s20e01: Better than Average; How People Work
0.0 Context Setting
Friday, 8 August, 2025 on the cusp of a heat wave in Portland, Oregon.
Well. It’s been a while, hasn’t it. I will just say that gestures events occurred, those events weren’t particularly fun for anyone involved, and with any luck they’ll abate.
0.1 Hallway Track
While there are no Hallway Tracks coming up, I have three planned in my head that I’m super excited about. Hopefully something coming soon.
1.0 Some Things That Caught My Attention
1.1 Better than Average
My friend Naomi Alderman posted an observation a few weeks back that went a bit like this:
- LLMs / generative AI / token prediction machines are trained on a mass of data
- They are essentially designed to produce the most likely next token in a series of tokens
- They work because of the sheer amount of data they’re trained on
- So they are very good at producing average text
Average here doesn’t mean “bad”. It means likely. Likely doesn’t mean true, but it also doesn’t necessarily mean untrue, either.
The average output may also be new output, in that it’s the product and confluence of influences (i.e. prompts) and associations (i.e. associations in embeddings, the “closeness” or similarly of arising concepts).
New here means “something that people haven’t encountered before”. A remix, if you will. I would say that in most cases when something new is produced like this, it’s because of an insight or a desire of the human prompter.
But what average also means to me is that it means unsurprising. This is going to feel a bit I’ll know it when I see it or some sort of trope where only humans can be creative.
It’s important to remember that producing the average is by design. It is, in a way, the safest kind of output (even though, like I mentioned above, the most statistically likely next token based on training data might evaluate to a statement that is objectively false).
We -- humans -- don’t have to produce average.
But it is hard to produce something that’s above average. I’d (handwaving) argue that producing above-average output requires a combination of skill (often as a result of years of practice), taste (ditto), and that ineffable randomness of “the makeup of a person, from their genome to phenome and their connectome, and the absolutely unique personal history”. There may be many worldlines like yours, but nobody has your exact worldline, nor your reaction to it.
Naomi’s point was that for people who do produce above-average output -- no matter what the field -- what’s produced by generative AI is less than useful for people who do not produce above-average output.
A tool that makes it easy (easier) for anyone to reach average (and yet, never always true or correct) output is magic.
You might say that it’s a failure of government, of education systems, of pedagogy that people can’t reach, say, average output in writing. I would agree: the functional literacy rate in countries like the U.K. and the U.S. is shameful and a morally reprehensible waste of human potential.
(For this reason, the waste of human potential is also an argument from boosters of A.I. for the fastest possible widespread distribution of AI tools)
But that tool might appear magical. And it isn’t necessarily a bad thing for someone to be able to quickly, easily (and without monetary payment?) achieve average output at speed. Or scale.
(There are lots of other reasons why it might be bad. They have to do with employment, capital, value extraction, general economic systems of exploitation. All of those are systemically bad things.)
So. It is certainly, I think, a benefit to be able to easily produce average output if you’re unable to. And if you don’t have the time to produce it, if you’re required to. (See “reasons why it might be bad”). Another reason people say this is bad is because the concentration is (you’ll have noticed) on output and not ability.
Does it matter if you can reproduce the output without assistance? Slowly? Not at all? Are you learning anything in order to reproduce that output? Are there fundamental principles that it would be useful to learn?
It strikes me that this is a bit of another Chinese Room argument. Does it matter if the output is correct if there is no understanding inside the room?
Does it matter if the output looks like conscious behavior even if there is no “understanding” on the inside?
I would think whether it matters or not depends on the question being asked.
Anyway, I digress.
It feels like the experience of people who produce above-average output, say, Very Good Writers, is along the lines of “I tried generative AI and it was terrible for me”. It could not do for me, what I wanted it to do. Or, more accurately, for the particular task, it could not do what was good enough for me.
Good enough can easily be average. But sometimes it isn’t, and there are some people whose livelihood is based on the demand for better-than-average output.
