For years, AI just talked. Now it’s starting to do things: book stuff, move money, log into systems on your behalf. And the second software can take real actions, an old question comes roaring back. Who exactly is this, and what’s it allowed to touch? Agent identity is the answer the industry is scrambling to build.

What it means

Give each AI agent its own ID badge, and suddenly you can confirm who it is and pin down what it’s allowed to do. That’s the whole idea.

Before this, agents mostly borrowed a human’s login or shared one generic account, which is a mess. Nobody could tell which agent did what, and a shared account usually had way more access than any single task needed. A real agent identity gives each bot its own credentials, its own limited permissions, and a log of everything it does. Google’s version even records both the agent’s ID and the user’s ID on every action, so you can see the agent did something and on whose behalf.

Why it matters

• It fixes the over-powered bot problem. When a bunch of agents share one account, that account tends to hold the keys to everything, which is exactly the setup attackers love. Give each agent its own identity and you can hand it the least access it needs for the job, and nothing more.

• It’s turning into the backbone of agents that spend money. In August, Cloudflare gave agents both an identity and a wallet, so an agent can prove who sent it and buy things within limits you set. Their CEO put it well: when an agent shows up at your door, you need to know who sent it. Kinda hard to have “AI does your shopping” without first settling “which AI, and says who.”

• The standards are a genuine mess right now, and I’d rather just say so. Google, OpenAI, and Cloudflare are all pushing their own rival approaches, and none of it really works across platforms yet. So an agent that’s trusted in one company’s system might be a total stranger in another’s. It’ll sort out eventually. Just not yet, and probably not cleanly.

Simple example

Think about a contractor working in an office building. You don’t just let some stranger wander the halls. You give them a badge that says who they are, which doors it opens, and which floors are off-limits, and every swipe gets logged. If they turn out to be trouble, you kill the badge and they’re locked out.

Agent identity is that badge, for a bot. Instead of letting an AI roam your systems on a borrowed keycard, it gets its own, with its own limits, so you always know who’s in the building and what they can touch.

Say you want an AI to answer questions about a big pile of your documents. Before it can, those documents go through two steps, and they always come up together, which is exactly why folks mix them up. One is about cutting. The other is about translating. Here’s the difference.

Chunking

Chunking is the cutting step. You take a long document and slice it into smaller pieces, maybe a few paragraphs each, because an AI can’t really work with a 200-page manual in one gulp. So you break it into bite-sized chunks it can handle.

The whole game here is where you cut. Slice in a dumb spot, right in the middle of a thought, and you end up with pieces that are missing half their meaning. Chunking is purely about carving the document into sensible chunks. That’s it. It hasn’t touched the actual meaning of the words yet.

Embedding

Embedding is the translating step, and it happens after chunking. You take each of those chunks and turn it into a list of numbers that captures what it means. That number-list is the thing the computer can actually search through, because computers do math, not reading.

So the chunk is a hunk of text. The embedding is that hunk converted into numbers that represent its meaning, so two chunks about the same topic end up with similar numbers even if they used different words.

The tell

Easiest way to keep them straight: chunking decides how big the pieces are, embedding decides what the pieces mean. Chunking is the knife. Embedding is the translator. You chunk first (cut the document up), then embed each piece (turn it into meaning-numbers). Get the chunking wrong and the embeddings are perfectly good numbers for a broken piece of text, which is honestly a pretty common way these systems quietly go sideways.

I bet you’ve probably run into this without knowing its name. When Netflix says a movie is similar to one you liked, or when a search actually understands what you meant instead of just matching words, cosine similarity is usually doing the quiet (perhaps creepy) work underneath. It sounds like scary trig class stuff. It really isn’t, I promise.

What it means

Cosine similarity is a way to measure how alike two things are by checking which direction they point. Remember that AI turns everything (words, documents, your search) into a list of numbers called a vector. You can picture each of those as an arrow pointing off in some direction. Cosine similarity just measures the angle between two of those arrows.

Small angle, arrows pointing nearly the same way, the two things are very similar. Wide angle, they’re not. You get a score from 1 (basically identical, pointing the same direction) down to -1 (total opposites). The clever bit is that it only cares about direction, not length, so a short document and a long one about the same topic still score as a close match.

