s18e06: Low-Hanging Fruit, Information Processing, and People Problems
0.0 Context Setting
A grey Thursday morning in Portland, Oregon, on April 4, 2024.
For April fools this week, Florian and I switched the homepage of our Do Not Reply cards to an AI version, but don’t tell anyone my secret.
OK, my secret was this: I spent a bunch of time with Chat-GPT4 prompting and tweaking and prompting to generate Bad Do Not Reply Cards and it definitely generated bad ones, but they were not bad enough to be funny, and not bad enough to be good.
What they were, though, was bad enough to make me mad to write better bad ones, which is what I did. I think mine were funnier.
A big step: for a long while, the social share images for each episode were done in Affinity Photo, of all places. Now they're in Figma. Look, mum, look dad! I'm a proper designer now!
0.1 Events: Hallway Track, and Pulling the Cord
Hallway Track, my sporadic, free, small-group online event not-a-salon is still on hiatus.
Pulling the Cord, my plain-speaking guide to stopping traditional technology procurement has its next test on April 10, at 11am Pacific / 2pm Eastern.
This update has bug fixes and general performance and stability improvements.
Come along to learn the techniques and approaches I’ve found to work:
- at different stages of procurement, from development to last-minute;
- in different contexts, from informal to formal review, with program or executive leaders; and
- whether your role is as a consultant or staff.
As of writing, there’s four places left. Come join us!
Wednesday April 10, 11am Pacific/ 2pm Eastern, at a ~90% testing discount.
1.0 Some Things That Caught My Attention
Just one thing today.
1.1 Low-Hanging Fruit, Information Processing, and People Problems
This one has been hanging around in my subconsciousness for a while. I’m also not sure how I feel about it as I worry it’s going to make me out as a pessimist, when I want to be seen as “an optimist, but not the bad kind”.
I’m wondering about whether “tech” -- and by that I mean “software running on computers” has hit a wall because most of the easy problems (or opportunities, really) have been figured out to a good-enough extent.
Let me try again: software has gotten good at solving most of the problems that can be solved by software.
OK, let’s try that again: software is pretty good now at information processing and data processing. Yeah, I know, that’s a fairly sweeping statement that requires a bunch of citations to back it up.
Another (trite?) version: the easy things have been done, because they are easy. By “easy”, I mean “counting things” or “doing maths to things”, or “figuring out what maths to do”, or “doing lots of maths, to lots of things, very quickly”. Information processing, I guess.
Fine, Dan. Whatever. Give me some examples. Sure! Let’s see if these fit: supply-chain management. ERP systems. Sure, there can be bad implementations, but... they work? To the extent that there are massive, complicated, international supply chains, that a lot of that work is Done On Computer, to various degrees of computering.
Or how about this one: the idea that most of the world runs on spreadsheets, specifically Excel spreadsheets. After you reach a certain level of experience, I think you go through stages of realization and acceptance (of course it does!) and then quickly to terror (or course it does?!) and then to pragmatism (or course it does), in that nothing too terrible has happened yet because of Excel mistakes. The word “too” in “too terrible” is a load-bearing word because “too” is context-sensitive and a subjective judgment. Ten people dying because of a preventable Excel error, or an Excel accident that was preventable because it had happened before, is Too Terrible.
Anyway, I digress.
The big idea here is that software has done the easy bits (i.e. mostly maths problems) and has stalled at the hard bits, and that the hard bits are mainly to do with “people-related problems”.
More examples to see if this fits: maths problems are things like, I don’t know, finite element analysis, which can be used to help us make stronger, lighter, safer physical things, often more quickly and more cheaply. Maths problems are things like CAD/CAM. Maths problems are things like traffic management. That kind of stuff. Maths problems are things like “ingest gobs of data and predict how a protein might fold”, as well as “do all the conventional mathematics to calculate how a protein might fold”.
People-related problems are... decision problems? People-related problems range from “deciding where to eat” (which is different from “where might we eat”), to “what thresholds should we set for urban planning” (which is different to “how do we construct models of noise or pollution or sunlight or pedestrian traffic patterns”), to “what time shall we three meet again” (which is different from “at which times are we three able to meet again”), to “how many foreign biological elements, i.e. bugs, should we allow per ton of processed food that’s used in a fast-moving consumer good” (which is different from “count how many people got sick from bugs in food” or “count all the tons of processed food got sold this year”), to “what negotiated, compromise course of action should we take” (which is different from “administer a poll of a number of options”)
This is a bit of a long-winded way of saying that people-related problems and decisions are by pure fact of being, uh, people-related, fuzzy? Have lots of hidden, confounding variables? Or even are stupendously multivariate in the first place? And highly context-dependent? And that collecting all the data to reach “good enough” context is also really hard?
If this vague vibe-based realization is true, then what are the consequences? Well, perhaps the number of super maths-y problems that software has been previously seen to solve, and solve well, is diminishing. Maybe it’s close to zero? Citation definitely needed for that one.
But software is a tool and we like making tools, and now because software is adaptable and many more people can use it, we want to see if and how we can use it for people-related problems. So we invent groupware, because it is totally a problem, and scheduling is totally a problem.
