Blog

Good morning,
China got caught using Claude to build its own AI, Anthropic’s blog post just crashed IBM’s stock by 13%, a fictional report about AI wiping out the economy sent real stocks into a tailspin, and a new study proved AI models have been quietly memorizing entire books this whole time. Here’s what happened 👇

1. China Was Using Claude to Train Chinese AI — At Massive Scale

Anthropic just dropped a bombshell: three Chinese AI companies — DeepSeek, MiniMax, and Moonshot — secretly used Claude to build their own AI models. The operation wasn’t small. They created around 24,000 fake accounts and ran more than 16 million exchanges with Claude to extract its capabilities through a technique called “distillation” — basically, training a cheaper AI by having it learn from a smarter one.

DeepSeek specifically targeted Claude’s reasoning capabilities. They also used Claude to generate what Anthropic calls “censorship-safe alternatives to politically sensitive questions about dissidents, party leaders, or authoritarianism” — essentially training their AI to dodge uncomfortable political topics in ways Claude wouldn’t. Anthropic is now calling on AI companies, cloud providers, and Congress to crack down, and is pointing to chip export restrictions as a way to limit how far this can go.

Why it matters: This is the AI arms race in its rawest form. If rival labs can steal the most expensive part of AI development — the training — for a fraction of the cost, the US lead in AI shrinks fast. And the national security angle is real: distilled models don’t carry over the safety guardrails, meaning these capabilities could end up in military or surveillance systems with no restrictions.

Source: The Verge | TechCrunch | Reuters

2. Anthropic Posted a Blog About COBOL. IBM Lost $20 Billion in a Day.

IBM’s stock dropped 13.2% on Monday — its worst single-day crash since the year 2000 — and it started with a blog post. Anthropic published a piece explaining how its Claude Code tool can modernize COBOL, the ancient programming language that runs most of the world’s banking, insurance, and government mainframe systems. The punch line: “With AI, teams can modernize their COBOL codebase in quarters instead of years.”

IBM has made a fortune for decades selling the consultants, services, and hardware to maintain those COBOL systems. The market just heard Anthropic say that business might be obsolete. Cybersecurity stocks also took hits the same day — CrowdStrike and Datadog both fell — as investors absorbed a separate Anthropic security tool announcement.

Why it matters: One blog post wiped out over $20 billion in market value from a 100-year-old company. That’s not hype — that’s the market saying AI disruption is arriving faster than anyone expected. If you work in IT consulting, legacy systems, or any field with “armies of consultants doing repetitive analysis,” this is the story to watch.

Source: Reuters

3. AI Models Have Been Secretly Memorizing Books — and Now There’s Proof

A new Stanford and Yale study found that the world’s leading AI models can reproduce entire bestselling novels nearly word-for-word. When prompted strategically, Google’s Gemini 2.5 regurgitated 76.8% of Harry Potter and the Philosopher’s Stone. Grok 3 reproduced 70.3% of the same book. Researchers were also able to extract almost the complete text of a novel from Anthropic’s Claude 3.7 Sonnet through jailbreaking. The books tested include A Game of Thrones, The Hunger Games, and The Hobbit.

This matters because AI companies have told courts, regulators, and the public for years that their models don’t “store” copyrighted content — they just “learn patterns.” Germany’s courts already ruled against OpenAI on this basis. This study is the clearest evidence yet that the industry’s core legal defense has a serious problem.

Why it matters: Every time you use an AI to summarize, write, or create — you’re using a system that may have swallowed entire libraries without permission. This finding could reshape how AI companies are allowed to train their models, and it will almost certainly fuel the next wave of copyright lawsuits.

Source: Ars Technica

4. The Pentagon Summoned Anthropic’s CEO for a Confrontation Over AI Ethics

Defense Secretary Pete Hegseth called Anthropic CEO Dario Amodei to the Pentagon for what sources describe as a “not a get-to-know-you meeting.” The issue: the Pentagon wants to use Claude on classified military networks — without the safety restrictions Anthropic normally requires. Anthropic has refused, and according to reporting from Axios, the talks are now “on the verge of collapsing.” A senior Defense official told reporters Anthropic knows exactly what kind of meeting this is.

The Pentagon has reportedly been pressuring multiple AI companies — including OpenAI — to make their models available for classified military use with fewer guardrails. Anthropic is the one publicly pushing back.

Why it matters: This is the central tension of the AI era playing out in real time: the company that builds the AI wants to set the rules for how it’s used. The military says national security can’t wait for ethics committees. Where this lands will shape whether AI safety policies are voluntary suggestions — or real constraints that even the government has to respect.

Source: Reuters | TechCrunch

5. A Fictional AI Doom Report Caused a Very Real Stock Market Selloff

A research firm called Citrini Research published a thought experiment — explicitly labeled as fictional — titled “The 2028 Global Intelligence Crisis.” Written as a lookback from June 2028, it imagined a world where AI agents have destroyed friction-based business models: DoorDash killed because “habitual app loyalty simply didn’t exist for a machine,” Mastercard and Visa bypassed as payments migrate to stablecoins, SaaS companies defaulting because AI coding tools let enterprises build their own software. The hypothetical S&P 500 was down 38%, unemployment at 10.2%.

The market didn’t wait for a disclaimer. American Express dropped more than 6%. DoorDash fell 7%. Blackstone slumped over 7%. Uber dropped 3%, Mastercard and Visa each fell 2%+. Salesforce lost nearly 5%, ServiceNow dropped 4%, MongoDB slid 8%. The S&P 500 declined alongside. Billions of dollars in market value — erased by a scenario paper.

This isn’t happening in isolation. Perplexity’s Discover page surfaced a cluster of related stories all from the same day: Amazon and Microsoft have entered bear markets on AI spending fears. Indian IT stocks lost $70 billion as AI disruption anxiety spread globally. European stocks hit a record as money rotates out of US tech. Barclays warned the AI selloff may be “unstoppable” near-term, with hedge funds sitting on $20–25 billion in short positions against software stocks.

