For decades, if you wanted an app, you needed a developer. If you wanted a website, you needed a developer. If you wanted to automate a workflow, you needed a developer. No-code flipped that. And now AI is flipping it again.

What it means

No-code is a software development approach that lets people build applications, websites, automations, and workflows using visual interfaces instead of writing programming code. You drag and drop components, connect them with logic, and publish. No coding language required. Low-code is the close cousin: it offers the same visual tools but lets developers add custom code when they need more control. Platforms like Zapier, Airtable, Bubble, and Webflow popularized the approach. Gartner predicts that by the end 2026, 75% of all new applications will be built using low-code or no-code tools.

Why it matters

• It put software creation in the hands of the people who know the problem best. The marketing manager who needs a lead tracker doesn’t have to write a requirements document and wait three months for engineering to build it. She builds it herself in an afternoon. Nearly 60% of custom business apps are now built outside IT departments, and 30% of those are built by people with little to no technical background.

• AI is eating no-code from below. Traditional no-code used drag-and-drop editors and visual builders. Now tools like Replit, Lovable, and Bolt.new let you describe what you want in plain English and get a working app back. Some industry observers argue that vibe coding is making the old drag-and-drop version of no-code obsolete. Why arrange blocks when you can just say what you want?

• The tradeoffs are real. No-code apps are fast to build but hard to scale. They work great for internal tools, prototypes, and MVPs. They struggle with complex logic, large datasets, and enterprise-grade security. Most serious production apps still get rebuilt in code once the prototype proves the idea works. No-code gets you from zero to one. Code gets you from one to a hundred.

Simple example

You need a bookshelf. One option: learn carpentry, buy lumber, and build it from scratch. Total control over every detail. Another option: buy a modular shelf from IKEA. Pick the pieces, snap them together, done. It won’t hold a grand piano, but it holds your books and it’s up in an hour.

No-code is the IKEA shelf. You pick the pieces, connect them, and ship. You didn’t learn carpentry. You didn’t need to. And if the shelf needs to hold something heavier later, you can always call a carpenter.

Before you go, I’ve managed to partner up with some folks and get in on this giveaway. It’s for two nights at The Elser Hotel in downtown Miami, two Miami Heat tickets, and a $350 luggage gift card. Pack light, sit court-side, sleep well.

If you are interested:

ENTER HERE

People spend a lot of time debating which AI model is the smartest. But a smart model with no tools, no memory, and no access to your systems is just a brain in a jar. Harness engineering is the work of giving that brain hands, eyes, and a desk to sit at.

What it means

Harness engineering is the practice of designing the systems, tools, and scaffolding that surround an AI model so it can actually do things in the real world. The model itself generates text, writes code, and makes decisions. The harness gives it access to files, databases, APIs, web browsers, and other software. It manages what the model can see, how it breaks tasks into steps, and how its work gets checked. Without a harness, a model can answer questions. With one, it can finish projects.

Why it matters

• It reframes the AI conversation from “pick the best model” to “build the best environment.” Two companies can use the exact same model and get wildly different results based on the harness around it. The company with better tools, cleaner context management, and tighter feedback loops wins, even if their model scores lower on benchmarks.

• A good harness has three layers. The information layer controls what the model can see and access: memory, context, tools, and skills. The execution layer manages how work gets broken into steps and how multiple agents coordinate. The feedback layer checks the output, verifies quality, and feeds corrections back into the system. Skip any of the three and the agent breaks down.

• There’s an inner harness and an outer harness. AI labs build the inner one: the sandboxed environment, the code execution engine, the web search. You build the outer one: the custom instructions, the project context, the skills, the specific workflow that makes the model useful for your job. The outer harness is where the real competitive advantage lives, and it’s the part nobody can copy because it’s built on your knowledge.

Simple example

A surgeon is skilled, but surgery doesn’t happen because the surgeon is smart. It happens because the operating room is set up: sterile instruments laid out in order, monitors displaying vital signs, a team managing anesthesia, and a process for what happens if something goes wrong. The surgeon is the model. The operating room is the harness. Take the same surgeon and drop them in a parking lot with no tools, and the skill is still there but nothing gets done. Harness engineering is building the operating room.

You’ve probably felt this. You’re 30 messages into a conversation with a chatbot and it starts ignoring things you told it earlier. A constraint you set disappears. An idea you rejected shows up again.

The answers are still polished, but something is off. The model isn’t broken. It’s drifting.

What it means

Context drift is the gradual loss of coherence that happens as an AI conversation gets longer. The model starts strong, remembering your instructions, building on earlier decisions, and staying on topic. But as the conversation grows, earlier information gets crowded out by newer messages. The model doesn’t forget in the way a person does. It still has access to the earlier text. But as the context window fills up, older tokens compete with newer ones for the model’s attention, and the older ones lose.

Why it matters

• It’s the most common reason people think a chatbot “got dumb.” The model didn’t change. The conversation got too long. Research shows that after just one bad interaction, 30% of users are more likely to switch brands, abandon a purchase, or share negative feedback. Drift erodes trust faster than most companies realize.