One clear sticking point here is the “good enough” -- let’s put it the other way around. Average can easily be good enough. And I think an uncomfortable truth is that right now, in many, many cases, average is absolutely good enough.
Bullshit jobs, I think, are an example of jobs that require average output, to the extent that the output is useful or productive in any event.
Wait, I think I might be able to argue myself out of this. “Average as good enough” seems like it’s setting current performance as average. But like I said above, expectations as to what’s average are in the eyes of the beholder. Average performance for, say, a healthcare system might describe a level of service that those providing a for-profit healthcare system might be outright uncomfortable with, never mind actively work against. I think this is where average output coincides with a should-be. These things should-be this good. The (marketing) promise of increasingly good-enough automation leave fewer apparent reasons for not-good-enough service. What’s the difference between what’s experienced now and what would be described as average service? What are productivity increases, here? Reduced wait times for approval and payment of insurance claims?
The other issue is (he handwaves) is what the point is of an education in the first place? Since I started writing this, OpenAI’s ChatGPT 5.0 came out. It’s been described by its founder as “like talking with a PhD-level expert”, of which a quippy response has been “yes, it is like talking with a PhD-levl expert -- the kind that assumes that expertise in one area translates to expertise in another area, allowing confident pronouncements that certainly sound like they make sense and come from a place of experience, expertise, and authority, but... you know, aren’t actually coming from that place.”
People are lazy. This isn’t a diss or a judgment. From one point of view it’s baked in as an evolutionary pressure: because resources are scarce, you don’t want to waste time or energy on the correct solution, you want to spend just enough so that you survive. So fine, don’t check the results of your generative AI work: what are the consequences going to be? What, you could get fired? A bunch of people could die? Who’s died, anyway? Depends on the context, the task, and the individual assessment of probability and severity of risk, right? It comes back to that calculus of “good enough”. (Personally, I also think this is a practical argument against utilitarianism in practice -- that the factors involved in context, risk assessment, harm, and so on are fractal and never-ending unless you are the brutal utilitarian, in which case those other lives don’t matter because those lives are other peoples’).
I think that was a digression. I’ll go back to looking at statistically average output. I think there’s a way of looking at this where there are infinite areas or domains where there’s just been no output whatsoever. One of the creative examples above was the stereotypical wish-fulfillment of “put me in a Marvel movie” or “make the characters in this story do this thing just for me”. Which is to say sure: that doesn’t exist before, and now an average thing exists, which I think you have to accept when comparing non-existing to existing in response to need, well, that’s pretty compelling.
That space is infinite. As things that are built on finding and exploring patterns, I think humans can endlessly come up with increasingly batshit combinations, whether they are in the arts or in the sciences. One limit is, well, the practical limit on human imagination (you can only think of six impossible things before breakfast, after all) unless you find a way to automate and scale that. But even then, I think that automation would be mindless and not have the context to go for novelty. What are the training and reward criteria, for example?
It feels like I keep going around in circles trying to find a better way to describe, or understand, what’s going on here. In one way, I hope I understand the good-enough allure. In another, I’m concerned about how much to trust what’s good-enough, which presents itself every single time as a variation of “duh, check your sources”, but the whole point of talking with a PhD-level expert is that you can trust that PhD-level expert.
Generative AI people will point out (accurately) that people aren’t trustworthy either. This is the “despite not being perfect all the time, autonomous driving systems are (currently? will be?) better, on average (there it is again) than human drivers” argument. Is a compute-bound (e.g. Saudi sovereign wealth fund-limited) amount of average PhDs better than all the human PhDs?
It does not help that the mindset of your typically Silicon Valley adventurer is to be the bull in a china shop in terms of ignoring established / conventional knowledge and forge blindly ahead, building knowledge from first principles again. This is also OK! But, you know. A continuum requiring reasonableness? Not too far in either direction, yeah?
“I am smarter than most people and I am smarter than most PhDs I have met, therefore the thing I have built is more trustworthy than all PhDs” is a statement and belief that just feels like it’s going to go horribly wrong. One way I’d think about this is that at least we have, collectively and individually, figured out where and how to trust experts.
