Science!! Fucking magnets, how do they work?

  • Guest, it's time once again for the massively important and exciting FoH Asshat Tournament!



    Go here and give us your nominations!
    Who's been the biggest Asshat in the last year? Give us your worst ones!
BrutulTM

BrutulTM

Good, bad, I'm the guy with the gun.
<Silver Donator>
14,762
2,644
There needs to be incentives for exposing fraud or even just mistakes. You get rich and famous for positive results, not for finding things that don't work, so there's always an incentive to cheat and the incentive for exposing the cheats is to be the asshole that no on likes because they ruined so and so's career and got the funding pulled from the university. Not hard to guess why people aren't lining up to do it.

I suspect that part of the problem is that the "easy" stuff has all been figured out. What's left is very complex and science isn't that good at dealing with extremely complex things. An ideal experiment only has one variable and you can directly compare the experiment to the control and see the results. In medicine, nutrition, and a lot of other fields it's never like that. If you're doing an experiment on actual humans there's a thousand variables for every person that you can't really control for. It's no wonder why we still can't even figure out what we should eat.
 
Sanrith Descartes

Sanrith Descartes

You have insufficient privileges to reply here.
<Gold Donor>
45,432
123,708
BrutulTM said:
There needs to be incentives for exposing fraud or even just mistakes. You get rich and famous for positive results, not for finding things that don't work, so there's always an incentive to cheat and the incentive for exposing the cheats is to be the asshole that no on likes because they ruined so and so's career and got the funding pulled from the university. Not hard to guess why people aren't lining up to do it.

I suspect that part of the problem is that the "easy" stuff has all been figured out. What's left is very complex and science isn't that good at dealing with extremely complex things. An ideal experiment only has one variable and you can directly compare the experiment to the control and see the results. In medicine, nutrition, and a lot of other fields it's never like that. If you're doing an experiment on actual humans there's a thousand variables for every person that you can't really control for. It's no wonder why we still can't even figure out what we should eat.
Click to expand...
The real problem is money. These govt grants can be serious cash and can fund for years. Questioning results if you are on the team gets you blackballed. So instead scientists just "remove their name from the paper" but keep their mouth shut. Any published work has to provide access to the data and method so it can be replicated independently.
 
Tuco

Tuco

I got Tuco'd!
<Gold Donor>
47,938
82,644
Sanrith Descartes said:
Falsifying scientific research for grant money and/or fame should be punished by a trip to the Gulag.

Science | AAAS

www.science.org www.science.org
Click to expand...
Great article, thanks for posting.

I'm glad my field (vehicular autonomy) mostly has easily falsifiable publications or expressions of ideas/software in paper form, instead of studies conducted on a given approach. Trusting a publication on their study is frightening to me, especially a result favorable to the financial backer or aligned with any kind of current-year social virtue.
 
  • 1Like
Reactions: 1 user
Sanrith Descartes

Sanrith Descartes

You have insufficient privileges to reply here.
<Gold Donor>
45,432
123,708
Tuco said:
Great article, thanks for posting.

I'm glad my field (vehicular autonomy) mostly has easily falsifiable publications or expressions of ideas/software in paper form, instead of studies conducted on a given approach. Trusting a publication on their study is frightening to me, especially a result favorable to the financial backer or aligned with any kind of current-year social virtue.
Click to expand...
Vehicular autonomy as in "self-driving cars" or as in "I live in NYC and dont need to know how to drive"?
 
wormie

Asshat wormie

2023 Asshat Award Winner
<Gold Donor>
16,820
30,968
Tuco said:
Great article, thanks for posting.

I'm glad my field (vehicular autonomy) mostly has easily falsifiable publications or expressions of ideas/software in paper form, instead of studies conducted on a given approach. Trusting a publication on their study is frightening to me, especially a result favorable to the financial backer or aligned with any kind of current-year social virtue.
Click to expand...
Yeah but in your field all publications start with "we have created a novel approach that is amazingly fresh, new and never been seen before for solving x and y that will change the world" and ends with "we have improved sota by 0.00002% on exactly one dataset". :D
 
Sanrith Descartes

Sanrith Descartes

You have insufficient privileges to reply here.
<Gold Donor>
45,432
123,708
Tuco said:
Self driving cars.
Click to expand...
Nice. Just curious on your take. I am currently playing with the Tesla Beta FSD for local roads. Got any thoughts on it? I have used the highway version crossing the country numerous times but am still playing with the local road version.

ps.. Long island and NYC with our shitty roads, shitty double parking and even shittier drivers ain't the best place for FSD in my opinion.
 
