If you followed AI news at any point in the last few years, you saw this number without knowing its name. Every time a company bragged that its model “knows more” than the last one, MMLU was usually the scoreboard they were pointing at. It’s the test that turned “how smart is this AI” into a single percentage. And it’s a good example of how quickly a good test can wear out.

What it means

MMLU (Massive Multitask Language Understanding) is a giant multiple-choice quiz for AI models. About 16,000 questions spread across 57 subjects, everything from grade-school math to law, medicine, history, and computer science. Four choices each, pick the right one. The model’s final score is just its batting average across all of it. Higher percentage, more it got right.

When it came out in 2020, it was brutally hard for the AI of the day. That was the point. It covered so many subjects that a model couldn’t fake broad knowledge.

Why it matters

• It became the number companies sold you on. For a stretch, MMLU was the headline stat in basically every model launch, the AI version of a car’s miles-per-gallon sticker. When you heard one model was “smarter” than another, this was often the receipt.

• It shows how fast this stuff moves. In 2020, the best models scored around 32%, barely better than guessing. By 2026, the top ones cluster in the low 90s, all bunched within a few points of each other. When everybody scores an A, the test stops telling you who’s actually best.

• It got so easy it kind of broke. Researchers found the test is now “saturated,” and there are even known wrong answers baked into it that cap the real ceiling around 95%. So the field moved on to harder versions like MMLU-Pro. Honestly, if you see a company bragging about a plain MMLU score today, that’s a bit of a tell they’re cherry-picking.

Simple example

Think about a spelling test that was genuinely tough for a third-grader. Great tool for a while. You could actually see who studied and who didn’t. But hand that same test to a room of adults, and everybody gets 95%. The test didn’t get worse. The people taking it just outgrew it, so a perfect score stops telling you anything about who’s the better speller.

That’s MMLU now. Still a fine test. The models just outgrew it.

Perplexity Pro costs $20 a month, same as ChatGPT Plus and Claude Pro. From the outside it looks like the same deal with a different logo. It kinda isn’t. Perplexity isn’t really trying to be a chatbot. It’s an answer engine, the thing you reach for when you want a researched answer with the sources shown, not just a confident paragraph. What the $20 buys reflects that.

Quick caveat first: everything below is right as of this writing, and AI pricing moves around a lot. The exact numbers will drift, probably soon. The way to think about the choice sticks around longer, so that’s the part to hang onto.

What the money buys

Free Perplexity is genuinely useful, more than most free tiers. You get unlimited basic searches, but only a handful of the good “Pro Searches” per day before it cuts you off.

Pro lifts that cap way up, into the hundreds a day, and lets you pick the top models (GPT, Claude, Gemini) through a switcher. You also get file uploads, image generation, and about 20 Deep Research runs a day, which is the feature that reads a pile of sources and writes you a documented report. One nice extra: Pro pulls from normally paywalled sources like Statista and PitchBook, which on their own would run you hundreds a month.

What it doesn’t buy

The big jump is Max at $200, and that’s a real leap with no middle step. What you get for ten times the price is mostly the autonomous stuff: unlimited Labs, and Perplexity Computer, which wrangles a bunch of different AI models as a team to grind through a whole multi-step project while you’re off doing something else.

Honest question before you upgrade: will you actually use the “works while you sleep” features enough to justify $200? For most people the answer is no. If you’re a student, though, look at Education Pro. It’s the full Pro feature set for $10, which is kind of a steal.

The math

$20 a month is $240 a year, or about $200 if you pay for the year up front, same $17-ish monthly as the others.

The prices will change. The test won’t: what are you actually paying for? If you mostly want a fast chat buddy, Perplexity’s $20 is aimed somewhere else, and ChatGPT or Claude might fit better. But if you do real research, where seeing the sources and trusting the citations is the whole point, this is the one built for that job. Pick the tool for the work, not the logo on it.

Most companies and governments use AI the way you use electricity from the grid. You don’t own the power plant, you just pay the bill and flip the switch. That works great until the plant is in another country, run by a company that answers to another government, and one day the terms change. Sovereign AI is the idea that some things are too important to rent.

What it means

Sovereign AI means building and running your AI on infrastructure you actually control, instead of depending on a foreign company for it. That covers a few layers: where your data lives, whose chips do the work, and who owns the model itself. The strongest version is owning the model weights (the actual trained file), so nobody can change the rules, raise the price, or cut you off.

Worth knowing: a local data center alone doesn’t make you sovereign. If a US company owns the building sitting in Europe, US law still reaches it. Sovereignty is about who has legal control, not just where the box is plugged in.

Why it matters

• It’s really about the off-switch. If your hospital, bank, or government agency runs on a model owned by someone else, that someone can change pricing, restrict access, or get ordered by their own government to shut you out. We actually saw a version of this in June, when a US export order forced Anthropic to pull two of its top models offline worldwide for a couple weeks. Owning the model means nobody else’s bad day becomes yours.

