Veedrac

32 ms·

361.
by Veedrac 5y ago
If you have a probability function over externalities rather than a single value, then just take the expectation. If you have latency in the system, account for that in the probabilities. Refusing to set a price, or using arbitrary techniqu
362.
by Veedrac 5y ago
> The authors are not kind to economists, who typically want to put a price on emissions and then let markets do the work. Economists have, the authors allege, skipped a chapter in the textbooks. They have focused on externalities, the d
363.
by Veedrac 5y ago
Nothing prevents it, no. Transformers are certainly capable of learning mathematical tasks; consider [1] as an example, which uses big but regular token lengths. Alternatively you could just scale 'till the problem solves itself. [1]
364.
by Veedrac 5y ago
> GPT-3 seems to have issues with large numbers. Moyix’s gist covers this in detail. GPT-3 tends to guesstimate an algebraic function instead of evaluating the numbers, so the answer is only correct to a certain approximation. There are
365.
by Veedrac 5y ago
The CPU is custom but the cores are off-the-shelf Arm cores.
366.
by Veedrac 5y ago
The first step they went wrong is using a slideshow in the first place. This was a time critical scenario, anyone who needed to be making decisions should have been part of the engineering discussions from the beginning. They should have be
367.
by Veedrac 5y ago
I don't think these numbers are meaningful. It's not far off from a degree-2 interpolation already, and I got a hidden layer size of 20 to an error of 7e-5 by just letting it optimise for longer and picking a seed that worked well
368.
by Veedrac 5y ago
> >> RMSE for NN approx: 0.00011776185865537907 Error is approximately 0.0001 because `tol`, the parameter that tells optimization to finish, is 0.0001. Set tol=0, and then beta_2=1-1e-15, epsilon=1e-30 to maximize stability f
369.
by Veedrac 5y ago
YouTube has different needs for different videos. The average barely-viewed video can use quick and dirty hardware compression, but a large fraction of views are on a small fraction of videos, which is the sort of thing that more intense
370.
by Veedrac 5y ago
I am objecting to the politics, not the practice.
371.
by Veedrac 5y ago
Abstinence-only climate education is such a depressing notion. There is plenty of clean energy. It is not running out. The idea that we should impoverish ourselves to protect against a problem we can solve without impoverishing ourselves is
372.
by Veedrac 5y ago
It'll hardly be the fuzziest line that the courts have to draw. They'll come up with a set of general principles to help decide which side of the line cases lie on, and then mediate the ambiguous cases. Some examples might be, is
373.
by Veedrac 5y ago
Your example suggests a fairly obvious line to draw; that a person should not be compelled to speech that they do not agree with, but that they may still be compelled to action that does not reasonably count as speech. As a less emotional e
374.
by Veedrac 5y ago
The null hypothesis is not your prior! https://en.wikipedia.org/wiki/Null_hypothesis
375.
by Veedrac 5y ago
If one considers oneself entirely unsure about what properties consciousness has, maybe one shouldn't be trying to refute twitter posts that hypothesis certain things might be conscious. Maybe one should do the more principled Bayesian
376.
by Veedrac 5y ago
So this article does not actually defend its claim, it just gets mad that some AI researcher expressed their opinion, makes an unsourced (albeit probably correct) claim that the criticized opinion is a minority position, and wishes really s
377.
by Veedrac 5y ago
Compared to oldspace they definitely do; Falcon 1 only flew a handful of times, and Falcon 9 has been continually upgraded. Musk has mentioned even the old F9 Block 5s (which are the newest block) are painful to refurbish versus the newer o
378.
by Veedrac 5y ago
Not only this, you have to have the ability to throttle while also being able to land if an engine fails, because Starship is going to carry people.
379.
by Veedrac 5y ago
I'm pretty certain that they'd rather try and fail to land than not try at all.
380.
by Veedrac 5y ago
2025 doesn't make sense. SpaceX has a Raptor 2 run-rate of 250/yr today, and they're ramping up, so you're talking about them sitting on around 750 engines, or about 18 full Starship stacks on the low end. The only way t
381.
by Veedrac 5y ago
As if human jokes aren't cherry-picked? AI's ability to generate endless off-beat variety is particularly valuable for generating humor, just not (yet) with zero effort. It's a subject of taste but a cherry-picked GPT-3 twitt
382.
by Veedrac 5y ago
I tried something like that for 3-digit multiplication with GPT-3 in another comment[1], successfully. You have to lay things out different manner than you did here, because GPT-*s have no sense of layout on a page; their byte-pair encoding
383.
by Veedrac 5y ago
I fiddled with this prompt Calculate 721*436 Step A: Multiplication table. 721*0=0 721*1=721 721*2=1,442 721*3=2,163 721*4=2,884 721*5=3,605 721*6=4,326 721*7=5,047 721*8=5,768 721*9=6,489 Step B: Multiply digits
384.
by Veedrac 5y ago
Apologies, I thought you meant ‘optimized’ in a different sense, not in terms of how list.pop is implemented, as AlphaCode wasn't involved in that. You are entirely correct that list.pop uses memmove. > Someone submitted this 1553D
385.
by Veedrac 5y ago
> AlphaCode's solution the "inner O(n) loop" is actually a memmove(), which is optimized to be insanely fast. Again, it is not. CPython does not do these things. The web page says, and this is corroborated in the paper, &g
386.
by Veedrac 5y ago
I was paraphrasing your argument. AFAICT, it remains faithful to what you are saying. > Accordingly, there is no doubt that neural nets can only represent what is in their training set. This is not true. If it were true that there was no
387.
by Veedrac 5y ago
Your comment is an excuse, not mine: “ignore how good the advertised model is, because a different, much smaller version without all the techniques merely does pretty well on an extremely tough problem set.” > Any claims to the contrary
388.
by Veedrac 5y ago
> Btw, APPS is not much of a benchmark. It evaluates code generation according to how close it resembles code written by humans. No, the metric used in this paper was the percentage of questions it could solve against the hidden tests. &
389.
by Veedrac 5y ago
The APPS benchmark was for a “small 1B parameter model”, fine-tuned “without using clustering, tags, ratings, value conditioning, or prediction”. > Overall, that's a brute-force, almost random approach that is ignoring entire decade
390.
by Veedrac 5y ago
It does not. Really the strings just never get long enough that O(n²) would be catastrophic; the maximum possible length is 2e5.

More ›