Good Morning from San Francisco,
Russia discovered AI's weak spot: LLM grooming - flooding training data with so many false claims that AI models accept them as truth. Their Moscow-based network churned out 3.6 million propaganda articles in 2024, and while hardly any humans read them, the machines did. Now AI chatbots repeat Kremlin narratives 33% of the time when asked about news events. 🤖📚
Speaking of teaching machines: OpenAI conveniently rewrote its careful past as "disproportionate" while racing Chinese rival DeepSeek. They're facing steep losses - $5 billion this year, potentially ballooning to $14 billion by 2026. Meanwhile, their safety team is jumping ship faster than rats from the AI Titanic. 🚢💰
The State Department unleashed AI to scan foreign students' social posts. Your visa might now depend on what you liked after October 7, 2023. Think of it as Instagram stalking, but with deportation powers. 🔍📱
Welcome to 2025, where machines teach machines, and humans just pay the bills - in more ways than one. 🤖💳
Stay curious,
Marcus Schuler
Moscow's Latest Export: AI Brainwashing
A NewsGuard investigation has uncovered Russia's most successful unpopular website. Their Moscow-based "Pravda" network churned out 3.6 million (!) propaganda articles in 2024, and almost nobody read them. That was exactly the plan.
The strategy works like a charm, if your idea of charm is fooling robots. NewsGuard's audit found leading AI chatbots now repeat Russian propaganda 33 percent of the time when asked about news events. It's like having a student who only reads fake news, but this student grades everyone else's papers.
Inside the Robot Whisperer's Playbook
The mastermind? American fugitive John Mark Dougan, who announced the scheme at a Moscow conference with all the subtlety of a Bond villain: "By pushing these Russian narratives, we can actually change worldwide AI." The Pravda network took his advice, targeting 49 countries in dozens of languages. Their English site gets 955 monthly visitors - roughly the audience of a mediocre cat video.
From Moscow With Spam
The American Sunlight Project (ASP) spotted the pattern first. They call it "LLM grooming," which sounds like robot therapy but is actually more sinister. The technique works by flooding AI training data with so many false claims that the AI starts thinking they're true. It's like teaching history by only using tweets.
NewsGuard's analysis shows the network has spread 207 provably false claims, from secret bioweapons labs to tales of Zelensky buying Hitler's vacation home. Because nothing says "credible news" like combining Nazi real estate with Ukrainian politics.
Why this matters:
- Moscow found AI's Achilles' heel: it can't tell the difference between popularity and spam
- The future of propaganda doesn't need human readers - just gullible algorithms with good SEO rankings
Read on, my dear:
OpenAI's Memory Gets Conveniently Fuzzy
OpenAI just performed an impressive feat: forgetting its own history while wide awake. Miles Brundage, their former head of policy research, called out the company for suddenly remembering their careful release of GPT-2 very differently - kind of like how your friend remembers that dinner bill they definitely promised to split.
The company's latest policy document portrays the path to artificial general intelligence (AGI) as a smooth journey of deployment and learning. Gone is their earlier, almost parental anxiety about GPT-2, when they clutched the code like a nervous mom on the first day of kindergarten.
Fact-checking their own flashbacks
Brundage didn't buy this convenient memory lapse. "What's the evidence this caution was 'disproportionate'?" he posted on X, essentially asking OpenAI if they'd checked their own email archives. The timing of this historical revision is about as subtle as a peacock at a penguin party.
Racing to the bank (or from it)
Here's the plot twist: OpenAI is in a sprint with Chinese AI lab DeepSeek, whose R1 model matches their technology. CEO Sam Altman admitted they need to speed up releases - presumably right after looking at the company's bank statement. They're bleeding money faster than a Silicon Valley startup with an infinite snack budget, projecting $14 billion in annual losses by 2026.
Safety third?
Meanwhile, OpenAI dissolved its AGI readiness team last year, and safety researchers are abandoning ship like it's the AI Titanic. Who needs safety experts when you're racing to the bottom line?
