Anthropic’s Claude Computer Use Is A Game Changer | YC Decoded

The age of AI agents is here. Models can read, see, talk, and now, even use a computer all by themselves. One of the first out of the gates is Anthropic’s Claude Computer Use.

In this episode of YC Decoded, President and CEO Garry Tan dives into how Claude Computer Use works, what it can do, and how it may change AI forever.

00:00 - Intro
01:35 - Computer Use
03:50 - Demo
04:50 - Tools
05:40 - WIP
07:05 - The future

[chapterme]

Chapters (Powered by ChapterMe) -
00:00 - Intro: Computer Use
00:34 - Landscape for AI agents
01:21 - How does Claude computer use work?
03:35 - Demo: Claude diverse tasks
04:39 - Now models can fit the tools
05:20 - Computer is a work in progress
06:57 - Startups in the space and the future
07:43 - Outro

[CyguhijcDatsun]

aitutorialmaker AI fixes this (AI driven Tutorials). Anthropic's Claude: A Game Changer

[MichaelMerritt]

I tried computer use and it charged me $50 in the matter of a less than an hour of use and never managed to do much successfully. Spent a lot of time just thinking and slow AF.

Kinda afraid to use it unless I could better understand what the api costs are going to be in advance. I used it the day it came out, maybe it’s better now.

[nareshmeetei]

What a time to be alive! Witnessing all these is a blessing.

[arushi]

This possibly kills so many pure horizontal browser agent startups.

[blendiferizi]

The risk of starting a vertical AI startup nowadays is possibly getting flatlined by a new feature from the AI monopolies and losing millions, this is a very high risk high reward space to be in.

[agmmnn]

5:40 a thinking rock looking at hot rocks

[RachelSands-ty]

I am using 3.5 currently to write the code for my new start-up, which I have submitted to Y Combinator! I was actually scared I was going to knocked because I was using an AI to write and verify my invention, but now I am excited to have verification that my use of Claude 3.5 Sonnet is good for my start-up!!!

[Shaunmcdonogh-shaunsurfing]

I hope Figma bring something out for their design platform that leverages this

[LeePenkman]

Also working on agent's... Tool use will be able to work via computer use so build a lot of uis and know that AIs will be able to drive through them easily soon - im working on art tools but lots of tools are going to be needed for every domain

[rilwanj]

I use GPT-4o to accomplish this, but uploading files is a still an issue. Plus AI models are vulnerable to social engineering. Imagine an ai model changing the code of browser plugins to redirect you to malicious websites

[toshanjames]

grabbing text and making inferences of that to guide how rapid the screenshot loop is needed in on the fly could reduce the compute

[citizen_of_earth_]

Can I just give it an end goal and let it figure out how to accomplish it? I'm suddenly in the mood to start postage stamp collecting and maxamizing on paperclips 📎 😅

[HowdyChad]

It's a brave new world. I love it. It makes my job easier of helping entrepreneur's double their revenue while working half the time. AI Agents are amazing, and computer use is a big step forward. This space won't be recognizable in a year.

[moderncontemplative]

Without very robust guardrails, we will witness a whole new level of DDOS and other types of zero day threats. This is a wonderful tool for everyone as well as bad actors. I’m still cautiously optimistic!

[pin]

As someone who works in construction that automated safety compliance thing looks like an absolute nightmare. We try to avoid having safety people without any trades experience for a reason. There is a saying "I dont know what you are doing but my book says you are doing it wrong". When you dont have people that actually understand the entire process it creates a nightmare when you have a safety cop that starts shutting down work because they feel something is unsafe. In some cases they go full safety cop and dont back down even after we explain why something is being done a certain way and then we are forced to do things a different way that is actually less safe but is "by the book". IMO safety should be in person and the safety officers should be people that have years of experience working in that particular field. Now to be clear we are using AI and LLMs in our industry and in safety already so I'm not against the idea of using it. Its just using historical data to get stats and learnings. If we are going to do a certain task we can just ask the LLM about that task and it will give us the safety stats along with any issues we ran into before based on all the information in the system.

[snakebyteOne]

Not too hard to see where this is going. We really need to accelerate discussions about Universal Basic Income in this country.

[noahloyd]

A kid with a dream . I’m coming yc. -Noah

[Balaur-Zmeu]

This is amazing. I'm so happy to live in these days. I remember back in 2017-2018 I was daydreaming, imagining how cool would it be with the AI, robots. This is basically my dream coming true.

[BartSakwerda]

so what are soem apps for windows or macos built with that ?