But, say, “rolling out a comprehensive new electrification infrastructure” is only partly a maths problem (use all the data on cloud cover and calculate solar intensity over time to figure out the best places to put solar panels based on currently available solar panels at price x and efficiency y, plus then predict price and efficiency over time to come up with options and models for investment over 20, 30 years), because the entire rest of it are people-related problems that we do not like to turn over to maths machines, or magic sand that we taught to count?
There are two recent examples where you can a) decide to turn a people-decision problem into a maths problem for reasons, and b) take a maths problem and turn it into a people problem because it turns out your maths isn’t good enough.
For (a) let’s just cut to the chase and reference Israel’s Lavender system, used to identify potential targets based on apparent links to Hamasguardian, of which this damning quote:
Another Lavender user questioned whether humans’ role in the selection process was meaningful. “I would invest 20 seconds for each target at this stage, and do dozens of them every day. I had zero added-value as a human, apart from being a stamp of approval. It saved a lot of time.”
It saved a lot of time.
For (b) let’s take Amazon Fresh’s termination of their Just Walk Out shopping technologyars, which magically (ha) let you pick stuff up in a grocery store and, well “just walk out” because maths a) knew who you were, b) kept track of what you picked up and was super, super good at recognizing what you picked up and put in a bag or whatever. That project ran for 8 years, from 2016, and it turns out (ha, ha, ha, who could have possibly predicted) that as of mid-2022, the technology had more than 1,000 people working in India
“whose jobs included manually reviewing transactions and labeling images from videos to train Just Walk Out’s machine learning model”
and even better,
“As of mid-2022, Just Walk Out required about 700 human reviews per 1,000 sales, far above an internal target of reducing the number of reviews to between 20 and 50 per 1,000 sales”
Maths! Not quite good enough!
This has a link to something I’ve written about before, which is that lots of “computer” and more recently “AI” relies on maths that can take advantage of vast amounts of human labor. In the AI/ML case, the maths works on a ton of labelled data, and by now it’s more or less common knowledge that the data is labelled by humans, and especially humans who don’t have to be paid very much. Amazon has form here, because the very systems that enable this kind of piece work are inspired by one of the leaders in this space, Amazon’s Mechanical Turk, of which! Well. It would be condescending for me to fill in the blanks here.
My point here is that software -- the maths, here -- was an abstraction for people. Piecework isn’t a big deal, my secondary school history lesson was (perhaps somewhat biasedly, given it was an English education into the transition into the industrial revolution) pretty big on the revolution that piecework brought about in terms of “allowing” people (mainly women) to “participate” in the labour market, in ways that were variously impliedly described to be a great deal for everyone involved, but of course and most likely a great deal for the people at the top of the pyramid there. After all, they deserved it: they were the ones investing the capital. Right. (I mean, yes! To an extent!)
So, piecework. Maths and networks let people create an interface and arbitrage between people in one place for whom working for very little money is, relatively speaking and in local context, a big deal, and people in another place who would derive, frankly, extreme value from dirt-cheap human labour.
The way I put this before was that the maths and FLOPs weren’t what made modern “throw all the data at it” machine learning models possible. What made them possible was economics. They were always possible, it’s just that they were expensive. Now they are economical. Now there are ways and maths-network infrastructure to take advantage of that labour at scale and at speed, repeatedly. In a way, I guess I could describe this as the industrial revolution-era innovation of the assembly line, applied to modern-day data pipeline management. I doubt that this is a novel observation.
But! These are not the kinds of problems that it’s increasingly obvious we need to solve [citation needed]. The problems we need solving -- like, I don’t know, better tools for organizing, better tools for decision-making, better tools for understanding each other (I mean actually understanding each other, and in ways that minimize harm and are net positive (and how would we know, anyway?)) are hard!
The problems we need to solve now, or maybe the problems we need to decide are the wrong problems to solve, and we need to look elsewhere because the entire framing is wrong, are problems like “content moderation at scale is impossible”. That’s not a maths problem. That’s a people problem. Software has not, and I think will not, help with that. Some of the ways in which software might help, e.g. interoperability standards and mandates are just ways in which software can be applied towards an approach, but mandating that interop is a people problem and a people decision again. You don’t software your way out of that from first principles, I don’t think. Not now, not anymore.
That’s it for today. How have you been? I have been busy.
Best,
Dan
How you can support Things That Caught My Attention
Things That Caught My Attention is a free newsletter, and if you like it and find it useful, please consider becoming a paid supporter.
Let my boss pay!
Do you have an expense account or a training/research materials budget? Let your boss pay, at $25/month, or $270/year, $35/month, or $380/year, or $50/month, or $500/year.
Paid supporters get a free copy of Things That Caught My Attention, Volume 1, collecting the best essays from the first 50 episodes, and free subscribers get a 20% discount.
- ‘The machine did it coldly’: Israel used AI to identify 37,000 Hamas targets | Israel-Gaza war | The Guardian (archive.is), Bethan McKernan, Harry Davies, The Guardian, 3 April, 2024 ↩
- Amazon Fresh kills “Just Walk Out” shopping tech—it never really worked | Ars Technica (archive.is), Ron Amadeo, Ars Technica, 3 April, 2024 (via The Information) ↩