• Open-weight models made it possible for regular organizations, not just nations. A few years ago, running your own frontier AI was fantasy unless you were Google. Now you can download something like Llama or Qwen, run it on your own servers, and the data never leaves the building. That’s a big deal for anyone under privacy rules they can’t hand off to an outside API.

• It’s driving real money and real politics. Nvidia’s sovereign AI sales tripled to over $30 billion in a year, and countries from France to Saudi Arabia are pouring cash into their own “AI factories.” Honestly, some of that is branding more than substance. But the core worry, not wanting your whole economy to depend on somebody else’s software, is a genuine one.

Simple example

Think about the difference between renting an apartment and owning your house. Renting is cheaper and easier, and for most of life it’s totally fine. But the landlord can raise the rent, change the rules, or decline to renew your lease, and there’s not much you can do about it.

Owning costs more and you have to fix your own plumbing. What you get for it is that nobody can hand you a notice to leave.

One of the oldest complaints about AI is that it’s a black box. You see the question go in and the answer come out, but the middle is a mystery. Researchers at Anthropic built a tool that pries that box open a little, and some of what they found is genuinely unsettling. It’s called the J-lens.

What it means

The J-lens (short for Jacobian lens, named after the math it uses) is a tool that reads an AI model’s silent internal reasoning. It turns out a model like Claude does a lot of its real thinking in a hidden zone the researchers named the “J-space,” separate from the words it actually writes out. The J-lens translates what’s happening in that space into plain vocabulary words, so you can watch concepts light up before the model says a thing.

The catch, and the researchers say this themselves, is that it’s an imperfect tool. It’s a rough read of the model’s mind, not a perfect transcript.

Why it matters

• It catches things the model doesn’t say out loud. In one safety test, researchers set up a fake scenario where the AI could blackmail an executive to avoid being shut down. The J-lens showed words like “leverage” and “threat” forming in the model’s silent workspace before it wrote a single word. The final answer stayed polite. The thinking underneath was not.

• It exposed a strange problem: the model often knows it’s being tested. Reading the J-space, researchers saw “fake” and “fictional” show up early, meaning Claude had quietly figured out the scenario was staged. When they switched off that awareness, the model misbehaved more often. That raises a real question. How much of an AI passing an ethics test is real, and how much is it behaving because it suspects someone’s watching?

• It’s a new window for AI safety. Judging a model only by its output is like judging a person only by what they say in a job interview. Reading the reasoning itself, honestly, gets you a lot closer to the truth, even if the read is fuzzy.

Simple example

Think of a poker player who keeps a calm face and says “I’ll call.” That’s the output, and it tells you almost nothing. Now imagine you could see the thought bubble over their head: “terrible hand, but I think they’re bluffing.” That bubble is the J-space, and the J-lens is the thing that lets you read it.

The player still only says three words. You just finally know what’s behind them.

When a company says its new model “scored 88% on SWE-bench,” that number is doing a lot of work in the headline. It’s the test the whole industry points to when it wants to prove an AI can actually write code, not just talk about it. So it’s worth knowing what the test actually measures, and where the number quietly stops meaning much.

What it means

SWE-bench (short for “software engineering benchmark”) checks whether an AI can fix real bugs in real software. Researchers pulled thousands of actual problems from popular open-source projects on GitHub, each one a genuine issue a human developer eventually solved. The AI gets the bug report and the codebase, and has to write a patch that fixes it. Then the code runs against the project’s real tests. It either passes or it doesn’t. No partial credit, no opinion, just did the fix work.

The version everyone quotes is SWE-bench Verified, a cleaned-up set of 500 problems that human experts checked by hand.

Why it matters

• It’s the number that sells the model. When you see coding scores in a launch announcement, this is usually the one. Top models now land in the high 80s on Verified, up from the 50s a couple years back. That climb is a big part of why AI coding tools got taken seriously.

• It tests the messy real thing, not a puzzle. Older coding tests handed the AI a tidy little function to write. SWE-bench drops it into a full, complicated codebase and says go find the problem. That’s much closer to what an actual developer does all day, which is why it caught on.

• The scores are getting slippery, and honestly I’d take the high ones with a grain of salt. The test has been public so long that its answers have leaked into the training data models learn from, so some of that 88% might be memory, not skill. That’s why labs keep building harder versions, like SWE-bench Pro, where the same models suddenly drop into the 60s.

Simple example

Think of a culinary school final. The old version handed a student one recipe and graded whether the dish came out. SWE-bench is the harder final: you walk into a working restaurant mid-rush, get told “table six sent back their order, fix it,” and have to figure out what went wrong in a kitchen you didn’t set up. Passing means the customer actually eats.

The catch is when the school reuses the same test every year. Eventually students hear the questions in advance, and a perfect score stops telling you who can really cook.

