I have alerted the team.

Perhaps in the future, you could put up the date in the thread title as well

Beth leaves next Sunday for a trip to grandkids-land for a week.

Give me a call on my cell and we'll make a plan.

Also, and this is kind of a strange idea...

Do you know the paper, Attention is all you need? Well the way they keep words from inflating themselves to the detriment of the whole, is they require each matrix row to sum to 1. Words are still typically more attuned to themselves than the rest of the words in the array/sentence, but they are attentive enough to the words around them that there is meaning. Have you tried a mathematical approach in that manner with your genetic style algorithms that are limiting their guesses to "game" the rules? I wonder if there is a linear algebraic solution to your issue. — Daniel Boals

I'm sorry, but I do not understand what you mean.

Perhaps we should put one of those lunches at the Atlantis real soon.

BTW, we ate at the Chinese restaurant next to the buffet on Friday night. While the menu is very limited, the food was phenomenal.

I take it you are using your own neural network in DET? I know you have experience with neural networks going back years, and when you started this project, there really weren't any open LLMs like we have today. I am no expert, but it seems like the neural nets in DeepSeek and Alibaba's Qwen are both pretty good and they have resources to train that you probably can't get without a billion dollar investment. Have you thought about using one of their models to check your DET? The new Qwen 3.6 29b is supposed to be really good. It is a dense model and it seems like what your are doing is more of a MoE, but then they do have Qwen 3.6 35b which is a MoE style logic under the hood. — Daniel Boals

First, I stopped using Neural Nets about 1994. They're not very good at "predictive AI." That is, they're not very good at accepting they can't get everything right.

(I write extremely advanced Genetic Algorithms.)

Second, when you upload racing data to an LLM, it absolutely DOES NOT use a Neural Network on it. Instead, it uses statistical methods.

Think of an LLM as a TRAINED tool. The LLM itself is not a NN. It is the OUTPUT of a NN.

Have you thought about using one of their models to check your DET?
Not for a second. LOL

...but it seems like the neural nets in DeepSeek and Alibaba's Qwen are both pretty good and they have resources to train that you probably can't get without a billion dollar investment.

The billion dollar investment could certainly have been used in other directions - but the goal was LANGUAGE, which is far different than making any kind of "investment prediction."

Yes, races #10 and #11 from last year. GPX only had 9 races today. They usually then card a few races from I believe Argentina. Looks like maybe they thought there were 11 races today, but the last two (2) were from Argentina, not GPX. — Jim Pommier

Just heard from Ryan at HDW.
He says that they repaired today's GP file to remove the wrong races.

FYI. Also, downloaded data files for GPX R10 and R11 from last year. I advised HDW. — Jim Pommier

Are you saying that they added 2 races from last year?

As always, Dave reached out immediately and we isolated the issue. Don't know if the guy ever sleeps, but his attention to his users is unmatched. Much appreciated! — Jeff Capper

LOL
Thank you.

It was like 8:30 when I saw the post.

Listen, guys... when you have a TRAFFIC STOPPER, do what Jeff did - make a quick post.
Many of you have my cell# - text me at any hour of the day or night - you will NEVER disturb me because I use DO NOT DISTURB when I'm off the air.

But at least I will see it instantly when I am back on air.

This morning I was actually up at 5:50am or so. I don't suddenly check forum messages, but I ABSOLUTELY DO check phone messages and email (I get email notifications of private messages.)

As an example, someone blow up their upgrade install yesterday. They sent me a long string of inbox messages. I saw them at 5am and told him to message me so we could zoom.

He's not yet read the message.

From Ryan at HDW
____________________

Ryan Smith
9:00 AM (4 minutes ago)
to me

Hey Dave,

We have just resolved that issue within the last 15 minutes. There was also an incorrect CD card from 5/3/2025 posted for a few hours yesterday evening that we resolved.

Ryan

Here's a look at Jeff Capper's screen.
Clearly it was an HDW error - which is now fixed - because that file shows as downloaded TODAY.

Perhaps they've corrected the download because this is what I find for today:

Yes, I saw it.

Funny thing was that I thought it was the 18.

WOW!

Social media said 14 1519 R the hot long shots.

In fairness, an AI engine - even if it trained on every KY Derby in the last 20 years would have... wait for it... twenty races.

That's such a narrow sample with crazy variances. It would basically create a backfit - which would be pretty much worthless.

Truth be told, this is one of the rare times when an LLM using social media opinions has the potential of shining.

Must ask the right questions for what you're searching for - based upon the Wisdom of the Crowd concept.

For example...
1. If you're interested in high hit rate, use a question like, "Who are the 3 most likely winners?"

2. If you're interested in crazy high exacta payoffs you could structure your question like, "Who are the potential big-priced horses that have a chance today?"


Please note that I put all of about 4 minutes thought into this.

Dave

Speed Ratings are changing.

In our current approach, there are several recent changes in the way things are done that have changed “what’s important.”

We’re seeing a trend away from “best of last 2” and “best 2of last 3” towards “best 2 of last 4.”

In other words, RECENT RACES are LESS IMPORTANT.

Do we need to update the par file in Determinator like in HSH? — Pete C

It's coming tomorrow.
I've been doing some testing before release.

(Also with HSH)

I understand what you are saying, thanks!

What I have difficulty with is comparing say a sprinters half mile time compared to a routers half mile time. Or say a sprinters 6F final time compared to a routers 6F time. I was thinking Quirin style pars such as yours might give me a ballpark or decent comparison.

I appreciate it, thanks! — jess boyd

Ah... not what I thought you were asking.
That is ALWAYS a problem.

The numbers themselves should do that for you.
The issue of whether or not a given horse will run to the front with a distance change is the biggest challenge.

IOW, the root problem isn't the adjustments of the ratings near as much as the adjustment of running style.

The only thing I would like is the distance equalization for each horse in todays race. — jess boyd

That's what Par Times are for.
There is not purely formulaic way to do it that actually works.

It's just like Beyer's idea - 2 points per length at 6f to 3 points at 9f.

That concept is (basically) "If I beat you by 2 lengths at 6f I'll beat you by 3 lengths at 9f."

The problem is that the average beaten lengths (between 1st and 2nd finishers) from 6f to 9f is not even close to 150%. More like 7%.

My belief is that the reason is horses don't run full speed for the entire race. They run fast to the 1st call and as fast as possible in the stretch.

Hi Dave, I recently learned that TFUS PP'S fractions are not adjusted Quirin style between horses at different distances to make head to head comparisons. The fractions reflect only what the horse ran at that particular distance, not projected to todays distance. Do you have any instructional video of how I could use the par times to convert each horse to todays distance? I'm not quiet sure if I would use the track pars, class levels for todays race, or what. Thank you! — jess boyd

First, I do have a video that shows how I use par times.

But your problem is different because you're getting the time of the horse rather than the race, so the approach would be different.

But there is good news...

When you look at a horse's paceline, just treat it like HIS TIME was the RACE TIME and ignore his beaten lengths.

It's actually easier.

Here's a link to that video

Agree completely.
Thank you for the great advice.