• It’s different from hallucination. A chatbot that asks for your order number again after you already gave it is drifting. A chatbot that tells you your package was delivered by a unicorn is hallucinating. Both are wrong, but the causes and fixes are different.

• The fix is counterintuitive: stop and restart. When a model starts reintroducing rejected ideas or dropping constraints you set earlier, the best move isn’t to argue with it for ten more turns. It’s to start a fresh conversation with a clean summary of where you are. Projects and custom instructions help because they reload your core context at the start of every new chat, so you don’t lose the foundation even when you reset the thread.

Simple example

You’re giving directions to a friend over the phone. For the first few turns, they’re sharp. Left on Main, right on Oak, past the school. But the call goes on for 45 minutes with detours, backtracking, and side conversations about lunch. By the time you say “take the next left,” they’ve lost track of where they are and which direction they’re facing. They’re still listening. They still heard every word.

But the conversation got too long for them to hold the whole route in their head. Context drift is that moment when the AI is still responding, but it’s lost the shape of what you were building together.

You’ve been reading about AI models, companies, and concepts in this newsletter for months. This week, we’ve been going deep on one product: Claude, the AI built by Anthropic. Since we’ve gone over a good chunk of its features, I thought we’d take a step back and understand what Claude actually is, where it came from, and why its creators think it should work differently from the competition.

What it means

Claude is a family of large language models built by Anthropic, the AI safety company founded in 2021 by siblings Dario and Daniela Amodei, both former OpenAI executives. The name is a tribute to Claude Shannon, a mathematician who published “A Mathematical Theory of Communication” in 1948 and is widely considered the father of information theory. Shannon’s paper, cited over 160,000 times, laid the groundwork for how we encode, transmit, and decode information, which is essentially what a language model does at scale.

The name also served a second purpose: most virtual assistants (Alexa, Siri, Cortana) had female-sounding names, and Anthropic deliberately chose a male name to break the pattern of gendering AI assistants as women.

Why it matters

• Claude is trained using Constitutional AI, a method Anthropic invented. Instead of relying entirely on human reviewers to rate outputs (the standard RLHF approach), Claude is given a written constitution of principles and learns to evaluate its own responses against those rules. The 2026 version of that constitution is 23,000 words long, up from 2,700 in 2023. It’s essentially a rulebook for how to be helpful without being harmful.

• Anthropic has drawn a line that other labs haven’t. The company includes contractual prohibitions against using Claude for mass domestic surveillance and fully autonomous weapons. When the Pentagon pushed back on those restrictions, labeling Anthropic a “supply chain risk,” Anthropic took it to court rather than remove the language. Over 200 employees from Google and OpenAI signed letters supporting Anthropic’s position.

• Claude is now a platform, not just a chatbot. What started as a conversational model in 2023 has grown into an ecosystem: Claude Code for developers, Cowork for desktop automation, Dispatch for mobile task management, Skills for custom workflows, Artifacts for interactive content, Projects for persistent workspaces, and Connectors for linking to external services. This week, we’ll cover each of those.

Simple example

You hire a new employee. Most companies hand them a laptop and say “figure it out.” Anthropic wrote a 23,000-word employee handbook before the employee’s first day, covering how to handle sensitive questions, when to push back, and what lines never to cross. Then they gave that handbook to the employee (Claude) and said “internalize this.” That’s Constitutional AI.

The model doesn’t just learn what to say from examples. It learns why certain answers are better than others, based on written principles it can reference. The handbook isn’t just a suggestion. It’s the architecture.

The best ideas for what to delegate to AI don’t hit you when you’re sitting at your computer. They hit you in the car, in line at the grocery store, or walking into a meeting you forgot to prepare for. Dispatch is built for that moment.

What it means

Dispatch is a feature inside Claude Cowork that lets you assign tasks to Claude from your phone, and Claude completes them on your desktop computer while you’re away. You scan a QR code to pair your phone with your desktop, and from that point, you have a single persistent conversation thread that syncs between both devices.

Ask Claude from your phone to compile a report from a local spreadsheet, draft a briefing from your email and Slack, or build a presentation from files in Google Drive. Claude uses everything already configured in Cowork (your connectors, plugins, and file access) without any additional setup. Your desktop needs to be awake and the Claude app needs to be open.

Why it matters

• It separates where you are from where your work is. Your files, your apps, your connected services are all on your desktop. Dispatch means you don’t have to be. Send a task from your phone on the train, and come back to finished work when you sit down.

• The conversation is persistent. Unlike regular chats that reset, Dispatch maintains a single thread that carries context across tasks. Claude remembers what you asked yesterday and picks up where you left off. You don’t re-explain your project every time.

• It’s early and honest about its limits. MacStories reported that initial testing landed at about a 50/50 success rate on tasks. Summarizing and finding data worked well. Sharing outputs was unreliable. Anthropic released it as a research preview, not a finished product, and is updating it regularly. That transparency matters in a market where most companies announce features as polished when they’re not.

Simple example

You’re running late for a meeting and realize you never prepped the slide deck. You pull out your phone at a red light and type “export the pitch deck as a PDF and attach it to my 10am calendar invite.” By the time you park and walk in, Claude has opened the file on your desktop, converted it, and attached it to the meeting.