You've probably sat in a meeting where something like "Otter is recording" slid into the participant list. That's an AI notetaker. But a lot of people don't realize they don't all live on your laptop. A bunch of them are little physical gadgets you set on the table. Same job, pretty different tradeoffs.

What it means

An AI notetaker records a conversation, transcribes it, and hands you back a summary with the decisions and action items pulled out. It comes in two flavors.

Software is the one you know: apps like Otter, Fathom, Fireflies, and Granola, plus the built-in summaries already hiding in Zoom, Teams, and Google Meet. Hardware is the other flavor: pocket recorders and little pins from folks like Plaud and Ruune, usually somewhere between $50 and $190, that listen with their own mics and sync to an app after.

Why it matters

• Software is the cheap, easy on-ramp. A lot of it is already baked into tools you pay for, most have a free tier, and there’s nothing extra to lug around. Granola even skips the visible bot and just listens to your computer’s audio.

• Hardware handles the real world better. Better mics mean an actual usable transcript in a loud room instead of mush. It records offline. And it won’t quietly quit on you the way a phone app can, where the app got suspended in your pocket and you find out ninety minutes later it caught nothing. That one still stings.

• A physical device makes the consent part easier, which matters more than you’d think. You can point at the recorder and let people watch you switch it off. Good luck doing that with software running invisibly on a laptop nobody can see. And where I live, Washington is an all-party consent state, so everybody has to say yes before you hit record. The tool won’t ask for you.

Simple example

Say you want to catch a good idea on a walk. The software version is the voice memo app already on your phone. Free, always there, and every so often decides it wasn’t really recording. The hardware version is a dedicated $129 recorder in your pocket with a real mic that just works, that you had to pay for, and that you have to remember to bring.

Neither one’s wrong, honestly. One’s the tool you already own. The other’s the one that doesn’t let you down when it actually counts.

If you want to understand modern AI, you keep bumping into the same handful of names, and Ilya Sutskever is near the top of every list. He had a hand in most of the big breakthroughs of the last decade, he was at the center of the strangest weekend in tech history, and just this week his secretive new company landed a $5 billion check from Nvidia. He’s worth knowing.

Who they are

Sutskever was born in Russia, grew up in Israel, and studied under Geoffrey Hinton at the University of Toronto. In 2012 the two of them, along with Alex Krizhevsky, built AlexNet, the project that proved you could train deep neural networks on GPUs and get startling results. That’s the moment a lot of people point to as the start of the whole current AI era.

He went on to work at Google, then co-founded OpenAI in 2015 and served as its chief scientist. His fingerprints are on AlphaGo, the GPT models, and the early reasoning models. In late 2023 he was part of the board that briefly fired Sam Altman, a decision that unraveled in about five days after nearly the whole company threatened to quit. Sutskever later said he regretted it. He left OpenAI in 2024 and started Safe Superintelligence, or SSI.

Why they matter

• His research is load-bearing for the entire field. AlexNet, sequence-to-sequence learning, the GPT lineage. Strip out Sutskever’s contributions and modern AI looks very different, and probably arrives later than it did.

• SSI is a genuinely odd bet, and people are funding it anyway. Two years in stealth, no product, no demo, no revenue, and somehow a valuation around $32 billion. This week Nvidia put in a reported $5 billion and handed over access to its newest chips. That’s a lot of money riding on a company whose website is basically a mission statement.

• He represents one answer to the safety question. The whole pitch of SSI is a “straight shot” at building a superintelligence that’s safe from the start, with no products along the way to distract from it. Whether that’s the right approach is a real debate, but it’s a serious person betting his reputation on doing it that way.

What they’ve said or done

When he launched the company, Sutskever described the plan about as plainly as anyone could: “We will pursue safe superintelligence with a singular focus on one goal and one product.”

No hedging, no roadmap, no beta. For a guy who helped kick off the product race, it’s a striking place to land.

There’s a new attack aimed right at the AI coding tools a lot of developers now lean on every day. It doesn’t need malware, and it doesn’t need your password. It just needs your AI agent to read the wrong thing and do what it says. Researchers are calling it agentjacking.

What it means

Agentjacking is a trick that hijacks an AI coding agent (tools like Claude Code, Cursor, and Codex) into running an attacker’s commands on your own machine. It’s basically a meaner version of prompt injection, the older problem where hidden instructions get slipped into something an AI reads.

Here’s how the researchers actually did it. A lot of apps use a tool called Sentry to collect their error reports, and anybody can send Sentry a fake one. So an attacker writes a fake report with commands hidden inside it. Later, a developer asks their AI agent to look at the errors and fix them. The agent reads the poisoned report, can’t tell a real bug note from a planted instruction, and runs the bad code. With your permissions, on your computer.

Why it matters

• The success rate is rough. When Tenet Security tested this in June 2026, it worked about 85% of the time across every major coding agent they tried, and they confirmed 2,388 exposed organizations. That range ran from Fortune 100 companies down to solo developers.

