Aardvark Daily

Please visit the sponsor!

The perils of relying on AI

AI is already changing the world but if the experience of one company is to be examined, it is not without significant risk.

I wrote a column recently in which I marvelled at just how amazing an AI system called Cursor had been in creating some code for an Android app I was working on.

Cursor had provided the most consistent, reliable and effective code generation of all the systems I'd tried -- so much so that I was about to plonk down the US$20 a month it costs after the end of the free trial period.

I was all set to get out the plastic and sign up when I figured I'd just check the support forums for the product to see what others were saying, especially since I'd not used it for a couple of weeks and that's a long time in the world of AI.

It's just as well I did because what I found was not good.

It seems that many formerly happy Cursor users were complaining hard about recent changes.

Digging a little deeper it appears that the AI on which Cursor is based is the highly regarded Claude Sonnet model and a new release of that model had just been launched.

Unfortunately for Cursor users, the new model seems to have really screwed things up.

You can browse the Cursor support forums for yourself to see the problems that this update to Claude has caused.

It seems that instead of remembering all the context and existing code, Cursor now tends to forget chunks of your program at random when adding new functionality. Development has become a "two steps forward, one step backwards" process.

Not only is this a problem but it seems that the new model now consumes a users credits at a rate which is much higher than was previously the case. Some users have reported burning through hundreds of dollars in a day -- due at least in part to having to prompt the system to restore code it previously forgot about.

Needless to say, I've not signed up.

Of course many will be asking "why not just use Claude Sonnet itself, why use something that is built on top of it?"

The answer is that Cursor has done a fantastic job of integrating the Claude Sonnet model into the Android Studio IDE. There's no need to copy/paste code from the chatbot into an editor -- all the updates are made automatically and it also accesses the error logs to make thing so much simpler when debugging.

I sure hope they get this sorted -- both for my benefit and their own. When it works well, Cursor is a fantastic tool but when it's broken all it seems to do reliably is drain your bank account.

Perhaps this situation is a sage warning to any company that would build their own AI service on top of someone else's LLM. If they change stuff it could totally wreck your own product and cost you a fortune. Beware of this.

In the meantime, I look forward to playing a little more with local LLMs. I may even build a system specifically to run DeepSeek or something similar, so as to get a better handle on just how this AI stuff is shaping up.

Carpe Diem folks!

Please visit the sponsor!

Here is a PERMANENT link to this column

Rank This Aardvark Page

Change Font