Embarrassing AI #41: Nano Banana gets squished

Things built: 18 People watching: 18 Cashflow:-$152

When it comes to LLMs, I play favourites.

My taste can change, but for each season, there is one that stands out. I’ve used Claude a lot and by far, as a tutor, instructing me on how to build, launch, update and tweak my directory, it’s been outstanding. The last week or two, I’ve been switching over to Gemini, both because it’s deep thinking is very impressive and because it is probably the best place to get SEO advise seeing as Google runs the cartel, more or less.

It’s also integrated into Nano Banana, for visuals. All of my Barry the Bee logos have come from running it through Gemini. The whole process just feels highly integrated and functional.

Or so I thought.

Finishing up the very final touches on bestfairlawn.com (you can click it now and it will take you there), I wanted to add a favicon. Claude does not do images, so instead I asked it for instructions I could give Gemini. Favicons are 32 by 32 pixels, which were in the prompt, which Gemini then handed over to Nano Banana, providing a great image in no time.

But something felt off. The whole thing looked too large, too detailed.

Being better at AI then I was a month ago, I asked Gemini if the image was indeed the right size, which it confirmed that it wasn’t. For some reason, it also completely failed to force the image model to stick to the instructions and create a result in the requested dimensions.

After a couple of tries it just gave up and declared it couldn’t do it.

I ended up going to favicon.io and creating one in 20 seconds.

Not that Claude is without its flaws. Or perhaps its my expectations that need adjusting. At no point did Claude suggest I change the password to my admin panel for the site. What’s more, it had written it on the login page (admin123).

I don’t know why I expected to have thought of this itself. Perhaps because it’s been so full of unprompted suggestions for things to do with the site that I’d anthropomorphized it (again).

It reminded me of a discussion about why LLMs are so much worse at learning than humans. Performance can look impressive, for instance the way it can pass a bar exam. But that changes when you take into account the amount of training data and compute is required to get the model there. Essentially its being trained to pass the test but when faced with an original problem outside of the test setting, LLMs often struggle.

They gave an interesting example to illustrate the point of how much better we still are at learning. A human teenager can learn how to drive a car with as little as 10 hours of training. A Waymo needs Billions (Trillions?) of hours training to do the same.

Here’s the conversation, if you’re interested.

Hooked up my domain to the site. Made a Favicon. Made an Opengraph. One with AI, the other without. Shoveled a lot of snow.

Tools I am using to fake being an expert.

1. Beehiiv (The Platform)

"I tried other platforms, but Beehiiv is the only one that makes growth feel like a game. It’s what runs this newsletter. The Deal: You get a free trial + 20% off (I get a commission to keep the lights on)."

2. Wispr Flow (The Time Saver)

"I hate typing. This tool turns my rambling voice notes into perfect text. It’s the only way I can write this much every day. The Deal: We both get a free month if you try it."

3. Emergent (The Coding Cheat)

"I am not a coder. Emergent is my 'vibe coding' partner that writes the messy parts for me. Essential if you want to build apps without a CS degree."

Last Word.

It’s easy to get wrapped up in the almost magical ability of AI, but the more I work with it, the clearer its current limitations become.

Making daily progress,

P.S. Did I do something totally backwards today? Or do you have a better prompt for this? Hit reply and tell me. I’m here to learn just as much as you are. (I read every single email).