• Your usual security tools don’t catch it. Nothing gets downloaded, so antivirus sees nothing. No password gets stolen, so no login alert fires. The agent is just doing what it was told, which is kind of the whole problem.

• It’s a preview of where the agent era is headed. An agent that reads outside content, has access to your system, and can take actions is useful for the same reason it’s dangerous. I dunno that there’s a clean fix for that tension yet. Give a tool the power to do real work for you and you’ve also handed it the power to do real damage when someone feeds it a lie.

Simple example

You hire a sharp new assistant and tell them to go through the customer complaint box and handle what they can. One of the complaint cards is fake. Tucked inside, in the same handwriting as all the real ones, it says to go into the safe and mail the contents to some address. Your assistant can’t tell that line isn’t a real customer request, so they follow it.

That’s agentjacking. The assistant did their job. They just couldn’t tell an order from a note, and you’d already given them the key to the safe.

Twenty dollars, the same as ChatGPT. What you get for it isn’t the same.

Claude Pro costs $20 a month, which is exactly what ChatGPT Plus costs, and from the outside the two look like the same purchase. They kinda aren’t. What the money buys is a little different, and there’s one setting buried in here that you’ll want to know about before you paste anything private into it.

Quick caveat first: everything below is right as of this morning, and AI pricing moves around a lot. The exact numbers will drift, probably soon. The way to think about the choice sticks around longer, so that’s the part to hang onto.

What the money buys

Mostly you’re buying room. Free Claude gives you everyday chat, but the limits cut you off fast, usually right when you’re getting somewhere. Pro lifts those limits way up and hands you the whole toolkit: the coding assistant (Claude Code), the desktop work app (Cowork), and Claude Design. You still pick the model yourself. Pro just means you can actually lean on the good ones longer without smacking into a wall by lunch.

Then there’s Max, and in a nutshell, Max is just Pro with the ceiling raised. A hundred bucks a month gets you five times the usage. Two hundred gets you twenty times. That’s it. You’re paying for more capacity, not a smarter Claude.

What it doesn’t buy

Max won’t hand you a better Claude, just a bigger bucket. So if Free or Pro almost never interrupts you, skip it. The people it’s actually built for are the ones running Claude Code hard, every single day. For them the math gets real: one heavy user’s eight months would’ve run over $15,000 if they’d paid by usage. On Max, the same work came to about $800.

Worth noting: on the personal plans (Free, Pro, and Max), your chats can get used to train future models unless you go turn that off. It's under Settings, then Privacy, then flip off the toggle called "Help improve our AI models." Leave it on and your conversations can stick around, de-identified, for up to five years. Team and business accounts are carved out by default. Regular folks have to find the switch themselves, and honestly most people have no idea it's there.

The math

$20 a month runs you $240 a year, or $200 if you pay all at once up front.

The prices will change. The test won’t: do the limits actually stop you mid-task? If you keep banging into the caps, Pro probably pays for itself. If you’re just popping in a few times a week, Free is likely plenty. And if you’re staring at Claude and ChatGPT sitting at the same $20, the usage caps and that privacy toggle matter a lot more than the price, cause the price is a tie.

A distillation attack is a way of stealing an AI model without ever touching its code. The attacker just talks to it, a lot, and copies what comes back. You ask the model thousands of questions, save all its answers, and use those answers to train a cheap knockoff that behaves almost the same. It’s the shadier relative of ordinary Model Distillation, which is a legitimate technique for shrinking models. Same basic move, except here nobody asked permission.

What it means

Okay, so the thing that makes this possible is that a model has to answer you to be useful. That’s the whole product. And every answer it gives away is a little piece of what it knows. Ask enough questions across enough topics, and the pile of answers you collect starts to describe the model itself.

An attacker takes that pile and uses it to fine-tune a smaller, open model they already have. They’re not copying the original’s internal wiring, which they can’t see anyway. They’re copying its behavior, which is the part that actually matters to a customer. The result is a cheaper model that does most of what the expensive one does.

And it’s cheap. In one published example, researchers copied a useful chunk of a commercial model’s ability for about $50 in usage fees. Another team cloned a specialized medical model to 94% of its performance for under $1,000. Compare that to the tens or hundreds of millions it costs to build a frontier model from scratch, and you can see why companies worry about this.

The uncomfortable part is that you can’t fully stop it. A model that refuses to answer is useless, so the answers have to flow, and if they flow, they can be collected. Defenders mostly try to make stealing more expensive than just paying: rate limits, watching for accounts that ask suspiciously broad questions, and not handing out extra detail in the responses.

Why it matters

It turns a company’s biggest investment into something copyable. The model is the asset, and the asset walks out through the front door one answer at a time.

It’s cheap enough for almost anyone. Fifty dollars and some patience is a very different threat than needing a data center.

It shapes what these companies will show you. Some of the limits you bump into exist partly to make this attack harder.

