AI models typically learn by memorizing patterns, then researchers bolt on reasoning as an afterthought. A new method called Reinforcement Pre-Training flips this approach—teaching models to think during basic training instead.
Good Morning from San Francisco, Meta's AI app turned users into accidental exhibitionists. 😬 People shared medical questions, legal
Meta users think they're chatting privately with AI. Instead, they're broadcasting medical questions, legal troubles, and relationship problems to the world through a public feed that many don't realize exists.
Good Morning from San Francisco,
Meta burned $10 billion chasing AI talent after losing its entire research team. Zuckerberg spent months claiming victory while 11 of 14 Llama researchers jumped ship to competitors. Now he's personally recruiting with nine-figure packages and rearranging office furniture to sit near new hires.
The Scale AI investment isn't expansion—it's expensive damage control. Meta's Llama 4 flopped internally. Legal troubles over pirated training data pile up. Meanwhile, the delayed "Behemoth" model won't arrive until late 2025.
Desperation has a price tag. Meta just set it at $10 billion.
Stay curious,
Marcus Schuler
Meta spent months claiming its AI was "crushing it." The truth was different. Now Mark Zuckerberg is investing over $10 billion in Scale AI and recruiting its CEO for a new "superintelligence" lab.
This isn't growth. It's damage control.
Meta lost 11 of the 14 researchers who created its Llama model. Most jumped to competitors like Mistral AI. The company's flagship Llama 4 disappointed everyone, including Zuckerberg himself. Meta delayed its most ambitious project, the "Behemoth" model, until late 2025.
Zuckerberg entered what insiders call "founder mode." He personally recruits 50 researchers with nine-figure pay packages. He rearranged office desks so new hires sit near him. He created a WhatsApp group called "Recruiting Party" for round-the-clock talent hunting.
The Scale AI deal shows Meta's real problem. The company can't build what it needs internally. Scale specializes in cleaning and processing data for AI training. Meta is paying $10 billion for services it should have developed itself.
Meta also faces a copyright lawsuit over using pirated books to train models. A judge questioned whether this qualifies as "fair use" when it might "obliterate the market" for authors' work. If Meta loses, it needs new data sources fast.
The AI race has reached unprecedented spending levels. Microsoft invested $13 billion in OpenAI. Amazon put $8 billion into Anthropic. But those companies build foundation models. Meta is paying $10 billion mainly for data services and talent.
Training costs grow by two to three times each year. The largest AI projects will cost over $1 billion by 2027. Meta tried organizing a funding consortium last year. Competitors declined to help.
Zuckerberg promises "hundreds of billions" in future AI spending. He tells recruits that Meta's advertising revenue can fund a multi-gigawatt data center. The message is clear: we have cash and we're desperate enough to spend it.
The new lab sits awkwardly with Meta's existing AI teams. It's unclear how this affects Yann LeCun, Meta's chief AI scientist and Turing Award winner. LeCun has long questioned the approaches other AI labs use.
Why this matters:
Read on, my dear:
Prompt:
a surreal hyperrealistic portrait of a woman with cracked marble skin like an ancient Greek statue, golden kintsugi veins running through the fractures, a real three-dimensional white bird with patterned wings flying out of her right eye, not a tattoo, soft golden tear dripping from her left eye, dramatic cinematic lighting, ultra detailed textures on skin, cracks, and feathers, ethereal atmosphere, symbolic of pain and liberation, fine art photography, close-up framing
OpenAI hit $10 billion in annual recurring revenue in June, nearly doubling from $5.5 billion at the end of 2024. The company reached this milestone in just six months, less than three years after launching ChatGPT.
The numbers tell a story of explosive growth built on massive spending. OpenAI lost $5 billion last year while chasing revenue targets. The company won't turn a profit until 2029.
Revenue comes from ChatGPT subscriptions, business products, and API sales. OpenAI now serves 500 million weekly active users and counts 3 million paying business customers.
OpenAI's $300 billion valuation puts it at roughly 30 times revenue. That multiple reflects investor faith in the company's ability to reach $125 billion in annual revenue by 2029. Getting there requires growth rates that border on the absurd.
The company raised $40 billion in March from SoftBank, Microsoft, and other investors. OpenAI has now raised nearly $58 billion total. Recent acquisitions show where the cash goes: $3 billion for coding tool Windsurf and $6.4 billion for Jony Ive's hardware startup.
Other AI companies follow the same pattern. Anthropic tripled its revenue to $3 billion this year. Cursor jumped from under $100 million to $500 million in 2024. None make money yet.
Signs of slowdown already appear. Business AI adoption stalled in May for the first time in 10 months. About 40 percent of US businesses now pay for AI tools, suggesting easy customers have signed up.
Why this matters:
Read on, my dear:
Google's AI tools are crushing news publishers by answering questions directly instead of sending users to websites. HuffPost and Washington Post lost half their search traffic in three years, while Business Insider cut 21% of staff after a 55% traffic drop.
Waymo yanked its self-driving cars from downtown Los Angeles after protesters torched at least five of its vehicles during weekend immigration raids demonstrations. The company says it wasn't specifically targeted but happened to be in the wrong place when things got heated.
SEC Chair Paul Atkins wants to create temporary exemptions that let crypto firms skip certain regulations while building new products. The former crypto lobbyist says this could help America become the "crypto capital of the planet" under Trump's vision.
Huawei's CEO says his company's chips trail American ones by a generation but they're using cluster computing and software tricks to bridge the gap. Ren Zhengfei told state media that Huawei spends $25 billion annually on research and doesn't need to worry about chip problems despite US export controls.
