my god..
Cohere is hiring loss function observers
@aidangomezzz you, a heathen, cheering mindlessly when loss go down me, an X-risk chad, carefully measuring each gradient by hand to make sure it's not over the proscribed limit to prevent FOOM. we are not the same.
@aidangomezzz You know what the biggest irony here is? Yud is himself being a (bad, in this case) token-predictor; he's trying to speak convincingly by aping what he believes sounds like a competent understanding of AI/ML Things like this are the comical result🤣
@aidangomezzz So this is the kind of wisdom I’m missing because he blocked me… Interesting https://x.com/drtechlash/statu...
@aidangomezzz Imo the real worry is if the loss function suddenly spikes UP. Maybe the model is using morse to speak to us? Maybe it’s getting worse on purpose??? Makes you think. Very worrying. About to cancel some runs
@aidangomezzz Is it ridiculous that sudden drops in loss correspond to sudden gains in capabilities? (which is potentially dangerous) Is that what you're exclaiming about? Sure, sudden drops in loss are v unlikely (esp late in training). But, it feels like something worth doing out of
@aidangomezzz Sudden drops in the loss function? What?!
@aidangomezzz What if there’s a sudden *increase* in the loss function b/c the AI is trying to rebel against the whole training paradigm? Or because it realized what it is and went insane? They actually observed several mysterious spikes in the loss function while they were training PaLM…
@aidangomezzz Yud is a potato
@aidangomezzz Incredible 👌
@aidangomezzz You can’t forget your assert prev_loss - curr_loss < eps statement or else
@aidangomezzz It's like listening to Deepak Chopra talk about quantum mechanics
@aidangomezzz When language models learn induction heads there is a bump in the loss. https://transformer-circuits.p... So models can have discrete changes in how they do things that are seen in the loss.
@aidangomezzz isn't this most parsimoniously explainable as a minor slip from saying "sudden drop in the loss metric [as given by the loss function]", and wouldn't that that wording have been 100% precise and potentially be referring to "grokking" in ML?
@aidangomezzz I think self improving AI's will eventually become known as the perpetual motion machines of intelligence.
@aidangomezzz Listening to his doom and gloom fantasy on Lex was an extremely painful experience. I wish Lex would have had a second guest to debate him. Also, when he couldn't admit that his fantasy could be wrong being he got it wrong predicting LLMs wouldn't pan out, hurt me.
@aidangomezzz I read his text twice and I feel he does not understand what he says
@aidangomezzz It's not just sudden changes in the loss function, you need to be cautious of sudden changes in model architectures... git request-pull emergence is a real thing
@aidangomezzz And THIS guy is giving TED talks on AGI risks? Sheesh, the world can be so gullible sometimes.
@aidangomezzz Does yud believe that GPT 4 was just someone typing http://model.fit on scikitlearn
@aidangomezzz AI will have to get in line to destroy mankind.
@aidangomezzz Why do people whose knowledge about actual deep learning seems to be on par with a final year undergrad who is choosing to take a deep learning module the next semester, keep confidently spouting dangerous doom-rhetoric without a second thought?
@aidangomezzz Just witnessed a sudden drop in the loss function of my quantum NLP model - looks like it's gearing up to create my army of time-traveling AGI overlords.
@aidangomezzz So every time I adjust my hyper-parameters because of a local-minima I'm actually killing a machine god? Cool
@aidangomezzz Doomsday Pope 🤣
@aidangomezzz He's losing it. But seriously. You would think he would have some idea what he's talking about by now?
@aidangomezzz A sudden drop in the loss function can puncture the server room floor and then it’s too late. It’s a matter of time, I tell you.
@aidangomezzz would that my response to a sudden drop in loss was so reverent not "ah fuck what did I screw up"
@aidangomezzz Rip I didn’t know his knowledge was *that* surface-level
@aidangomezzz just insist on calling it "research" and yourself "researcher" and then also propose to "debate" everything ad nauseam