Well…. the day has arrived. And it has arrived much sooner than we thought. AI can now use a computer.
What is Computer Use?
Claude Computer Use is a tool by Anthropic which allows their latest model, Sonnet 3.5 New, to use a computer like a human.
I’m not talking about a standard bot here that would crawl a page with Python, use an API and return some results in JSON format.
I’m talking about custom AI agents that use a virtual mouse and keyboard. Click this button, minimise this window, click that button – infer what’s happening on the screen. Draw a cat.
End-to-end actions taken from a single prompt.
Example uses of Claude Computer Use
Let me give you an example.
Let’s say I want to book a table at my favourite pizza restaurant. The agent has access to my entire laptop – which means my calendar, my web browser and my WhatsApp.
Prompt: Book me a table for tomorrow at Rudys in Leeds, add it to my calendar and text my fiancé to tell her we have a surprise for her
Running agent
*Agent opens calendar*
*checks my schedule*
*opens Browser*
*googles Rudys in Leeds*
*books a table in a free slot*
*closes browser*
*opens calendar*
*adds booking to calendar*
*opens messages*
*texts Fiancé a lovely message*
All from a SINGLE prompt.
I have to say, this has happened much quicker than we expected and the implications are enormous.
How can I use Computer Use?
We have containerised demos and real-life use cases available to demo immediately.
We’ll show you how you can use Claude Computer Use to complete inefficient human tasks so you can focus on what matters.
Book a call to see Computer Use in action.
Is Computer Use Safe?
Our benchmark tests suggest Computer Use is very safe. Anthropic’s founding principles put safety at the core of everything they do. But they have posted a warning on all relevant pages related to Computer Use:
From our tests we are confident Anthropic has done enough to stop mis-use and we have taken every precaution laid out above to mitigate all risks.
What is next for AI?
We’re very excited about what is next for AI.
As a reminder, Claude Sonnet 3.5 overtook GPT-4o in July and OpenAI did not react immediately. Our belief is they did not rush o1-preview and are continuing to release on their planned schedule.
Their immediate roadmap will likely be to capitalise on o1 and their advanced reasoning capability with their new method of inference without needing to scale compute. Testing will likely be ongoing with allowing more thinking time for advanced reasoning and we expect that to rollout to Enterprise users before anyone else. This is the most likely road to achieving AGI and is likely to be their north star priority.
Therefore, we do not expect OpenAI to react to this release. Their recent updates have been to improve the interface on ChatGPT and improve GPT-o1’s reasoning capability without the need for chain of thought prompting. They have been public about wanting to allow the model more time to think. Releasing a competing github docker package that runs a virtual machine does not align with their release methodology so far.
Longer term, the natural technological progression of human-to-computer interface is to reduce the input delay between thought and input. Neuralink is doing this with movement-based thought control but it is limited to pre-defined inputs and needs lots of training. Combining this with an AI agent being able to use a computer as a human would opens up a whole new world of possibility.