Tuco

Tuco

I got Tuco'd!
<Gold Donor>
47,938
82,644
wormie said:
Yeah but in your field all publications start with "we have created a novel approach that is amazingly fresh, new and never been seen before for solving x and y that will change the world" and ends with "we have improved sota by 0.00002% on exactly one dataset". :D
Click to expand...
Right. There are basically two-classes of papers I consume professionally, those that are easily quantifiable and those that are basically software platforms that do fancy things and use a paper to communicate it.

For the first class, if you look at the current top of COCO test-dev ( Papers with Code - COCO test-dev Benchmark (Object Detection) ) you see the best result is DINO which can be evaluated from code GitHub - IDEACVR/DINO: Official implementation of the paper "DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection" or implemented independently to provide a very modest increase over the previous best.

1658772108885.png


For the second class, https://arxiv.org/pdf/2008.08294.pdf is a good example used by Apollo apollo/jointly_prediction_planning_evaluator.md at master · ApolloAuto/apollo . Whether you just use the coded implementation or implement your own there is only a moderate level of trust placed in their execution on datasets like [1911.02620] Argoverse: 3D Tracking and Forecasting with Rich Maps . The ideas and principles are useful and if you want to buy in you have to integrate/implement and figure out all the unique problems that go along with it. There's much less impact caused by them subtly falsifying results because the cost to reproduce the results is much less.
 
  • 1Like
Reactions: 1 user
Tuco

Tuco

I got Tuco'd!
<Gold Donor>
47,938
82,644
Sanrith Descartes said:
Nice. Just curious on your take. I am currently playing with the Tesla Beta FSD for local roads. Got any thoughts on it? I have used the highway version crossing the country numerous times but am still playing with the local road version.

ps.. Long island and NYC with our shitty roads, shitty double parking and even shittier drivers ain't the best place for FSD in my opinion.
Click to expand...
People whine about Tesla's FSD system being delayed and professionals talk shit but I'm glad they're out there aggressively trying. Every year so many people die in car accidents that it's unsafe to be cautious.
 
Sanrith Descartes

Sanrith Descartes

You have insufficient privileges to reply here.
<Gold Donor>
45,432
123,708
Tuco said:
People whine about Tesla's FSD system being delayed and professionals talk shit but I'm glad they're out there aggressively trying. Every year so many people die in car accidents that it's unsafe to be cautious.
Click to expand...
I probably have 10k miles on the highway version and it works. Still have some issues? Yes, especially the cameras interacting with distinct changes in highway color. Examples like certain overpasses where the shadows make the road too dark compared to the lighted road and in Orlando they have new elevated highways where the very tops of the over passes are almost white compared to the tan-ish road. Aside from those issues, I have been impressed with it.

Having the car stop at stop signs/lights, make turns etc on its own on local roads is still taking some getting used to. One thing i noticed is it accelerates really quick from a stop sign or red light. Much heavier foot than I use.
 
Leaton

Leaton

Trakanon Raider
133
92
I'm in agreement with BrutulTM, all the easy stuff has been figured out.

My solution is mainly to push for better AI and computing alorgrithms/power for researchs and new inventions which then will accelerate more breakthroughs that would otherwise be impossible or would have taken a very long time.

For example, modelling protiens is not something a scientist can just sit down and do, but computing simulation is now able to do that since a few years ago which is an absolutely maassive break through.

About tons of different variables for different people, computing will eventually feature out in what order the millions or so of genetic codes does what in every possible arrangement and have the answer for each individual for drugs, etc.

When you think about that, yeah humans cannot do that by themselves anymore.....
 
Sanrith Descartes

Sanrith Descartes

You have insufficient privileges to reply here.
<Gold Donor>
45,432
123,708
But isn't "easy" relative? Figuring out aeronautical engineering problems may be "easy" today, but i bet it wasn't so easy 100 years ago. I think we constantly push the envelope of scientific research and tech advances just keep moving that envelope farther out.
 
Tuco

Tuco

I got Tuco'd!
<Gold Donor>
47,938
82,644
Yeah I don't know when aeronautics, chemistry or biology was ever easy. In this conversation's context there has never been a time where these sciences were easy to the point where fraud was obvious and easily caught and handled. Humans are just fancy apes who are good at making mistakes and tools and we will always try to cheat the system, whatever it is.
 