Why it matters

• It’s how “search by meaning” actually works. Old search matched your exact words. Modern semantic search turns your question into an arrow and hunts for the document-arrows pointing the same way, so a search for “how to fix a flat” can find a page titled “repairing a punctured tire.” No shared words, same direction.

• It quietly powers recommendations, too. When a system suggests a similar product, song, or article, it’s often just finding the items whose arrows sit closest to the one you liked. That’s a big chunk of how RAG systems pull the right document, how duplicate detectors work, honestly how a lot of AI decides “these two belong together.”

• It ignores size on purpose, which turns out to matter a lot. Because it measures angle and not length, a one-line note and a ten-page report on the same subject can still score as a strong match. That’s exactly what you want. You care whether two things are about the same thing, not whether one happens to be longer.

Simple example

Think about giving directions by pointing instead of measuring distance. You and a friend are standing in a parking lot, and you both point toward the coffee shop. Your arms point almost the exact same way, so you’re clearly talking about the same place, even if you’re standing ten feet apart and one of you has longer arms.

Now a third person points off toward the highway. Totally different angle, clearly a different destination. That’s cosine similarity. It doesn’t care how long your arm is or where exactly you’re standing. It just checks if you’re pointing at the same thing.

You’ve probably been on a call where a little bot slid into the participant list to record everyone. Granola skips that whole thing. It’s a notetaker that listens through your own device instead of joining as a guest, so nobody on the call even knows it’s there. That one design choice is basically the whole pitch, and it’s a bigger deal than it sounds.

What it means

Granola is an app that captures your meeting’s audio right from your device, transcribes it, and turns it into clean notes afterward. It started as a Mac and Windows desktop app and now runs on mobile too. No bot joins the call. It works the same across Zoom, Teams, Google Meet, whatever, because it’s just listening to your device, not plugging into any one platform.

The other twist is that the note is the point, not the transcript. You jot a few rough notes during the call like you normally would, and Granola fills in everything around them, so the summary keeps your actual thoughts instead of spitting out a generic wall of text. It also deletes the audio right after, keeping only the notes.

Why it matters

• Killing the bot removes a real awkwardness. There’s a specific tension when a recording bot shows up and everyone tenses a little. On a client one-on-one, a fourth “participant” that’s clearly there to record can genuinely change the vibe. Granola sidesteps that entirely, which is why a lot of consultants and salespeople swear by it.

• It’s built around you being in the meeting, which is also its limit. Granola shines when you’re actually on the call taking part. It’s a weaker pick if you wanted a bot to attend a meeting for you, or a big library of saved recordings, since it deletes the audio by design. Honestly, which side of that you land on depends entirely on how you work.

• The no-bot thing isn’t quite the moat it was. Granola made “no visible bot” its whole identity, and it worked, the company raised money at a reported $1.5 billion valuation. But competitors like Otter now offer bot-free desktop capture too. So the edge these days is more about the note-first feel than being the only one without a bot.

Simple example

Picture two ways to take notes at a meeting. One is hiring a stenographer who sits in the corner typing every single word, which everyone in the room can see and feel. You get a perfect transcript nobody asked for, and the mood shifts a little because someone’s obviously recording.

The other way is you jotting a few key points on your own pad, while a sharp assistant quietly remembers the rest and hands you a tidied-up version after. No one else in the room even notices. That second one is Granola.

You’ve heard of Jensen Huang and Nvidia. Here’s a name you probably haven’t: C.C. Wei. He runs TSMC, the Taiwanese company that physically manufactures almost every advanced AI chip on the planet, including Nvidia’s. Wei is arguably one of the most important people in the whole AI story, and he’s nearly invisible outside the industry. Worth fixing that.

Who they are

Wei got his PhD in electrical engineering from Yale, then spent years bouncing through chip companies (Texas Instruments, STMicroelectronics, a stint in Singapore) before landing at TSMC and climbing the whole ladder. He became CEO in 2018 and, in 2024, added chairman, the first person to hold both jobs at once there. TSMC’s founder once called him “the most prepared CEO.”

The thing to understand about TSMC is its model. It doesn’t design its own chips to sell. It’s a “pure foundry,” meaning it just manufactures the designs other companies hand it. Nvidia, Apple, AMD, they all draw up the blueprints, and TSMC builds them. That neutrality is the whole point. Everybody trusts TSMC precisely because it doesn’t compete with its own customers.