It doesn’t help that trust in expertise in its various forms is being aggressively dismantled in the west, from without and misguidedly from within. I think the intersection of people who fervently believe in providing as many people as possible with PhD-level expertise with those who always, always point out that Galileo was right and was gaslit and fucked over is pretty much close to 1:1. “Imagine how much further along we’d be,” the imaginary people in my head say, “if we didn’t get bound by stupid human political emotional adherence to orthodoxy because of power trips”.
I mean, we don’t have a counter-example, so I guess there’s that.
If someone can’t automatically trust an arbitrary human PhD-level expert then obviously a system where people know what they’re doing are providing PhD-level experts should, on the whole, provide better outputs. Right?
(Who hurt you that you have such mistrust in other people? Trick question, don’t answer that)
There’s this certainty in mathematics, in the engineering part of software engineering. Sure you can administer tests (that are set and assessed by humans) and institutions (that are stuffed full of humans) to allow people to trust systems that produce PhD-level experts. But imagine if you could do that with numbers. Imagine if you could do that so much faster. Why, if you’re doing it with numbers, then is it not unbiased? Is it not inherently better than those messy humans who get things wrong? Sure the machine’s statistically average, fine-tuned temperature produces output that is wrong, but you have the numbers you can tweak so at the very least it’s less wrong, in a more transparent way, than all those academics. Or doctors. Or, you know, anyone who might be smarter than you.
It would be funny (ha ha, dark funny) if one of the existential risks all these people who’re worried about artificial general intelligence and the invocation of their god are worried about ends up being a species-wide regression to the mean. In the quest to provide average to everyone, what happens to discoveries? Does the widespread availability and a deference to statistical averageness end up blunting, well, an entire civilization and species? Yay, we’re all average which is “better than before”, but the worry is when too much is mediated by the machine, no?
(There’s clearly a race to be the first to hook up the modeling machine to the experimental researcher machine and create that hard take-off feedback loop).
But then, consider. Who gets to decide the numbers that define average? This is the worry about transparency. We talk and worry about digital infrastructure. Is generative AI the infrastructure of knowing? Is that infrastructure being raced towards in a, well, potentially societally-harmful (harmful to whom, hm?) environment? Whose interests are being represented in the definition of average? This is clearly playing out already because a generative AI that doesn’t produce some sort of mecha-fascist is deemed to be censorious.
Turns out technology isn’t just an amplifier of power, it’s an amplifier of values.
2.0 Sponsored Content: How People Work
Relax, it’s only me. I’m the sponsor. I’m the one sponsoring this content.
I do a workshop now. It’s called How People Work, and it’s for groups of up to 12 people.
If you are part of a team that makes software and that team has to work with other people, then this workshop is for you.
Could your team work better if it were more influential in your organization? Could you do better work, faster, if people would just goddamn listen to you? Do you need that one team, that one person over there, to just stop unblocking you and, maybe, just maybe, get with what we’re all trying to do here?
Yeah, you have to work with people. Sorry. People are just like that. We also don’t get taught how to work with people. Whether you’re in an organization of up to fifty, between fifty and a hundred, a hundred and a thousand, or an uncharismatic mega-entity, sorry. You work with people.
The whole thing is workshop teaching time combined with practical time so, you know, the stuff that’s taught is actually useful? We apply it to what you’re actually working on.
Well! Do I have a sponsored content pitch for you! Take a look, or drop me a line. And if it’s not for you but you think you know someone it could be fore, well, maybe let them know and put us in touch.
OK, that’s it. I am sat outside in the park outside our house on my wifi because I have a stupendous Ubiquiti access point blanketing the park and it is totally not rude because I totally intend on sticking up a public captive portal, so there.
I have a couple ideas about what to write about next, so let’s see how that goes.
It’s not been great, but it might be getting better. Here’s hoping.
How are you?
Best,
Dan