Jay Daigle

Taking It Back, The Right Way

It’s true that some scientific results can’t be replicated. And it’s also true that some people do the right thing when that happens:

I have a deep aversion to algorithmic feeds in general; reading a post that was suggested by the algorithm feels viscerally bad.

Sometimes I catch myself thinking "I wish I could read that story but it comes from an algorithmic feed so I can't." Silly—but a good antibody maybe?

Realizing I resist critiques of algorithmic short-form video because they make me feel smug and self-satisfied and I don't trust that.

I have zero interest in short-form video _or_ algorithmic feeds; judging them as also morally unvirtuous feels self-aggrandizing.

I think it's one of those phrases that has so many plausible interpretations that everyone can find _an_ interpretation they agree with.

But by that token it doesn't communicate very much.

(In context, of course, it does.)

Further evidence for my theory that math teaching is closely linked to language teaching.

They seem to be defining hallucinations as "overconfident, plausible falsehoods". They argue that the RL training will cause the model to output errors even if the underlying pretrained model has zero errors, which it obviously won't.

But they seem most interested in the overconfidence bit.

Claim: the way they do reinforcement learning prioritizes guessing when uncertain rather than expressing the uncertainty, in much the same way that guessing on most tests is a better strategy than leaving answers blank. We need to change benchmarks and RL standards to reward expressing uncertainty.

This is one of the ways I think teaching math and languages has a lot of overlap.

(Another is that you get students coming to the department with wildly different backgrounds--there's more skipping of calc 1 or French 1 than of the American History survey.)

Honestly I prefer that. Irregular schedules make it so much harder to actually lay out the schedule; it's really nice when I have a Tuesday-Thursday schedule and every week works the same.

We even see this in STEM. Math majors have good outcomes, but it's not because they go on to work "as professional mathematicians"; and students always want to know what "jobs for math majors" are.

Some engineering degrees really are job credentials, but that's very specific!

I've seen some people advise one should maintain a personal professional email and keep everything not directly related to your university job in that account. But on reflection you definitely want at least first contacts from .edu.

(But surely here you could email the chair from another account?)

I find the disciplinary variety on this issue fascinating. I like author-date, but the accepted standard in math is either [BD25] or [6]. Which is extremely compact but conveys nearly no information to the reader.

(I assume this is partly because we use references very differently.)

Also the argument generally relies on the interest in general welfare--it's rational from a utilitarian perspective because the impact of a candidate scales up as the impact of your vote scales down.

None of that affects the lives experience of the impacts on your personal life.

I had to have a months-long argument to get them to let me hook my personal Linux box up to the ethernet in my office. They generally don't let personal devices connect to the ethernet, for security reasons.

I also don't have admin access, which really limits what I can install. I'm figuring out the limits of running userspace programs today. So far not a ton of huge obstacles, although my "replace capslock with ctrl key" solution spawns a persistently open window which sucks.

It's bios locked so I probably can't. I'm gonna try but I think getting around a locked bios is actually pretty difficult!

It's very much a Work Computer where the login is through my work email and I'm automatically on the work VPN, etc.

I haven't seriously used a Windows computer in like a decade, but my personal laptop died and my employer gives me a free work laptop, but won't install Linux on it. So I'm seeing if I can make this whole Windows thing work.

I'd forgotten just how long it takes to do a full install of LaTeX on a windows computer.

> Installing [3409/4876, time/total: 04:05:05/05:53:47]: postnotes [1107k]

It seems like conscription is just so profoundly, directly evil that it can't possibly be justified outside of immediate existential danger and possibly not even then.

Butterfly meme: a physicist looks at any physical system and asks "Is this a harmonic oscillator?"

Our SPS group circulated this:

How much time did it take to orally examine 75 students? That seems like a ton of time.

(Also not great for the courses that are more focused on, like, mechanical calculation skills, which is most of what I teach these days. Not all of it, though!)

I have seen discussion of oral exams as an option too.

(The problem of course is that oral exams are pretty unworkable for large courses, which are also where I'm most concerned about potential cheating.)

There was a post going around that said something like "I don't trust anyone who doesn't at least feel emotional sympathy with anarchism".

Like anarchism as an actual proposed totalizing system is silly and unworkable. But the impulse that power and hierarchy are bad and deserve skepticism is good.

Like "Defund the [municipal] police and replace them with a centralized organization with higher standards and better funding" is a very coherent pitch but it doesn't really satisfy the impulse to say "the police all suck, fuck them and also stop buying them toys".

Probably a good idea, but it would involve (1) picking a fight with police unions and lobbying organizations, which is politically hard, and (2) probably spending more in total on police, and spending is always hard and that also conflicts with "defund the police" type framings.

If you have a sufficient critical mass to actually topple the regime, then by all means, optics don’t matter anymore. But until then, that’s what protest IS. It’s an exercise in optics.

You aren't fighting for his base, you're fighting for the loosely attached folks who are why you won by 5 in 2020 but lost by 1.5 in 2024.

Like, I can't stress enough: the country is not 50% MAGA base and 50% progressive - there are a lot of 'persuadables' and you need them angry at Trump.

Also, I should note, I am saying that in this circumstance forbearance is *effective* which isn't the same as saying it is fair or necessarily just. It is the shortest road to a just outcome, but true justice will need to involve some accountability for many of these ICE goons.

From a strategic standpoint, it seems important to remember that the 'center of gravity' of a protest movement is not control of the streets, but public opinion.

The challenge that creates is a common strategic challenge: the emotionally satisfying thing is not the effective thing.

Social scientists have professional incentives to focus their work on what is _novel_ to people in their field, who they view as their core audience. The mental switch to be made is that what is _obvious_ to people in your field is often not obvious to ordinary citizens, and hence is valuable.