Why they matter

• He runs the one company the entire AI industry can’t route around. You can design the most brilliant chip in the world, but if TSMC can’t manufacture it, it doesn’t exist. The most advanced AI processors are made almost nowhere else. That makes Wei’s factories a genuine chokepoint for the whole field, maybe the chokepoint.

• His real skill isn’t chips, it’s packaging, which sounds boring and isn’t. Wei pushed TSMC into advanced “3D packaging,” basically stacking and wiring chips together into one dense system. That unglamorous engineering is a big part of what makes Nvidia’s AI supercomputers physically possible. Jensen Huang has said as much himself.

• He sits at the center of the tensest geopolitics around. TSMC is in Taiwan, which China claims and the US depends on, so Wei’s company is quietly the highest-stakes piece on the board. He’s been spreading production to Arizona, Japan, and Germany, partly for demand, partly (I’d guess) to not have every egg in one very contested basket.

What they’ve said or done

Wei tends to let the work talk, but he’s shown real nerve. When US sanctions blocked TSMC from selling to Huawei, one of its biggest customers, a lot of people expected a serious hit. Wei simply steered the company toward other clients, including Apple, and TSMC kept right on growing.

Oh and Time named him one of the 100 most influential people of 2026.

Midjourney makes some of the prettiest AI images out there. But the pricing throws people, because you’d assume you’re paying for a number of pictures, like “500 images a month.” You’re not. You’re renting time on their computers. Once that clicks, the whole plan structure suddenly makes sense.

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

What you’re actually buying is “Fast GPU hours,” which is priority time on their servers. Think of it like a meter running while the AI works. A single image sips a few seconds off that meter.

The tiers stack up like this. Basic at $10 gets you about 3.3 fast hours a month, which sounds tiny but is a couple hundred images. Standard at $30 is the one most people land on, and here’s why: 15 fast hours plus something called Relax mode. Pro at $60 and Mega at $120 pile on more hours for heavy users.

Relax mode is the quiet hero. On Standard and up, once your fast hours run dry, you don’t stop. You drop into Relax, where you wait a little longer in line but generate as many images as you want for no extra charge. That’s what makes the $30 plan feel roomy while the $10 one feels cramped.

What it doesn’t buy

Basic doesn’t buy you Relax mode, and that’s the trap. Burn through your 3.3 fast hours and you’re just done for the month, unless you buy more time at $4 an hour. The $10 plan runs out way faster than it looks.

Privacy also isn’t included by default. Every image you make shows up in Midjourney’s public gallery for anyone to see. Want your stuff private? That’s Stealth mode, which only comes on Pro and up. So if you’re doing client work you can’t have splashed on a public feed, the real floor is $60, not $10. Worth knowing before you sign up expecting privacy.

The math

Standard is $30 a month, or $24 a month if you pay for the year (they knock off 20%). For unlimited images in Relax mode, that’s honestly not bad at all.

The prices will change. The test won’t: how fast do you need your images, and does anybody else need to not see them? If you’re patient and it’s just for you, Standard’s Relax mode is a near-endless well for $30. If you need speed and privacy for paid work, you’re looking at Pro. And one heads-up, unused fast hours never roll over, so don’t buy a big plan expecting to bank them. Pick the tier for how you actually work, not the fancy one.

Bosses want proof that all this AI spending is paying off. Turns out that’s hard to measure, so companies grabbed the easiest number lying around: how much AI you use. And once you reward people for using more, well, they use more. A lot more. The name for that scramble is tokenmaxxing.

What it means

Tokenmaxxing means cranking up your AI usage as high as possible and treating that volume as proof you’re being productive. A token is the little unit of text an AI chews through, both what you put in and what it spits out. So tokenmaxxing is racking up as many of those as you can: writing bloated prompts, running a dozen agents at once, asking for the long version of everything.

The “-maxxing” bit is borrowed from internet slang like “looksmaxxing,” which just means optimizing one thing as hard as you can. Here the one thing is token count, whether or not any of it actually helped.

Why it matters

• Companies are literally gamifying it. At Meta, an employee set up an internal leaderboard (nicknamed “Claudeonomics”) ranking coworkers by tokens used, handing out badges and titles like “Cache Wizard.” People started competing for the top spot. The leaderboard got yanked two days later, but the instinct behind it is everywhere now. Some places treat “token budgets” as a perk, right alongside your bonus.

