I think perspective is important: even in industries flush with money (think semiconductor industry), it is hard to hire people who have a background in ML and have subject matter expertise. And those people are really hard to find and not cheap. When I worked in the semiconductor industry, we were constantly understaffed and customers wanted progress yesterday. The biggest two problems of my former employer were to ship orders and to find qualified staff to meet their growth goals.
My impression from the outside looking in is that compared to its competitors TR seems to be at the top when it comes to basing its ML features on a solid footing (i. e. backed by a good and sufficiently large dataset). At least what I hear from competitors does not fill me with confidence in that regard (I remember one competitor’s feature being “validated” on 6 athletes and about 100 workouts …).
I have avidly followed a few threads on competitors’ products and it seems none have cracked it. (There is also the issue of customer service, and in my experience, TR’s customer service has been top notch, on par with the best brands out there.)
“Everything is inaccurate and imperfect.” The question is whether and how much it matters. Only people at TR know the answer to that one at the moment. Very often things matter that laypersons would not think about while others don’t.
Some of the disagreement in this thread revolves around how much utility most athletes would derive from a fully featured implementation of WLv2 that includes changes in PL. If my conjecture is correct, then WLv2 would not have as big an impact on the majority of athletes than e. g. RL/GL. Of course, nobody outside of TR knows, but to me it sounds like the most plausible theory (and knowing something about machine learning and handling expensive data sets).
Of course, I could be wrong as I base my hypothesis on circumstantial evidence. (It is very easy to find post hoc justifications that are entirely wrong.) If someone else has another theory as to why WLv2 is such a hard nut to crack, I’d really love to hear your opinions 
There are two interesting aspects, one is corporate communication — something I know nothing about. Was it wise to pre-announce the vision to such degree? No idea. Likely, @Nate_Pearson was very optimistic that he’d deliver on that promise much faster. I’d just add that Nate’s style of communication does not just have downsides, i. e. there are situations where this really pays off. (Maybe he is reading this thread and this very post for inspiration?)
The other question we can speculate about is what TR should have done: imagine you have a project that the company has invested a lot of money into (sunk cost fallacy?), but has hit a wall. Hard. Several times. Pressure is mounting. What should you do if you were in the product manager’s shoes?
- Lower your quality threshold so that this product can be released after all, and hope to improve it over time.
- Keep going until it is ready.
- Can the project outright and give up on the idea?
- See what parts of the project can be re-used and shift focus on something else, in the hope you can revisit it?
I’m glad TR did not opt for the first option. I remember hearing that some incarnation of WLv2 was “close”, so much so that “close” has apparently become a forbidden word (this was a joke on the latest podcast episode). From that I infer that TR could have released “something” and hoped to improve it as time goes on.
Instead for several years they went with option 2. And now they have decided for option 4: From public communication we know that RL/GL came from “WLv2 tech” (not sure about the exact wording). If I were in charge, I’d look at all the features I could build with bits of WLv2 and then focus to build those features:
- Custom volume training plans are likely enabled by what underlies RL/GL, for example. If my hypothetical is correct, RL/GL would likely give you a lot of the benefits of WLv2 (as a fully featured WLv2 would include a fatigue meter).
- Inclusion of others sports in the calendar and training plan, especially strength training. Given that @Nate_Pearson has announced they will be releasing a new calendar as public beta soon, it seems likely TR is laying the groundwork for that, too. (Edit: the new calendar is live as an early access feature.)
- Strength training features are a big one. Nate seems to focus on strength training these days, which means he he’ll have some first-hand knowledge on that subject. I hope he will hire someone like Derek Teal for training plans. Imagine applying TR tech to strength training. That market is quite big and if TR leveraged its tech and experience, it could become a player, too.
- One more thing about dynamically generated workouts: maybe it would be useful to turn things on its head and not focus on WLv2 to (also) build dynamically generated workouts. But instead, directly try to build dynamically generated workouts instead and see what tech you actually need to make that a reality. (I. e. you narrow your problem space and approach it from the application end.)
Personally, I would ask: has the product gotten better (≠ perfect) consistently in ways that matter to me? (Obvious counterexample: Strava.)
What do you think?