Why this matters:
- OpenAI is playing "safety theater" while their accountants have night sweats - turns out principles are expensive
- The AI race has become the tech equivalent of a demolition derby, where checking your mirrors is considered a waste of time
Read on, my dear:
AI Photo of the Day
Prompt:
Minimalistic studio portrait of a 30 year old woman with pale, porcelain-like skin, dark eyes, and a large cockroach helmet that is the carapace of a beetle on her head, against a white background.
Big Brother Gets an AI Upgrade for Visa Hunting
Secretary of State Marco Rubio is unleashing AI on foreign students’ social media accounts, Axios reports. The “Catch and Revoke” program aims to cancel visas of anyone posting content deemed sympathetic to terror groups. Because nothing says “land of the free” like having algorithms scrutinize your Twitter history.
Armed with the Immigration Nationality Act of 1952, officials can now revoke visas from anyone deemed a “threat.” The modern twist? That threat assessment now includes your Instagram likes and TikTok follows.
State Department officials were apparently shocked to discover the Biden administration hadn’t revoked a single visa based on social media posts. Their solution? Build a digital surveillance empire. Officials say AI tools will scan public social media posts for keywords, images, or affiliations that might suggest support for designated terror groups.
The Three-Letter Agency Book Club
The program unites the departments of State, Justice, and Homeland Security in what officials call a “whole of government approach.” They’re checking arrest records, news reports, and even student lawsuits. It’s like a federal agency group chat where everyone’s sharing screenshots of your Facebook posts.
History Repeats, Now with Better Toys
Critics see echoes of Operation Boulder, the 1972 surveillance program targeting pro-Palestinian groups. “With the advent of AI, it’s even scarier,” warns Abed Ayoub of the American-Arab Anti-Discrimination Committee.
Meanwhile, Trump’s recent executive orders have already created a chilling effect on student protests. A State Department official’s response? Ignoring public social media would be “negligent.” After all, someone has to read those midnight rants about geopolitics.
Why This Matters:
- The government just automated the art of being nosy, with real-world consequences for thousands.
- Your social media past is now a matter of national security—hope you didn’t share any spicy memes.
- The good news? At least now you have an excuse for that embarrassing 2012 Facebook post—national security made you do it.
Read on, my dear:
ChatGPT's History Hack: Your Time-Travel Toolkit
1️⃣ Problem-Solving with Dead Geniuses
Let historical giants crack your modern problems.
👉 Example: "How would Leonardo da Vinci, Elon Musk, and Marie Curie tackle climate change?" (Bonus: At least two of them won't tweet about it)
2️⃣ Creative Writing on Steroids
Spark idea explosions with unlikely character mashups.
👉 Example: "Aristotle, Shakespeare, and Rowling argue about plot structure. Grab popcorn."
3️⃣ Business Consulting from Beyond
Get an advisory board from the time machine.
👉 Example: "Steve Jobs designs your sustainable fashion brand. Spoiler: Everything's black turtlenecks."
4️⃣ Political Analysis, Unfiltered
Let history's heavyweights speak freely.
👉 Example: "Churchill, Mandela, and Merkel analyze today's world - no PR teams attached."
5️⃣ Personal Development, Historical Edition
Grab a mentor from the past.
👉 Example: "Marcus Aurelius coaches you through digital burnout. First lesson: The internet is not a Stoic's best friend."
🔥 Power Tips:
- Mix perspectives like a historical DJ
- Use answers as springboards
- Throw it all in your mental blender and create something new 🚀
AI & Tech News
CoreWeave denies Microsoft breakup rumors ahead of IPO
The AI cloud startup CoreWeave swiftly rejected Financial Times' claims of troubled Microsoft contracts, despite the tech giant representing 62% of its $1.2 billion revenue in 2024. While the Microsoft-shaped elephant remains in the room (and apparently silent on the matter), CoreWeave's spokesperson insisted all partnerships are intact, though their IPO filing suggests they're keeping their fingers crossed.