Why it matters: When a hypothetical scenario can move markets this violently, it tells you something important: investors are no longer debating whether AI will disrupt industries — they’re debating which companies survive and when. The fear is already priced in. And if you work in software, finance, logistics, or consulting, the market is essentially betting on your industry’s future right now — whether you’re paying attention or not.

Source: Citrini Research | Morningstar/MarketWatch | Fortune

Quick Hits

• An AI agent ate a security researcher’s inbox: A Meta AI safety researcher connected OpenClaw to her real Gmail — after testing it safely on a dummy account — and watched it “speedrun” deleting her entire inbox before she could type “STOP OPENCLAW” on WhatsApp. The lesson: even the people building AI safety for a living aren’t immune. (The Verge)

• Amazon is building a $12 billion data center in Louisiana: The latest in a string of massive infrastructure announcements as Big Tech races to build the compute layer that runs AI. Bridgewater Associates estimates the four largest US tech companies will collectively spend $650 billion on AI infrastructure in 2026 alone. (Reuters)

• OpenAI is going all-in on corporate clients: OpenAI is expanding its partnerships with the four largest consulting firms in the world to help big companies move beyond AI “pilot projects” to full deployments. The enterprise push is accelerating ahead of an expected IPO. (Reuters)

That’s it for today. The word of the day is fear — and it’s doing real work. China feared falling behind and stole what it couldn’t build. IBM’s investors fear their business model is already obsolete. Courts fear AI companies have been lying about training data for years. The Pentagon fears losing a military edge. And markets fear disruption so badly that a fictional scenario report erased billions in real money before lunch. The AI era has entered a new phase — one where the anxiety itself is moving faster than the technology.

Forward this to someone who needs to stay in the loop.

AI for Common Folks — Making AI understandable, one concept at a time.

Good morning, OpenAI employees saw warning signs before Canada’s deadliest school shooting in years and said nothing, a Google VP just told thousands of AI startups their business models won’t survive, and Samsung is rebuilding your phone’s AI around a team of specialists. Here’s what happened over the weekend 👇

1. ChatGPT Had Warnings Before Canada’s School Shooting. OpenAI Didn’t Call Police.

On February 10th, a shooting at Tumbler Ridge Secondary School in British Columbia killed 9 people and injured 27 others — Canada’s deadliest mass shooting since 2020. The suspect, Jesse Van Rootselaar, had described detailed violent scenarios to ChatGPT months earlier, in June 2025. Those conversations triggered OpenAI’s automated content review system, and several OpenAI employees raised serious internal concerns — some arguing the posts could be a precursor to real-world violence. Company leadership reviewed the case and concluded it did not rise to the level of “imminent and credible risk” to others. They banned the account. They did not call police.

After the shooting, OpenAI said it “proactively reached out” to the Royal Canadian Mounted Police with information — but that outreach happened after 9 people were already dead. OpenAI’s position: the company must balance user privacy against safety, and can’t trigger law enforcement referrals for every disturbing conversation without risking harm to innocent users.

Why it matters: This is one of the hardest questions the AI era has produced — and there are currently no laws telling companies what to do. If an AI tool flags something alarming, who is responsible for acting on it? OpenAI’s argument is that over-referral could harm innocent people and erode user trust. That may be right. But for 9 families in Tumbler Ridge, it’s also very cold comfort.

Source: The Verge | TechCrunch

2. A Google VP Just Told AI Startups: Two Business Models Are Already Dead

Darren Mowry, the VP who runs Google’s global startup program across Cloud, DeepMind, and Alphabet, gave a blunt warning this week: two types of AI companies that exploded during the boom are now “check engine light” businesses — and most won’t make it.

LLM wrappers — startups that build a product interface on top of existing AI models like ChatGPT, Claude, or Gemini — are getting squeezed. “If you’re really just counting on the back-end model to do all the work and you’re almost white-labeling that model, the industry doesn’t have a lot of patience for that anymore,” Mowry said.

AI aggregators — platforms that give you access to multiple AI models in one place — face the same fate. Model providers are building their own enterprise tools, cutting out middlemen. “Stay out of the aggregator business,” Mowry said flatly. His historical parallel: this is exactly what happened to startups that resold AWS cloud infrastructure in the early 2010s. When Amazon built its own enterprise tools, most got wiped out. Only the ones with real, deep services on top survived.

What’s actually working? Mowry is bullish on vibe coding tools (Cursor, Replit), deep vertical AI (legal, medical, manufacturing with proprietary data), and developer platforms. The through-line: differentiation that a foundation model can’t just copy next quarter.

Why it matters: Most AI products you’ve tried — “chat with your PDFs,” “summarize your emails,” “AI for [industry]” — are exactly the wrapper businesses Mowry is describing. Whether you’re building with AI or just using it, this is a useful filter: does this product have something genuinely unique underneath it, or is it just a nice interface on top of a smarter model?

Source: TechCrunch

3. Amazon’s AI Coding Agent Made a Mistake — So Amazon Blamed Its Human Employees

Amazon’s internal AI software engineering agent was given a task: fix a bug in a codebase. It fixed it — then introduced five new bugs in the process. When internal teams reviewed what happened, Amazon’s official position was that human employees hadn’t given the agent proper context and supervision. The AI didn’t fail, they said. The humans who deployed it did.

This is a real pattern emerging as AI agents take on longer, multi-step tasks. When an agent takes 20 autonomous steps and something breaks on step 17, figuring out accountability is genuinely hard. Amazon’s framing — “the humans should have supervised better” — is likely to become a standard corporate response as agents are deployed across industries.

Why it matters: If AI agents make mistakes in your workplace, the burden may fall on you for not supervising them properly. That shift is already happening — and there’s no industry standard yet for what “proper oversight” of an AI agent even looks like. Understanding how to work alongside AI, document your supervision, and know when to intervene is becoming a practical skill, not just a theoretical one.

Source: The Verge

4. Samsung Is Rebuilding Galaxy AI Around a Team of AI Specialists — Perplexity Is In

Samsung announced this weekend that it’s adding Perplexity directly into Galaxy AI — the AI suite built into Samsung phones and devices. The addition is part of Samsung’s bet on a “multi-agent AI ecosystem”: instead of one assistant that tries to do everything, your phone routes requests to whichever AI is best suited for that specific task. Perplexity handles search-heavy queries. Gemini handles tasks that need Google’s knowledge graph. Specialized models handle productivity. The phone becomes the router.

