by 189% and betting $500B on building ASI.

A thread🧵 So get this.

Vision Fund losses:
2022: $27.4 billion destroyed
2023: $32 billion destroyed

$59B gone in 24 months. For context, that's more than the GDP of Slovenia. The largest venture capital failure ever.

And one investment was the main culprit.

ChatGPT, Gemini, and Claude don't train on data owned by Big Tech.

They train on 250 billion web pages scraped by a nonprofit nobody knows exists.

It's free. It's fragile. And it's about to break.

How it fuels AI (and what happens if it stops)🧵 First, the scale:

- 9.5 petabytes of web data since 2008
- 3-5 billion new pages every month
- 64% of all large language models use it.

Without this non-profit, ChatGPT wouldn't exist.
The founder previously built Google's money printer.

Meta, Google, and Microsoft all use encryption built by the same 50-person nonprofit.

Zero revenue from 2 billion users. The founder uses a fake name. And when the FBI subpoenaed them, they only provided 2 pieces of data.

Here's how a non-profit secures the internet🧵 In 2013, a tiny nonprofit had just 3 developers building encrypted messaging that no government could crack.

Their leader went by "Moxie Marlinspike" - not his real name.

WhatsApp founders were tracking their journey.

Everyone laughed at Nadella’s “crazy” $26B LinkedIn purchase.

Now it drives more SaaS revenue than some entire Fortune 500 firms.

$200B+ in value from one move.

Here’s how Satya Nadella quietly made Microsoft’s smartest acquisition ever: 🧵 2014: Satya Nadella became Microsoft’s 3rd CEO.

The company was worth $300B
Windows sales were flat
Cloud growth was slow

He needed a bold bet to pull Microsoft out of its comfort zone.

And it would come from the last place people expected…

In 1936, an airplane engineer discovered a formula that would predict 99.6% collapse in solar panel prices.

It's the same formula that explains why EVs will be 50% cheaper than gas cars by 2030, and why AI training costs are dropping 65% a year.

Here's how it works🧵 Theodore Wright was studying airplane manufacturing in 1936 when he noticed a pattern.

Every time total production doubled, labor costs dropped by 10-15%.
Not sometimes. Every single time.

He had no idea he'd just discovered the formula that would predict every tech cost curve for the next century.

Out of 9,000 companies at YC, only 164 (1.82%) went on to become unicorns.

Whereas the hit rate for Thiel Fellowship is 13.79% (40 out of 290 companies became unicorns).

These companies together control >$220 Billion in market cap. The question is how?

Thread First, look at some of the companies founded during Thiel Fellowship:

• Ethereum: $450 Billion
• Figma: $68 Billion
• Ramp: $16 Billion
• OYO Rooms: $9 Billion
• Scale AI: $7 Billion

Tech debt almost killed Notion. With only a few weeks' runway left, the founders left SF, moved to Kyoto, & rewrote the entire app.

No launch. No hype. Just a quiet resurrection. Today, Notion's worth $10B.

Their monk-mode product strategy every founder should study:
Notion was dead in 2015.

The app crashed constantly. Users lost their work. The code was so broken they couldn't ship new features. Investors walked away.

The founders fired everyone to save money. Cash was almost gone.

Then they made a decision that changed everything...

Before AWS existed, one company ran the servers for Twitter, LinkedIn, and Facebook's entire app ecosystem.

They owned Node.js, invented containers 8 years before Docker, and Peter Thiel even backed them.

Then something happened... In 2004, a cancer researcher turned entrepreneur named Jason Hoffman started a cloud company called Joyent.

While Amazon was still figuring out AWS, Joyent was already hosting the internet's hottest startups.

Their client list would make your jaw drop.

Facebook once bought a VPN app for $120M and turned it into a surveillance tool that spied on 33M+ users' entire phones for years.

This app helped Zuck buy WhatsApp for a whopping $19B and break Snapchat's encryption.

Thread The name of this Israeli app was Onavo.

It promised to “secure your data” and reduce mobile data usage.

When Facebook bought it in 2013, Zuck said the app would help them connect more people to the internet.

Facebook even promised to keep Onavo running as a standalone brand.

You know those AI apps that always work? They all use JSON prompts.

Started using the same format for my AI prompts and my outputs became shockingly consistent.

Here's exactly how to write JSON prompts (with code, screenshots, real examples): What you'll learn in this thread:

→ What JSON is & why LLMs love it
→ Basic JSON prompt structure
→ API modes that guarantee valid output
→ Templates for extraction, generation, analysis
→ Production patterns & error handling
→ Real automation examples

Let's start:

AI just made experienced developers 19% SLOWER.

Not faster. SLOWER.

New study: 16 developers, 246 coding tasks, best AI tools available.

Everyone predicted 20-40% productivity boost.

The reality Silicon Valley doesn't want you to see🧵 These weren't random coders. Average 5 years on their specific repositories. 1,500 commits each. Working on codebases with 1.1 million lines of code.

The kind of developers who know their code so well they can navigate it blindfolded.

What happened to them was... unexpected.

BREAKING. An ex-OpenAI engineer published the most revealing insider account of how the $100B AI giant ACTUALLY operates.

They tripled to 3,000 employees in ONE year.
No email. No roadmap. Run entirely on Slack.

Highlights from the most important tech blog of the year: 🧵 The blog is written by Calvin French-Owen, who sold Segment for $3.2B (back in 2011), worked extensively with OpenAI, built Codex, then quit.

He wrote a blog sharing his first-hand experience at OpenAI.

Here's what he revealed:

OpenAI was all set to pay $3B for Windsurf.

So how come Google stole Windsurf for a lesser price ($2.4 billion)?

The actual villain is Microsoft.

How a secret Microsoft clause on OpenAI's IP is helping Google win the AI race: 🧵 Microsoft (OpenAI's biggest investor) has a clause: they get access to any IP that OpenAI acquires.

On the other hand, Microsoft owns GitHub Copilot, which is a direct competitor of Windsurf.

Do you see the problem?

JUST IN: Grok 4 drops after the wildest 48hrs in the history of X:

- Nazi posts
- CEO resigns
- X Worth $11B less than paid

Musk's now claiming Grok 4 is PhD-level in every subject + scores 45% on impossible tests.

Here's everything Musk revealed at the Grok 4 launch🧵⬇️ 1. Grok 4 scored 45% on the "impossible" AI test

Humanity's Last Exam has 2,500 expert questions across 100+ disciplines. GPT-4 gets 3%. Gemini gets 21%.

Grok 4 hit 45% - more than double the previous best. Think PhD-level questions in physics, medicine, philosophy.