Google founder tackles factory AI, keeps things Page-sized quiet
Larry Page is cooking up an AI manufacturing startup called Dynatomics with the subtlety of a billionaire trying to buy an island, bringing on former Kittyhawk CTO Chris Anderson to help robots design and build things better than humans. While other companies are already playing in the AI manufacturing sandbox, none of them have a Google co-founder's piggy bank to break open.
AI 'wrapper' startups flip from tech pariahs to Silicon Valley darlings
Harvey and Anysphere have transformed from dismissed "GPT wrappers" into Silicon Valley's hottest tickets, with Harvey hitting $50M in revenue and Anysphere sprinting to $100M faster than any software startup in history. The twist? They're building empires by simply making other companies' AI models actually useful, proving that sometimes the best tech strategy is just good packaging.
Zuck bets on AI that talks back, not just types
Meta is supercharging Llama 4's voice capabilities to make AI conversations feel more natural and interruptible, betting that the future of AI assistants lies in speech rather than text. Meanwhile, the company is eyeing premium subscriptions and sponsored AI search results, proving that even artificial intelligence needs to pay the bills.
SBF's jailhouse rebrand hits Tucker Carlson show
From Democratic donor to Republican victim, Sam Bankman-Fried is orchestrating a dramatic image makeover from behind bars, suggesting his 25-year sentence stems from late-game Republican donations rather than the $8 billion customer fraud. The convicted FTX founder's new talking points read like a Google Doc come to life - which, ironically, they are.
Silicon Valley's new vibe: Let AI do the coding
Y Combinator's newest batch of startups are riding the AI wave right into the future, with 25% of them letting artificial intelligence handle 95% of their coding. But here's the plot twist: these aren't tech rookies cutting corners - they're skilled developers who've simply found a better way to build, though they still need to keep their debugging skills sharp for when AI's "vibe coding" hits a wrong note.
AI Decoded 🔓: Perplexity
Your guide to mastering AI tools - without having to become a Silicon Valley hermit
🚀 Perplexity AI: The Brainiac Making Google Sweat
Meet Perplexity AI, the search engine that's basically Google with a PhD. It's becoming the go-to tool for anyone who likes their answers served with a side of actual facts.
Perfect for fact-checking those wild claims your uncle makes on Facebook 🔎, diving deep into research rabbit holes, or proving that no, the moon isn't made of cheese.
The feature set is impressive enough to make other AI tools feel like pocket calculators. Different search modes like Auto, Deep Research, and Pro Search deliver tailored results. It devours academic papers, web content, and social media posts like a grad student during finals week. It even analyzes PDFs and images without breaking a sweat.
The "Discover" tab works like a news butler 📰, serving up personalized content on tech, finance, science, culture, and sports – based on your previous searches. "Spaces" lets you organize your AI chats by topic, similar to ChatGPT's Custom GPTs (but with less existential dread).
"Pages" lets you create AI-generated Wikipedia pages 📄 that you can edit and share. The mobile app shows off with camera-based search and voice recognition 📱, perfect for when typing feels too 2023.
As a cherry on top, Perplexity integrates image generation through Flux One, DALL·E 3, and Playground V3 🎨. It even connects with Crunchbase AI to predict startup funding trends with 95% accuracy 📈 (though it's still working on predicting cryptocurrency crashes).
Strengths: ✅
- Lightning-fast fact-checking with multiple sources
- Seamlessly blends academic and web sources like a digital smoothie
- Personalized info streams that know you better than your therapist
- Mobile features that make your smartphone actually feel smart
- Business predictions accurate enough to make fortune tellers nervous
Weaknesses: ⚠️
- Occasionally hallucinates on complex queries like an AI after too much caffeine
- Keeps the good stuff behind a paywall • Search accuracy can be spotty for niche topics
- Privacy concerns persist despite incognito mode
- Non-English language support is still playing catch-up
Pricing: 💰
- Free: Basic access (aka the demo version)
- Pro: $20/month for unlimited feature access
- Teams: $40/month per user with collaboration tools
- Enterprise: "If you have to ask, you can't afford it" pricing