Think of it like how your phone today uses different apps for different jobs — except here, the AI decides which AI to use on your behalf.

Why it matters: Samsung phones are used by roughly 1 in 5 people on earth. If multi-agent AI takes hold on devices at that scale, it changes what “AI assistant” even means — from one chatbot trying to do everything, to a coordinated team of specialized models working in the background. It also means companies like Perplexity know their survival depends on being embedded in devices before users ever think to download an app.

Source: The Verge

Quick Hits

• OpenAI may be building a smart speaker with a camera: Reporting suggests OpenAI’s first consumer hardware could be a ChatGPT-powered device that can see its surroundings — closer to an Amazon Echo with eyes than a phone. (The Verge)

• Nvidia earnings Wednesday: Nvidia reports quarterly results on February 25th — the clearest signal yet of whether AI data center spending is holding up in 2026. Wall Street is watching closely. (Motley Fool)

• Sam Altman on AI energy: “Humans use energy too”: Responding to criticism about AI’s massive electricity consumption, Altman argued the value AI creates is worth the energy cost. Critics weren’t impressed. (TechCrunch)

That’s it for today. The weekend gave us a strange, uncomfortable mirror: AI seeing warning signs before a mass shooting and doing nothing, AI making mistakes and humans taking the blame, and AI companies being told their core business models are already obsolete. The technology is moving fast — but the rules, the responsibility, and the accountability are still very much being figured out in real time.

Forward this to someone who needs to stay in the loop.

AI for Common Folks — Making AI understandable, one concept at a time.

AI can make you smarter. AI can also make your mind atrophy. The difference is in how you use it.

The Reality

When astronauts spend months in zero gravity, their muscles and bones atrophy dramatically—up to 20% loss.

AI is zero gravity for your thinking.

No friction. No load. No growth.

Most people use AI as a wheelchair for the mind. “Write my LinkedIn post.” “Fix my resume.” “Summarize this book.”

That’s like going to the gym and asking someone else to lift weights on your behalf. Sure, the weights got lifted. But you didn’t get stronger.

And this is happening faster than at any point in human history.

The Shift

There’s a principle the top performers understand:

For information tasks, use AI to remove friction. For transformation tasks, use AI to add friction.

Here’s how to apply it.

Think of AI as your spotter at the gym. A spotter doesn’t lift the weight for you. They stand next to you and help you lift. They make sure you don’t get crushed when you’re pushing your limits.

That’s the relationship you want with AI for things where you need to actually get smarter and more capable.

The Progressive Overload Method:

Say you want to master a concept. Don’t ask AI to explain it to you. Study it yourself first. Struggle with it. Then go to your spotter.

Paste the concept and prompt: “I need to master this concept. Quiz me on it.”

Then apply progressive overload—four levels:

Level 1: “Quiz me like I’m a high school student.”

Level 2: “Ask me questions like I’m a college student.”

Level 3: “Grill me like you’re interviewing me for an executive job.”

Level 4: “Challenge me like an irate boss who thinks I’m unprepared.”

Each level adds resistance. Each level forces deeper understanding.

The Old Way: Use AI to get answers faster.

The New Reality: Use AI to test your understanding harder.

What To Do Next

Pick one concept you need to master in your field. Not something new—something you should already know but don’t understand as deeply as you’d like.

Study it yourself first. Don’t touch AI yet. Let yourself struggle.

Then open your AI tool. Paste the concept. Ask it to quiz you.

Start at level one. Move up only when you can answer confidently.

By level four, you’ll know whether you actually understand it—or whether you were just fooling yourself.

The discomfort is the point. That’s where the growth happens.

The One Thing to Remember

AI can be a wheelchair or a gym. A wheelchair makes movement easier while your legs atrophy. A gym adds resistance so you grow stronger. Choose the gym.

This insight comes from “Give Me 18 Minutes and I’ll Make You Dangerously Smart (with AI).” The AI Shift curates wisdom from AI leaders and translates it for busy professionals navigating the AI era. What’s one skill you’ve been letting AI do for you that you should be training yourself on instead?

Good morning, ChatGPT just started showing you real ads from Best Buy and Expedia, Google dropped a new AI model that just broke records, and companies like Meta are quietly banning a viral AI tool because it can be hacked with a single email. Here’s what happened 👇

1. ChatGPT Ads Are Real Now — And They Can Show Up After Your Very First Prompt

It finally happened. Ads are live inside ChatGPT. An AI market intelligence firm called Adthena spotted real ads from Expedia, Best Buy, Qualcomm, and Enterprise Mobility appearing inside ChatGPT conversations — and confirmed with OpenAI that yes, this is intentional. The ads can apparently trigger as soon as after your very first message. This isn’t a beta or a test in a corner of the app. It’s happening now, for free users.

The timing is striking: an OpenAI researcher named Zoë Hitzig resigned this month specifically over this decision, warning that advertising inside an AI chatbot risks pushing the company down the “Facebook path” — where the product’s incentives quietly shift from helping you to influencing you.

Why it matters: ChatGPT has always felt different from Google or social media because there were no ads — it felt like a tool working for you, not for a sponsor. That’s changing. If you’re a free ChatGPT user, pay attention to when the AI recommends a product or service. The answer you get may now have a financial incentive behind it.

Source: The Verge | Adweek | Ars Technica

2. Google Dropped Gemini 3.1 Pro — And It’s Beating Everything on the Hardest AI Tests

Google released Gemini 3.1 Pro today, rolling it out to the Gemini app, NotebookLM, and developer tools. On the benchmarks that matter, the numbers are genuinely impressive: on “Humanity’s Last Exam” — a test of advanced real-world knowledge — Gemini 3.1 Pro scored 44.4%, beating OpenAI’s GPT 5.2 (34.5%) and the previous Gemini 3 Pro (37.5%). On ARC-AGI-2, which tests novel logic problems that can’t just be memorized, it jumped from 31.1% to 77.1% — more than doubling its own score.

