Beeminding outcomes [weight]

Hey all. I’ve been experimenting with using Beeminder to lose weight (https://www.beeminder.com/sohum/weight) and had my first derailment in a while, which prompted me to write this up :slight_smile:

We’ve all heard the issue with beeminding outcomes instead of actions. The problem I’ve had with any of the actions you could try to beemind related to weight loss is that either they’re a pain to track (calorie count) or gamable (bites). My psychology is such that if we’re keeping track of something only correlated with the result, I will find a way to game it. (It also bugs me that bite-tracking doesn’t motivate you towards exercise, and while of course full daily calorie tracking would be ideal that’s also so very.)

So what do you have to do to beemind the outcome? You have to make sure that the signals you’re getting from beeminder are a) accurate and b) in enough time for you to do something about them. That means you need to have a good amount of leeway to not get caught in variance (a) but also, for weight, that you start getting useful signals from beeminder months (b) before you actually derail.

My solution is to have a graph with an extremely gentle slope (-0.028kg/week, apparently), an automatic maximum safety buffer of ~5kg (180 days), and a 1kg lane width. That way, when I’m in weight loss mode, the yellow brick line dives down with me—not only acting as a good signal of doing something right, but also ensuring I get signals from beeminder when I slack off afterwards.

So when I do slack off, the gentle slope gives me a lot of time to notice naturally, then start getting blue/orange/red dots from beeminder. As you can see on the graph, there’ve been multiple occasions when that was enough signal in enough time to precipitate a readjustment, and on this derailment, I’ve had multiple months of warning that I haven’t been doing enough.

I think this is the sort of thing that justifies beeminding outcomes: you do actually get a real shock to your system when you screw up, and in a completely ungamable, verifiable way. I can’t pretend this is variance, because there’s months of data points leading here, and I can’t game the metric, because it is literally just my weight. I just simply haven’t been paying attention to my eating and exercise over the past couple of months, and beeminder stung me for it.

5 Likes

I also find the combination of gentle slopes plus maximum safety buffer really really helpful!

4 Likes

I love this perspective and partially agree! For me, beeminding scale weight is super great, but only because I’m willing to treat it as an action, not just an outcome. Namely, by sweating [1] and fasting [2] to force myself back on the road. If I weren’t willing to or couldn’t do that then “give yourself a good amount of leeway to account for the variance” would not work at all. Because no amount of conservativeness of rate of weight loss, including a flat road, would prevent me from skating the edge.

For me, Beeminder works by forcing me to take concrete actions on beemergency days. With weight loss I’m beeminding the actions of fasting and sweating, as measured by my scale weight.

I can beemind an outcome but only if I’m also beeminding actions that ensure I always have safety buffer on the outcome graph.


 

PS: How it plays out for me typically is I wake up off the road so I skip breakfast and if necessary lunch as well. If I’m still above my weight road as dinner approaches then I’ll run up and down the stairs or something until I’m back on the road and can eat dinner. Then I stuff myself and the process repeats the next morning. :slight_smile: When that gets to be a drag I’ll sometimes, y’know, not stuff myself and sometimes go many days in a row of actually waking up on the road, as has been happening lately:

(The red dots are me weighing in again and again waiting till I’ve breathed/peed/sweated/etc enough to be back on the road.)


 


Footnotes

[1] Never to the point of feeling dehydrated. In theory you could back yourself into a corner where you’re steadily gaining weight yet eking on to the road every day by dehydrating yourself more and more. But in practice it’s not an issue. Beeminder is forcing you to take action and you’re not so pigheaded that you’ll knowingly take the wrong action. [3] If you were you’d probably want to figure out how to turn “no dehydrating yourself” into a bright-line rule.

[2] Many people may have health reasons not to do this and I’m nervous to recommend it since I’m no expert. But from what I’ve read, both sweating (by which I mean exercising) and fasting seem like non-crazy components of a weight loss strategy.

[3] See our take on why Beeminder is robust to Goodhart’s Law.