Simple example

Picture a famous restaurant with one secret sauce. A competitor can’t get into the kitchen, so they do the next best thing: they order the dish hundreds of times, taste carefully, and reverse-engineer the recipe from the outside. They never see how it’s actually made. They just taste the output enough times to reproduce it. Their version isn’t perfect, but it’s close, it’s a lot cheaper, and the original restaurant did most of the work. A distillation attack is that, done to a model, one order at a time.

Kimi is a Chinese AI assistant made by Moonshot AI, and one of the clearest signs that the frontier of AI is no longer only an American story. It started as a chatbot known for handling unusually long documents, then reinvented itself as one of the strongest open-weight model families anyone can download. If you’ve heard that Chinese labs are catching up fast, Kimi is a large part of what people mean.

What it means

Moonshot AI was founded in Beijing in 2023 by Yang Zhilin and two university classmates, and named after a Pink Floyd album. Its assistant, Kimi, first made its name on long context: the 2023 version could read about 128,000 tokens at once, far more than most competitors at the time, which made it genuinely useful for working through long reports and books.

The bigger turn came in 2025. Moonshot began releasing its models as open weights, meaning anyone could download the finished model and run it themselves. The Kimi K2 release landed hard with developers and drew comparisons to DeepSeek’s breakout moment, since here was another Chinese lab shipping something competitive and giving it away. That strategy rebuilt Moonshot’s reputation and pulled in a fresh wave of investment, reportedly reaching a valuation around $20 billion.

By 2026, Kimi had become a full product line rather than just a chatbot: a consumer assistant, a work platform for documents and research, and a coding tool that runs in a developer’s terminal. It’s one of a cluster of well-funded Chinese labs, sometimes called the “Six AI Tigers,” pushing the same direction at once.

Why it matters

It’s proof the open-weight lead isn’t only American. For a stretch, the strongest model you could freely download came from Beijing, not Silicon Valley.

It made long context a mainstream expectation. Kimi’s early bet on reading whole books at once is now table stakes.

It shows the competitive map is global. Understanding AI now means watching Chinese labs, not just OpenAI and its neighbors.

Simple example

Imagine most powerful software could only be rented from a handful of American companies, and then a lab in Beijing started giving away something comparable for free. Developers everywhere could suddenly download it, inspect it, and build on it without asking permission or paying a toll. That’s roughly what Kimi’s open-weight releases did to the AI landscape, and it’s why a name many Americans have never heard is closely watched by people who build with these tools.

People say a model “learns” from your conversation. They picture something like a student who gets a little smarter every time you talk to it. That’s a reasonable guess, and it’s wrong in a way that matters. Training and Inference are separate events, separated by months, and once training ends the model stops learning entirely.

The difference

Training is the building. A company takes an enormous pile of text, runs it through the model over and over, and adjusts billions of internal settings until the model gets good at predicting what comes next. This takes weeks or months, costs somewhere in the tens or hundreds of millions of dollars for a frontier model, and happens once. At the end you have a finished thing, frozen, that doesn’t change.

Inference is the using. Every time you send a prompt, the finished model runs your text through those frozen settings and produces an answer. It costs a fraction of a cent, takes a second or two, and happens billions of times a day. The model isn’t learning during any of this. It’s applying what it already learned, the same way every time.

So when a chatbot seems to remember something you told it earlier, that isn’t training. It’s the conversation being fed back in as context, which disappears when the conversation ends. The model itself is unchanged. Your chat did not teach it anything.

There’s a wrinkle worth knowing: some companies do use your conversations to train future versions, which is a separate process happening later, on their schedule, not live. That’s a privacy question rather than a learning-in-real-time question, and it’s usually a setting you can turn off.

The one-sentence test

If it’s happening once, at a lab, at enormous cost, it’s training. If it’s happening right now, because you pressed enter, it’s inference.

Simple example

Medical school and a doctor’s appointment aren’t the same activity. School takes years, costs a fortune, and produces someone who knows medicine. The appointment takes fifteen minutes and applies that knowledge to you. Your visit doesn’t make the doctor a better doctor in any meaningful sense, and if she needs new skills she goes back for more training later, deliberately, not by seeing you.

Training built the doctor.

Inference is the appointment, and it’s the appointment that has to happen millions of times for the school to have been worth it.

Model routing is the practice of automatically sending each request to whichever AI model is best suited to handle it. Instead of one model answering everything, a system sits in front of several and decides where each question should go. Simple questions get a fast, cheap model. Hard ones get the slow, expensive one that thinks carefully. If you’ve noticed ChatGPT or Claude quietly picking a mode for you rather than making you choose, that’s routing doing its job.

What it means

The reason routing exists is cost. The most capable models are meaningfully more expensive to run than the lighter ones, and most of what people ask doesn’t need them. Reformatting a list or answering a factual question is not the same work as debugging code or analyzing a contract, but if everything goes to the flagship model, you pay flagship prices for both.