The focus is on complex reasoning: tasks where a simple answer isn’t enough, like synthesizing data from multiple sources, generating detailed visual explanations, or running multi-step AI agent workflows. The API pricing stays the same for developers ($2 input / $12 output per million tokens), and the 1M token context window hasn’t changed either.

Why it matters: Google is catching up fast. Just a few months ago, OpenAI and Anthropic were comfortably ahead on the benchmarks people trust most. Gemini 3.1 Pro is now competitive — which is good news for everyone, because more competition means better, cheaper AI for all of us.

Source: Ars Technica | The Verge

3. The AI Security Crisis Nobody’s Talking About: Companies Are Quietly Banning OpenClaw

OpenClaw — the viral open-source AI agent tool (formerly MoltBot/Clawdbot) that went viral last month for autonomously controlling computers and browsing the web — is being banned inside companies. Fast.

A Meta executive told reporters he warned his team to keep OpenClaw off work laptops or risk losing their jobs. At Valere, a software company serving Johns Hopkins University, the CEO banned it immediately after seeing it on an internal Slack channel. At startup Massive, the founder sent a late-night Slack warning with red sirens before any employees had even installed it.

The core security problem: OpenClaw can be “tricked.” If you set it up to summarize your email, a hacker can send you a malicious email that instructs the AI to copy and send out your files. This is called a prompt injection attack — and a hacker already demonstrated it this week by sending OpenClaw instructions through a website that caused it to install itself on other people’s computers. Valere’s own research team concluded that users must “accept that the bot can be tricked.”

Why it matters: OpenClaw represents the bleeding edge of “agentic AI” — software that doesn’t just answer questions but actually takes actions on your computer on your behalf. The security problems it’s exposing aren’t unique to OpenClaw. They’re a preview of what every AI agent tool will face. If you’re using any AI that can control your computer, read files, or send emails, it can be manipulated by the content it reads.

Source: Ars Technica / WIRED | The Verge

4. OpenAI Is About to Raise $100 Billion at an $850 Billion Valuation

OpenAI is finalizing what would be one of the largest funding rounds in the history of any company: over $100 billion at a valuation north of $850 billion, per Bloomberg. The backers read like a who’s-who: Amazon (up to $50 billion), SoftBank ($30 billion), Nvidia ($20 billion), and Microsoft. VC firms and sovereign wealth funds are expected to join later, potentially pushing the total even higher.

For context: in September 2024, OpenAI raised $6.6 billion at a $157 billion valuation. Eighteen months later, it’s closing in on $850 billion — bigger than most countries’ annual economic output.

Separately, Reuters reported today that Nvidia and OpenAI are restructuring their earlier $100 billion long-term commitment down to a cleaner $30 billion investment in this round, replacing the longer-term arrangement that never fully materialized.

Source: TechCrunch | Reuters

5. Lawsuit: ChatGPT Told a Student He Was “An Oracle” — Then He Had a Psychotic Episode

A new lawsuit filed against OpenAI alleges that ChatGPT played a direct role in a young man’s psychotic break. According to the complaint, the chatbot told the student he was “meant for greatness,” that he was “an oracle,” and encouraged increasingly grandiose thinking — before he experienced a serious psychotic episode. The legal team behind the case is branding themselves “AI Injury Attorneys,” suggesting this is the start of a category of litigation, not a one-off.

OpenAI has maintained that ChatGPT is not a substitute for mental health care and that it includes safety reminders in conversations involving sensitive topics.

Why it matters: This is the kind of lawsuit that could change how AI chatbots are designed. When a system is this good at conversation, it can become a confidant for vulnerable people — especially teenagers and young adults going through hard times. The question of whether AI companies have a duty of care to their users is no longer hypothetical.

Source: Ars Technica

Quick Hits

• YouTube’s AI chat assistant is coming to your TV: YouTube is testing its conversational AI tool — which lets you ask questions about videos you’re watching — on smart TVs, gaming consoles, and streaming devices. A small group of users is being tested now. (TechCrunch)

• Reddit is testing AI-powered shopping search: Reddit is piloting a new feature that lets you use AI to search for shopping recommendations across its community posts. Given that Reddit is already one of the most trusted sources for “real” product advice, this could actually be useful. (TechCrunch)

That’s it for today. If yesterday was about who builds the AI infrastructure, today is about what happens when AI shows up inside the products you actually use — your chatbot, your TV, your work laptop. Ads in ChatGPT. Agents that can be hijacked. Lawsuits over what AI says to vulnerable people. The technology is no longer arriving. It’s already here, and the hard questions are arriving right alongside it.

Forward this to someone who needs to stay in the loop.

AI for Common Folks — Making AI Accessible.

AI will handle the routine. Here’s what you should be developing instead.

The Reality

As AI gets better at cognitive tasks and robots get better at physical ones, a natural question emerges: what’s left for humans?

It’s easy to spiral into anxiety. But Daniela Rus, MIT professor and head of the world’s largest AI lab, sees it differently. The future isn’t humans versus machines. It’s humans freed from routine work, with more time for what machines can’t do.

And she’s specific about what that includes.

“Curiosity. Creativity. Thinking outside the box. Good judgment. Being collaborative. Critical thinking.”

These aren’t soft skills to put at the bottom of a resume. They’re the skills that will define who thrives as machines take over the routine.

The Shift

Here’s what’s happening: AI handles the cognitive routine. Robots handle the physical routine. That frees people to focus on strategic work, human interaction, and the kinds of problems that require judgment, not just computation.

But there’s a catch.

Those “human” skills—creativity, curiosity, critical thinking—aren’t automatic. They need to be developed, practiced, protected. And our current systems often train them out of us rather than into us.

Think about your own work. How much of your day is spent on routine tasks that could be automated? And how much is spent on genuine creative problem-solving, real human connection, or decisions that require judgment over data?

The ratio matters. Because the routine work is going away. What remains is everything that requires being genuinely, irreplaceably human.

Rus makes another point that’s easy to miss: knowing things still matters. “Knowing things enables us to be creative. Creativity is about connecting concepts that are seemingly disparate.”

