petters

22 ms·

301.
by petters 5y ago
If 16-bit floating-point numbers are used, it can presumably encode all tokens. In theory. It would not be very easy to work with.
302.
by petters 5y ago
Training data has 0.339T tokens, less than the number of training parameters. A model like that could store all of the training text with 100B+ parameters left for computation.
303.
by petters 5y ago
Judea Pearl writes about these news: https://twitter.com/yudapearl/status/1447596402670903304?t=X... Edit: downvotes suggest some context may be needed. Pearl win the Turing award for casual inference but has been
304.
by petters 5y ago
Solved problem or not, the important thing to realize is that it is a problem with zero marginal cost. That is, we don't make the storage problem harder by continuing.
305.
by petters 5y ago
Deepmind can pick whatever problem they want and use enormous effort to solve it. Other teams at Google must solve problem that Google has... which can be harder.
306.
by petters 5y ago
No need for electricity. Just make a grill of angled metal bars, alternatingly going into hot and cold water below.
307.
by petters 5y ago
If a ≤ b or b ≤ a. That does not always have to be the case. For example, x ≤ y could mean that x is a descendant of y in a tree.
308.
by petters 5y ago
For partial orderings, we can have neither a ≤ b nor b ≤ a. Perhaps it's that?
309.
by petters 5y ago
select_related and prefetch_related are absolutely essential when coding Django.
310.
by petters 5y ago
No, erasing all that debt world not affect the amount of money in circulation. It would only affect the balance sheets of the two institutions. (The money is already in circulation)
311.
by petters 5y ago
In Sweden, much of the national "debt" is between the government and the central bank. It could all be erased by decree and no one would be affected. The national debt number truly does not matter. > You can't pay your gro
312.
by petters 5y ago
> I've suffered with 2.5mg before for long enough to understand that it doesn't work for me, no matter how many studies you have with 95% CL that the dose is much lower. The 5mg tablets work much better for me. Well, have you t
313.
by petters 5y ago
Wasn't he a bit lucky here? Those two cities are almost directly above each other in the north-south direction. If they had been differently located, it sounds like the measurement would have way off?
314.
by petters 5y ago
> I guess what rubs me slightly the wrong way about this list is it seems focused on "who said what, and when" and not necessarily what's true. Right, I'm just glad we don't have that in math, which is closely re
315.
by petters 5y ago
It can find the global optimum for a function of a few variables.
316.
by petters 5y ago
Hopefully the tax authorities takes half of it, but $800k clean money is pretty sweet for a couple of years in jail of you are a criminal. Hope they owe the amount to someone.
317.
by petters 5y ago
There is a C backend for LLVM. If/when that works, that should solve most of the portability issues.
318.
by petters 5y ago
It does not show up in 32-bit mode. It does not show up on Macs. Therefore, it should probably be fixed.
319.
by petters 5y ago
We're asking because that is often a legal requirement for production use.
320.
by petters 5y ago
> And that is, unsurprisingly (and quite banally), used by many detractors to dismiss it all There is a stronger argument against crypto: the main use case (outside speculation) seems to be illegal activities. If what you are doing is le
321.
by petters 5y ago
I love SQLite as much as anyone here on HN, but the type system is not its strongest point. I can declare a column of type POINT and that will work fine. The result will be a column of integer affinity because it contains the string INT. ?!
322.
by petters 5y ago
Follow-up: Earth is at a tipping point right now. So far life has been governed by evolution. But within a short time (≤10,000 years, say) it will likely be possible to simulate a human brain quite cheaply. Then things will get weird quite
323.
by petters 5y ago
> there are certain things we can still say about them with reasonable certainty. Topping the list: They evolved. This is far from certain. They can be designed by someone who was designed by someone who evolved, for example. The evolved
324.
by petters 5y ago
The automatic rendering of "3x9" into "3×9" is not a good default IMO. The page says it is usually enabled by default. The other contextual changes look good.
325.
by petters 5y ago
Not sure about the definition, but a "cubicle" without high walls is not really a cubicle, right? I have worked at Google and a cubicle would have been better than the open source we had.
326.
by petters 5y ago
I'm guessing something else needs to be imported first.
327.
by petters 5y ago
Agreed, couldn't "I hereby relinquish my ability to bring other-origin content into this process without their opt-in" be a JS API call (one-way)? HTTP headers should not affect app functionality in this way.
328.
by petters 5y ago
> Google is a for profit company. On the side the offer products and services. Their profits come from selling your data to others. Google does not sell the data to others. (In fact, much of Google's profits would remain intact even
329.
by petters 5y ago
Can I make full stops between sentences? This is a serious question.
330.
by petters 5y ago
I have used the internet for 25 years and I learned that ending sentences with a period is/can be rude like two years ago. That surprised me a lot. But I never used chat a lot. I think on Usenet etc., full sentences were the norm.

More ›