@sasuke___420@dileeplearning Dileep’s talking about the LZ algorithm, not gzip. Idealized LZ provably converges to the entropy rate with no constant gap; which is exactly what makes it asymptotically optimal and cat not (cat’s rate never approaches H)
@l0tkas@captgouda24 That’s literally exactly what it does (disclaimer that I’ve taken linear algebra and econometrics courses before AI could handle them)
SAE interventions can be unreliable. 🧠🔒
We show that even when features are clamped, bad behaviors can still return through alternative residual-space directions. 🧩↩️
Feature control ≠ behavior control. 🚨
Paper: https://t.co/6soSFXKI5N
Page+Code: https://t.co/tvg1lEky07
@the_art_of_Li@Hesamation In a more general sense, minimizing KL-divergence is the act of approaching the “true” distribution of the document/collection of documents, which is directly equivalent to better compression of those documents. LLM loss minimization is essentially better compression.
@pl_zeng I’d imagine this would be the case for any identity element of an operation. Do you guys intend on testing multiplication with a holdout on 1?
“Believe all women” applies to individuals making claims. Doubting the 250k number isn’t doubting women who’ve claimed to be SAed, it’s doubting the extrapolation (which is clearly smudged to anyone with half a brain).
"There's no way 250k girls were raped by Pakistanis over decades" says the side that believes 1/4 women have been raped.
It's "believe all women" until the women say they perputrator was brown.
Exhibit a: (wait until the end)
@RighttoTryGuy@AndyMasley They want a global slowdown. It doesn’t work if countries like China can continue researching while we’ve stifled development domestically.
I immediately called that the next up from Anthropic will be ID Verification to use a closed AI model
Your freedoms will be taken one thing at a time if Local and Opensource AI doesn't win