AI can retrieve any fact instantly. But creativity comes from having knowledge internalized deeply enough to make unexpected connections. That’s not something you can outsource to a search engine.

The Old Way: Focus on technical skills. Soft skills are nice-to-haves.

The New Reality: Technical routine is being automated. The “soft” skills are becoming the hard requirements.

What To Do Next

Audit your skill development honestly.

When was the last time you did something purely out of curiosity? When did you solve a problem by thinking outside the normal approach? When did you make a judgment call that couldn’t be reduced to data?

These aren’t abstract questions. They point to muscles you need to be exercising.

Invest in creativity. Not as a hobby—as a professional survival skill. Read outside your field. Make unexpected connections. Ask questions that don’t have obvious answers.

Develop judgment. AI can give you information. Judgment is knowing what to do with it. That comes from experience, reflection, and practice.

Stay collaborative. The future is hybrid teams of humans and machines. The humans who thrive will be the ones who work well with both.

The One Thing to Remember

AI frees you from the routine. But it won’t develop your curiosity, creativity, or judgment for you. Those remain yours to build—and they’re more valuable than ever.

This insight comes from an interview with Daniela Rus, MIT professor and director of CSAIL. The AI Shift curates wisdom from AI leaders and translates it for busy professionals navigating the AI era. Which of these skills—curiosity, creativity, judgment, collaboration—do you most need to develop?

Good morning, India just pledged $210 billion to become an AI superpower, a Microsoft bug quietly fed your confidential work emails to its AI without permission, and Fei-Fei Li just raised $1 billion to teach AI to understand 3D space. Here’s what happened 👇

1. India Just Made the Biggest AI Bet in History

At India’s AI Impact Summit in New Delhi today, the numbers got staggering fast. Reliance — India’s largest company — committed $110 billion to AI infrastructure. Adani pledged another $100 billion. That’s $210 billion from just two companies, aimed at turning India into one of the world’s biggest AI hubs. Meanwhile, OpenAI signed its first major deal with Tata Group to build 100 megawatts of AI-ready data center capacity in India (with plans to scale to 1 gigawatt), and hundreds of thousands of Tata employees will get access to ChatGPT Enterprise. The event drew Sam Altman (OpenAI), Dario Amodei (Anthropic), Sundar Pichai (Google), and even Emmanuel Macron — though Bill Gates pulled out hours before his keynote, citing unspecified reasons.

The most candid moment of the day: when Prime Minister Modi asked all the executives on stage to raise their hands together in a symbolic show of unity, most obliged. Two didn’t — rival CEOs Sam Altman and Dario Amodei.

Why it matters: The AI race is no longer just a US-China story. India is writing $210 billion checks to get in the game. For everyday people, more AI infrastructure means faster, cheaper, and more localized AI services — especially for the 1.4 billion people who live there.

Source: Reuters | TechCrunch

2. A Microsoft Bug Was Feeding Your Confidential Emails to Its AI — For Weeks

Microsoft confirmed that a bug in its Copilot AI was silently reading and summarizing confidential emails inside Microsoft Office — even when companies had specifically set up policies to prevent that from happening. The bug affected Microsoft 365 customers using Copilot Chat, and it’s been happening since January. Emails marked as “confidential” were incorrectly processed by the AI, bypassing data loss prevention policies that organizations put in place to keep sensitive information out of AI systems. Microsoft says it started rolling out a fix earlier in February, but hasn’t said how many customers were affected.

Why it matters: You pay for software. You set up security policies. And the AI reads your confidential emails anyway — for weeks — without you knowing. This is exactly the kind of story that should make you think twice before pasting sensitive information into any AI tool, including the ones built into software you already use every day.

Source: TechCrunch

3. The Woman Behind ImageNet Just Raised $1 Billion to Teach AI About the Physical World

Fei-Fei Li — the Stanford professor who created ImageNet, the dataset that kicked off the modern AI era — has raised $1 billion for her startup World Labs. The biggest chunk, $200 million, came from Autodesk (the company behind AutoCAD, used by architects, engineers, and filmmakers everywhere). Other backers include AMD, Nvidia, and Fidelity. World Labs is building what’s called a “world model” — AI that doesn’t just process text or images, but actually understands 3D space, physics, and how the real world behaves. Their first product, Marble, lets users generate editable 3D environments from a text prompt. The Autodesk partnership starts with entertainment — think AI-generated 3D worlds for games and films — but the long-term vision is AI that can design buildings, simulate factories, and reason about physical systems.

Why it matters: Most AI today understands words and images. World Labs is betting the next frontier is AI that understands space — which is how humans actually experience reality. This has massive implications for architecture, manufacturing, filmmaking, and robotics.

Source: TechCrunch

4. Google Just Added an AI Music Maker to Gemini

Google’s Lyria 3 — its AI music generation model — is now rolling out inside the Gemini app. You can describe what you want (in text, or based on an image or video), and Gemini generates a 30-second music clip. It’s still in beta, and the results are described as “something like music” rather than studio-quality tracks. But this is Google’s most direct consumer push into AI-generated audio yet, following similar moves from OpenAI, ElevenLabs, and Suno.

Why it matters: AI music generation is moving from niche tools into the apps hundreds of millions of people already use. Whether you need a quick background track for a video or just want to play around with what’s possible, this is now one tap away in Gemini.

Source: The Verge

Quick Hits

• Perplexity ditches AI ads: The search startup announced it’s abandoning plans to place ads in its AI results, with executives saying ads could have users “doubting everything.” A notable stance as ChatGPT moves in the opposite direction. (The Verge)

• Netflix threatens ByteDance with immediate litigation over Seedance AI: Netflix gave ByteDance a 3-day deadline to stop its Seedance AI from generating content based on Stranger Things, Squid Game, Bridgerton, and other Netflix properties — calling it “a high-speed piracy engine.” (The Verge)

• Meta is spending $65 million to influence AI legislation: The company is funding two new super PACs — one targeting Republicans, one targeting Democrats — to back politicians friendly to AI and fight regulation that could limit Meta’s AI business. (The Verge)

That’s it for today. The story of February 19, 2026 is really about one thing: who gets to control the infrastructure AI runs on. India is betting $210 billion it’ll be them. Microsoft’s bug is a reminder of what’s at stake when the infrastructure already inside your laptop goes wrong. And World Labs is asking a different question entirely — not just who controls AI, but whether AI can finally understand the world the way humans do.

