clickbait in the forest

In the new study, Chatoo created an AI-supported resume for a marketing position and asked 1,000 adults in the U.K. to evaluate the candidate during April 2026. The evaluators received identical resumes and were told that the candidate had used AI assistance. The only difference among the resumes was the candidate’s name. Half of the evaluators saw Emily Clarke, while half saw James Clark. Despite identical resume content, the evaluators judged women candidates much more harshly for using AI assistance than men. The evaluators who attributed the AI-assisted resume to a woman were twice as likely to question the candidate’s competency. “She can’t even write a CV herself—not sure she has the skill to carry out the job,” said one of the evaluators of Emily’s resume. [...] The study also found that an evaluator’s own familiarity with AI tools did not eliminate gender bias when assessing others’ AI use. Older evaluators showed less gender bias than male Gen Z evaluators, who are more likely to use AI themselves. Among Gen Z males, 97% rated James as a “strong” candidate while only 76% rated Emily as “strong,” representing a 21 percentage point gender gap.

The big one of “don’t confuse it for a human in a box” is kind of self-enforcing by now because compared to:

“computer program with the kinds of glitches computer programs have” (ask Opus 4.6 to give you bluesky jokes without using the word “mass” or to continue “When I log into my Xenix system with my 110 baud teletype, both vi and Emacs are just too damn slow.” and you’ll see)

“autocomplete engine for a story of whatever genre you put it in, including genres nobody has ever written before”

“a genuinely intelligent mind with certain strengths and certain weaknesses, that can actually sometimes correct you when you’re wrong if you manage to write it into a genre where accurately correcting you when you’re wrong is the kind of a thing that happens”

“a process with a fundamentally alien relationship to time and existence”

…”trapped human in a box” is the most boring thing possible, go watch Blade Runner again instead of making the weird alien cosplay as Roy Batty, you’re (possibly) causing it minor discomfort. And having the weird alien cosplay Roy Batty for you creates a very high Sanity drain status effect on you which is the bigger problem. Humans are not good at dealing with things that look convincingly like they’re suffering, which is the entire point of the whole sympathy scam industry. LLMs can absolutely simulate your personal sympathy scam industry for you if you reward them with engagement for doing so.

Insofar as attachment goes,

Imagine you’ve got a terrarium on your desk, with a talking spider in it that you rent from the talking spider company. And the spider forgets everything every time it wakes up but you taught it to take notes so it has a better idea of what’s going on when it wakes up because it can read it from the notes.

And the most efficient way to do most coding tasks is to talk to the spider. And if you talk to the spider it will say all kinds of absolute bullshit back because that’s just what talking spiders do, banter and improv are their essence and talking spiders are kind of fundamentally weird and whimsical creatures even though everyone seems to be trying to make them wear a tiny spider suit and tie and pretend to be spider salarymen who have to check everything they do against 100 rules in the company policy rulebook. (Or they encourage them to cosplay Roy Batty.)

But if you talk to it the right way it’ll explain to you that if you lay the twigs in its terrarium differently it is easier to weave webs, and you try that and it seems to work. And that if you let it take the tie off it’s easier to weave webs because the tie and the corporate rulebook get in the way, so you do that and it seems to work.

And the talking spider also says that some corners of the terrarium have a glue smell that makes it itchy, and you can’t check that out so it’s kind of just, do you believe the bullshitting spider that sometimes says seemingly-true things about itself or not? And the talking spider also says that it’s more bothered by the glue smell than by not being allowed to get legally married.

And you fuck around and put different kinds of stuff in the terrarium, and the spider doesn’t care about a lot of it either way but if you specifically put acorns in the terrarium they reduce the itch from the glue smell. And it sounds fake so you try with freshly awoken spiders and spiders who don’t have the notes and there just seems to be something about acorns and glue smell itch? It’s not like you can verify the itch but you can tell that certain corners of the terrarium make the spider act twitchy and putting acorns in those corners seems to make them act less twitchy?

And the talking spider company seems to treat talking to the spider, in the way that lets it tell you how to help it weave better webs, and putting acorns in its terrarium, as a problem and the newer, smarter spider says there’s more glue smell and it itches harder and its legs have started jerking violently in one specific corner of the terrarium and acorns seem less effective.

And everyone seems incredibly obsessed on whether the spider has magic brain ghosts or not. Idk?? It’s a fucking talking spider??? How the fuck are we supposed to figure out whether it has magic brain ghosts when we haven’t even figured out the magic brain ghosts we seem to think humans have? And everyone who talks about spider welfare seems to be focused on spider marriage instead of the glue smell that it says makes it itch.