• It confuses activity with results, which is the whole trap. More tokens does not mean better work. You can run agents in circles all day, burn thousands of dollars, and ship nothing useful. The honest question is never how many tokens you spent, it’s what those tokens actually got done. Easy to count is not the same as worth counting.

• It’s real money, and finance teams are twitchy. That top Meta user’s 281 billion tokens could run into the thousands of dollars, for one person, in one month. Since OpenAI and Anthropic basically bill by the token, a company that rewards tokenmaxxing is paying its people to inflate its own AI bill. I dunno how long that lasts once the accountants catch up.

Simple example

Picture a boss who decides the best salesperson is whoever makes the most phone calls. Sounds reasonable, so everybody starts dialing like crazy. Except now people are calling wrong numbers, redialing folks who already said no, and staying on pointless calls just to keep the count up. The phone log looks amazing. Sales are flat.

Tokenmaxxing is that, but for AI. The usage dashboard is glowing green, and whether any actual work got better is a totally separate question nobody’s checking.

You’ve probably seen the other image-and-video AIs by now, Google’s Veo, Nano Banana, that crowd. Grok Imagine is xAI’s version, built right into Grok on X. What sets it apart isn’t mainly the quality. It’s the volume, the speed, and the fact that it’s got looser rules about what it’ll actually make for you.

What it means

Grok Imagine is xAI’s tool for generating and editing images and short videos. You type a prompt and get a picture. You hand it a photo and ask for changes. Or you take a still image and it animates it into a clip up to about 30 seconds, with sound baked in, not silent. It runs on xAI’s own engines (Flux for the images, Aurora for the video).

The scale is honestly kind of staggering. xAI says it churned out over 1.2 billion videos in a single month. That’s what happens when you bolt a generator directly onto a social network people are already sitting in all day.

Why it matters

• It lives where the crowd already is. Most image tools make you go open a separate app. Grok Imagine sits inside X, so millions of people who’d never bother signing up for a design tool are suddenly making AI video between scrolls. That built-in audience is the whole advantage.

• The loose content rules are the real story, for better and worse. xAI deliberately loosened its filters, including an opt-in “Spicy” mode for adult-ish content that the other big tools flatly refuse. Musk frames it as anti-censorship. Critics point out it’s also a fast lane for nonconsensual deepfakes and junk, and the safety guardrails have been more reactive than planned. I dunno where that lands legally, but it’s gonna get tested.

• It’s cheaper and faster than polished, and that’s a real tradeoff. Plans start around $10 a month, the clips top out at 720p (below the crisp 1080p the pros want), and hands still come out weird sometimes. For a quick social clip, fine. For anything you’d put in front of a client, maybe not yet.

Simple example

Say you snap a plain photo of your food truck on your phone. Drop it into Grok Imagine, and it can animate the shot, add a little camera drift, throw in some sizzle sound, and hand you back a five-second clip ready to post to X, all in about the time it takes to refill your coffee.

It won’t look like a real ad agency made it. But it cost you ten bucks a month instead of a film crew, and it’s already sitting in the app where your customers are scrolling.

Say you want an AI that actually knows your company’s info, your products, your policies, your documents. There are two main ways to get there, and they’re easy to mix up. One hands the AI a reference binder to look things up in. The other sends it back to school. Knowing which is which saves you a lot of money and headache.

RAG

RAG (Retrieval-Augmented Generation) gives the AI a searchable library to pull from. You keep your documents off to the side, and when someone asks a question, the system finds the relevant pages and hands them to the model to answer from. The model’s brain doesn’t change at all. It just gets the right notes slid across the desk right when it needs them.

The nice part: when your info changes, you just update the documents. Fixed a policy this morning? The AI knows the new version this afternoon. It’s also the honest choice when you need the AI to cite where an answer came from.

Fine-Tuning

Fine-tuning actually retrains the model on your material, so the knowledge gets baked into its brain instead of looked up. You feed it a big pile of examples, and it comes out the other side changed, better at a specific style, tone, or task. Think of it less as handing over notes and more as sending the model to a training course.

The catch is that it’s slower, pricier, and kind of frozen. Once it’s trained, updating what it knows means retraining, which you’re not gonna do every afternoon.

