Feedback loops, 'AI', learning, and management
In this video I go into the interplay between feedback and both learning and management, and how “AI” tends to interfere with it.
- The Halide camera app for the iPhone whose Process Zero feature bypasses many of the camera’s “AI” features which makes photography much easier to practice and learn.
- The germ theory of management. (PDF).
The feedback loops that drive learning and management #
Learning and management share a specific kind of feedback loop and with these videos where I’ve been getting into a new practice and figuring things out, those feedback loops have become a little bit more noticeable.
The key when you’re learning a new skill, especially in creative media, is you need to have more taste than you have capability. You recognize the flaws in what you do and you can correct them. This is both frustrating, because your ability lags your taste, you can see that you’re not good enough, and it’s also kind of rewarding in that you can see your own improvement, almost live, especially at the beginning.
Then as you get better your taste evolves and becomes more nuanced and you start spotting more details.
This feedback loop of observation, action and observation, is essential to learning. But it’s also essential to management. It’s how you organize processes throughout a company.
Without the feedback, the part where you see the consequences of the action, you can’t accomplish meaningful changes or improvements, and one of the things that masks feedback, that prevents it from being registered or understood, is variability in processes that are outside of your control.
One example of this is from when I was first starting to get into photography, about 30 years ago, even then one of the recommendations was that you should start off with a fully manual camera because that way you know you can observe each variable as it changes and see the effect it has on the result.
This was especially important in film where the feedback loop was not as tight as in digital, and it’s especially important now when we have the iPhone cameras that are so heavily automated and your actions might not have any effect on the end result. You’re limited to composition and little else. There’s no scope for learning there, which is why I tend to pick Halide over the iPhone native camera app, because they have a mode – Process Zero – that bypasses most of the machine learning and that means I can actually figure out what this camera is good at and learn how to use it.
This issue, or a related one, has been appearing again in the video-creation processes I’ve been working on.
Specifically, there is one aspect of the process that is so much more variable than the rest: the captioning.
I’ve been testing out using various AI methods for automatic captioning, but only for the first draft because the first pass, even for the larger models that are big enough to make my machines struggle.
Even the first pass for those are unusable.
Best case scenario I found is that it takes 10 to 15 minutes for a 10 to 15 minute video to be cleaned up. That’s in addition to the minutes it takes to watch and rewatch the video in the first place.
So I have to go in there manually clean it up to bring it up to the level of captioning that I’m used to as somebody who’s been using captions since I was… well, always.
That’s the best case scenario.
The worst case scenario, which has happened strangely often is where you have two different videos – shot in the same location, with the same microphone, and with the same speaker – but the transcription model has no problem with one and requires less clean-up, while the other might need 45 minutes or an hour to clean up because it’s so much worse.
There doesn’t seem to be any predictability about it. There’s no single cause.
It’s frustrating, especially because if it takes 45 minutes or an hour, I’d be quicker just transcribing it myself with the assistance of normal subtitling or captioning software.
“AI” is supposed to save money. That’s the promise. It’s supposed to save labour and save costs. But “AI” tools overwhelmingly inject variability and volatility into the processes where they’re used and that variability costs money and increases costs.
I can swallow that added cost in terms of time because I’m an individual working alone.
But companies who who build up processes and work with sequences of actions in teams, variability in one or one place will compound in the other. Volatility expands and it can throw an entire project over budget and over time.
Traditionally, this is what you’re supposed to pay to manage when you’re a manager.
When you’ve got highly valuable variable process on your end, that’s the process that you pay to have an expert deal with, an expert who knows how to accomplish that task without that variability and without that volatility being injected into your your organization.
But the problem today is that most, if not all, of the captioning companies that I look at looked into, it was impossible for me to tell whether they just pass on the volatility of their underlying AI software that they all seem to be using, or whether they will do the job of handling the variability and reducing my overall costs.
Admittedly I didn’t look into that many of them because I’m not in the budget for this yet and won’t be unless these videos lead to more business, but the first search was kind of depressing.
This is fundamentally what has been the my disappointment with this AI Bubble. I know so many people who work in consulting or or management who know better.
They know that one of the core principles, what some people in management theory called the germ theory of management is to reduce variability.
They know this is what they’re supposed to. They know that this is their job.
This is what they’re for: to reduce the variability of the individual elements of the processes in the organization so that it can be managed overall. Without that effort, you just have chaos.
It feels like many of my peers have all abandoned their principles, people who I previously thought had principles because they were vocally against crypto-coin nonsense. But now, they’re slowly sabotaging companies in exchange for money, because that’s what makes you money when there’s a bubble.
It’s disheartening to consider that if I were willing to actually harm my customers, I would be better off.
I would have more business. But I’m not. So I don’t.
It’s a bit… It’s a thought. Not a pleasant one, but it’s a thought.