Roblox named Paramount's Naveen Chopra as its new CFO after longtime finance chief Michael Guthrie announced plans to step down. Chopra brings experience from Amazon's devices business and helped guide the gaming platform through its 2021 public debut and 89% revenue growth to $3.6 billion.
X's new encrypted messaging system keeps user private keys on company servers, letting the platform decrypt any message. The company uses a "Juicebox" system that splits keys across three servers, but security experts say this offers little protection since X controls all the servers and doesn't use hardware security modules.
Amazon will spend at least $20 billion building data centers in Pennsylvania, creating 1,250 jobs as tech companies pour money into AI infrastructure. The investment follows Amazon's recent $10 billion commitment to North Carolina and $5 billion for Taiwan cloud facilities.
Britain's media regulator opened investigations into 4chan and seven file-sharing services for possible breaches of online safety laws. Ofcom received complaints about illegal content on 4chan and potential child abuse material on the file-sharing platforms, which could face fines up to $24 million or 10% of global revenue.
IBM says it will build the world's first large-scale, fault-tolerant quantum computer at its New York data center before 2030. The machine, called Quantum Starling, would have 20,000 times more computational power than today's quantum computers by using new error-correction techniques to solve the technology's biggest problem.
El Capitan at Lawrence Livermore National Laboratory remains the world's fastest supercomputer with 1.742 exaflops of processing power. Europe made strong gains in the latest TOP500 rankings, capturing 11 of the top 20 spots including Germany's new JUPITER Booster system at number four.
Tencent Music will buy Chinese podcasting startup Ximalaya for $1.3 billion in cash and stock to build China's answer to Spotify. The deal combines Tencent's music apps like QQ Music with Ximalaya's 303 million monthly users, creating a dominant audio platform in the world's second-largest economy.
Video game actors and major publishers reached a tentative deal to end an 11-month strike over AI protections. SAG-AFTRA says the agreement includes guardrails for AI use and consent requirements, though members must still ratify the contract before returning to work.
Apple used WWDC 2025 to announce new AI features while most promises from iOS 18 remain broken. The company revealed a Foundation Models Framework for developers and integrated ChatGPT into Xcode. But these additions can't hide Apple's poor track record with AI delivery.
Siri still can't see what's on your screen reliably, despite being promised over a year ago. Apple Intelligence writing tools produce mediocre results compared to ChatGPT. Photo searches miss obvious objects that Google identifies correctly 85% of the time.
The new Foundation Models Framework gives developers access to Apple's on-device AI models. But these models only work on Apple Silicon chips, excluding millions of older devices. The framework also delivers worse results than cloud-based alternatives because phone processors can't match data center computing power.
Apple's decision to integrate ChatGPT into Xcode reveals gaps in its own capabilities. The company demonstrated Swift Assist coding tools last year but never released them. Partnering with OpenAI admits Apple can't build competitive AI features internally.
Live Translation and expanded Visual Intelligence sound impressive in demos. Real-world performance will likely disappoint, based on Apple's pattern of overpromising and underdelivering. The company delayed major Siri improvements by eight months after iOS 18's launch.
Apple chose privacy over performance with on-device processing. This protects user data but creates AI that works worse than competitors. Google, Microsoft, and Meta ship functional AI while Apple perfects marketing demos.
Why this matters:
Read on, my dear:
Exa builds semantic search infrastructure that actually understands what you're asking for. The San Francisco startup ditches keyword matching for transformer-based retrieval that serves AI agents, not humans clicking blue links.
The Founders 🧠 Will Bryk and Jeff Wang launched in 2021 after meeting at Harvard. Seven employees now, up from the original duo who burned through a million bucks on GPUs before most people knew what neural search meant. Based in SF because where else would you reinvent Google?
The Product ⚡ • Web Search API that converts queries into mathematical embeddings • Custom-trained link prediction model (think PageRank meets GPT) • Handles complex, multi-faceted searches that break traditional engines • Websets tool compiles research into structured datasets • SOC 2 certified with zero data retention by default • 144 NVIDIA H200 GPUs powering their own index
The Competition 🥊 Perplexity chases consumers with chatty answers. You.com builds another Google clone. Exa targets developers who need perfect recall over pretty interfaces. Google owns 90% of search but can't risk breaking their ad machine. Microsoft's Bing integration feels like duct tape on a spaceship. Exa owns the hard queries nobody else solves.
Financing 💰 $22M total: $5M seed plus $17M Series A led by Lightspeed. NVIDIA's venture arm joined because Exa devours their chips. Valuation whispers hover around $100M. Y Combinator doubled down through their Continuity fund.
The Future ⭐⭐⭐⭐ AI agents need smart search infrastructure. Exa could become invisible plumbing that powers every AI copilot. Risk: giants might build competing APIs overnight. But Google's tied to ads, Microsoft's tied to Bing's legacy, and Exa's got a head start indexing the web for machines. 🤖
Perfect search beats fast search when robots don't mind waiting.
Good Morning from San Francisco, Meta's AI app turned users into accidental exhibitionists. 😬 People shared medical questions, legal
Good Morning from San Francisco, Meta just dropped $14.8 billion for half of Scale AI 💸. That's Silicon
Good Morning from San Francisco, ☀️ OpenAI cut o3 model prices by 80%. Input tokens dropped from $10 to $2 per
Good Morning from San Francisco, Apple researchers just shattered AI's reasoning hype. Those fancy "thinking" models
Get tomorrow's intel today. Join the clever kids' club. Free forever.