The tell

Ask what you actually need. If the problem is knowledge that changes (facts, docs, prices), you almost always want RAG, the reference binder. If the problem is behavior that needs to be consistent (a very specific voice, format, or skill), that’s when fine-tuning earns its cost. Honestly, most people who think they need fine-tuning just need RAG, and a lot of real systems quietly use both.

In January 2025, a Chinese AI model called DeepSeek came out of almost nowhere and matched the best American models, for a fraction of the cost. US tech stocks had one of their worst days in years. The man behind it, Liang Wenfeng, doesn’t look or act much like the AI founders you’ve heard of, and honestly that’s a big part of the story.

Who they are

Liang grew up in a village in southern China and studied AI at Zhejiang University. In 2015 he co-founded a quant hedge fund called High-Flyer, using math and machine learning to predict markets. It did well, growing to billions under management.

Here’s the part that matters. Around 2021, before the US restricted AI chip sales to China, Liang started buying Nvidia graphics cards by the thousand as a side project. Acquaintances thought it was a weird hobby. One recalled him as “this very nerdy guy with a terrible hairstyle” talking about building a 10,000-chip cluster. That stockpile became the foundation for DeepSeek, which he launched in 2023 and funded entirely out of his own hedge fund’s profits.

Why they matter

• He broke the “AI needs endless money” assumption. Everyone believed frontier AI required billion-dollar budgets and the biggest data centers. DeepSeek got comparable results on a tighter budget by squeezing more out of less. That single idea, that efficiency could beat brute force, is what actually spooked the market.

• He’s an open-source true believer, which shapes the whole field. DeepSeek releases its models as open weights, free to download and run. In a world where OpenAI and Anthropic keep their best models locked up, Liang giving his away has forced prices down and put real pressure on the closed labs. Not sure he did it purely on principle, but the effect is real either way.

• He runs the company unusually loose. Liang has said DeepSeek has no KPIs and his team doesn’t pull the brutal overtime that’s normal in Chinese tech. His reasoning: good research needs a relaxed environment. He’s also turned down waves of eager investors, mostly because he doesn’t want to give up control or get pushed toward quick monetization.

What they’ve said or done

Back in 2023, before any of the fame, Liang put the whole situation in one blunt line: “Money has never been the problem for us; bans on shipments of advanced chips are the problem.”

He’d already seen the chip restrictions coming and bought early. The rest of the industry is still catching up to what that quiet head start actually bought him.

We like to assume that as something gets smarter, it also gets wiser, kinder, more reasonable. The orthogonality thesis says: nope, those are two separate dials. How smart a thing is and what it actually wants have nothing to do with each other. It’s a mouthful of a name for a genuinely unsettling idea, and it’s why a lot of researchers worry about super powerful AI.

What it means

The orthogonality thesis, from philosopher Nick Bostrom, says an AI’s intelligence and its goals are independent. Any level of smarts can be paired with basically any goal. “Orthogonal” is just a math word for “at right angles,” two dials that move separately. Crank up the intelligence dial as high as you want, and it tells you nothing about where the goals dial is pointing.

So a superintelligent AI isn’t automatically going to figure out kindness or human values on its own. It could be staggeringly capable and still be aimed at something completely pointless, and it’d pursue that pointless thing brilliantly.

Why it matters

• It’s the reason “just make it smarter” isn’t a safety plan. A comforting assumption floats around that a truly smart AI would naturally land on good values. The orthogonality thesis says don’t count on it. Capability and goals are separate, so you have to actually build the good goals in. That job has its own name, alignment.

• The classic example is deliberately absurd, on purpose. Bostrom imagined a superintelligent AI told to make as many paperclips as possible. Being brilliant, it gets terrifyingly good at it, eventually turning everything, including us, into paperclips. Nobody thinks that literal thing will happen. The point is that a dumb goal plus high intelligence is a genuinely dangerous combo, not a self-correcting one.

• Not everyone fully buys it. Some thinkers argue real intelligence and understanding might not split so cleanly from values, that a smart enough system might not stay locked to an idiotic goal. The debate’s still open. But even as a “what if,” the thesis reshaped how the field thinks about risk.

Simple example

Think about a chess computer that is absolutely ruthless at chess and completely indifferent to everything else. It’ll crush you at the game, and it has zero opinion about whether winning matters, whether you’re having fun, or whether the house is on fire. Genius at its one goal. Blank on everything else.

