I'm not a software developer, was literally no code 6 months ago, (tech background ttho), but I'm now using cursor with anthropic in bash and um holy crap, combined with aider, I having a huge time. Thanks to your direction, I'm moving in leaps and bounds.
I think Opus needs additional training with the updated features they accomplished with computer--use and possibly more features. They probably want to rebrand the next big release also. 👍I like the name 'GUI Agent' for 'computer using AI agents'. 👍 The industry hasn't standardized on the name yet.
I think that sonnet 3.5 new is distilled fine tuned version of opus 3.5. this allows them to offer a slightly better version than the standard sonnet 3.5, because the performance decreases with distillation and inference speed increase (compared to standard opus model), and they got a cheaper internal model that has better performance than the sonnet 3.5. And maybe they use Opus 3.5 for generating syntetic data for future models.
They intially released 3 sets of model haiku, sonnet and opus. Sonnet was long context and faster than haiku, haiku was good at templating and was almost more performant than opus on long-shot prompting. And opus was good at novelty.
I think opus is a BIG model , i dont think they rebranded opus to sonnet as sonnet is pretty fast and opus would be much slower given its size. they are probably focusing on sonnet 3.5 as they have a huge competitive advantage at the moment and can iterate much faster with sonnet, then take their learnings and apply them to opus if they need later. opus always felt very big, models of that size would take a LOT of computer and time to make
I think the same. Opus is the bigger model and has (had?) much more parameters. Probably that was needed when they developed Claude 3 to compete with GPT-4. Now their technology has advanced and they don't need so many parameters any more to be competitive. Sonnet has a very attractive price point and TBH I wouldn't use Opus as much because of the price. Sonnet's knowledge is enough for the use cases of the majority of users. A bigger model with more world knowledge is simply a waste of resources/money for most of us.
Doing something similiar with that loop for my Red Team agent. Only difference is im scaling this function up with machine learning/GPU's to get that speed, to build attack vectors fast. This whole game is about to get real spicy LOL Btw you see google is doing this now as well, cat is out the bag.
Computer management or bash built into the model, are very interesting solutions. However, I personally prefer to create my own tool base, with appropriate restrictions. And it is these tools that my assistant will be able to use. There is no way that at this stage of development I would give him the ability to fire background generated bash commands :)
didnt even know anthropic released a bash tool and a text editor tool! thanks for bringing that to our attention
what up techfren
@@remsee1608 yoooo 👋🏽👋🏽
I'm not a software developer, was literally no code 6 months ago, (tech background ttho), but I'm now using cursor with anthropic in bash and um holy crap, combined with aider, I having a huge time. Thanks to your direction, I'm moving in leaps and bounds.
I like what you are building, keep up the great videos!
We all knew the computer use video was coming. Hahah. Thank you.
Text to Action exactly. Btw thanks for showing the tokens cost + $
Thanks for this excellent video!. I made a fork to run this with dockerized container just to reduce possibility to wipes out my entire disk.
I think Opus needs additional training with the updated features they accomplished with computer--use and possibly more features.
They probably want to rebrand the next big release also.
👍I like the name 'GUI Agent' for 'computer using AI agents'. 👍
The industry hasn't standardized on the name yet.
I think that sonnet 3.5 new is distilled fine tuned version of opus 3.5. this allows them to offer a slightly better version than the standard sonnet 3.5, because the performance decreases with distillation and inference speed increase (compared to standard opus model), and they got a cheaper internal model that has better performance than the sonnet 3.5. And maybe they use Opus 3.5 for generating syntetic data for future models.
They intially released 3 sets of model haiku, sonnet and opus. Sonnet was long context and faster than haiku, haiku was good at templating and was almost more performant than opus on long-shot prompting. And opus was good at novelty.
I think opus is a BIG model , i dont think they rebranded opus to sonnet as sonnet is pretty fast and opus would be much slower given its size. they are probably focusing on sonnet 3.5 as they have a huge competitive advantage at the moment and can iterate much faster with sonnet, then take their learnings and apply them to opus if they need later. opus always felt very big, models of that size would take a LOT of computer and time to make
I agree. I suspect they made the business decision to prioritise the inference compute required to run Opus elsewhere.
I think the same. Opus is the bigger model and has (had?) much more parameters. Probably that was needed when they developed Claude 3 to compete with GPT-4. Now their technology has advanced and they don't need so many parameters any more to be competitive. Sonnet has a very attractive price point and TBH I wouldn't use Opus as much because of the price. Sonnet's knowledge is enough for the use cases of the majority of users. A bigger model with more world knowledge is simply a waste of resources/money for most of us.
Doing something similiar with that loop for my Red Team agent.
Only difference is im scaling this function up with machine learning/GPU's
to get that speed, to build attack vectors fast.
This whole game is about to get real spicy LOL
Btw you see google is doing this now as well, cat is out the bag.
Oooh, I need to try it for multiple json edits!
IndyDevDan is the goat
@@daburritoda2255 🤝🔥🔥
Computer management or bash built into the model, are very interesting solutions. However, I personally prefer to create my own tool base, with appropriate restrictions. And it is these tools that my assistant will be able to use. There is no way that at this stage of development I would give him the ability to fire background generated bash commands :)
opus was the GOAT in this playground everyone knows it. Hopefully they bring it back
Yes, Claude is great, but I use other models (GPT-4 and Gemini) as well. All models have their pros and cons :
Awesome content
It’s like Anthropic has a plan to eliminate SaaS by AI-zing the entire dev & tooling stack and converting software by coding to software by AI.
you have any github repos that you share ??
Sir, Codebase link not in the description
This is a much bigger deal than I realize, isn't it?
Claude 3.5 Opus will be trained using Nvidia's brand new Blackwell chips
Maybe they are getting ready for opus 4
Will Anthropic be able to see all the files and personal data the models retrieves from your computer?
There is around 1942 bas commands
So what are the odds Claude will not rm -rf /
Please test it running adobe, Unreal, software
firrssst lesgoo