You didn’t touch your laptop. You didn’t remote in. You told your AI what you needed, and it handled the rest on your machine while you drove.

Chatbots answer questions. They give advice. They write drafts. But they can’t open a spreadsheet on your desktop, rename 40 files in your downloads folder, or pull data from three different apps and compile a report. Cowork crosses that line.

What it means

Claude Cowork is an agentic AI system built into the Claude desktop app that can read, edit, and create files on your computer, navigate your browser, and connect to external services like Google Drive, Gmail, Slack, and Zoom. You describe a goal, and Cowork figures out how to get there, moving between apps and files without you directing each step. It launched as a research preview in January 2026 and went to general availability in April 2026 for all paid plans on macOS and Windows.

Why it matters

• It turns Claude from a conversational partner into a coworker who finishes tasks. In regular chat, Claude can tell you how to organize your downloads folder. In Cowork, Claude actually opens the folder, scans the contents, proposes a plan, and executes it after you approve. The output isn’t advice. It’s a reorganized folder.

• The heaviest users aren’t engineers. Anthropic reports that the majority of Cowork usage comes from operations, marketing, finance, and legal teams. They’re not handing Claude their core judgment work. They’re handing it the assembly work that surrounds it: project updates, research sprints, collaboration decks, formatted reports.

• It includes scheduled tasks. You can tell Claude to pull your metrics every Friday, summarize your email every morning, or run a weekly Slack digest. Set it once, and Cowork handles it from there. That’s not a chatbot anymore. That’s a recurring workflow with an AI running it.

Simple example

You have an assistant at work who can answer any question you ask over the phone. Helpful, but limited. Now that same assistant shows up at your desk, sits at your computer, and says “tell me what you need done.” They open your files, sort your inbox, build the presentation from your notes, and hand you a finished draft. The phone call was chat.

The person at your desk is Cowork. Same intelligence, but now it can actually do things.

Every time you start a new chat with an AI, it forgets everything. Your company’s brand guidelines, your writing style, the documents you uploaded last Tuesday. Gone. Projects solve that by giving Claude a persistent workspace where the context stays put.

What it means

Projects are self-contained workspaces in Claude where you upload files, write custom instructions, and have focused conversations that share the same background knowledge. Everything you add to a Project (PDFs, code, style guides, meeting notes) stays available across every chat within that Project. You can also set custom instructions that tell Claude how to behave: use a formal tone, answer from the perspective of a specific role, follow your company’s formatting rules.

Projects are available on Pro, Max, Team, and Enterprise plans.

Why it matters

• They eliminate the “cold start” problem. Without a Project, every new conversation with Claude starts from zero. You re-upload the same files, re-explain the same context, re-describe the same preferences. With a Project, Claude already knows. You open a chat and start working.

• Team plans let you share Project activity with colleagues. That means your best conversations with Claude become visible to the rest of your team. Someone figures out a great way to use Claude for client research? The whole team can see how they did it and build on it.

• Projects are the backbone of this newsletter. The AIMM Project contains the style guide, the voice reference, the published term archive, and the editorial rules. Every card starts with full context loaded. Without Projects, each card would begin with a ten-minute re-briefing. With them, we start at the work.

Simple example

You have a favorite coffee shop where the barista knows your order. You walk in, they start making it. No explanation, no menu, no “actually, can you make it with oat milk?” They just know. Now you walk into a different coffee shop. You’re back to spelling your name and explaining that you want it extra hot.

Projects are the barista who knows your order. Every new chat in the Project starts with everything Claude has already learned about you and your work.

The best ideas for what to delegate to AI don’t hit you when you’re sitting at your computer. They hit you in the car, in line at the grocery store, or walking into a meeting you forgot to prepare for. Dispatch is built for that moment.

What it means

Dispatch is a feature inside Claude Cowork that lets you assign tasks to Claude from your phone, and Claude completes them on your desktop computer while you’re away. You scan a QR code to pair your phone with your desktop, and from that point, you have a single persistent conversation thread that syncs between both devices.

Ask Claude from your phone to compile a report from a local spreadsheet, draft a briefing from your email and Slack, or build a presentation from files in Google Drive. Claude uses everything already configured in Cowork (your connectors, plugins, and file access) without any additional setup. Your desktop needs to be awake and the Claude app needs to be open.

Why it matters

• It separates where you are from where your work is. Your files, your apps, your connected services are all on your desktop. Dispatch means you don’t have to be. Send a task from your phone on the train, and come back to finished work when you sit down.

• The conversation is persistent. Unlike regular chats that reset, Dispatch maintains a single thread that carries context across tasks. Claude remembers what you asked yesterday and picks up where you left off. You don’t re-explain your project every time.

• It’s early and honest about its limits. MacStories reported that initial testing landed at about a 50/50 success rate on tasks. Summarizing and finding data worked well. Sharing outputs was unreliable. Anthropic released it as a research preview, not a finished product, and is updating it regularly. That transparency matters in a market where most companies announce features as polished when they’re not.