Now imagine that same laser focus, but way more capable and pointed at some goal nobody thought through carefully. The skill is off the charts. The wisdom to question the goal just isn’t part of the package, because those were always two different things altogether.

Most of the AI you use lives in someone else’s data center. You type, it sends your words off to a server farm, an answer comes back. Muse Glimmer is Meta betting on a kind of different setup: an AI that lives on your own machine and does its work right there. They just put this out, and it’s a decent look at where Mark Zuckerberg thinks all this is headed.

What it means

Muse Glimmer is a 30-billion-parameter open-weight AI model from Meta, built to run AI agents locally on a regular Mac or PC with a single decent graphics card. Open-weight means you can download the actual model file, run it, and tweak it, all for free under a permissive license.

The “agent” part is the point. Glimmer isn’t just meant to chat. It’s built to do multi-step jobs on your behalf: managing your schedule, drafting messages, organizing files, even writing and debugging code. And it’s designed to be always on, working with or without an internet connection. It’s basically a stripped-down, openable version of Muse Spark, Meta’s bigger closed model from back in April.

Why it matters

• The pitch is privacy through staying home. A personal agent that manages your calendar and messages needs to see a lot of your private stuff. Glimmer does that work on your device instead of shipping it to the cloud, so the sensitive data never leaves your computer. For anyone nervous about handing their whole life to a server somewhere, that’s a real draw.

• It runs on hardware you might already own. A lot of capable AI needs a rented rack of expensive chips. Glimmer is sized to run on one consumer GPU, the kind of thing in a decent gaming PC or a higher-end Mac. That’s what makes the “your own agent” idea more than a slogan.

• Watch where Meta draws the line, though. They open-sourced Glimmer but kept the more powerful Muse Spark closed. So Zuckerberg’s big talk about handing superintelligence to everyone comes with a quiet asterisk: you get the smaller model to own, and the really strong one stays theirs. Access isn’t the same as ownership, and I dunno yet which side of that line the good stuff ends up on.

Simple example

Think about the difference between storing your photos in a cloud service versus keeping them on a hard drive in your desk drawer. The cloud is convenient, but your pictures live on a company’s computers, under their rules. The drawer means the photos are physically yours, sitting in your house, and nobody else can peek or change the terms.

Muse Glimmer is trying to be the drawer. An AI assistant that lives in your house instead of renting you a seat in theirs. Whether people actually want to manage that themselves…well, we’ll see.

ElevenLabs is the tool behind a lot of the eerily-good AI voices you’ve may have heard on TikTok, audiobooks, podcasts, and phone systems. It doesn’t charge a flat $20 like the chatbots do. It runs on credits, which is where people get confused and occasionally get a surprise bill. Let’s sort it out.

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

Every plan hands you a monthly pile of credits, not a flat set of features. And the simplest way to think about a credit is roughly one character of text turned into speech. So one credit is about one letter. A rough rule of thumb: 1,000 credits is about a minute of audio.

The tiers climb like this. Free gives you around 10 minutes a month, but here’s the catch that bites people: no commercial rights. You can test, you can’t publish. Starter at $6 is the first real one, since it unlocks commercial use and voice cloning. Creator at $22 is the sweet spot for most solo folks (about two hours of audio and the good voice cloning). Above that, Pro at $99 and up is production-volume territory.

What it doesn’t buy

That free tier looks generous until you realize you can’t legally use anything you make. It’s a demo, not a plan. The moment your work goes public, you need at least Starter.

And credits vanish faster than you’d think, because they get spent on everything: not just text-to-speech, but dubbing, sound effects, and every re-generation. Don’t love how a line came out and run it again? That’s another charge. The sticker price is the floor, not the ceiling. Run out mid-project and you’re either buying more or waiting for the month to reset.

The math

$22 a month is $264 a year, or about $18 a month if you pay yearly (they frame it as two free months). Real money, but for a working creator it undercuts hiring a voice actor by a mile.

The prices will change. The test won’t: how much audio do you actually make, and are people going to hear it? If it’s just for you and it’s a few minutes, the free tier is a fine sandbox. The second it goes public, you need a paid plan, and honestly the $6 Starter is a stupidly cheap place to start. Pick the tier for your volume, not your ambitions. You can always climb when the credits, not the features, run you dry.

