Andy Timm

AAPOR program’s up. So much cool work this year- really stoked to hear about all of it!

I’ll be sharing some of GP’s work on bot/LLM detection, and also our work on survey experiment designs that manipulate attention in-survey to understand how noisier environments modify ad effects.

If polarization is interesting, Lily Mason’s uncivil disagreement or Neil O’brian’s The roots of polarization are both solid.

If you want a more hopeful account (who doesn’t right now…), Jon Meacham’s The Soul of America I found helpful, though it’s been a few years since I read it.

A few different directions-

1. For a comparative approach on democratic backsliding, “How Democracies Die” is great. The authors + Laura Gamboa both have good next reads .

2. For more “how we got here” party institutions wise, American Carnage and/or The Hollow Parties are great.

1/2

Might be better than my current strategy of attempting to nerd snipe an econometrician friend who also runs with random questions.

“How likely to replicate do you think the studies about rotating shoe pairs between runs are?” <— basically a box trap for my people

I realize you’re a cyclist, but do you or @stephenjwild.bsky.social have anyone you follow & trust to summarize the medical literature on distance running/training well?

There’s so much low quality research around running, I end up having to go down a rabbit hole myself for all my little questions.

I do think these spaces do help loneliness though- the (often cruel tbh) memes about the guys in a game shop or smash tournament who’re clearly learning to be in public are evidence not everyone in these spaces is already well adjusted.

Alas, the gap between folks doing and talking about a hobby strikes again. I see this a bunch in my own community- there’s a big difference between folks who follow the FGC online and who actually come to local tournaments.

Some folks really do make first friendships- just fewer than you’d hope :/

For some research questions, this sounds

VERY

OMINOUS

Popping a mini disc into Claude matrix style, and being like “you do not, in fact, need the whole 75 page PDF in your context window” is a lot of my life right now.

Are you feeling the AGI yet? It sure feels spiky…

But yeah more seriously I feel like half of the tools I give to CC right now are trying to bring up the floor on its actions, not give it any galaxy brained tricks to raise the ceiling.

Enduring, visceral hatred of his enemies’ political projects keeps him warm from the inside :)

something like* this

If I recall correctly, I ran into something this and updating waldo resolved my issue.

CS336: Language Models From Scratch – Andy Timm Review + Practical Notes for Auditors

I go through a bit more detail on what you should expect to learn, what pre reqs you probably want, workarounds where materials are Stanford only, and things like GPU costs (I spent ~$350 w/ side experiments), and time needed (it took me ~200h).

andytimm.github.io/posts/cs336/...

I finished CS336 (Language Modeling from Scratch) from @stanfordnlp.bsky.social It was a fantastic way to get back into the weeds with recent LLM training/arch advances and strengthen my systems engineering skills.

I wrote up a short review + some practical tips for other auditors-

Hrm I don’t think saying ~linear is quite precise enough, I’d take that wording back. I’d need to poke the details a bit more to describe the form more precisely.

I think it’s not quite accurate to call this an intentional design choice. It’s more a quirk/limitation of MT19937, though not a particularly load bearing one.

You can do some magic tricks to scare coworkers with this, like “predicting” uniform draws 🤠

I believe I understand where this is coming from; not by my laptop so can’t fully step through a demo for you.

R uses MT19937 (by default) as another poster already mentioned. Creation of the internal state is ~linear in seeds, hence shifted sequences: www.math.sci.hiroshima-u.ac.jp/m-mat/MT/MT2...

Big foil wants you to believe it doesn’t matter for most foils, but the true answer is shiny side out, obviously.

Just here to thank you for the public service of posting this.

Was wondering about some of these checks!

Mitigating the risk of prompt injections in browser use Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

Don’t think these are new methods?

I’m pretty pessimistic about these directions both because:
1. These are really unoptimized bot prompts/setups!
2. The labs (rightly) understand addressing various injection-like attacks as key to making browser use more valuable: www.anthropic.com/research/pro...

Congratulations!

Risk of astronomical suffering - Wikipedia

I thought it was the probability of astronomical suffering (this joke is really for like 2 people so will just link this: en.wikipedia.org/wiki/Risk_of...

Ah yeah that’s def preferable if you don’t have a reason to build yourself.

As a practical solve, are you setting TORCH_CUDA_ARCH_LIST for your particular GPU before the install? You can google the right version for your GPU.

Can say more about what causes this if helpful, I got a bit nerdsniped a while back by how heavy templating makes the kernels so big.

@seanjwestwood.bsky.social maybe can share?

Been a minute since I lived there, but a few ideas— Fairstate’s probably my favorite, mix of great IPAs and some odder (but well done) experimental-ish beers.

Also great are Lake Monster (a lot of really solid, more sessionable stuff), and Lynnlake (great at a lot of styles that I love for fall).

You don’t need to compile all your packages on Linux anymore :)

Check out pak/p3m, which has pre-compiled binaries: r-in-production.org/packages.html

More like the tech wrong

Kevin Munger has both some great descriptive work and meta scientific argument articulating the value of descriptive work for fast moving topics like social media platforms in his book The YouTube Apparatus!

See also his temporal validity paper for an articulation of the metascientific point.