4 HARD Challenges for Claude Computer Use: Very Promising Results for AI Agents!

[@Monuccestduletpou]

What I would personnaly like to see is testing the AIs with real professional use cases, using same long and detailled prompts. The mail test seems the closest thing to that, but I had in mine things like a full website, a tiny e-commerce store, a food delivery app, a set up of multiples docker images organized with Swarm or Kubernetes… The benchmarks with standardized prompts like to-do app or tetris Game are interesting to compare the theorical intelligence, but I can’t prevent myself from thinking that beside benchmarking for the sake of it, the use cases that we see in general (not only this video) are not meant for the professional world. But that’s just my personnal opinion.

[@Phonognomiks]

Exactly

[@andreinikiforov2671]

You are doing cutting-edge content, as always! By the way, 93.7% is 123 IQ points -pretty good!

[@John-il4mp]

It is 127 the real number ;) even better.

[@John-il4mp]

Being in the top 6.3% corresponds to about the 93.7th percentile. Using a normal distribution table or calculator, this percentile roughly aligns with an IQ score of 127.

[@nathank5140]

Amazing. Love to see someone showing what’s possible.

[@herramientak]

¿Cuál fue el precio total de las cuatro pruebas?

[@rjackstheartofwealth6152]

How much did it cost??????

[@lyeln]

"AI has no owner" Jokes asides impressive quality, thank you for sharing this experiment!

[@sirrobinofloxley7156]

Amazing stuff, really nailed it there, though I'm surprised Firefox doesn't have a dark theme?

[@godonholiday]

You should test if it can pass the google ‘not a robot’ security tests were you have to select all the pictures of cars etc..

[@derrelecte]

I am trying to experiment with Claude using your tutorials. I want Claude to create a video snippet for me. I'm trying to get Claude to download and install lightworks but I'm running into tons of issues. Do you have any advice?

[@JNET_Reloaded]

you didnt show the cost at the end :/

[@fearhand]

Couldn't something like Make or Zapier do something like this more efficiently through API calls? Or even have the agent itself use API calls instead of GUI web based interactions.

[@Mookummockup]

Yes but you don't have to deal with as many syntax issues this way. Probably less efficient if zapier etc can handle it but way more flexible

[@Newsinrealestate]

Are you using something ADDED on Claude??

[@Phonognomiks]

“Computer Use”?

[@JNET_Reloaded]

good job its not outl;ook it would be doing spammy stuff rn lol

[@dewijones92]

awesome

[@noviceartisan]

That percentile number = an IQ of 123

[@noviceartisan]

For you next challenge, use ONSHAPE (browser based 3d modelling program) to create a 3D model of some complexity ;)

[@ElishaGefen-t5w]

First