js8

42 ms·

751.
by js8 2y ago
The argument that this is 90% of what matters in ML seems a bit bold. AFAICT it is completely missing reinforcement learning, which has been source of some mindblowing results (from Deepmind) in the past decade. It is also missing other stu
752.
by js8 2y ago
But was it Amiga's chipset the noncompetitive part? My recollection from that era was that graphics/sound capabilities of the era were far ahead of the PC, but PC had better CPUs, which was a problem of Intel vs Motorola. Anyway,
753.
by js8 2y ago
For analytics, I am a fan of type minimalism, like in SAS. You only really need to distinguish numbers and strings (each with missing value), and table as an overarching structure. They do everything else with input and output "formats
754.
by js8 2y ago
Compared to ISPF, some people also loved XEDIT or Roscoe for their ability to integrate adhoc scripting with TUI.
755.
by js8 3y ago
I love solving other people's problems. If I succeed, they are usually happy. If I fail, it is, after all, someone else's problem. It's a proposition you cannot lose.
756.
by js8 3y ago
That's the banality of evil.
757.
by js8 3y ago
Last year, China built roughly 4 times more renewables than coal. They argue that they build coal because renewables are unreliable, but surprisingly it was hydro what was unreliable (thanks to droughts brought by climate change), not solar
758.
by js8 3y ago
There definitely is a trade-off between memory size and how quickly it can be accessed. IIRC IBM z/Arch processors (AFAIK they are internally similar to POWER) have clock limited to around 5 GHz or so, so that L1 cache lookup costs onl
759.
by js8 3y ago
Again, you miss the point, there will be a paradigm shift. Yes, symbolic maths still exists, but nobody is looking up a special function to solve an equation. They just plug the numbers (and the state machine approximation of the numerical
760.
by js8 3y ago
I think you miss the point. I am not disputing FOL being undecidable. What I am saying is that, once we know how to solve SAT well enough, most mathematics will become obsolete (especially for applications). It will be obsoleted the similar
761.
by js8 3y ago
I know you're responding to someone else, but funny you say that. I suspect modern NN's are just iteratively solving some really large instances of SAT (similar to WalkSAT), to find a representation of a circuit that best matches
762.
by js8 3y ago
> Do you not realize that SAT might just be O(2^n) at best? Yes, but that's the worst case on all possible instances. Even with this worst case performance, there might be an algorithm that works on par with other algorithms that sp
763.
by js8 3y ago
> Even is SAT is P that doesn't imply a solution can feasibly be found. It does, because in this case you also assumed that running numerical algorithm was feasible. So if your SAT solver fails to find the answer, it is either subop
764.
by js8 3y ago
I think what's more fundamental than just answering P vs NP, is what exactly makes boolean functions hard to (pseudo)inverse? And it's the latter where answering (unless the answer comes from an oracle) the former will help. And t
765.
by js8 3y ago
> Do you not realize that P vs. NP might have a negative outcome (extremely likely) The outcome will matter less than having an answer. And I do think it will be a negative outcome either way, because it will obsolete lot of beautiful ma
766.
by js8 3y ago
Before there was AlphaGo, many top Go players believed that understanding Go is somehow special in helping us to understand the universe. The AlphaGo was a shock, because it turned Go into "just another game", which can be beaten
767.
by js8 3y ago
I disagree that converting from an equation to a solution algorithm is a straightforward process. It is very much where mathematics helps, for example just to ensure convergence of the method. Also, the way you model reality (choose the equ
768.
by js8 3y ago
My point is, the need for semantics is a human construct, it's not required to solve practical problems. Practical value of MVT is that we can apply it in the process of describing a solution, i.e. an algorithm. But if you have a SAT a
769.
by js8 3y ago
IMHO it's even worse. Once we figure out P vs NP, we will understand what it means to invert boolean functions algorithmically, and all mathematics will be replaced with an automated process. Today, most of the math is figuring out how
770.
by js8 3y ago
I think the best analogy for money is blood. From the perspective of a household (a cell), or a company (an organ), money (blood) are a limited resource, which constraints what you can do. If you don't have that resource, you will die.
771.
by js8 3y ago
With whom? Anyway, usable local AI will probably lead to fully automated luxury communism, so I am not very worried.
772.
by js8 3y ago
Interesting topic. I fondly remember, when I was doing geometry at school, I propagated a lot of errors.
773.
by js8 3y ago
> a sudden change in perf characteristics when one additional usage of some feature tips a heuristic over This already happens with humans, changing features will change how the product is used and thus performance characteristics change
774.
by js8 3y ago
> we are really good at code sharing of small things like vectors, and of large things like operating systems. Its the middle-sized problems we are bad at. Interesting. But I am not sure we are good at sharing small things - every progra
775.
by js8 3y ago
Another data type that would be quite useful is a table (like in a database). For the same reasons, too many design choices. Anyway, that being said, I have felt that progress will be made in programming languages if the compiler gets to ch
776.
by js8 3y ago
I think the real problem is we don't know what these LLMs SHOULD do. We've managed to emulate humans producing text using statistical methods, by training a huge corpus of data. But we have no way to tell if the output actually ma
777.
by js8 3y ago
> should we also be considering if math textbooks are soul prisons? Why not? There are even philosophical theories that the reality itself is just a manifestation of mathematical laws.
778.
by js8 3y ago
Well, in reinforcement learning, wrong outputs are inherently painful for the AI. If the system expects getting negative feedback regardless what it does, I would say it is suffering. I would also define torture as giving it negative feedba
779.
by js8 3y ago
My question to Varoufakis is this (and I used to be skeptical of his distinction because of this, but I am less so today). If these systems of power are maintaining themselves through cultural hegemony, as Gramsci has described, what is the
780.
by js8 3y ago
> One of the nice things about writing new tests by hand is you've got someone who can judge if it's the system being stupid or if it's the test. This is an instance of a more general problem, which I call "the probl

More ›