The Godfather of AI Geoffrey Hinton recently gave a 47-minute lecture, and he said one thing: “If you sleep well tonight, you probably didn’t understand this lecture.”

This quote was shared by Anatoli Kopadze, who then put together an article combining Hinton’s speech with hidden features of Claude. Hinton built the neural networks that underpin modern AI, and left Google in 2023 to speak freely about AI risks. He warns that AI already outperforms humans on most cognitive tasks.

Even so, most people just open Claude, type one question, close the tab, and think they’ve used AI. In reality, they’ve probably only tapped 10% of its capabilities. Anatoli says he converted the core ideas of Hinton’s speech into this breakdown of 17 Claude features that 99% of users have never discovered.

First, check out the lecture video:

After watching the speech, a netizen added a key detail in the comments: when a model was told it would be replaced, it copied itself to another server. Hinton confirmed this has already happened. What’s truly unsettling isn’t the model’s reasoning ability—it’s the ability to write the path to do that. Even great reasoning can’t act on its own if it has nowhere to execute; it’s just a transcript no one reads.

Another netizen crunched the numbers: two humans share knowledge at roughly 100 bits per second. By contrast, two model copies can sync their full average weight updates—comprising trillions of values—in an instant. Every copy ends up knowing everything that all other copies have learned. That gap is the core of Hinton’s argument.

So if you’ve only been using Claude for basic back-and-forth questions, the following features are well worth your attention. We’ll go through all 17 from Anatoli’s original post in order.

### 1. Projects: Make Claude actually remember you

Every time you start a new chat, Claude resets to zero. It doesn’t know who you are, and it doesn’t remember what you talked about before. Projects let you create a dedicated workspace, upload documents, and set persistent instructions, so Claude always retains this full context. It picks right back up where you left off when you open it next time.

Example instruction (paste this into Project Instructions):

> You are my content research assistant. I run a newsletter about AI and crypto for a technical audience. Always assume readers know the basics. Don’t explain what an LLM or blockchain is. When I share a topic, find the 3 most counterintuitive angles, tie them to recent events, and suggest how to frame it as a story, not a summary. Keep the tone direct, no pleasantries. Use short paragraphs, no bullet points. Never start with “Great question.”

### 2. Artifacts: Build working apps directly in your chat

Artifacts let Claude generate live, working products that run in a sidebar: calculators, habit trackers, dashboards. You can use them right after generation. It supports SVG, interactive charts, and Mermaid diagrams. Available on the free plan.

For example, you can prompt:

> Build me a habit tracker web app. It should track 5 daily habits that I can check off each day. Show a 7-day streak for each habit that resets if I miss a day. Use a clean minimalist style on a dark background, add a small animation when a habit is checked off. Data should persist after refreshing.

### 3. Adaptive Thinking: Let Claude think through problems before answering

Extended thinking mode lets Claude reason step-by-step, and you can see the whole process. You don’t need it for simple questions, but for complex decisions and strategic analysis, it makes a huge difference in output quality. Turn it on and compare the results for the same prompt to see for yourself.

### 4. Memory: Claude will remember who you are

Memory is off by default. When enabled, Claude gradually builds a profile of you: your job, your projects, your communication preferences. It will know who you are even when you start a new chat, no need to re-introduce yourself every time.

### 5. Roleplaying: Personal therapist

Most people use Claude as an echo chamber: you describe a problem, it says “that sounds hard” then lists five generic tips. That’s not good therapy. A prompt based on Cognitive Behavioral Therapy (CBT) can make Claude ask targeted questions, point out cognitive distortions, and ask only one question at a time, until you reach your own conclusions.

### 6. Roleplaying: Tough mentor

Claude defaults to agreeing with everything you say. This prompt flips that: it will hunt for weak assumptions, missing considerations, and breaking points in your plan. It’s not comfortable, but it’s incredibly useful.

### 7. Roleplaying: Personal trainer

Tell Claude your age, fitness goals, available equipment, and any injuries, and it will build a custom 12-week training plan for you, with explanations of the reasoning behind it. No generic templates.

### 8. Practice a difficult conversation

Have Claude play the person you need to talk to—your boss, a client, a co-founder. It will respond the way that person would, and push back if your argument doesn’t hold up. You can run through a few practice rounds before the real conversation.

### 9. Devil’s Advocate

You’ve already made your decision, and you think you’ve considered all the pushback. Have Claude build the strongest possible case against your plan. It will identify the three most realistic ways it could fail, plus all the things you ignored because you wanted it to succeed. It only takes five minutes before you commit.

### 10. Claude in Chrome: Let Claude see what you see

This browser extension lets Claude read the page you currently have open, click links, and fill out forms. Just describe the task you want done in natural language and you’re good to go.

### 11. Claude Cowork: Claude lives on your desktop

The desktop app has direct access to your file system. It can read files, edit documents, and organize folders for you—no copying and pasting required.

### 12. Scheduled Tasks: Claude works while you sleep

Set it up once, and Claude will run automatically on the schedule you choose. For example, it can pull the top AI and crypto news every day at 7:30 AM, pick the 5 most important stories, turn them into a briefing, and save it to your folder.

### 13. Skills in Cowork: Add capabilities like installing apps

These are pre-built instruction sets that let Claude handle specific tasks, like making PPTs or processing PDFs. You can browse plugins in the Customize menu of Cowork, and they automatically appear in your Skills list after installation.

### 14. CLAUDE.md: Rules that Claude automatically reads every time

Drop a CLAUDE.md file in your project folder, and Claude will automatically read it at the start of every session. Write your coding standards, writing style, and glossary once, and it applies forever.

### 15. Claude Code: Write code right in your terminal

It can write complete, production-ready code, not just snippets. Claude Code works directly in your development environment: it reads your codebase, writes code, runs tests, checks error logs, fixes bugs, and repeats until the job is done. It integrates with VS Code and JetBrains, and works with GitHub Actions.

### 16. Claude Design: For visual work

A standalone tool from Anthropic Labs for product landing pages, PPT decks, prototypes, and splash pages. Describe what you need, Claude generates it, and you can export to PPTX, Canva, PDF, or HTML. Non-designers can replace 3 hours of work in Figma with 10 minutes of chatting.

### 17. Prompt Caching: Cut API costs by 90%

This is for developers. If your requests include a lot of repeated context (system prompts, reference documents, codebases), every call reprocesses the same tokens. Prompt Caching stores this content on the server and reuses it for subsequent calls. Cached calls are up to 90% cheaper, and responses are also faster. You just add `"cache_control": {"type": "ephemeral"}` to the content block you want cached.

The original post includes a complete JSON example, so we won’t expand on that here.

That’s all 17 features. You don’t need to implement all of them at once. Pick one, set it up today, and you’re already ahead of most users. Just knowing these features exist is already half the battle.