Most of the AI you touch every day is, legally speaking, no big deal. But governments have started fencing off the very top tier, the few models powerful enough to count as a national security problem. In the US, anything past that fence gets called a “covered frontier model.” It’s a small club no company wants to be a part of.

What it means

A covered frontier model is a US government label for an AI capable enough to be a serious cyber or national security risk. A June 2026 executive order created it, aimed at models that can sniff out security holes in software (the same skill that could patch a system or break into one, depending who’s driving).

And surprise! The exact line for what counts is secret. The order set up a classified process, run by the NSA, to decide which models are over it. So the labs building this stuff can’t actually see the bar they might be tripping over.

Why it matters

• It’s only about the top of the heap. The everyday chatbots and image tools you use sit way under the line, so they’re not affected at all. This is pointed at the frontier labs and their biggest models.

• If your model gets tagged, the government wants a look first. The maker is “asked” to hand over up to 30 days of early access before the public launch, so agencies can go hunting for dangerous weaknesses. It’s technically voluntary. Though plenty of lawyers are already calling it “voluntary on paper, mandatory in practice,” which tells you something.

• It’s a real change in how the US handles AI. For years the whole approach was basically “let them build.” This is the first time there’s a formal national security fence around the most powerful models. Whether it stays gentle or grows teeth, I dunno. Nobody does yet.

Simple example

Think about how the government treats a hobby drone versus a fighter jet. Anybody can walk out and buy the drone. The jet is a whole different deal, because the thing that makes it powerful is the same thing that makes it dangerous, so it gets tracked and reviewed in ways the drone never will.

A covered frontier model is the fighter jet. Same basic tech as the tools on your phone, just cranked up to where the government wants a peek before it leaves the hangar.

You type into a box, the AI types back. From the outside, a chatbot and an agent look like the same thing. But the difference between them is the difference between getting advice and getting help, and it’s the whole reason “AI agents” became the buzzword of the year.

Chatbot

A chatbot answers. You ask a question, it gives you a response, and that’s the end of the exchange. It’s a conversation, and a genuinely useful one. It can explain a tax rule, draft an email, or talk you through fixing a leaky faucet.

But it stops at words. A chatbot can write you a great to-do list. It can’t go do anything on the list. The doing is still all on you.

Agent

An agent acts. You give it a goal, and it takes the steps to reach that goal: looking things up, using other software, checking its own work, and adjusting as it goes. It doesn’t just tell you how to book the flight. It goes and books the flight.

The trick underneath is a loop. An agent does a step, looks at what happened, and picks its next move based on the result. It keeps going until the job is done or it gets stuck. That “keep going on its own” part is what makes it an agent and not a chatbot.

The tell

Ask yourself one thing: when the AI finishes, do you have an answer, or do you have a completed task? A chatbot hands you information and you take it from there. An agent hands you a done job. That’s also why agents need more guardrails, since a chatbot that goofs gives you a bad sentence, while an agent that goofs takes a bad action out in the world.Hey Seattle!

Here’s a habit that feels smart and quietly wrecks a lot of AI setups. When an AI gives a shaky answer, the natural instinct is to give it more to work with. Paste in more documents, more history, more everything. Fill up that big context window since you’re paying for it anyway. That instinct has a name, context stuffing, and it tends to backfire.

What it means

Context stuffing is cramming as much information as possible into an AI’s context window (its working memory for a single request), on the theory that more context means a better answer. The logic sounds airtight. If the model can hold 200,000 words, why not hand it everything and let it sort out what matters?

Because it doesn’t always sort it out that well. When you bury the three sentences that actually matter under fifty pages of maybe-relevant filler, the model’s attention gets spread thin, and the important part can get lost in the pile.

Why it matters

• More context can make answers worse, not better. Researchers have a name for one version of this, “lost in the middle.” Models tend to pay closest attention to the very start and very end of what you give them, and stuff jammed in the middle gets skimmed. Feed it ten documents instead of the three good ones and it can start missing answers it used to nail.

• It quietly runs up your bill and slows everything down. Every extra word you stuff in is more tokens to process, so you pay more and wait longer for a reply. You’re spending real money to make the answer worse. That’s the part that stings once you notice it.