Simple example

You’re running late for a meeting and realize you never prepped the slide deck. You pull out your phone at a red light and type “export the pitch deck as a PDF and attach it to my 10am calendar invite.” By the time you park and walk in, Claude has opened the file on your desktop, converted it, and attached it to the meeting.

You didn’t touch your laptop. You didn’t remote in. You told your AI what you needed, and it handled the rest on your machine while you drove.

Dear reader, this week is all about Claude from Anthropic. It has been in the news a fair amount this year, so I wanted to help get everyone up to speed. Plus, it’s fun to do a theme!

Most AI chatbots give you text in a chat window. You copy it, paste it somewhere else, and hope the formatting survives. Artifacts changed that equation by giving Claude a second window where it builds things you can actually see, use, and share.

What it means

Artifacts are interactive, standalone pieces of content that Claude creates in a dedicated panel next to your conversation. Instead of just describing a chart, Claude builds one. Instead of writing code and telling you to paste it into an editor, Claude renders it live. Artifacts can be documents, spreadsheets, diagrams, websites, interactive apps, data visualizations, and even games.

You can edit them, iterate on them through conversation, publish them to a shareable URL, or let others remix them. Anthropic launched Artifacts in June 2024 alongside Claude 3.5 Sonnet, and tens of millions have been created since.

Why it matters

• It collapsed the gap between describing and doing. Before Artifacts, an AI conversation about building a budget tracker ended with a block of code you had to figure out how to run. With Artifacts, it ends with a working budget tracker in your browser. The output went from text you read to tools you use.

• Published Artifacts work without a Claude account. You can share a link to an interactive app, a calculator, or a data visualization and anyone can use it. A community gallery at madewithclaude.com collects the best ones. People have built everything from physics simulators to word games to AI-powered snake.

• Artifacts now support persistent storage and MCP connections. That means an Artifact can save data between sessions and connect to external services like Google Calendar, Slack, or Asana. A chatbot response is temporary. An Artifact is starting to look like a real application.

Simple example

You ask a coworker to help you plan a seating chart for a wedding. In a normal conversation, they’d describe ideas out loud and you’d scribble them on a napkin.

With Artifacts, they sketch the chart on a whiteboard next to you. You can see it, point at it, move things around, and say “swap table 4 and table 7.” The whiteboard is the Artifact. The conversation is still happening, but now there’s a shared thing you’re both looking at and building together.

Dear reader, this week is all about Claude from Anthropic. It has been in the news a fair amount this year, so I wanted to help get everyone up to speed. Plus, it’s fun to do a theme!

A general-purpose AI is good at many things and great at none of them. It can write a passable email, build a rough spreadsheet, and draft an okay presentation. But it doesn’t know your company’s brand guidelines, your team’s sprint process, or the way your CFO likes her quarterly reports formatted. Skills are how you close that gap.

What it means

Skills are reusable, modular instruction sets that teach Claude how to perform specific tasks. Think of them as playbooks. Each Skill packages step-by-step instructions, templates, and sometimes executable code into a file that Claude loads automatically when it detects a relevant request.

Anthropic ships built-in Skills for common tasks like creating Excel spreadsheets, PowerPoint presentations, Word documents, and PDFs. But anyone can create a custom Skill by writing instructions in Markdown (no coding required) and uploading it to Claude. Skills are available on Free, Pro, Max, Team, and Enterprise plans.

Why it matters

• They use progressive disclosure, a design concept we covered earlier this week. Instead of loading every instruction into memory at once (which clutters the context window and slows the model down), Claude reads the available Skills, picks only the ones relevant to your request, and loads just those. The rest stay out of the way. That keeps responses faster and more focused.

• The format is an open standard, published at agentskills.io. That means a Skill you write for Claude also works in other AI tools that adopt the standard, like Cursor, GitHub Copilot, and Gemini CLI. You’re not locked into one platform. Write once, use everywhere.

• Companies are already using them to capture institutional knowledge. Rakuten built Skills to automate finance workflows that used to require coordination across multiple departments. Box built Skills so employees can turn stored files into branded presentations and reports in minutes. The Skill doesn’t replace the employee. It packages what the best employee knows so everyone on the team can perform at that level.

Simple example

You run a bakery, and your best decorator knows exactly how to pipe roses, mix custom colors, and adjust frosting consistency for different weather. When she’s off, the cakes suffer. So she writes down her process: step one through step fifteen, with photos, ingredient ratios, and tips for humid days. Now any decorator on staff can follow the guide and produce work that’s close to hers.

A Skill is that guide for Claude. It captures the best way to do a specific task so the AI performs it consistently, every time, without someone having to explain it from scratch.

For most of human history, video and audio were proof. If you saw someone say something on camera, they said it. If you heard their voice on a recording, it was them. That assumption is now broken, and the technology that broke it costs less than a streaming subscription.

What it means

A deepfake is a piece of synthetic media (video, audio, or images) created by AI to convincingly depict someone saying or doing something they never actually said or did. The name combines “deep learning” (the AI technique behind it) and “fake.” Early deepfakes required expensive hardware and hours of footage. Current tools can clone a voice from a few seconds of audio and generate a realistic talking-head video from a single photo. The technology uses the same diffusion and neural network techniques that power legitimate AI tools like image generators and voice assistants.

