Majromax

16 ms·

211.
by Majromax 2y ago
> What good it's it to train a model basically on itself? If the model generates data of variable quality, and if there's a good way to distinguish good data from bad data, then training on self-generated data might "boots
212.
by Majromax 2y ago
> My take: it's a mix of brand bundling and lack of data. They're roughly equivalent but shorts is bundled with youtube which has its own brand perception and reels are bundled with IG/FB and have their own brand perceptio
213.
by Majromax 2y ago
> Calling back the possibly spoofed number Don't call back the number possibly being spoofed (i.e. using your Caller ID as the source of the callback number). Call an independently-listed number for the company, such as the phone n
214.
by Majromax 2y ago
> Let's say you are given the declaration but not the implementation of a function with the following prototype: > const char * AskTheLLM(const char prompt); > Putting this function in charge of anything, unless a restricted
215.
by Majromax 2y ago
I hate to just drop a book recommendation, but Jason Schreier's Play Nice covers the history of Blizzard in comprehensive detail. The overall situation was multifacted, but my takeaway is that Blizzard's recent failures come dow
216.
by Majromax 2y ago
Not necessarily. A "good predictor" could still result in non-Gaussian performance for at least two reasons: 1. The prediction could be on a relative rather than quantitative basis. If IQ(A) > IQ(B) always implies Perf(A) &gt
217.
by Majromax 2y ago
> And aluminum has more benefits than these AIs. It's not like these data centers or the power to run them are free, or being paid for out of public generosity. If AI proves to be over-hyped, then the private-sector backers of these
218.
by Majromax 2y ago
More than the residential use of some entire cities. This is a misleading comparison. "AI data centers" aren't a residential use, they're much closer to an industrial one, and they should be considered on that basis.
219.
by Majromax 2y ago
What you propose is a harder AI safety scenario. You don't need a 'vastly more competent AI overseeing its own training' to elicit this potential problem, just a malicious AI researcher, looking for (e.g.) a model that's
220.
by Majromax 2y ago
But the LLM isn't "using next-token prediction" to solve the problem, that's only how it's evaluated. The "real processing" happens through the various transformer layers (and token-wise nonlinear networks
221.
by Majromax 2y ago
> One of the best understandings of inflation is to use the mathematical concept of "equality" on those. They're two ways of phrasing the same thing. They're not quite the same thing. All other things equal, if a pri
222.
by Majromax 2y ago
> In this case large bureaucracies are everywhere in the public sector as well, where, at least to first order, price mechanisms, profit maximization etc. are not the driving force. I'd say that large bureaucracies are endemic to t
223.
by Majromax 2y ago
Not necessarily. To be inexpressible in software, all it has to do is be unpredictable; it can still be boring. Let me define "local snoozing time" (LST): it's set to my local standard timezone as of today, but every time I
224.
by Majromax 2y ago
> This always seemed wildly implausible to me. A very large number of communicable meanings, sure, but infinite? This is "trivial" in the boring kind of way. With just digits, we can communicate an infinite set of distinct num
225.
by Majromax 2y ago
> The bar for self driving isn’t «does it never kill anyone», it’s «does it kill people less than manual driving». Socially, that's not quite the standard. As a society, we're at ease with auto fatalities because there's
226.
by Majromax 2y ago
> how do you exempt those companies from civil liability and make it impossible for victims/their families to sue the manufacturer? I don't think anyone in this thread has talked about an exemption from civil liability (sue f
227.
by Majromax 2y ago
> 99% of the same old things ... tyres, brakes Aren't brakes one shared system where electric vehicles have an advantage? With the widespread use of regenerative braking, most of the energy of motion is bled off without friction on
228.
by Majromax 2y ago
> It's unclear to me what "convex hull" means though. The convex hull ( https://en.wikipedia.org/wiki/Convex_hull ) of a set is the smallest convex shape that includes that set. Geometrically, it'
229.
by Majromax 2y ago
The point of the experiment isn't the sound itself, however, it's carefully timing the sounds to the allegedly correct part of the alpha brainwave cycle. That's why the participants were wearing EEG headbands for the study.
230.
by Majromax 2y ago
> During the intervention week, participants wore an electroencephalogram (EEG)-enabled headband that delivered acoustic pulses timed to arrive anti-phase with alpha for 30 min (Stimulation). During the Sham week, the headband silently r
231.
by Majromax 2y ago
> Math.sin(x)/x (the sinc function) for 7 terms over [-3,3] gives coefficients c0...c6 that are all NaNs. Is this a bug? That wouldn't exactly be a bug. The code is undoubtedly calculating the Chebyshev coefficients by evaluat
232.
by Majromax 2y ago
Your point and the grandparent's point are not contradictory. 'car' and 'cdr' can be opaque pieces of jargon that make the language slightly but noticeably more difficult for new learners, and they also can be sta
233.
by Majromax 2y ago
> The cc-by-nc-4.0 license applies to the network weights. I'm not even sure if network weights are copyrightable independently of the code and data used to generate them. In my personal (not a lawyer) view, the weights of a neural
234.
by Majromax 2y ago
For an extensive discussion of similar issues, see Paul Krugman 1978, "The Theory of Interstellar Trade:" https://www.princeton.edu/~pkrugman/interstellar.pdf .
235.
by Majromax 2y ago
Good point, I retract that comment about detectability. I was too concerned with magnitudes to consider the rest of it.
236.
by Majromax 2y ago
> Could one derive the Earth's rotation? Yes, existing gyroscopes are sensitive enough for that. In 1966, one of the first Soviet attempts to launch the Soyuz failed on the pad because the guidance system remained active after a sc
237.
by Majromax 2y ago
It looks like the compared-to values are hard-coded into the generated function, so you would be calling compile once per unique query per process.
238.
by Majromax 2y ago
> Is the interview for an engineering position or for sales? Either way. The coin-flip example and Ballmer's binary search game could apply with simple extensions to complicated processes like SLAs on cloud services. > The data
239.
by Majromax 2y ago
Note that the attack described in this article is a passive attack, so sufficiently long calls that were previously recorded might be vulnerable. I doubt this will matter for most people, but mass interception for later decryption was one
240.
by Majromax 2y ago
> That degree of moderation, for that purpose, would make you liable if you left illegal or tortious content up i.e. if you moderate, you're a moderator, and your first duty is legal. I'm not sure that's quite it. As I re

More ›