lowki jammin

Man, I'm probably not going to win this; the gatekeeping tactic is simple and effective exactly because the mundanes in the audience don't know and can't trust that there *isn't* some deep and macro-relevant arcane science that *you* don't know about, if you're not wearing the appropriate medals to be a credentialed expert. They can say "Oh well I bet Eliezer has never implemented a network or written a single line of Python" and I can reply "Actually I went out and implemented and trained a simple transformer net from scratch just in case there was something surprising to be learned that way, which there wasn't" and then they instantly move the goalposts to "but you haven't trained one of the billion-dollar systems" and again for all the audience knows there could actually be some deep technical thing they know that I don't. It just bugs me that they're getting away with pulling the gatekeeper card on what's just... really, really not very complicated computer science. It really is not hard to understand how a kqv attention layer works, if you're otherwise good at understanding simple math, and the one piece that is clever and requires knowing one very basic trigonometric fact to understand (the reason behind the sin and cos waves in the positional encoding) doesn't reflect up to debates about AI alignment in any way I can see. The big takeaway for macro issues is the O(N^2) cost of the context window, if OpenAI hasn't already broken it; and that there's still only 200 serial steps of computation being carried out in a 200-layer network, regardless of the length of context window, again assuming that still even holds for GPT-4; and believe it or not, all this is just NOT VERY COMPLICATED COMPUTER SCIENCE objectively speaking, in the part that's relevant up at the macro layer. There's a story, which I sure hope is apocryphal, that when Euler met the atheistic philosopher Diderot at the court of Catherine the Great, Catherine asked him to debate Diderot on atheism, and Euler opened with "(a+b^n)/n = x, therefore God exists", and Diderot, not knowing algebra and unable to refute this, gave up and returned to Paris. The story is about a hugely asshole move, which I hope Euler never tried in real life; and anyone who actually pulls something like that is burning all the trust that actual technical thinkers sometimes legitimately have to request from a lay audience who actually can't follow a technical argument. What I'm trying to say here is: I know the incredibly simple math that underlies deep learning, I implemented a simple transformer network just in case there was something to learn that way, and while I do not know the arcane alchemical details with learning rate schedules and layer normalization and the eLUs that you use instead of ReLUs, as are required to make a big damn training run actually work, I am incredibly skeptical that the knowledge there has huge macro implications for AI alignment which just can't be explained to somebody as ignorant as I am and which are impossible to even gesture at. It's like they're trying to pull "(a+b^n)/n = x, therefore God exists" only they won't actually write down the fake equation. But, you know, maybe the reason why they keep it up is that pulling "therefore God exists" and refusing to name the equation is just fundamentally effective as a gaslighting tactic if the audience is willing to believe they know deep mathematical and engineering secrets that can't possibly be explained to me or the audience. But, really - I think any scientist who actually knows a technical thing, and who is not burning the trust of people who actually need it - should be able to gesture to the audience as to what secret the supposed layman doesn't know. Provide a reference, provide a citation, maybe even the grandma-explanation of it. It will at least filter out the half-Eulers who want to say "therefore God exists" and not even write down the equation where somebody could check it.