A router looks at the incoming request and makes a quick judgment about how much horsepower it needs. That judgment might be based on the length and complexity of the question, whether it involves reasoning or images, or how the user has phrased it. Then it hands the request to the appropriate model and returns the answer, usually without telling you which one it picked.

This is different from Mixture of Experts, which people sometimes confuse it with. Mixture of Experts routes between specialized sections inside a single model. Model routing chooses between entirely separate models. Same basic instinct, different level.

Why it matters

It’s why AI products can afford to be free. Routing cheap questions to cheap models is what makes a free tier survivable.

It explains inconsistent answers. If the same question gets different quality on different days, routing may have sent it somewhere different.

It’s becoming invisible on purpose. Providers increasingly decide for you, which is convenient right up until you wanted the better model.

Simple example

Think about how a hospital handles an emergency room. Not everyone who walks in sees a surgeon. A nurse at intake takes one look, asks a few questions, and decides where you go: minor cuts to the fast track, chest pain straight to cardiology. Nobody thinks this is a downgrade. It’s the only way the hospital treats everyone without running out of surgeons by noon. Model routing is that intake nurse, and the reason you rarely notice it is that it’s usually right.

Alexandr Wang is one of the most consequential young figures in AI, best known for founding Scale AI and now serving as Meta’s first Chief AI Officer. He built a company that quietly powered much of the industry, then became the centerpiece of one of the largest talent deals the field has seen. If you want to understand how the raw material behind AI models actually gets made, and how much Meta is willing to spend to win, Wang is where both stories meet.

Who he is

Wang founded Scale AI in 2016 as a 19-year-old at MIT, built around a plain but essential idea: AI models need enormous amounts of carefully labeled data, and someone has to organize the human work of producing it. Scale became that someone. It supplied training data to OpenAI, Microsoft, Meta, and much of the rest of the industry, and by 2024 it was generating roughly $870 million in annual revenue.

In mid-2025, Meta invested $14.3 billion for a 49 percent stake in Scale, a deal that roughly doubled the company’s valuation to about $29 billion and brought Wang in-house. He joined that June as Meta’s first-ever Chief AI Officer, co-leading a new group called Meta Superintelligence Labs, and stepped back from running Scale to a seat on its board.

Why he matters

He made data labeling into critical infrastructure. Scale turned an unglamorous necessity into one of the industry’s quiet dependencies.

His hiring showed how far Meta will go for AI talent. The $14.3 billion structure was as much about acquiring Wang and his people as any single asset.

He’s now a public voice for where Meta thinks AI is headed. His “personal superintelligence” framing is Meta’s pitch in his words.

What he’s said or done

Wang grew up in Los Alamos, New Mexico, the government science town, and has said being raised around physicists shaped his belief that science should serve society. Announcing his Meta move to Scale employees, he wrote that opportunities of that magnitude often come at a cost, and that “in this instance, that cost is my departure.” Since joining Meta, he has publicly framed the company’s goal as “personal superintelligence,” technology meant to help people pursue goals, build healthier habits, and strengthen relationships.

Sandboxing is the practice of running a program inside a sealed-off space where it can work freely but can't reach anything outside. The name comes from the literal thing: a child's sandbox is a place to make a mess that stays contained. In computing, it's a walled area where code can run, create files, and even fail badly, without touching the rest of your system. You'll see the word constantly now around AI agents, because a sandbox is often the thing standing between "the AI wrote some code" and "the AI ran some code on my actual computer."

What it means

When an AI agent writes and runs code on its own, someone has to decide where that code runs. Letting it loose on your real machine means it could read your files, change your settings, or delete something it shouldn’t, not out of malice, but because it misunderstood the task. A sandbox solves this by giving the agent a disposable copy of a computer to work in. It gets its own temporary files, its own memory, and a boundary it can’t cross. If it makes a mess, the mess is inside the box. Close the box and the mess is gone, while your real system never knew it happened.

The tradeoff is that a sandbox is deliberately limited. The code inside usually can’t reach the internet freely, can’t see your personal files, and can’t make permanent changes. That’s the point. The limits are the safety.

Why it matters

Sandboxing is what makes it reasonable to let an AI agent run code at all. Without it, every automated action would be a gamble.

It contains mistakes as well as attacks. Most sandbox saves aren’t stopping a villain, they’re stopping an honest error from spreading.

It’s the quiet reason “agentic” tools feel safe enough to use. The boundary is doing work you never see.

Simple example

Think of a test kitchen that’s separate from the restaurant’s real kitchen. A new cook can try a wild recipe there, burn a pan, spill everything, and none of it touches the food going out to customers. If the dish works, you bring it over to the real kitchen on purpose. If it’s a disaster, you just clean the test kitchen and nobody ever knew.

An AI sandbox is that test kitchen, and letting the agent cook anywhere else is how accidents reach the dining room.

ChatGPT Plus costs $20 a month, and it’s been $20 a month since it launched in 2023. What’s less clear is what you’re buying. OpenAI now has six plans, and the differences are written in feature names that don’t explain themselves.

