Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Reality has a surprising amount of detail (2017)

johnsalvatier.org

Post reply on HN
Print view View original

91–100 of 120 posts

Re: Reality has a surprising amount of detail (2017)

#91

Post by Zenbit_UX »

amasad wrote: post #30

I really like the essay but I think there is a form of Stockholm syndrome playing out especially with the following statements: > using a Haskell package for the first time, and being frustrated by how many annoying snags there were > If you’re a programmer, you might think that the fiddliness of programming is a special feature of programming, but really it’s that everything is fiddly, but you only notice the fiddli…

Despite the tool you work on which solves (maybe) that one very specific problem the author posed, I'd say his point still stands about 99% of other things in this universe.

Re: Reality has a surprising amount of detail (2017)

#92

Post by semerda »

c3534l wrote: post #53
combatentropy wrote: post #8

This is why I think people should still learn their job through hands-on apprenticeship, as opposed to a lot of book-learning in the classroom. In fact I think people do still learn their jobs that way. It's just delayed by several years of book-learning that has little return on investment. When you finally land your job, your first few months are when you really start learning enough to become useful. Disclaimer: I…

You need both. It's unfortunate that we separate out the two. It wasn't until the second semester of my third year of my accounting degree that we actually had a project where we would get fake checks we had to enter into the books, fill out, and put in a folder called "mailed." This should have been day 1 and we should have learned what it was we were doing as we were doing it. And this should have prepared us for s…

Ditto. Learning theory without real-life application makes it boring and people forget. Learning application without theory leads to hacks. Often the real-life application happens in the last year when it's too late for some. Maybe the interim assignments should be more real-life examples so students could see application of and be inspired to learn deeper.

Re: Reality has a surprising amount of detail (2017)

#93

Post by cookie_monsta »

omneity wrote: post #18

Related essay: http://blog.omarkama.li/the-market-decides-your-project-scop...

Given the subject matter I had to smile at the way the author bungled the detail within the first sentence:

> "A while ago, I read this amazing piece about building a ladder."

Re: Reality has a surprising amount of detail (2017)

#94

Post by ganzuul »

ssivark wrote: post #79
ganzuul wrote: post #63

Earlier quoted context omitted.

Embodied intelligence seems to break this mold, in principle. I believe the principle to get good priors is called reinforcement learning, and that currently the problem you describe is called inductive bias. Or maybe I'm just pushing your goalposts to something I can talk about... There are some really brave ideas being kicked around among the researchers who focus on their passion of AI, as opposed to those who pos…

Embodied intelligence, Inductive bias, etc are the right words but we rarely know how to build them into our models in all but the simplest problems :-) Eg: Our reinforcement learning models are largely stupid and without any great inductive biases. That’s why it takes a crazy amount of training and reward tweaking to even do anything simple. > There are some really brave ideas being kicked around among the researche…

Tensor networks come from quantum physics but are finding use in ML research.

The papers on Neural Turing Machines, and Neural Tangent Kernels, which is related to Gaussian Processes, are getting a lot of citations. There is work on making these kernels composable.

There are some papers combining some of these approaches but not yet all of them, that I have seen.

If I have understood correctly:

TN in the context of ML seems to be a pre-optimization step which lets you deal with inductive bias.

The NTM learns to use its own memory through an attention mechanism.

The NTK builds compressed representations which should be of great interest to anyone needing interpretable models.

Gaussian Processes are a very different perspective on ML which lets you talk about composition of probability distributions.

It seems to me that these components could be assembled into an ML algo capable of general intelligence (not 'strong AI' IMO) given a curriculum and enough compute. It wouldn't compete with the smartest humans but at least it should make for useful robots, because of the cost of training.

Re: Reality has a surprising amount of detail (2017)

#95

Post by Izkata »

jasode wrote: post #4

A related concept: tacit knowledge : https://en.wikipedia.org/wiki/Tacit_knowledge

Stuff like this is why we always encouraged our new hires to add to documentation anything they had to ask us about, even for what should already exist as a good set of explicit steps. It was pretty much a guarantee something would be missing.

Re: Reality has a surprising amount of detail (2017)

#96

Post by Izkata »

earenndil wrote: post #57
post_below wrote: post #45

Earlier quoted context omitted.

