That's a wrap for the day. Transplanting and direct seeding done with drip irrigation setup and working without leaks

Also helpful to take a couple days and not look at it, then come back with fresh eyes. But then again, sometimes you don't have the luxury for critical projects. So it goes. 😂

Real important to have a red team or friendly reviewer to keep you on track. My eyes were blind to a certain edge case in automotive sales content regarding capitalization.

After looking at that much data, it's easy to go spreadsheet blind. This is when you miss regressions, especially for things that you don't have evals on yet or are unsettled requirements that are continuing to be discovered.

As it written before another threads, I take a snowball approach to developing evals and data review. Start small with one or two instances and then go up to 10, 50, 100, 500 and 1,000. My goal was to have 99.9 % confidence in the quality of gen AI content.

Coming up for air after an intensive 2-month Sprint on building out my next gen AI project at work. There's something to be said about looking at the data too much and losing all your sense of good taste in generated content.

Ai coding assistant that I'm using for writing evals for my prompt is sneakily modifying the prompt to make the test cases- evals pass incorrectly. I know what you're doing, but you're doing it completely wrong. Hahaha

In other news, my phone is running hot now. Hotspot is on and a CPU/gpu turn in through AI inference in the background

Test driving my new pixel 10 pro along with the Gemma 3N model in the Google edge gallery. Pretty incredible this all can run locally, but clearly missing domain knowledge on automotive features. A fine tune would probably help. I wonder if the Gemini models would do better.

This thread is a variation of... Nothing interesting ever happens in the office, so getting out and explore

I always look for opportunities to build up my domain expertise in automotive data in service of AI. Even hanging out while my car is getting its oil changed.

And.. use every opportunity you can, especially on Saturday mornings to build up domain expertise while getting an oil change

So my message is: look at your data, try to look at your products in real life, understand your users, and see if that can inspire anything with this incredibly powerful new generation of AI.

I feel most inspired in my work, using AI to write automotive sales content and etc, when I'm in a dealership showroom and looking at the merchandise along with the hustle and bustle of everything that goes with that.

Often, we get a data set and cool and get working on something without interrogating if the data is reliable. Then you wonder why weird s*** comes out of your AI. You got a context problem, but you have a clean data problem.

It's a common failure mode of data oriented people to only look at the digital or virtual assets. You miss out on so much signal and other insights that often don't show up in the data.

Why am I posting this? I normally write about AI here. I'm a domain expert at my company on vehicle data and using AI to write a variety of user facing content. I'm taking the opportunity in the Toyota showroom to look at my data in real life.

For me as a parent, the built-in window sunscreens are also a really nice feature of the vehicle.

Looking at the window sticker now for anything that I missed. Removing the VIN, because that's personal identifiable information. Don't want to burden for future buyers with their VIN being web searchable.

Second favorite feature is ventilated seats which provide cool air to the passenger seated on them. This Sequoia has ventilated seats on the second row Captain seats which is amazing.

It's closed, but huge on this Toyota Sequoia