I agree! I was surprised by the 306 ftp, but definitely agree it’s an estimate that I proved wrong. During the ride I could feel that 320 + felt high and around 310± felt right for ftp setting.
I don’t know how to play with intervals to well, when I look at the ride it breaks everything down in intervals. I don’t know why when I hit the lap button at the start and finish.
At the bottom of the page showing your ride data, click on “Actions” and choose “Use laps.”
Within the context of a “one off” testing protocol, I’m not disagreeing with you, and let me preface my next statement by saying I haven’t read through this ENTIRE thread, but I think it’s unfair to try and compare systems that likely don’t work in the same way… in this case “one off” tests vs AI FTPD… they simply don’t work the same way… as I mentioned before, AFAIK AI FTPD works off a series of data points… so if a user provides it the following data…
30m @ 323w
30m @ 285w
30m @ 290w
30m @ 310w
etc… i wouldn’t be at all surprised that AI FTPD spit out a much lower value than expected… HOWEVER, if a user gave it a series of data points that was much more CONSISTENT…
30m @ 323w
30m @ 319w
30m @ 322w
30m @ 325w
etc… I would then expect AI FTPD to spit out a MUCH CLOSER value due to the consistency in the repeatable data it was fed…
Remember what TR folks always talk about… consistency… if athletes consistently hit power targets in each workout, over time AI FTPD will dial the athlete in to where they need to be…
A challenge to @ChefAcB, if you feel your FTP is closer to what intervals and/or strava gave you (and you weren’t simply having a GREAT testing day during your 30m effort), manually enter that FTP value into TR and use it for your base phase… if you can consistently hit the power targets it provides for each session and not fail workouts, the next time you run AI FTPD, it won’t be 306 anymore…
This is such a great point. Let’s say for the sake of argument that you have half a dozen GREAT days a year when the stars align and you’re the very best version of yourself.
Do you really want to hold yourself to that standard for the other 360 days of the year?
I don’t think AIFTP takes an average. We don’t know how it works, but it’s not algorithmic and it’s not looking at mean maximal efforts, but they must go into the calculation. It’s also unclear what duration it’s meant to represent.
323 doesn’t mean you could hold 323 for 60 minutes, it means you can hold it for 30 minutes. Your 60 minute MMP could be 306. Depends on your power curve. Maybe TR is picking up a high anaerobic component. If you’re using TR, why not use 306? If the workouts are easy, TR should suggest harder ones.
I wasn’t trying to imply that AI FTPD takes an average, just that feeding the machine consistent data (over a period of time… whatever that period may be) is likely to produce a more accurate result.
Like you said, we don’t know how IT works, but we also shouldn’t be be surprised by the value it suggests based upon the results of ONE data point, which is where the crux of my argument began.
We shouldn’t be comparing various testing protocols vs AI FTPD… IMHO, that’s like comparing apples to bananas.
This is where Coggan would come in and say PPP or something like that. “The best predictor of performance is performance itself”
I get what you’re saying, my Ai ftp 4 weeks ago was 316 and since I have done less volume and took a week and a half easy. It thinks my ftp dropped, I thought it did too. I have no problem lowering it if I fail workouts.
This past effort I beat out my 20 minute power from a month ago and was 2 watts under my all time best 30 minute power. Maybe I needed the rest and my fitness is expressing itself.
I think it’s interesting why it’s lower compared to when it’s normally the higher estimate.
It could have been a great day and vibes were high because I was excited to go hard. I’ll find out in the next month when I use the 315 “estimate”.
Is he related to Coggan?
We do know the basics of how it works:
And I listened to the podcast where Nate provided a few more high level details. I’m paraphrasing and oversimplifying, but it’s pretty straightforward and I could build a model having dabbled in modeling and machine learning since the early 1990s (learned it from 3 Stanford PhDs and an IEEE Fellow) 
Not saying my programming skills are up to snuff for production quality code…. I’m a sales engineer not an actual engineer.