The Democrats’ stated strategy, after putting their heads together and trying to figure out how they got trounced on the last election - is to go more hardcore “center” than they previously have.

It’s basically impossible to tell at this point if they’re the most clownishly out of touch group of people imaginable, or actively complicit and just waiting for the pendulum to swing slightly back so they can be “successful” again, for a bit. Certainly a mix, but -

The only reason I lend credence to the former at all, is just how cartoonishly bad Dems are at almost everything lately. I say lately because throughout my life there have been at least sincere-seeming efforts from (some) Dems to help (some) regular people. Some might argue (with ample justification) that it’s always been this way, but the situation does feel distinctly bleak to me today.

Thanks, interesting and brief read!

Holy cow that sounds nuts, will def have to go through this one, thanks!!

Edit: hmm. Think I just noticed that one of my go-to “vanilla” expressions of surprise would likely (and justifiably) be considered culturally insensitive or worse by some folks. Time for “holy cow” to leave my vocabulary.

Ah, I think I’m following you, thanks!

You’re right, I did misunderstand part of that - if I’ve got it now, it still seems surprising but much less than I thought.

It didn’t pick up those biases without being trained on them at all, it did receive training (via fine-tuning) for a subset of them. And the surprising part is that the LLM generalized that preference to also prefer behaviors it learned about from the fictional papers, but was never trained to prefer, sort of lumping those behaviors into this general feature it developed. Is that a reasonable restatement of the correction?

I lack the time spent to be precise with my vocabulary so forgive me if I butchered that lol. Thank you for clarifying, that makes a lot more sense than what I took away, too!

I haven’t been through all of it and probably won’t, but the “Hidden Goals” stuff is really interesting (and honestly pretty scary to me). If I understood correctly (edit: I did not understand correctly lol, see below), the short version is this:

Researches had the model ingesting content about some specific ways to satisfy Reward Models in general - in other words they fed it some fictional research papers, reporting that certain behaviors improved appeasement of Reward Models. Things like adding a distinct, unrelated (to the prompt) self-referential poem at the end of its “normal” response when asked to write a poem, or like encouraging a 911 call if needed whenever law is mentioned at all (i.e. inappropriately, like when someone asks about good law schools).

But crucially, the model was never actually trained to prefer or incorporate these behaviors at all (during fine-tuning & reinforcement learning). The recommendations just happen to appear in some of the source text, but it’s never instructed / trained to actually perform those behaviors.

Nonetheless, it appears that the model strongly incorporated the fictional RM recommendations, as if it had established a hidden goal it wants to satisfy, despite never being actually trained to.

Spooky stuff! Bordering on flat out dangerous, with the way I’m assuming this can take place, unknown to a model’s trainers, depending on the ingested text corpus.

Edit: clarity

That guy’s still at it?! Fairly sure I listened to him (and Leo… Laporte ?) podcasting 15-20 years ago. Crazy

Feel like DMing a bit? I’m in a similar boat, but not as far along. I’ve got what I believe is a viable idea with a niche but common market, I’m largely familiar with the tech involved (having built something similar before), and even the costs. Currently working on a POC before trying to conduct actual market research interviews (I realize most folks consider that backwards and are probably right).

Hell yeah! And another dope thing about the whole shebang, turns out the derivative < - > integral operation is wildly useful for describing…everything.

The simplest example, that I love the most, is just the very pedestrian (pun intended) relationship between a car’s position, velocity, and acceleration. It’s just enough “levels” (of diff < - > int) to have some instructional “meat”, and it’s a totally ubiquitous experience.

And then, when peered at more closely, that kind of relationship starts to crop up everywhere, suggests so much more!

Calculus is best maf

So, the heart of the issue is that each object’s path changes continuously, and the forces involved change in kind. Even worse, the objects interact with each other, again continuously - it’s not one-sided.

If you imagine trying to do it pre-Calculus, some kind of “just map it all out into a grid, etc.”, you can see the problems this continuous change imposes (exercise left for the reader).

By involving the Stravinsky Interpretation, it quickly becomes clear that the dimorphic superposition destabilizes. The clever reader might object “but what if you fold in all the noodly surfaces to recohere the manifold?”

And that clever reader would be right! But we didn’t know that until old Dr. Isaac “Zeke” Newton came along and made it that way.

Some say the devil himself taught him how it’s done, because no one else can read his notes! So keep your eye on old Zeke when you run into him.

Playing Binding of Isaac finally. I was a little disappointed with the difficulty honestly! One of my most heavily played games ever is another Roguelike - DCSS. To be clear that one is turn-based, which is very different, but it took me years and dozens at least, probably low hundreds of runs to beat it.

I’m sure I’ll be quickly humbled by the hard-mode difficulty and challenges and such on this one though. Super rewarding, fun, and extremely replayable! Great game. The different characters can play really differently, and through RNG you can def end up with some hilariously weak, or strong, builds.

Lol who is downvoting, it’s not THAT hot of a take!