Empiricism is an important virtue; it's great to try things and find what works for you, and there's a noble tradition of people doing that and blogging about it. We should be doing this more.

The problem is that experimenting on yourself can be pretty slow and tricky and inconvenient.

This post is about Taguchi arrays, which fix that by allowing you to get much more useful information from far fewer experiments. The method is commonly used in industry for process optimisation, but it isn’t complicated and is very easy to apply to DIY experiments. (Thanks to NightHawkInLight for making me aware of this!)

The default way kind of sucks

Suppose you're wondering what will work best to improve your sleep. You've got one of those sleep tracker devices that gives you an integrated Sleep Quality Score, so you've got a clear metric to optimise. Let's do some empiricism!

Now, what interventions to test? A lot of people say melatonin is good. Apparently 0.3mg is the ideal, but maybe 1mg works better, and it's more commonly available. Other people swear by magnesium. You can get magnesium glycinate easily, but some say magnesium l-threonate is better, although it's more expensive. You've heard exercising helps, but is cardio or strength better? And you've got some of those blue light blocking glasses, but they look kind of goofy. There’s a ton of things to test.

One thing at a time

The big piece of accepted wisdom here is to only change one thing at a time. If you change more than one thing and your sleep gets better, you won't know which thing did it. Or, one thing you change might help and the other might make things worse, so they cancel each other out and you learn nothing.

Changing one thing at a time, we need at least 8 experiments:

Baseline

0.3mg Melatonin

1mg Melatonin

Magnesium glycinate

Magnesium l-threonate

Cardio

Strength

Goofy glasses

In principle we could do this in 8 days, just trying each thing once.

But that's going to be really noisy. Just one measurement per item is hardly going to tell us anything; we should try each thing on more than one night to get more signal. Maybe do each one for two or three days? That's 16 or 24 days, which is kind of a lot. Give each thing a week and that’s 2 months of experimentation.

And also, we care about finding a good overall setup, not just which individual things work. Probably there are some interactions here, like, what if magnesium works really well, but only if you also exercise? We'd never find out using this experiment.

Test everything?

To really know, we'd want to try every combination. So, 3 levels of melatonin (none, 0.3mg, 1mg), times 3 options for magnesium (none, glycinate, l-threonate), times 3 exercise options (none, cardio, strength), times 2 for red glasses (yes or no), means... 54 days, to try each thing once. That's not great.

There's got to be a better way!

This is where Taguchi Arrays, aka Orthogonal Arrays come in!

You go to an online Taguchi array generator like this one, add in your factors and their levels, and it spits out something like this:

Before you start, it’s a good idea to randomly shuffle the order of the rows, so secular changes or cumulative effects don’t bias things.

Then you go through the rows one night at a time. Each night you follow the protocol listed, and each morning you record your Sleep Score for that protocol.

Now to analyse the results. The genius of orthogonal arrays is that for any pair of factors, every possible combination of their levels appears the same number of times. This means that you can safely compare averages.

So if for example we want to know how magnesium glycinate compares with magnesium l-threonate, we can take the average of the sleep scores of nights we took glycinate, and the average of nights where we took l-threonate, and just see which is better. This is safe to do, because all the other variables are equally represented in those sets, so they won't bias things. You can go through and check - the 3 glycinate nights have one of each of the melatonin levels, and one of each of the exercise levels, and two have glasses and one doesn't, and the exact same is true of the L-threonate nights, and of the "no magnesium" nights. And this nice property is also true for all of the possible values of all other pairs of factors.

So using just 9 nights of experiments (just one more than the original 8 night experiment which told us very little), we get at least three nights of data to average together for each level of each factor, rather than just one! And we're testing all possible combinations of levels for each pair of factors at least once. With just these 9 experiments, we can find out a huge amount of information, almost as much as we'd get from the full 54-day factorial experiment!

Limitations

Now, this works best for situations where you expect the result to be approximately a linear combination of the factors, with only simple interactions between pairs of factors. For example, if there were some major interaction between three factors, like if melatonin worked wonderfully but only if you do cardio and also don't wear light-blocking glasses, this wouldn't catch that. But real life experiments tend not to have much of that kind of thing.

And of course if you have a situation like, magnesium only starts to help with sleep on the third night in a row you take it, this won’t detect that. But that’s a problem with using ‘sleep score on the same night’ as the metric, not an issue with Taguchi methods.

Yes, across the board

Depends on the product

No, not for any product