cause you can attend to them later but for now they don't change what you're doing. if I'm understanding right. also iiuc what you're describing is just a single dimension which doesn't give you a lot to work with, so maybe most of this space-scrunching is just done boringly with nonlinearities
it seems like similarly behavior-impotent directions in the residual stream would be useful then for storing planning information multiple tokens ahead, like in the poetry example here: transformer-circuits.pub/2025/attribu...
it's a simpler kind of null space but this is reminding me of neuro work on preparatory activity in motor cortex lying in the null space of muscle readouts (e.g. pmc.ncbi.nlm.nih.gov/articles/PMC...)
goofy regulations leading to a country full of geniuses in a datacenter running on corn ethanol...
well, maybe, you know. but some people just wanna hang out
I discovered the xenofeminist manifesto through a tag on the gender acceleration blackpaper (vastabrupt.com/2018/10/31/g...) back in college. that was a fun one. the main mark it left on me though is just the "unix was named that because it's a castrated multix" which is apparently maybe apocryphal
I behold the mouse. I see it the way a hungry kitten would. I see it the way a laboratory technician would. Finally, I see it through mouse-eyes, wreathed in a halo
yeah, and he's claiming that making them flatter, by shrinking the highest logit, on 0.4% of tokens for MMLU, and otherwise doing nothing, gives 72.5% accuracy at that temp. really I think even if the steering was magical that low a fraction lets you dismiss this out of hand
I think it's just nonsense. it implies throughout that activations are actually being steered towards the truthfulness manifold, or I think that's how it's easily read if you're skimming. but the only actual steering it describes is this which cannot possibly do what the paper claims
though to be fair to claude this reads more like GPT to me
and that's it! no way that does anything! come on!
like there's all this gobbledygook about the mahalonobis distance between the "truthfulness manifold" and that's like weighted against (a function of) the shannon entropy of the pre-temperature logits but all of that just goes into a computing a scalar. if that scalar is low enough you do this:
to be a dick, given the obvious LLM prose and my own experience with using Claude as a research assistant (he is very eager and excited about positive results and doesn't look too hard, like others with no research experience), I think this guy probably just has a bug somewhere
as he's described it that steering would be effectively increasing the temperature anyway, making the top logit less likely, no?
calling bullshit on this. I'm not 100% sober but after staring at this for some time, afaict the only "steering" that happens is a slight reduction of the top logit. this happens on 0.4% of tokens on MMLU, the rest of which are untouched. I really don't think that's gonna give you 72.49% at temp 3.0
more relevant correspondence to draw is ~5ish output tokens per second for a human. clock cycles and minimum ISIs aren't really relevant computational units. token thing has issues too (input tokens more complicated, also have high-bandwidth physical output, internal reasoning), but it's ballpark ok
frank had some bangers
there's a limited amount of information processing you can do in a day, and it's not worth it to most people to process that information (or even possible to do it on the scale of 20+ options). if that information processing becomes dirt cheap, the picture really could change. but we'll see...
yeah it's very speculative, and I think an unreasonably short timeframe. but I do think a lot of the search friction is just tedium, like, I don't give enough of a shit to even compare two food delivery apps against each other. but LLMs don't get bored and that could really change the picture.
I don't know if I'd bet on it but it doesn't sound totally crazy, at least
"Agents accelerated both sides of the destruction. They enabled the competitors and they used them... An agent doesn’t have a home screen. It checks DoorDash, Uber Eats, the restaurant’s own site, and twenty new vibe-coded alternatives so it can pick the lowest fee and fastest delivery every time."
I think a key part of the citrini story was wide use of personal shopping agents by consumers and of app-management agents by drivers, which they could have emphasized more. I think it's plausible that those 1) can soon exist and 2) can largely overcome network effects and other non-coding barriers
Do watchers keep overshadowing your illusions? You may be able to disclaim vantage! Call 1-800-MARINADE today
when it's applicable "always visualize everything" has saved me a lot of regret (and caused me a lot when I've ignored it)
"Hello babies. Welcome to Earth. It's hot in the summer and cold in the winter. It's round and wet and crowded. On the outside, babies, you've got a hundred years here. There's only one rule that I know of, babies-'God damn it, you've got to be kind.'"
that's this: arxiv.org/abs/2503.20020 really would like a followup on it but there hasn't been one afaik
I feel a bit like this too but I'm not sure if it wouldn't be trivially trainable for when they get around to it. I remember the gemini robotics paper from last march showing that 2.0 Flash additionally trained on some more (vaguely described) spatial reasoning tasks outperformed 2.0 pro handily...
so we beat on, boats against the current, borne forward ceaselessly into the future
i have to be the most fuckable person being converted to computronium by the swarm of diamondoid bacteria
