scarmig

19 ms·

331.
by scarmig 2y ago
Hoping he speaks a lot of truth to power.
332.
by scarmig 2y ago
Hinton's most recent paper on forward-forward acknowledges Peter Dayan explicitly for his feedback on the paper, and cites a paper they cowrote together back in the 90s. Dayan being the author of the canonical textbook on theoretical
333.
by scarmig 2y ago
He does get a reference in the Nobel announcement, so he can console himself with that at least.
334.
by scarmig 2y ago
Schmidhuber enters the room and has lots of things to say.
335.
by scarmig 2y ago
This prize is more of a settler colonialist land grab by physicists. ML is just a subfield of physics (like every other field), so let's make sure that everyone knows that it's in our domain. Speaking as a onetime physicist now in
336.
by scarmig 2y ago
McCulloch and Rumelhart were psychologists as well.
337.
by scarmig 2y ago
Nonlinearity somewhere is fundamental, but it doesn't need to be between each layer. You can, for instance, project each input to a higher dimensional space with a nonlinearity, and the problem becomes linearly separable with high prob
338.
by scarmig 2y ago
Clippy, I've got a lot of TPS reports here. Can you get me as many paperclips as you can?
339.
by scarmig 2y ago
On the other side of things, imagine OAI purchased something critical to their success from a near-monopoly producer. If they suddenly are devouring massive amounts of the economy, it will rationally attempt to increase its margins. Not to
340.
by scarmig 2y ago
Because users flagged it. That pushes the question to "why did users flag it," but it's not hard to speculate reasons why. A dedicated minority roll their eyes at discussions of AI safety. On top of that, although the article
341.
by scarmig 2y ago
It remains highly interesting, and indicative of something, that researchers in other fields orient almost entirely on the "give us money to improve our technology," not "give us money to research how to fix our technology, a
342.
by scarmig 2y ago
They do have particular insight into the trajectory of capabilities; less insight into how the newfound capabilities will interact with the world; and no particular insight into values. The usual failure mode with experts is that they overs
343.
by scarmig 2y ago
UIs in general are pretty hard unless you have a set of primitives that do everything you want. HTML + CSS used to be lacking (and still are in some respects). A thin layer of vanilla js on top of them has a tendency to turn into a mess of
344.
by scarmig 2y ago
> I bet we could throw away half of publications tomorrow and see no effect on the actual pace of progress in science. It might actually improve the pace of science, if the half eliminated were not replicable and the remaining half were
345.
by scarmig 2y ago
You should be able to build an entire career out of replications: hired at the best universities, published in the top journals, social prestige and respect. To the point where every novel study is replicated and published at least once. Un
346.
by scarmig 2y ago
You announce a set of boundaries, imply that crossing them shifts the system to some new, disastrous attractor state, and get a lot of media and donor attention. For a couple years you get lots of mileage out of it, until you need to come u
347.
by scarmig 2y ago
Cheapness has a quality all its own. Gemini is substantially cheaper to run (in consumer prices, and likely internally as well) than OpenAI's models. You might wonder, what's the value in this, if the model isn't leading? But
348.
by scarmig 2y ago
I came across "Learning XOR: exploring the space of a classic problem" other day: https://www.maths.stir.ac.uk/~kjt/techreps/pdf/TR148.pdf Even something with three units and two inputs is nontrivia
349.
by scarmig 2y ago
For the ride hailing side of things, it's already happened; there is no question of plausibility. I've probably done 50 rides in 2024, and only two had human drivers.
350.
by scarmig 2y ago
The third way is that companies find hiring workers, with all their complaints and needs and desire for dignity, too inconvenient, so they replace them with things that have none of those. Particularly relevant since this paper is primarily
351.
by scarmig 2y ago
Infinity intelligence is a very vague and probably ill-defined concept; to go to an impossible extreme, if you're capable of modeling and predicting everything in the universe perfectly at zero cost, what would it even mean to be more
352.
by scarmig 2y ago
Anterograde amnesia is a significant disruption of plasticity, and yet people who have it are still intelligent. (That said, I agree plasticity is key to the most powerful systems. A human race with anterograde amnesia would have long ago g
353.
by scarmig 2y ago
Another is the Nakano-Amari-Hopfield model, which is based on Ising. Hopfield himself was a trained physicist.
354.
by scarmig 2y ago
Is there a proof that collections of human neurons are capable of carrying out proofs in first order logic, in full generality?
355.
by scarmig 2y ago
Different audiences. 3B1B can be very effective for intuition. It does not, however, prepare you for proofs or anything rigorous or thinking about things in an axiomatic way. For example, it isn't going to provide a basis for anything
356.
by scarmig 2y ago
I'm guessing this was posted because it's the relatively new 4th edition. Tensors now included!
357.
by scarmig 2y ago
Challenge: find other sequences that you can use to convert, using integer coefficients in the recurrence. For miles to kilometers, you have to get a better approximation than phi. I wasn't able to for miles to kilometers, but for poun
358.
by scarmig 2y ago
Everything is a wrapper. The questions with GPT-wrappers are 1) how much would it would cost you to create an equivalent or better wrapper and 2) how much would it cost OpenAI etc. to replicate the same wrapper, either explicitly or through
359.
by scarmig 2y ago
It doesn't need to be perfectly random; you just need every sequence to have a nonzero probability of being generated. "to be or not to be" may be more likely for a monkey on a keyboard to generate than a random string, actua
360.
by scarmig 2y ago
Every organization will have incompetent individuals. A healthy organization is able to remove them; a typical one will blunt their effects; and an unhealthy one will allow them to reproduce themselves and seize power. We want it all to be

More ›