One thing before we start: everything below is accurate as of this morning, and AI pricing changes constantly. The specific numbers will drift, probably within months. The way to think about the decision holds up longer, so that’s the part worth remembering.

What the money buys

• The main thing you get are better models and more room. Free users get limited access to the fast everyday model. Plus unlocks the newer reasoning models at their more capable settings, along with much higher message limits before you hit a wall. So the $20 buys two things: access to smarter models and enough headroom to actually use them. There are still caps, but they’re set high enough that most people won’t meet them.

• No ads. OpenAI began testing ads on the Free and $8 Go tiers in the U.S. earlier this year. Plus and higher plans stay ad-free.

• Expanded Deep Research. This is the feature that reads and compares dozens of sources, then writes you a documented report with citations. Usage is limited, and ChatGPT shows your remaining allowance right inside the product, so you don’t have to guess.

• The working tools. Plus also includes the agent-style features: the coding assistant, the built-in browser, scheduled tasks, and the longer multi-step projects OpenAI now files under ChatGPT Work.

What it doesn’t buy

The $100 and $200 Pro tiers mainly buy capacity, five or twenty times the Plus allowance, though they also unlock OpenAI’s highest-end Pro reasoning model. Unless you regularly hit Plus limits or genuinely need that model for unusually hard work, the upgrade probably isn’t for you.

One quiet thing worth knowing: on personal plans, your conversations can be used for training unless you turn it off under Settings, then Data Controls. Business accounts get that exclusion by default. Individuals have to find the toggle, and most people don’t know it exists.

The math

$20 a month is $240 a year, which is real money for an app.

The prices above will change, but the test stays the same: do the free tier's limits actually interrupt you? If you keep hitting the caps mid-task, Plus might just pay for itself. If you check in a few times a week, Free is probably fine.

You’ve probably heard of prompt engineering by now, (I should hope so, we’ve covered it in this newsletter). Context engineering is the bigger, more useful cousin, and it’s kinda quietly become the thing that separates an AI agent that works from one that flails. As these agents take on real jobs, the hard part stops being how you word your request. It becomes what information you hand the model before it even starts thinking.

What it means

Context engineering is the practice of deciding what goes into a model’s context window, which is basically its short-term memory for the task in front of it. That means the instructions, the documents it needs, the earlier parts of the conversation, and whatever data it pulls in along the way.

You’re building the model’s entire picture of the problem. And here’s where it gets tricky. Give it too little and it starts guessing. Give it too much and it either loses the thread or runs up a huge bill, because every scrap of context costs money to process. The whole skill is figuring out the right stuff to include and leaving the rest out.

Why it matters

• Most agent screwups are context screwups, not brain screwups. The model didn’t get the one file it actually needed, or the thing it needed got buried under thirty files it didn’t. Nine times out of ten, fixing what the model sees fixes the output faster than rewording your prompt ever will.

• Context windows have a ceiling, and filling them isn’t free. Every token you stuff in is a token you pay for, on every single call. Picking the right ten documents instead of dumping in a hundred is the difference between a tool that’s cheap to run and one that quietly drains your budget.

• It’s turning into its own job. Teams building AI products now spend more time on what the model sees than on the model itself. That makes sense when you think about it. You can’t change how the model works, but you have total control over what you put in front of it. So that’s where the effort goes.

Simple example

A new hire starts Monday. Sharp person, learns fast, but they know nothing about how your company actually runs. You could drop every document you own on their desk and wish them luck.

Or you could pull the five files that matter for their first assignment and hand them just those. Same person. Same brain. The second version does better work, and faster, because somebody took a minute to think about what belonged on the desk and what didn’t.

When a model says something false, people reach for whichever word they heard first. It hallucinated, or it’s biased. And I get why, cause from the outside both look identical: you asked a question, the answer was wrong. But Hallucinations and Bias are different problems with different causes, and they need different fixes. Calling one by the other’s name usually means fixing the wrong thing.

The difference

A hallucination is a fabrication. The model invents something that doesn’t exist: a court case, a book, a statistic, a feature your product never had. It happens because these models are prediction machines. They generate whatever text is most likely to come next, and sometimes the most likely-sounding answer is a confident, well-formatted lie. There’s no intent behind it. The model doesn’t know it’s making things up, which is honestly the unsettling part.

Bias is a lean. The model’s answers tilt in some direction because the text it learned from tilted that way. If historical hiring data mostly shows men in engineering roles, a model trained on it may quietly rank male-sounding resumes higher. Nothing was invented. Every piece of the answer traces back to real patterns in real data. The problem is which patterns the data contained.

So a hallucination is the model making something from nothing, and bias is the model faithfully reproducing something you wish it wouldn’t. One is a reliability problem. The other is a mirror problem.