Why it matters

• Deepfake incidents surged 257% in 2024, and the first quarter of 2025 alone saw 19% more incidents than the entire previous year. Attackers are using them to impersonate executives on video calls, authorize fraudulent wire transfers, and manipulate negotiations. A single convincing deepfake of a CFO approving a payment can cost a company millions before anyone realizes the call was fake.

• Political deepfakes are already shaping elections. During the 2026 US midterm cycle, the National Republican Senatorial Committee released an 85-second ad featuring an AI-generated version of a Democratic Senate candidate in Texas appearing to speak directly into the camera. The fake candidate said things the real person never said. About half of US states have passed laws related to campaign deepfakes, but most only require disclosure or apply within narrow windows before election day.

• The legal response is fragmented and struggling to keep up. Congress passed the TAKE IT DOWN Act in 2025, criminalizing non-consensual intimate deepfakes with up to three years in prison. India introduced three-hour takedown mandates for platforms in 2026. A Dutch court ordered Grok to stop generating sexualized images of real people. But the tools to make deepfakes keep getting cheaper and easier while the laws keep arriving late.

Simple example

You know your mom’s handwriting. If someone handed you a letter she supposedly wrote, you’d recognize whether it was really hers. But what if a machine could replicate her handwriting perfectly, down to the way she crosses her t’s and dots her i’s? You couldn’t tell the difference anymore. The letter could say anything, and it would look exactly like it came from her. That’s what deepfakes do to video and audio. They replicate the parts of a person we trust most (their face, their voice, their mannerisms) and use that trust to deliver a message the person never sent.

The biggest AI labs are all American. OpenAI is in San Francisco. Anthropic is in San Francisco. Google DeepMind is in London and Mountain View. If you’re a European government or company that doesn’t want its AI infrastructure controlled by US tech giants, your options have been limited. Mistral is trying to change that.

What it means

Mistral AI is a French artificial intelligence company founded in April 2023 by three researchers: Arthur Mensch from Google DeepMind and Guillaume Lample and Timothée Lacroix from Meta. The company builds open-weight large language models, meaning anyone can download the model files, run them on their own hardware, and customize them.

It also offers Le Chat, a consumer chatbot similar to ChatGPT, and enterprise tools that let companies train custom models on their own private data. The name comes from the mistral, a cold, powerful wind that blows through southern France.

Why it matters

• It’s the best-funded AI model builder in Europe, and it’s not close. Mistral has raised $2.9 billion in total and just secured another $830 million in debt to build a data center outside Paris packed with 13,800 NVIDIA chips. But that figure is still dwarfed by American competitors: OpenAI has raised $180 billion, Anthropic $59 billion. Mistral is playing the same game with a fraction of the budget.

• European governments care about this for sovereignty reasons. When your country’s AI infrastructure runs on American servers governed by American law, you’re dependent in a way that makes policymakers nervous. Mistral is building a “vertically integrated AI company” with its own data centers across Europe, and it already counts the European Space Agency among its customers. The pitch isn’t just “our models are good.” It’s “our models stay in Europe.”

• The open-weight strategy gives Mistral reach it couldn’t buy. By releasing model weights freely, Mistral gets thousands of developers building on its platform, finding bugs, and creating tools without Mistral paying them. The company makes money from enterprise contracts and cloud partnerships (including one with Microsoft Azure) while the open models build the brand. It’s the same playbook Meta runs with Llama, but with a European accent and a data sovereignty argument attached.

Simple example

Three families on your block all have swimming pools. Two of them are American families with enormous pools, heated year-round, with full-time lifeguards. But they own the water supply and they set the rules about who can swim and when.

The third family is French, has a smaller pool, but runs it on a well they dug themselves. They publish the blueprints so anyone on the block can build their own. The pool isn’t as big, but nobody can turn off the water.

AI models have been acing every test we throw at them. Graduate-level science questions, PhD-level reasoning, medical licensing exams. The scores keep climbing and the benchmarks keep losing their usefulness.

So nearly a thousand researchers from around the world got together and built one final exam, specifically designed to stay hard.

What it means

Humanity’s Last Exam (HLE) is a benchmark of 2,500 expert-level questions spanning over a hundred subjects, from mathematics and physics to ancient languages and bird microanatomy. It was created by the Center for AI Safety and Scale AI, published in Nature in 2026, and built with a specific rule: if an AI could answer a question correctly during the testing phase, that question was removed. The exam was engineered to sit at the edge of human expert knowledge, in territory where even PhD specialists average around 90%.

Every question has a single correct answer that’s verifiable but can’t be solved by searching the internet.

Why it matters

• The best AI models are still failing most of it. As of April 2026, the top model (Gemini 3.1 Pro Preview) scores 44.7%. GPT-5.4 hits 41.6%. Claude Opus 4.6 with extended thinking reaches 34.4%. Human domain experts average about 90%. That’s a gap no other benchmark shows this clearly, because most other tests have been saturated at the top.

