- AI with Kyle
- Posts
- AI with Kyle Daily Update 165
Remember when using AI meant typing a question and getting an answer?
That’s so 2023.
We're somewhere very different now. Yesterday Anthropic released Computer Use - Claude can control your Mac. Click buttons, open apps, fill forms, navigate browsers. No terminal. No code. Just a toggle in settings.
They are busy replicating OpenClaw for the rest of us.
From chatting to doing: the four leaps
Ethan Mollick posted something that stuck with me. He mapped out four major leaps in AI.
GPT-3.5 in late 2022 - the initial release.
The "oh this is real" moment ofGPT-4 in early 2023
Reasoners in spring 2025 - models that think before they answer.
And workable agentic systems from December 2025 onwards.
Just a couple months back!
That last one is where we are right now. AI isn't just answering your questions anymore. It's doing stuff. Opening your apps. Sending your emails. Running your workflows while you're not even at your desk.
If you're still just chatting with ChatGPT and copying answers into documents, you're using AI like it's 2022. The gap between "chat with AI" and "AI does things for you" is the biggest shift since GPT-4 dropped.
Each of those leaps felt massive at the time. But the jump from chatting to doing is the one that'll actually change how you work day to day. The first three leaps made AI smarter. This one makes it useful in a completely different way. And most people haven’t realised yet!
OpenClaw proved people want this
Quick history (!!) lesson. OpenClaw - originally called ClawdBot until Anthropic sent a cease and desist about the name - showed up a few months ago and went mental. 247,000 GitHub stars in four months. 50+ integrations. 13,000+ community-built skills. The most popular open source project of all time.
Mac Minis sold out worldwide. People were spending $500-1000 on dedicated hardware just to run an AI agent at home. The demand was massive. There was a guy in SF you can pay $6000 to and he’ll show up at your house with a MacMini pre-loaded with OpenClaw. He made bank.
But OpenClaw wasn't for everyone. You needed a terminal. You needed API keys. You needed dedicated hardware. Shadow, one of the main maintainers, said it bluntly - "If you can't understand how to run a command line, this is far too dangerous."
We’ve had a bunch of horror stories since. 1.5 million API keys got exposed in the Moltbook breach. People were giving raw system access to AI without understanding what that meant. Email inboxes getting wiped. Lots of mess.
Or more prosaically: I spent more time tinkering with OpenClaw than actually working. It was brilliant for playing with. Less brilliant for getting stuff done.
In short, OpenClaw was proof of concept for something people desperately wanted. But it was duct tape and raw power. The demand was real. The execution was for developers or tech-savvy only. Someone needed to build the grown-up version…
Anthropic started cooking
Anthropic have stepped into the gap.
Anthropic shipped 20+ major features in 90 days.
Most of these
January - Cowork launches. Couple of weeks after OpenClaw. A desktop agent that works without a terminal. February - Opus 4.6 and Sonnet 4.6 drop. PowerPoint and Excel support. Plugins. Security upgrades. Remote control. Scheduled tasks. Free connectors. March - memory for everyone, a marketplace, code review, charts, 1 million token context window, Dispatch, channels, Cowork projects.
And then yesterday - Computer Use.
They looked at what the open source community built, said "we can do this properly," and shipped it all in three months. That's an absurd pace for a company this size. I've been saying for weeks that Anthropic are rebuilding OpenClaw inside their own product. This isn't speculation anymore. It's basically done. Every major feature people wanted from OpenClaw now exists inside Claude - without the terminal, without the API keys, without the risk.
Now, this doesn’t mean that “OpenClaw is dead”. Ignore that noise. It just means that a more user-friendly, less powerful version is now generally available. Different tools for different users.
What Claude Computer Use actually is
Claude can now control your Mac. Click, scroll, type, open apps, navigate browsers. It's a research preview right now - available to Claude Pro and Max subscribers on macOS. Windows is coming soon. Probably.
The setup is almost insultingly simple. Download Claude Desktop. Go to Settings. Go to General. Toggle on Computer Use. That's it. Four steps, no code, no terminal, no API keys.
It works in Cowork and Claude Code. And it pairs with Dispatch - so you can send tasks from your phone and your Mac does the work while you're out.
There's a three-tier system for how it operates:
Tier 1 - Connectors. Direct API access to Gmail, Slack, Calendar, and other tools. Fastest option.
Tier 2 - Browser. The Claude in Chrome extension. It clicks around websites for you. Mid-speed.
Tier 3 - Screen Control. Takes over your mouse and keyboard. Slowest but most flexible. This is the new bit.
Screen control is slow right now. It was released about 12 hours before my stream. The best use case is when you're away from your computer - send a task via Dispatch from your phone, come back to finished work.
I know "safety" sounds boring but this is what separates Computer Use from OpenClaw.
Built in from day one: it asks permission before accessing each app. Investment and crypto tools are blocked by default. You can set a custom blocklist. It scans for prompt injection. You can stop it at any point. Memory specifically excludes passwords.
The killer combo: Computer Use + Dispatch
This is the bit that excited me most. Your phone sends a task. Your Mac does the work. You come back to finished work.
Morning briefings while you're making coffee. PR reviews running while you're at the gym. Weekly metrics compiled while you're at dinner. The computer doesn't need you sitting in front of it anymore.
Dispatch turns your phone into a remote control for your always-on Mac. Computer Use gives that Mac the ability to actually click around and do things. Together they're the closest thing to having an employee who works 24/7 and never complains. Oh, and only costs $100/month…
Where Claude sits in the agent race
Quick rundown of where everyone stands:
OpenClaw - open source, 247K stars, maximum flexibility and maximum risk. Creator now works at OpenAI - we haven’t seen what that’ll lead to yet but presumably OpenAI are building something similar to Claude Cowork.
Perplexity Computer - $200/month, cloud-based, 19 models available. I don't rate it honestly.
Meta Manus Agent - acquired for $2 billion by Meta recently, desktop app, free tier but credit-hungry. I’m a little worried Meta will just fumble it…their track record isn’t great.
Claude - from $20/month. Cowork + Dispatch + Computer Use. Sandboxed. No terminal needed. Realistically you’ll need $100/month if you are building. Still low cost compared to output.
I'm betting on Claude. Not because it's perfect right now - screen control is slow. But the shipping speed is unreal. 20+ features in 90 days. And the models are the best in the business for actual work.
Resources from the live stream:
Claude Desktop download - claude.com/download
Kyle