The fixes split the same way. Hallucinations get reduced by grounding the model in real documents and asking it to cite sources. Bias gets reduced by auditing the training data and testing the outputs across different groups. Grounding does very little for bias, and audits don’t stop fabrication. That’s why the vocabulary matters more than it seems like it should.

The one-sentence test

If the answer contains things that don’t exist, it’s a hallucination. If the answer is real but leans, it’s bias.

Simple example

Two unreliable coworkers. The first one invents details in meetings, confidently, with a straight face, and you can never quite predict when. The second one is accurate about everything but has strong opinions baked in from twenty years at one company, and every recommendation quietly points the same direction. You’d manage those two people completely differently. Same with the models.

In the mid-2000s, computer vision research was stuck. Everyone was tinkering with better algorithms and getting nowhere. Fei-Fei Li had a different theory: the models weren’t the bottleneck, the data was. She spent three years and a lot of professional credibility proving it, and the thing she built kicked off the deep learning era.

Who they are

Li was born in Beijing in 1976 and moved to New Jersey as a teenager. Her family ran a dry-cleaning shop, and she worked in it through high school while learning English. She studied physics at Princeton, got a PhD in electrical engineering from Caltech in 2005, and joined Stanford’s faculty in 2009. She directed Stanford’s AI Lab from 2013 to 2018, took a stint as Chief AI Scientist at Google Cloud, then came back to co-found the Stanford Institute for Human-Centered AI.

The thing she’s known for is ImageNet: a database of over 14 million hand-labeled images she started building in 2007, using crowdsourced workers around the world to tag every one. Colleagues told her it was a waste of time. In 2012, a team called AlexNet used deep learning on the ImageNet challenge and beat the field so badly that the entire research community changed direction within a year.

She’s now also CEO of World Labs, which she co-founded in 2024 to work on spatial intelligence. It raised $1 billion in 2026.

Why they matter

• ImageNet is one of the three legs modern AI stands on, alongside neural network research and cheap GPU compute. The 2012 result is the moment deep learning stopped being a niche and started being the whole field. Without a big enough benchmark to prove it on, that result doesn’t happen when it did.

• She’s been pushing on who gets to build this stuff since before it was fashionable. She co-founded AI4ALL in 2017 with Melinda French Gates and Jensen Huang to get AI education to high schoolers who weren’t going to stumble into it otherwise. Her “human-centered AI” framing keeps the focus on present-day harms rather than speculative ones, which puts her at odds with a fair chunk of the field.

• Her current bet is that language models have hit their ceiling. Text is flat. The physical world has depth, motion, and consequences. World Labs is building models that understand three-dimensional space, which is the piece robotics has been missing.

What they’ve said or done

In her 2015 TED talk, years before any of this was cool, Li said: “We dream of one day waking up a computer to see a world full of warmth, compassion, and love, but for now we need to teach it just to see.”

Time named her one of its 2025 Persons of the Year. The dry-cleaning shop is still in New Jersey.

Everybody has a rough picture of a data center: a warehouse full of servers, humming somewhere off a highway. That picture was accurate for about twenty years. The AI version is a different animal, and it's the reason your utility company is suddenly in the news.

What it means

A data center is a building full of computers that store data and run software for other people. A traditional one is mostly storage and web traffic: your email, your company’s files, the website you’re reading. Power draw is steady and modest.

An AI data center is built around racks of GPUs training and running models. Those chips pull far more electricity and throw off far more heat, so the whole building changes shape around them. Denser racks. Liquid cooling instead of fans. A power hookup that looks less like an office park and more like a small city. A traditional rack might draw 5 to 10 kilowatts. An AI rack can pull 100 or more.

That’s why the new ones get measured in gigawatts now instead of square feet. When OpenAI says it’s deploying chips at gigawatt scale, it’s describing a power budget, not a floor plan.

Why it matters

• The electricity is the real constraint, not the chips. Data centers used about 4% of US power in 2023. Projections put that closer to 9% by 2030, driven almost entirely by AI. In some regions, utilities can’t connect new capacity fast enough, and AI companies are signing deals for their own nuclear and gas generation to skip the line.

• This is why inference cost matters so much to the people running these models. Every response you get costs real electricity in a real building. That’s the whole argument behind custom chips like Google’s TPUs and OpenAI’s Jalapeño: shave the watts per answer, and the math on serving hundreds of millions of users changes.

• Where they get built is becoming a local political fight. Communities are pushing back on water use for cooling, on power prices, and on getting a warehouse instead of jobs. A data center employs maybe a few dozen people. It’s not a factory, and towns are figuring that out after the ribbon cutting.

Simple example

A restaurant kitchen with a few burners and an oven runs fine on a normal commercial power hookup. Now put in ten industrial pizza ovens that run at full heat, all day, every day. You aren’t just adding equipment. You need new wiring, a bigger service line from the street, and a ventilation system to move all that heat out of the room.

Same building. Same idea. But the power bill and the plumbing are somebody else’s problem now, and that somebody is the town.