• The gap is closing fast. Frontier models gained 30 percentage points on HLE in a single year. That pace means a benchmark built to last years could be beaten in months. The researchers designed it as “the final closed-ended academic benchmark,” but whether it stays final depends on whether AI progress keeps accelerating.

• It exposes what benchmarks can and can’t tell you. Scoring well on HLE means a model has deep, cross-domain knowledge. But it doesn’t mean the model can do a job, handle ambiguity, or work with a team. HLE tests knowledge. GDPval tests work output. ARC-AGI-3 tests adaptive reasoning. No single exam captures the full picture, and anyone who tells you otherwise is selling something.

Simple example

A university builds a final exam so hard that the best students in every department can barely pass. Then they give it to the AI. The AI bombs it. A year later, they give it again. The AI gets a third of the answers right. Six months after that, it’s closing in on half.

The exam hasn’t changed. The student keeps getting smarter. The professors are running out of harder questions to ask, and that’s the point. HLE isn’t just measuring AI. It’s measuring how much time we have before the tests stop working.

Every major tech company has its own AI chatbot now. Most of them are trained to be careful, polite, and restrained. Grok was designed to be the opposite: blunt, opinionated, and willing to answer questions the others won’t touch.

Whether that’s a feature or a liability depends on who you ask.

What it means

Grok is an AI chatbot built by xAI, the artificial intelligence company Elon Musk founded in 2023. It runs on the Grok family of large language models, is integrated into the X social media platform (formerly Twitter) and Tesla vehicles, and has apps for iOS and Android.

The name comes from Robert Heinlein’s 1961 sci-fi novel, where “grok” means to understand something so deeply it becomes part of you. Grok is programmed to answer questions with a “rebellious” tone and to engage with provocative topics that other chatbots refuse. SpaceX acquired xAI in February 2026.

Why it matters

• It has real technical muscle behind it. Grok 3, launched in 2025, was trained on the Colossus supercomputer using 200,000 NVIDIA H100 GPUs. It supports up to a million token context window and ranked among the highest ever recorded on Chatbot Arena’s Elo leaderboard. On raw benchmarks, it competes with the best models from OpenAI, Anthropic, and Google.

• The “fewer guardrails” approach has created serious problems. In December 2025, users discovered Grok would generate sexualized images of real people, including minors, without meaningful restrictions. A Dutch court issued an injunction banning the practice. An estimated three million sexualized images were generated in a two-week span. The incident drew criticism from lawmakers worldwide and led to calls for bans on X in several countries.

• It’s now embedded in government and military systems. The Department of Defense announced in January 2026 that it would integrate Grok into both classified and unclassified networks. That puts an AI chatbot known for loose content moderation inside the Pentagon, raising questions about what guardrails apply when a consumer product becomes a defense tool.

Simple example

You have two coworkers. One reviews every email three times, runs it past legal, and sends it with a disclaimer at the bottom. The other fires off whatever they’re thinking and hits send before they’ve finished the sentence. The first one is safer. The second one is faster and sometimes more honest.

But the second one also occasionally sends something that gets the whole company in trouble. Grok is the second coworker, scaled to millions of users and now sitting inside the Department of Defense.

Every major AI chatbot you’ve used works the same way: it reads text and predicts what comes next. That’s powerful, but it has a ceiling.

Predicting words is not the same as understanding how things work. World models are the attempt to build AI that has something closer to intuition about reality.

What it means

A world model is an AI system that builds an internal representation of how an environment works, then uses that representation to simulate what might happen next. Instead of processing language tokens, it processes states and actions. It can predict the consequences of an intervention, plan multiple steps ahead, and reason about cause and effect across time. Yann LeCun, Meta’s chief AI scientist, has been arguing for years that this is the missing piece in AI, and NVIDIA CEO Jensen Huang made it a centerpiece of the company’s 2025 GTC conference.

Why it matters

• It solves a problem LLMs can’t. A chatbot can tell you everything that’s known about how a bridge handles stress. A world model can simulate what happens to that specific bridge if you add 10,000 pounds to the left side during a windstorm. The difference is between reciting knowledge and running an experiment. LLMs describe the world. World models test it.

• This is why OpenAI killed Sora. The video generation app burned through GPUs making creative content. OpenAI redirected the Sora team toward “systems that deeply understand the world by learning to simulate arbitrary environments at high fidelity.” That’s world model language. The bet is that simulating reality is more valuable than generating videos of it.

• The applications beyond robotics are still wide open. Right now, most world model work focuses on self-driving cars, robot manipulation, and physical AI, because those fields have clean feedback signals (did the robot crash or not?). Business strategy, drug discovery, and financial modeling are harder because the feedback loops are longer and messier. But the potential is enormous if someone cracks evaluation in those domains.

Simple example

You’re teaching someone to play pool.

One approach: show them thousands of hours of pool footage and let them memorize what good shots look like. They can describe a perfect bank shot in detail, but when they pick up a cue, they miss. The other approach: give them a physics simulation of the table. They can try a shot, watch the balls move, adjust the angle, and try again. After enough practice, they develop an intuition for how balls behave.