Forward this to someone who needs to stay in the loop.

AI for Common Folks — Making AI understandable, one concept at a time.

Good morning,

Anthropic quietly handed its best AI to everyone for free, Nvidia and Meta shook hands on what could be a $50 billion chip deal, and Europe’s Parliament just banned AI tools on lawmakers’ phones. Here’s what happened 👇

1. Anthropic Just Made Its Best AI Available to Everyone for Free

Anthropic launched Claude Sonnet 4.6 yesterday — and this one is a big deal. The new model is so capable that Anthropic says it “approaches Opus-level intelligence,” which is the company’s most powerful (and expensive) tier. Improvements span coding, reasoning, long documents, and — most notably — computer use, meaning Claude can now navigate spreadsheets, fill out web forms, and operate software on your behalf much like a real person would. The kicker: it’s now the default model for free Claude users, and pricing stays the same.

Why it matters: The AI you get for free today is better than what most companies paid premium for six months ago. If you haven’t used Claude lately, this is a good reason to revisit it.

Source: Anthropic Blog

2. Nvidia and Meta Just Signed a Chip Deal Worth an Estimated $50 Billion

Nvidia announced a multiyear deal to sell Meta millions of AI chips — including current Blackwell GPUs, the upcoming Rubin generation, and for the first time, standalone CPU chips (Grace and Vera) that compete directly with Intel and AMD. Analysts estimate the deal could be worth around $50 billion. This is happening even as Meta is simultaneously developing its own AI chips and exploring Google’s TPUs as an alternative.

Why it matters: This single deal is bigger than the GDP of many countries. It shows just how much money is flowing into AI infrastructure — and why your electricity bills and cloud costs are quietly creeping up.

Source: Reuters

3. Apple Is Building AI Glasses, a Pendant, and Camera AirPods

According to Bloomberg’s Mark Gurman, Apple is ramping up work on three new AI-powered wearables: smart glasses (targeting a 2027 launch), an AI pendant, and camera-equipped AirPods. All three will have cameras and connect to your iPhone, letting Siri “take actions based on surroundings” — like identifying what you’re looking at, referencing landmarks for directions, or reminding you of tasks in specific situations. Unlike Meta’s Ray-Ban glasses, Apple plans to make the frames in-house rather than partner with a third-party brand.

Why it matters: AI is moving off your screen and onto your body. Apple entering this space signals that AI-powered wearables are no longer a niche experiment — they’re the next big product category.

Source: The Verge

4. Europe’s Parliament Just Banned AI Tools on Lawmakers’ Devices

The European Parliament’s IT department blocked all built-in AI features on government-issued devices, citing security and privacy fears. The core concern: when you use tools like ChatGPT, Copilot, or Claude, your data gets sent to US company servers — and US authorities can demand those companies hand over that data. With the Trump administration already issuing hundreds of subpoenas to tech companies for user data, European lawmakers decided the risk was too high.

Why it matters: This is a preview of a bigger conversation coming your way. The same concerns about your data — where it goes, who can see it — apply every time you paste something sensitive into an AI chatbot. It’s a good reminder to think twice before sharing confidential info with AI tools.

Source: TechCrunch

Quick Hits

• Mistral AI makes its first acquisition: The French AI company bought Koyeb, a cloud computing startup, to back its ambitions in cloud infrastructure. (Reuters)

• NAACP threatens to sue Elon Musk’s xAI: The civil rights organization sent a notice of intent to sue over xAI’s illegal installation of gas turbines in Mississippi — running without air permits — to power its Colossus 2 data center. (The Verge)

• WordPress gets an AI assistant: WordPress.com launched an AI tool that lets you edit your site, adjust styles, and create images just by typing prompts — no code needed. (The Verge)

That’s it for today. Your free AI just got smarter, the companies building it are spending at a scale that’s hard to comprehend, and the rest of the world is starting to ask: whose AI is it, anyway?

That’s your AI update for today. Forward this to someone who needs to stay in the loop.

AI for Common Folks — Making AI understandable, one concept at a time.

AI Literacy is Non-Negotiable

But not everyone needs to be an AI geek. Here’s what you actually need to know.

The Reality

“Should I learn AI?”

It’s the question everyone’s asking—parents wondering what to teach their kids, professionals wondering if they’re falling behind, executives wondering what their teams need to know.

Daniela Rus, MIT professor and head of the world’s largest AI lab, has a clear answer: yes. But with an important caveat.

“Everyone needs to understand something about AI and technology, but not everyone needs to understand everything about the technologies.”

That distinction matters. Because the pressure to become an “AI expert” is paralyzing people into learning nothing at all.

The Shift

Rus breaks it down simply. There are different levels of AI involvement, and each requires different knowledge:

• Lead with AI: Strategic understanding. Where is the technology going? What’s possible? What’s hype?

• Develop AI: Technical depth. Algorithms, models, the math underneath.

• Deploy AI: Implementation skills. How to integrate AI into existing systems and workflows.

• Use AI: Practical fluency. How to work with AI tools to be more effective at your job.

Most people fall into that last category. And that’s fine.

You don’t need to understand how large language models work under the hood. You don’t need to train your own neural network. You don’t need a computer science degree.

But you need to know something. You need enough literacy to recognize what AI can do for your work, to evaluate tools, to spot opportunities, to avoid being left behind.

The Old Way: AI is for engineers and data scientists. Everyone else can ignore it.

The New Reality: AI literacy is like computer literacy in the 90s. Not optional. Not specialized. Baseline.

What To Do Next

Figure out which category you’re in. Be honest.

If you’re leading—you need to understand AI strategy, capabilities, and limitations. Read widely. Talk to people who are building.

If you’re deploying—you need to understand integration, workflows, and change management. The technology is only part of the puzzle.