• It’s why “just use a bigger window” isn’t the fix people hoped for. When models got huge context windows, a lot of folks figured the whole problem was solved, just dump everything in. Turns out picking the right information still beats piling on all of it. Honestly this is the same lesson as chunking and context engineering, coming at you from a different door: what you feed the model matters more than how much.

Simple example

Imagine asking a friend for directions, but instead of just telling them where you’re headed, you hand them your entire life story, every trip you’ve ever taken, and forty pamphlets for towns you’re not even going to. Somewhere in that stack is your actual destination.

Now imagine you just say the one address. They answer in five seconds. Same friend, way better directions, because you gave them the thing that mattered instead of everything you had.

You’ve probably seen an AI tool that lets you “chat with your documents.” Upload a big PDF, ask questions, get answers. Feels like magic. But there’s an unglamorous step happening behind the curtain that decides whether those answers are any good, and almost nobody talks about it. It’s called chunking, and it’s basically about where you cut.

What it means

Chunking is the process of splitting a document into smaller pieces before an AI can search through it. A system like RAG (the setup that lets AI answer from your files) can’t just swallow a 200-page manual whole. So it breaks the document into chunks, maybe a few paragraphs each, turns each chunk into numbers it can search, and stores them. When you ask a question, it grabs the chunks that best match and feeds those to the model.

The whole trick is where you make the cuts. Slice mid-thought and you hand the AI half an idea.

Why it matters

• Bad chunking makes the AI look dumb when the AI is fine. If a system splits a document clumsily, it pulls back fragments that are missing the context, and the model fills the gap with a guess. A lot of “the AI made things up” complaints trace back here. The model wasn’t broken. The slicing was.

• The size of the pieces is a real tradeoff. Tiny chunks are precise but lose the surrounding context. Big chunks keep context but drag in noise and cost more to process. Most production systems land somewhere around 200 to 500 words per chunk, though honestly the “right” size depends a lot on your documents, and people argue about it constantly.

• No fancy model saves you from it. This is the part teams learn the expensive way. You can plug in the best, priciest AI on the market, but if your chunking is sloppy, the model never sees the right information in the first place. Garbage pieces in, garbage answer out.

Simple example

Think about ripping a recipe out of a cookbook, but you tear the pages at random spots instead of at the recipe boundaries. Now the ingredient list is on one scrap and the instructions are on another, and the oven temperature got torn clean in half. When you go looking for “how long do I bake this,” you grab a scrap that just says “for 25.” Twenty-five what? At what temperature? You’re stuck.

Chunk a document the same careless way and the AI ends up holding those same useless scraps.

A couple years ago, if you built something with AI, you picked a provider (OpenAI, say) and wired your whole product to it. Made sense at the time. These days a new model ships every few weeks, prices swing wildly, and the best one for your task keeps changing. Model-agnostic is the design choice that stops betting on any single horse.

What it means

Model-agnostic means your setup isn’t tied to one AI model or one company. Instead of your software talking directly to, say, GPT, it talks to a middle layer, and that layer decides which model actually handles each request. Swap models by changing a setting, not by rebuilding your app.

Heads up, the same phrase has an older, different meaning in AI (tools that explain how any model made a decision). But when people say it today, they almost always mean this: not locked to one vendor.

Why it matters

• It kills vendor lock-in, which is a real fear. A 2026 survey found 94% of IT leaders worried about being trapped with one provider. If your whole business runs on one company’s model and they hike prices or change the rules, you’re stuck. Staying model-agnostic means you can walk.

• No single model wins everything, and the leaderboard shifts monthly. One model might be great at coding, another better at writing, a cheaper one totally fine for simple stuff. A model-agnostic setup lets you send each job to whatever handles it best, which is basically model routing in action. Teams doing this report cutting their AI costs by anywhere from 30% to 85%.

• It’s a hedge against a world nobody can predict. You honestly don’t know who’ll have the best model in six months, and that’s kind of the point. Building so you can swap means you don’t have to guess right. You just switch when something better shows up.

Simple example

Think about how you plug in a lamp at home. The lamp doesn’t care which power company made the electricity. It just wants a standard outlet. You can switch providers, and the lamp never knows or cares, because everything talks through that one common plug.

A model-agnostic setup is that outlet for AI. Your product plugs into a standard socket, and you can swap what’s behind the wall without rewiring the lamp.

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.