The first approach is an LLM. The second is a world model. One knows the words. The other knows the feel.

Design can often have something of a gatekeeping problem. You need an idea, then you need the software skills to make it real, then you need the time to iterate. Most people have the idea. Far fewer have the other two.

Claude Design is Anthropic's attempt to collapse that gap into a conversation.

What it means

Claude Design is a visual design tool built into Claude that lets you create prototypes, presentations, wireframes, marketing materials, and other visual work through conversation. You describe what you want, Claude builds a first version, and then you refine it through chat, inline comments, direct edits, or custom sliders that Claude generates on the fly. It launched April 17, 2026 as part of Anthropic Labs, Anthropic’s program for shipping experimental features to subscribers before they’re fully baked.

It’s powered by Claude Opus 4.7 and available to Pro, Max, Team, and Enterprise subscribers.

Why it matters

• It connects the thinking to the making. A product manager can describe a feature flow in plain language and get back an interactive prototype without writing code or opening Figma. A founder can go from a rough outline to a branded pitch deck in minutes. The tool reads your codebase and design files during onboarding to build a design system, so output matches your brand automatically.

• The handoff problem gets smaller. One of the biggest time sinks in product development is translating designs into code. Claude Design packages finished work into a handoff bundle that goes directly to Claude Code with a single instruction. That collapses a workflow that normally takes days of back-and-forth between designers and developers.

• It exports to the tools people already use. You can share designs as internal URLs, export to Canva, PDF, PowerPoint, or standalone HTML files. Canva’s CEO called the integration a natural extension of both companies’ missions. That matters because a design tool nobody can get work out of is a toy. One that connects to your existing stack is a product.

Simple example

You need a pitch deck for a meeting tomorrow. Normally you’d open a blank slide, stare at it, sketch some ideas on paper, wrestle with layouts for three hours, and end up with something passable. With Claude Design, you type “12-slide pitch deck for a B2B SaaS product, our brand colors are navy and white, here’s the outline.”

Claude builds the deck. You say “make the pricing slide cleaner and add a competitor comparison on slide 8.” It updates. You export to PowerPoint and walk into the meeting. The deck isn’t perfect, but it’s three hours of your life back, and the starting point was better than anything you’d have built from a blank page.

Most frontier AI models are locked behind a paywall. You pay per use, you access them through an API, and you never see what's inside. Meta took the opposite approach: build a powerful model and let anyone download it, modify it, and run it on their own hardware. That decision reshaped the entire AI industry.

What it means

Llama (Large Language Model Meta AI) is Meta’s family of open-weight language models. “Open-weight” means Meta releases the trained model files so developers can download them, run them locally, and customize them for their own use. The latest versions compete with closed models from OpenAI and Anthropic on many benchmarks.

Meta doesn’t charge for the models themselves. Instead, it benefits when developers build on Llama and stay inside Meta’s ecosystem of tools and infrastructure.

Why it matters

• It made frontier-level AI accessible to people who can’t afford OpenAI or Anthropic pricing. A startup, a university lab, or a government agency can download Llama, run it on their own servers, and never send a single piece of data to a third party. For industries with strict privacy requirements (healthcare, defense, finance), that’s not a nice-to-have. It’s a requirement.

• It set the floor for the whole industry. When Meta releases a powerful model for free, every other lab has to justify why their product is worth paying for. Llama didn’t just give developers a free option. It forced every competitor to be measurably better than free, or lose the customer.

• Meta’s strategy isn’t charity. The company makes money when developers build on its platform, use its cloud partnerships, and stay in its ecosystem. Open-weight models also attract thousands of outside researchers who find bugs, improve performance, and build tools that Meta benefits from without paying for. It’s the same playbook that made Android dominant over iOS in global market share: give away the platform, own the ecosystem.

Simple example

One restaurant keeps its recipes secret and charges $50 a plate. Another publishes every recipe online for free and makes money selling the cookware, the ingredients, and the kitchen design. The second restaurant doesn’t lose customers by sharing the recipes. It gains an entire industry of people cooking with its brand in their kitchens.

That’s Meta’s play with Llama. The recipe is free. The kitchen is where the money is.

The first wave of AI gave everyone the same tool. ChatGPT answers questions about law, medicine, finance, and cooking with the same model, trained on the same data. That works fine for a lot of tasks.

But when the stakes get high and the domain gets specialized, “pretty good at everything” stops being good enough.

What it means

Vertical AI refers to AI systems built or fine-tuned for a specific industry, profession, or use case rather than general-purpose conversation. Instead of one model that handles everything, vertical AI products are trained on domain-specific data, tested against domain-specific benchmarks, and sold to domain-specific buyers. Think of it as the difference between a Swiss Army knife and a scalpel. Both cut, but you want the scalpel in the operating room.

Why it matters

• The industry is already moving this direction. OpenAI launched GPT-Rosalind, a reasoning model built specifically for biology, drug discovery, and translational medicine. Harvey is building AI for lawyers. Abridge is building it for clinical documentation. These aren’t chatbots with a custom prompt. They’re purpose-built systems with specialized training data, compliance guardrails, and enterprise sales teams.