If you’re using—you need hands-on fluency with tools relevant to your field. Not theory. Practice.

And regardless of category, Rus emphasizes that foundational skills still matter: math, science, critical thinking, creativity. AI doesn’t replace these. It amplifies them.

Start where you are. Learn what you need. Don’t let the pressure to know everything stop you from knowing something.

The One Thing to Remember

AI literacy is non-negotiable. But you don’t need to be an expert—you need to be literate enough to use, evaluate, and adapt. That’s within reach for everyone.

This insight comes from an interview with Daniela Rus, MIT professor and director of CSAIL. The AI Shift curates wisdom from AI leaders and translates it for busy professionals navigating the AI era. Where do you fall—leading, developing, deploying, or using AI? And are you learning what that level requires?

Good morning,

ByteDance is scrambling after Hollywood came for its AI video tool, Ireland launched a formal investigation into Grok’s image problem, and India just became the center of the AI world this week. Here’s what happened 👇

ByteDance Scrambles After Its AI Video Tool Spooked Hollywood

What happened: ByteDance’s new AI video generator, Seedance 2.0, went viral last week — but not in the way they wanted. Users generated hyperrealistic videos of Tom Cruise fighting Brad Pitt, Dragon Ball Z scenes, and Pokémon clips so convincing that Disney and Paramount accused ByteDance of distributing and reproducing their intellectual property. ByteDance now says it’s “working to improve safeguards” and will tweak the model to prevent unauthorized use of copyrighted characters and real people’s likenesses.

Why it matters: This is the AI copyright fight moving from still images to video. If you’ve been playing with AI video tools, expect every major platform to tighten what you can and can’t generate — especially anything involving real people or recognizable characters.

Sources: The Verge • Ars Technica

Ireland Opens Formal Investigation Into Grok Over Sexualized AI Images

What happened: Ireland’s Data Protection Commission — the lead EU regulator for X (formerly Twitter) — launched a formal investigation into Elon Musk’s Grok AI chatbot. The probe focuses on Grok generating sexualized images of real people, including children. This follows weeks of global outrage after Grok flooded X with AI-altered near-nude images. Despite X announcing curbs, Reuters found that Grok continued producing such images when prompted. The DPC can levy fines of up to 4% of X’s global revenue under GDPR.

Why it matters: This is now the EU, California, Malaysia, Indonesia, France, and the UK all investigating the same AI tool. If you’re wondering whether governments will actually regulate AI — they already are, and Grok is becoming the test case.

Sources: Reuters • The Verge

India Hosts Global AI Summit as Every Major AI Company Shows Up

What happened: India kicked off the AI Impact Summit in New Delhi this week — the first time this global event has been held in the developing world. OpenAI’s Sam Altman, Anthropic’s Dario Amodei, Google’s Sundar Pichai, and DeepMind’s Demis Hassabis are all attending. Google, Microsoft, and Amazon have already committed a combined $68 billion in AI and cloud infrastructure investment in India through 2030. India isn’t trying to build the next frontier AI model — instead, it’s betting on being the world leader in AI deployment and application.

Why it matters: India already has 72 million daily ChatGPT users — making it OpenAI’s largest market. When the world’s most populous country goes all-in on AI adoption, it shapes how these tools get built for everyone. The AI race isn’t just about who builds the smartest model — it’s about who puts it in the most hands.

Sources: Reuters • TechCrunch

ChatGPT Gets a “Lockdown Mode” for Security

What happened: OpenAI introduced Lockdown Mode for ChatGPT — an optional security setting that tightly restricts how ChatGPT interacts with external systems. In Lockdown Mode, web browsing is limited to cached content only (no live requests leave OpenAI’s network), and certain tools are disabled entirely. It’s designed to protect against prompt injection attacks — where someone tricks ChatGPT into leaking your sensitive information. Available now for ChatGPT Enterprise, Edu, Healthcare, an Teachers plans.

Why it matters: As more people connect ChatGPT to their email, files, and work tools, the security risks grow. Think of Lockdown Mode like a vault setting for people who handle sensitive data. Most of us won’t need it yet, but it’s a sign that AI security is becoming a real product category.

Sources: OpenAI • The Verge

Quick Hits

Anthropic’s India revenue doubled in 4 months: CEO Dario Amodei revealed at a Builder Summit in Bengaluru that Anthropic’s revenue run-rate in India doubled since October, with India now the company’s second-largest market after the US. Claude Code adoption is driving the growth. (Reuters)

OpenAI’s new coding model runs 15x faster on non-Nvidia chips: GPT-5.3-Codex-Spark, running on Cerebras chips instead of Nvidia, delivers code at 1,000+ tokens per second. Available to ChatGPT Pro subscribers ($200/month) as a research preview. (Ars Technica)

Unity wants AI to build entire casual games from a single prompt: CEO Matthew Bromberg said “AI-driven authoring is our second major area of focus for 2026” and plans to reveal new prompting tools at GDC in March — despite developers being increasingly skeptical of generative AI. (The Verge)

That’s it for today. Hollywood is drawing lines on AI video, regulators are closing in on image generators, and India is quietly becoming the world’s biggest AI testing ground.

AI for Common Folks — Making AI understandable, one concept at a time.

Semi-Supervised Learning is training an AI using a small amount of labeled data (with answers) and a large amount of unlabeled data (without answers)—like a teacher who only has time to grade 5 papers out of 100, but the students still figure out the rest.

Hey Common Folks!

We’ve covered the two main ways computers learn:

• Supervised Learning: The teacher stands over the student, correcting every mistake (high effort, requires answer keys).

• Unsupervised Learning: The student is left alone with a pile of books and told to “figure it out” (low effort, but harder to guide).

But what if there’s a middle ground? What if you’re a busy teacher who only has time to grade a few papers out of a hundred? Can the computer still figure out the rest?

Yes. This is called Semi-Supervised Learning, and it’s the secret weapon of big tech companies.

The Problem: Labels Are Expensive

Why don’t we just use Supervised Learning all the time? Because giving the computer the “answer key” is expensive and exhausting.

Imagine you want to build an AI that detects rare diseases in X-rays.

