Todd Geist

YouTube video by FileMaker Training Videos FileMaker 22 and AI on Pro Todd Geist - Day 1

Day 1 of my FileMaker AI Livestream

www.youtube.com/watch?v=t9dr...

AI loves Tanstack Query

Honestly, I felt lucky to be able to do it. How many times do you get to help somebody so directly?

I donated a kidney 12 yrs ago. It was a great experience over all. I feel lucky that I was able to do it and I have had no detectable side effects. I would do it again except well you know… I need the one I have left 🤪

Then I joined Ronnie and the RCC FileMaker Tv crew for a recap and Q and later that day.

www.youtube.com/watch?v=5Vwk...

Claris FileMaker 2025 is out with some pretty cool new AI features. I joined the Claris team to talk about some of the new features and show some of the stuff we have been working on at Proof+Geist

community.claris.com/en/s/externa...

I think I’d still pay it. There is number where I’d stop paying it but I am not sure what that is, yet.

We are probably getting closed to 100
Per dev per month with the usage fees. Still feels like a good deal.

I also think we maybe over using opus 4. It is very good. But if you just yolo on that for a few hrs you can burn some serious tokens.

Sonnet 4 is plenty good enough for most stuff

I dunno maybe.

I am happily paying, quite a bit, for our whole team to use it. It still feels more than worth it to me

How long before LLMs become LTMs?

Large Thought Models

It’s all about the IO. How fact can I get thoughts into the model and how fast can I get them out

The strange little side effect of using LLMs is that I no longer even wanna type. I just wanna speak to the applications that I’m using and let the LLMs fix all of the punctuation errors and grammar errors.

It’s just way faster. High IO for the win!

“I think I am the sand, man”

Ritchie
The Bear

I think that’s my point. Most of the code that gets written is a fairly trivial remix of what came before.

I hope so anyway, because I do worry about this issue. My relationship to the code is changing for sure. I am not yet sure if that is a good or a bad thing

And isn’t possible that the tools evolve to also improve understanding?

But isn’t possible that most code written today is just a slight recombination on code that was written before. It’s a remix of you will.

And what articles like the above are saying is that there already is better solution, now. And that is to let the AI remix it?

You can charge for maintenance and upkeep of the house. But you can’t charge for knowing how to how to build in the first place.

They can still charge something, but it will be for the compute and other infra to run the sales force application. But they can’t charge for old work.

It’s like imagine if once you built a house, you couldn’t charge for it. Because anyone can build a house now.

Here is another way to think about what AI is doing to the software industry.

Salesforce and all other SaaS companies have been charging rent on work that they did in the past. You paid a fee to get to use the particular combination of bits that make up salesforces source code.

Those days are over

And changing the relationship to the code isn’t all downside. One thing I have noticed is that the code is less precious. Starting over and refactoring are easier. I am less attached to the code and more likely to toss out a bad design and take another swing. This is immensely valuable!

I Worry about this too. It does change your relationship to the code. You didn’t type the keys.

But I think I can say now that we might able to “master the plan”. “The plan” is a higher level abstraction. It’s still very leaky, but it will probably get better.

You will be able to tell if the agent understand you. Then once it does let it loose.

You will still have to correct it from time to but once it knows the larger picture of what you want and where stuff is, it will do pretty well.

The planning is key. This was hard for me to get at first. I was so use to thinking through a problem by coding.

This doesn’t work with an agent.

Spend time iterating with the agent on the plan before you begin. You’ll learn to think through a problem or a feature by chatting with the AI.

I use cursor. I toggle back and forth between agent mode and Ask mode. I use ask mode for brainstorming and planning. Then once there is a plan I switch to agent and say “go”

Claude 4 sonnet is the best right now I think. Gemini 2.5 pro was what used before Claude 4 came out.

As you can tell by this thread there are still quite a number of coders who don’t think AI is any good. This writer has some as friends as do I. This funny piece is targeted at that phenomenon. Not the broader arguments about AI, of which there are many.

My usual flow is to spend some time brainstorming a plan with the AI. Then once I am happy with the plan I let the agent have at it. It run tests, and I’ll review the code.
Then I tell the agent to update the plan with progress

Rinse and repeat.

Things that took weeks take days or just hours

🎉👻

My much much smaller company has also had great success.

I think what people miss is that you do have to prepare your project with rules and plans and guides. You have to a lot more planning up front.

But you can use AI for that too

My AI Skeptic Friends Are All Nuts My smartest friends have bananas arguments about LLM coding.

This is really good. It’s a way better, funnier write up of the way I have been feeling.

AI assisted coding is extremely good right now, & getting better all the time. If you are a software dev who doesn’t think so, I am sorry I don’t believe you have actually tried it.

fly.io/blog/youre-a...

It is now faster and easier to write a node.js script to remove the white background from an image from scratch than it is to remember how to use an app with a GUI that has that as a feature.

I wonder if we see a decline in ORM use in this era of code generation. 🤔