I don't personally think the theory that we likely live in an artificial sim is true. But I also don't think it's 'obviously' untrue. Yes, the level of detail is mind boggling. But quite a lot of that detail follows naturally from the framework itself: The complexity of an object moving through space and time, interacting with heat and light and atmosphere and etc. is huge. However once the framework is in place (phy…

Absolutely, it could be that we are living in a simulation. It's just that there's no reason to believe that as yet. FWIW I think that a 'biological interpreter' is rather unlikely; it's probably just emergence from atomic interactions. (It could be planted there, however.) The problem with your argument is that it makes too many assumptions. We know nothing about the nature of the simulator, nor do we have any reaso…

I'd like to propose two pro-simulator things that have been bouncing around my head for a while:

* Plank length/time: The granularity the simulation is capable of.

* Quantum entanglement: A deep copy/shallow copy bug (that is, shallow copy was accidentally used where deep copy should have been used, hence why quantum entanglement holds across large distances).

Re: Reality has a surprising amount of detail (2017)

#97

Post by ncmncm »

hinkley wrote: post #52
stickfigure wrote: post #48

What a great read! But the staircase example seems overwrought. Reality often doesn't need to be that precise. If you're a few degrees off in your angle, it won't matter. The floor and/or walls are probably not perfectly flat and may not even be perpendicular. Stairs will work anyway. This angle problem has been solved in tooling. You set the ( If you need precision, work with metal.

Notching angles into boards is the reason framing squares exist. I've learned and forgotten this three times in my life, the last not so long ago. Stairs are for human legs, and they care about the rise, the run, and the depth of the stair (tread). And all those can be done with a framing square. Angles are incidental to this (google around and you'll see a lot of stringers being cut with a circular saw). A 3/5 ratio…

Maximum ratio. In the ("whole") rest of the world it is probably 15:30 to 17:26.

Crappy tooling, and crappy software in general, is our legacy from Microsoft. Before Microsoft, people had expectations that software would work right. It didn't always, but it was a big deal when it didn't. Microsoft, single-handedly, got people used to rolling their eyes and rebooting, and not demanding a refund, partly by never ever refunding anything. It worked for them, and here we are.

The avoidance of detail, enabling its occasional rediscovery today, is equally new. People took pride in attention to detail, and demanded it. Watch any detective movie from the 30s or 40s and you get a glimpse of how attention to detail was respected.

Re: Reality has a surprising amount of detail (2017)

#98

Post by xorfish »

combatentropy wrote: post #8

This is why I think people should still learn their job through hands-on apprenticeship, as opposed to a lot of book-learning in the classroom. In fact I think people do still learn their jobs that way. It's just delayed by several years of book-learning that has little return on investment. When you finally land your job, your first few months are when you really start learning enough to become useful. Disclaimer: I…

This is mostly the system we have in Switzerland.

Only around 20% of teens go directly from school to an university. The rest makes a three or four year long apprenticeship, starting at age 15-16. Apprentices get compensated in the range of 500 to 2000 chf per month.

You can combine the apprenticeship with a secondary education that allows you to go to universities of applied science to study engineering, computer science or whatever it is you are interested in.

So most Swiss with a tertiary education got some practical experience before they stated their tertiary education.

Re: Reality has a surprising amount of detail (2017)

#99

Post by ncmncm »

selestify wrote: post #36
earenndil wrote: post #9

Earlier quoted context omitted.

FWIW I think that simulation theory is fairly obviously bunk, but that's not really an argument. We have no frame of reference for the nature of the simulator, including its constraints and the level of detail it can manage.

Why do you think it’s “obviously bunk”?

It's obvious projection. Centuries ago we lived in a dome, in the 18th century we lived in a clock, now we live in a computer, soon enough in a quantum computer. Give it 50 years and we'll live in the new fad, and giggle at the old fools who thought we lived in a simulation. So primitive!

Re: Reality has a surprising amount of detail (2017)

#100

Post by ncmncm »

cookie_monsta wrote: post #93
omneity wrote: post #18

Related essay: http://blog.omarkama.li/the-market-decides-your-project-scop...

Given the subject matter I had to smile at the way the author bungled the detail within the first sentence: > "A while ago, I read this amazing piece about building a ladder."

Hint: a staircase is the same thing as a ladder, just nailed down. On ships they call staircases ladders.

Mind blown

Post reply on HN