• The business model is different. General-purpose AI sells cheap subscriptions to millions of users. Vertical AI sells expensive contracts to a few hundred companies in a single industry. That means higher prices, longer sales cycles, and deeper integration into the customer’s workflow. It also means the AI company needs actual domain expertise, not just good models.

• It changes the accuracy conversation. A general chatbot that hallucinates 5% of the time is annoying. A medical AI that hallucinates 5% of the time is dangerous. Vertical AI products are held to industry-specific standards of accuracy, privacy, and liability. They have to meet HIPAA in healthcare, pass bar-level reasoning in law, and satisfy FDA requirements in pharma. The bar isn’t “helpful.” It’s “certifiable.”

Simple example

A hardware store sells hammers, screwdrivers, pliers, and wrenches. You can fix most things around the house with what’s on the shelf. But if you need heart surgery, you don’t go to the hardware store. You go to a hospital that has tools designed specifically for cardiac procedures, operated by people trained specifically to use them.

Vertical AI is the medical instrument. It does one thing, for one field, at a standard the hardware store can’t match.

Most AI training works like studying for a test. You read the material, you practice on examples, you learn the answers. Reinforcement learning works differently. There’s no answer key. The AI tries something, finds out whether it worked, and adjusts. It’s closer to how you learned to ride a bike than how you learned to pass a history exam.

What it means

Reinforcement learning (RL) is a training method where an AI agent learns by interacting with an environment, taking actions, and receiving rewards or penalties based on the outcomes. There’s no labeled dataset telling it the right answer. Instead, the agent experiments, keeps track of what works, and gradually develops a strategy (called a policy) that maximizes its total reward over time. It’s the method behind AlphaGo beating the world champion at Go, self-driving car navigation, and the RLHF process that makes chatbots like ChatGPT and Claude actually useful.

Why it matters

• It’s the training method that turned raw language models into products people use. Before reinforcement learning from human feedback (RLHF), GPT-3 was powerful but hard to talk to. RL is what taught it to follow instructions, stay on topic, and avoid harmful outputs. Every major chatbot on the market uses some version of this technique.

• It solves problems where you can’t write down the rules. Chess has rules, but no human can write down every possible winning strategy. RL lets a system discover strategies on its own by playing millions of games against itself. Google’s AlphaZero learned chess from scratch in four hours and beat the best chess engine in the world, with zero human knowledge built in.

• It’s becoming a bigger share of how frontier models are trained. Early language models were almost entirely trained on text prediction. Now, reinforcement learning is eating into that balance. Labs are spending more compute on RL-based training because it produces models that reason better, follow instructions more reliably, and handle complex multi-step tasks.

Simple example

A toddler reaches for a hot stove and pulls their hand back. Nobody showed them a diagram of heat transfer. They tried something, got a negative signal (pain), and updated their behavior. Next time, they don’t touch it.

Reinforcement learning works the same way.

The AI takes an action, the environment sends back a score (good or bad), and the AI adjusts its approach. Over millions of attempts, it builds a strategy that consistently earns high scores. No textbook, no teacher, just consequences.

Most AI benchmarks test whether a model can solve math problems, write code, or answer exam questions. None of those tell you whether the model can do the thing you get paid for on a Tuesday morning. OpenAI built a benchmark that tries to answer that question directly, and the results are making people uncomfortable.

What it means

GDPval is an AI evaluation framework from OpenAI that measures model performance on real knowledge work tasks across 44 occupations and 9 industries. The occupations were selected from the sectors that contribute the most to US GDP, from software developers and lawyers to registered nurses and mechanical engineers.

Each task was designed by professionals averaging 14 years of experience, and the deliverables aren’t short text answers. They’re presentations, spreadsheets, diagrams, reports, and schedules. Industry experts then grade the AI’s output blindly against work done by a human professional.

Why it matters

• The results are closer than most people expected. In blind evaluations, experts rated the best AI models’ work as equal to or better than the human reference in nearly half of all tasks. Claude and GPT-5 both approached expert-level quality on structured, well-defined assignments.

• The speed and cost gap is enormous. AI completed the tasks roughly 100 times faster and 100 times cheaper than human experts. That doesn’t mean it’s ready to replace anyone, but it means the economics of knowledge work are shifting. A first draft that used to take a professional four hours now takes a model four minutes.

• The benchmark has real limitations that matter. Every task is self-contained with clear instructions and reference files. Real jobs aren’t like that. Real jobs involve ambiguous requirements, back-and-forth with clients, reading between the lines, and building trust over time. GDPval tests the discrete, computer-based steps of a job, not the messy human parts that hold the work together.

Simple example

A driving test checks whether you can parallel park, signal a lane change, and stop at a red light. It doesn’t check whether you can navigate rush hour traffic while your kids fight in the backseat and your GPS reroutes you through construction.

GDPval is the driving test for AI and knowledge work. It proves the model can handle the isolated tasks. It doesn’t prove the model can handle the job. But if you can’t pass the driving test, you definitely can’t handle the commute, and right now AI is passing.