Wow just used this for the project I kept having new files created for and mesing up and it did it in 1 minute vs the past 2 hours fighting it haha Thanks!
I agree that results are inconsistent, but I'm puzzled by the supposition that it is cursor itself that is the cause of the problem you identified at the end with the model. Isn't that code being generated by the model on the back end and then sent back to cursor? So if it got the wrong model, wouldn't that be the LLMs fault, not cursor's? Or is cursor intercepting the token flow from the model and purposely injecting the wrong model? I would think that it's the model failing to follow the instructions consistently, rather than cursor intercepting and screwing the results. Not sure but that would be my first guess. I also agree that you really need to be a developer already to work effectively with these tools because they are inconsistent and will lead inexperienced developers down torturous paths if they don't already know programming well. For example, I asked cursor (sonnet) to create a somewhat complex t-sql routine. It created something extremely complicated because it didn't realize it could use a pre-existing t-sql function that comes shipped with SQL Server. But I knew. So I told it that its solution was way more complicated than necessary, and mentioned the function. It replied, "You're absolutely right!" and went ahead and reduced the 30 lines of code it had created to the 3 lines it actually needed. Etc, etc, etc, etc. But I don't blame cursor for this. It is the LLMs (sonnet, gpt4, etc) that cannot produce reliable results. Lastly, despite the flaws of the LLMs, because I am an experienced programmer who knows how to avoid the pitfalls presented by the LLMs, I am able to use the tools and save a tremendous amount of time. So they are good. It's just that you have to already be an expert to actually gain a real advantage from them. Newbies are likely to fall into every pit the models spit out at them, simply because they don't know better and will assume the model is doing things correctly. Bad assumption, sorry. Anyway, great video. I was wondering about agent. I think I'll give it a test drive. Thanks.
Learn how to build with AI in my new courses here: dair-ai.thinkific.com/
Use code BLACKFRIDAY to get a 35% off. Offer expires 11/29
Wow just used this for the project I kept having new files created for and mesing up and it did it in 1 minute vs the past 2 hours fighting it haha Thanks!
The Composer agent may have skipped coding the MD conversion step initially because of the prompt to focus on scraping.
Good point
I agree that results are inconsistent, but I'm puzzled by the supposition that it is cursor itself that is the cause of the problem you identified at the end with the model. Isn't that code being generated by the model on the back end and then sent back to cursor? So if it got the wrong model, wouldn't that be the LLMs fault, not cursor's? Or is cursor intercepting the token flow from the model and purposely injecting the wrong model? I would think that it's the model failing to follow the instructions consistently, rather than cursor intercepting and screwing the results. Not sure but that would be my first guess.
I also agree that you really need to be a developer already to work effectively with these tools because they are inconsistent and will lead inexperienced developers down torturous paths if they don't already know programming well. For example, I asked cursor (sonnet) to create a somewhat complex t-sql routine. It created something extremely complicated because it didn't realize it could use a pre-existing t-sql function that comes shipped with SQL Server. But I knew. So I told it that its solution was way more complicated than necessary, and mentioned the function. It replied, "You're absolutely right!" and went ahead and reduced the 30 lines of code it had created to the 3 lines it actually needed. Etc, etc, etc, etc. But I don't blame cursor for this. It is the LLMs (sonnet, gpt4, etc) that cannot produce reliable results.
Lastly, despite the flaws of the LLMs, because I am an experienced programmer who knows how to avoid the pitfalls presented by the LLMs, I am able to use the tools and save a tremendous amount of time. So they are good. It's just that you have to already be an expert to actually gain a real advantage from them. Newbies are likely to fall into every pit the models spit out at them, simply because they don't know better and will assume the model is doing things correctly. Bad assumption, sorry. Anyway, great video. I was wondering about agent. I think I'll give it a test drive. Thanks.
Why conda instead of poetry/uv?
I think it doesn’t matter tbh
cool
❤