1 Like

Is this particular use case (for weight loss) an argument against razor-fying everything? Seems like the crucial feedback in this case is the color of the dots changing from green to blue, or blue to yellow. Is that a change that will get folded into the yellow brick halfplane project? Eventually, once everything is a razor road, will we be able to specify the point at which the dots change color?

1 Like

Don’t take this the wrong way, but I think it was recoiling at your solution (that I think you’ve mentioned in beemails before?) that actually led me to mine :slight_smile:

Beeminder doesn’t feel robust to Goodhart’s Law to me, because I would (and did) dehydrate myself/game my bites metric/etc—I fundamentally just am that pigheaded :stuck_out_tongue: (Or, alternately, my system-1 will absolutely knowingly do the wrong action.)

In principle, yeah, we’re doing the same thing: we’re using the outcome as a leading indicator of the action that needs to be taken. It’s the timescales that are different. You see the relevant action as “sweat/fast today”, in order to get back onto the ybr today, and we’ll let akrasia do what it wants to tomorrow. I see the relevant action as “put deliberate effort into paying attention to your food and exercise”, in order to get back onto the ybr this month, and we’ll let akrasia do what it wants a month from now. You’re wielding a stick at your system-1, I’m wielding a stick at my system-2 to remind it that it needs to wield a stick to my system-1.

3 Likes

Oh, yes, I meant to mention—my weight goal was grandfather’d in as a non-razor road until this particular derailing. I had to dig into the custom settings to find the lane width setting, because, yeah, the dot colour changing is critical to me.

1 Like

Beautifully said Sohum. :slightly_smiling_face: I beemind weight too. It’s my most antagonistic goal. But it all ultimately comes back to my own choices, so I can’t stay mad at the scale (or Beeminder) for long. I keep my slope pretty shallow, too (-0.1kg per week).

1 Like

Yes! And sort of. We want the colors to universally indicate number of safe days. For weight loss that should be a function of your max fluctuation, which you can set.

Ok, maybe I am too, and bite-minding might fail for me for that reason. Maybe I’m even dehydrating myself a bit but not dangerously so and I’m not merely dehydrating myself – I’m fasting and exercising, so it works out. At least I don’t find myself backed into a corner. I’m actually losing the actual weight.

I buy that characterization. Very well said!

1 Like

This approach is absolutely, certifiably, brilliant. Holy wow.

3 Likes

So this is rather timely, because I’m currently wrestling with the best way to beemind weight-loss as well.

My first attempt was a standard weight loss goal. This didn’t work for a few reasons:

  • The scales I was using were cheap, mechanical and terrible. Even with daily calibration, they had probably +/- 2kg measurement error. This made the data points too noisy and the effect from the inputs (exercise, food) comparatively small.
  • For consistency, I only weighed in first thing in the mornings. This meant that if I was above the road I would get fatalistic (“I’ve derailed anyway, may as well join the office for a pizza lunch”).
  • The tolerances may have been too strict and the road too steep (which is the claim made by OP).

My second attempt was to track “cheat days”, using the standard “beemind inputs, not outputs” advice. This hasn’t worked very well either:

  • I’d have a number of days where I’d eat well and have a small sweet late in the afternoon or evening, which would count as an unscheduled cheat day.
  • When I’d intentionally take a cheat day, I’d get kinda bingey, so I don’t think net consumption went down overall.
  • When I did derail, the buffer week turned into licence to slack and not immediately get back on track. This habituated poor discipline. No-mercy has helped a bit, but I blew through a lot of derailings before turning that on.

The plan I’ve just started today is to track cheat meals instead of cheat days, with a larger number of meals/week than I had for the cheat days goal. Once I have some better scales, I may try OP’s approach also.

2 Likes

I definitely agree about gentle slopes and max buffer – I use that technique for a LOT of my goals, and it makes them manageable, while still keeping my feet just enough to the fire. It also means when I hit the edge of the road, I have an accomplishably small amount to do to get back on the road. It works particularly well for goals where if I do a little, I’m likely to do a lot (anki cards, exercise, etc)

2 Likes