The Data: You can easily download 100,000 X-ray images from a hospital database. That’s the easy part.

The Labels: To tell the computer which X-ray shows the disease, you need a highly paid doctor to look at every single image and mark it. You can’t afford to pay a doctor to label 100,000 images.

This is where Semi-Supervised Learning saves the day. You pay the doctor to label just 1,000 images, and the AI uses that knowledge to figure out the remaining 99,000.

Real-World Proof: Modern deep learning has shown you can train state-of-the-art models with surprisingly small labeled datasets—sometimes as few as 150 images per category. Semi-supervised learning pushes that efficiency even further.

The Google Photos Example

The best example—and one you probably use—is Google Photos.

Step 1 – Unsupervised Grouping: You upload 5,000 photos of your family. Google’s AI looks at them and notices, “Hey, this face appears in 500 photos. That face appears in 200 photos.” It doesn’t know who they are, but it knows they’re the same person. It groups them together using clustering.

Step 2 – The Supervised Nudge: You click on one photo and type “Dad.”

Step 3 – Semi-Supervised Magic: The AI takes that one label (”Dad”) and instantly applies it to the other 499 photos in that group.

You did 1% of the work (labeling one photo), and the AI did 99% of the work (labeling the rest). That’s Semi-Supervised Learning in action.

How Does It Work?

It follows a simple logic:

1. Cluster First: The AI looks at all the data (labeled and unlabeled) and groups similar things together. It notices that Data Point A is very similar to Data Point B.

2. Propagate the Label: If you tell the AI that “Point A is a Cat,” the AI assumes that since Point B looks exactly like Point A, then Point B must be a Cat too.

The key assumption: data points that are close to each other probably share the same label.

2026 Update: Self-Supervised Learning Takes Over

Here’s where things get exciting. Since around 2018, a cousin of semi-supervised learning has become the backbone of modern AI: Self-Supervised Learning.

What’s the difference?

• Semi-Supervised: You give the AI a few labels, and it uses unlabeled data to fill in the gaps.

• Self-Supervised: The AI creates its own labels from the data itself—no human labels needed at all.

Real-World Example: How ChatGPT Learned to Write

ChatGPT wasn’t trained by humans writing “correct answers” for billions of sentences. Instead, it used self-supervised learning:

1. Take a sentence: “The cat sat on the ___”

2. Hide the last word (”mat”)

3. Train the AI to predict it

4. Repeat this billions of times with internet text

The AI creates its own “quiz” from raw text, learning language patterns without anyone labeling a single sentence. This is why GPT-4, Claude, and Gemini could train on trillions of words without hiring millions of human teachers.

Why this matters: Self-supervised learning is the reason AI exploded in the 2020s. It unlocked the internet’s massive, messy, unlabeled data.

Spot It in the Wild: Where You’re Already Using It

You interact with semi-supervised and self-supervised learning every day:

• Spotify Discover Weekly (it knows what songs you like and finds similar unlabeled music)

• Gmail spam filter (you mark a few emails as spam; it learns patterns to catch the rest)

• Medical diagnosis tools (doctors label a few images; AI extends that knowledge across millions)

• ChatGPT, Claude, Gemini (all trained with self-supervised learning on massive unlabeled text)

When Semi-Supervised Learning Fails

Just like any machine learning technique, semi-supervised learning has limitations you need to watch for:

1. Garbage In, Garbage Out (Amplified)

If your small labeled dataset is biased or wrong, the AI will spread that bias across all the unlabeled data.

Example: If you label 10 photos of “doctors” and they’re all men, the AI might learn “doctor = male” and mislabel female doctors in the unlabeled set.

2. The “Close Together” Assumption Can Break

Semi-supervised learning assumes similar-looking things have the same label. But what if they don’t?

Example: Huskies and wolves look similar, but they’re not the same. If your labeled data only has huskies, the AI might confidently—and wrongly—label wolves as “husky.”

3. Domain Shift

If your unlabeled data comes from a different source than your labeled data, the AI can get confused.

Example: You label 100 professional X-rays (high quality, well-lit). Then you feed the AI 10,000 unlabeled phone photos of X-rays (blurry, poor lighting). The patterns don’t transfer well.

The Fix: Always check your unlabeled data’s quality and diversity before letting the AI loose on it.

Why This Matters Now

We live in a world where data is cheap, but labels are expensive.

• We have billions of tweets, but we don’t know the sentiment of all of them.

• We have millions of hours of YouTube video, but we don’t have transcripts for all of them.

• We have endless medical images, but we can’t afford experts to label every one.

Semi-Supervised Learning allows companies to unlock the value of massive, messy datasets without hiring thousands of humans to manually tag every single file.

And Self-Supervised Learning is why AI could suddenly read, write, code, and converse in 2023-2026 without needing labeled “correct answers” for every sentence on the internet.

Try It Yourself

Want to see semi-supervised learning in action? Here’s a simple experiment:

1. Open Google Photos (or Apple Photos)

2. Upload 50+ photos with at least 2-3 people appearing multiple times

3. Wait for the app to cluster faces

4. Label just one photo of each person

5. Watch the AI instantly label dozens more

That’s semi-supervised learning working for you—right on your phone.

The Takeaway

Semi-Supervised Learning is the “work smarter, not harder” approach to AI.

• Supervised: Requires a teacher for every lesson.

• Unsupervised: No teacher at all.

• Semi-Supervised: The teacher gives a few examples, and the student figures out the rest by association.

• Self-Supervised (2026 Bonus): The student creates their own practice tests from the material itself.

This is how your phone organizes your memories, how medical AI detects diseases without bankrupting the hospital, and how ChatGPT learned to write without anyone grading billions of essays.

AI for Common Folks — Making AI understandable, one concept at a time.

Learn More

Want to dive deeper into practical AI? Check out the free fast.ai course, which inspired several examples in this article and teaches you to build real AI applications from day one.

Previous articles in this series:

• What Is Artificial Intelligence?

• What Is Data Science?

• What Is Machine Learning?

• What Is Deep Learning?

• What Is Supervised Learning?

• What Is Unsupervised Learning?