Sanrith Descartes

Sanrith Descartes

You have insufficient privileges to reply here.
<Gold Donor>
45,432
123,708
Tuco said:
Yeah I don't know when aeronautics, chemistry or biology was ever easy. In this conversation's context there has never been a time where these sciences were easy to the point where fraud was obvious and easily caught and handled. Humans are just fancy apes who are good at making mistakes and tools and we will always try to cheat the system, whatever it is.
Click to expand...
When I was in grad school, one of the things that got drilled into us was replication of results. I am not defending anyone who commits research fraud (quite the opposite), but it is so damned easy to end up down a rabbit hole doing your research that you sometimes get tunnel vision and your objectivity can get unintentionally biased. Its why having others check your results is so critical. Also - You can end up learning a ton by replicating other's work. One semester I dedicated to nothing but replicating Michael Ross' initially body of work for "Does Oil Hinder Democracy".
 
Tuco

Tuco

I got Tuco'd!
<Gold Donor>
47,938
82,644
Sanrith Descartes said:
When I was in grad school, one of the things that got drilled into us was replication of results. I am not defending anyone who commits research fraud (quite the opposite), but it is so damned easy to end up down a rabbit hole doing your research that you sometimes get tunnel vision and your objectivity can get unintentionally biased. Its why having others check your results is so critical. Also - You can end up learning a ton by replicating other's work. One semester I dedicated to nothing but replicating Michael Ross' initially body of work for "Does Oil Hinder Democracy".
Click to expand...
My most valued papers are typically in two groups:
* Ground breaking work done by absolute legends in the field that pioneer the future of humanity
* Grad students who replicate the previous group's findings
 
  • 1Solidarity
Reactions: 1 user
Leaton

Leaton

Trakanon Raider
133
92
Yes, I meant "easy". It may have been very hard in the past, but mark my words, eventually the majority of improvements and break throughs will come via AI or AI assisted research. There is really no way around it.

That is what will create the hockey curve graph in the explosion of further science break throughs. Most likely by 2045 or thereabouts. Humans (even very smart legends ) will not be able to keep up with out the help of AI.
 
wormie

Asshat wormie

2023 Asshat Award Winner
<Gold Donor>
16,820
30,968
Tuco said:
Right. There are basically two-classes of papers I consume professionally, those that are easily quantifiable and those that are basically software platforms that do fancy things and use a paper to communicate it.

For the first class, if you look at the current top of COCO test-dev ( Papers with Code - COCO test-dev Benchmark (Object Detection) ) you see the best result is DINO which can be evaluated from code GitHub - IDEACVR/DINO: Official implementation of the paper "DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection" or implemented independently to provide a very modest increase over the previous best.

View attachment 424061

For the second class, https://arxiv.org/pdf/2008.08294.pdf is a good example used by Apollo apollo/jointly_prediction_planning_evaluator.md at master · ApolloAuto/apollo . Whether you just use the coded implementation or implement your own there is only a moderate level of trust placed in their execution on datasets like [1911.02620] Argoverse: 3D Tracking and Forecasting with Rich Maps . The ideas and principles are useful and if you want to buy in you have to integrate/implement and figure out all the unique problems that go along with it. There's much less impact caused by them subtly falsifying results because the cost to reproduce the results is much less.
Click to expand...
Makes sense that in industry you check and test the research papers claims. My comment was about academia. Most papers are written and then never checked after. People just chase the tiny improvements so they can get published in some non junk conference and then off they go for the next paper that will produce more "novel" results. But the unchecked/unvalidated papers continue to be cited as reference, with new ideas being built on top of such references. Academia needs to do better than "publish or perish". Whole thing is poop.
 
  • 1Like
Reactions: 1 user
Tuco

Tuco

I got Tuco'd!
<Gold Donor>
47,938
82,644
wormie said:
Yeah but how do you choose which to read? You arent validating all of them obviously.
Click to expand...
Oh that's impossible. In my current position for very common algorithms like object detection I'm able to just coattail ride off of open source software systems written by people who DO have to separate the wheat from the chaff on the bleeding edge and build something on an approach that will be obsolete before they release their first version.

I wasn't responding to you with that video btw, I agree with everything you said. I just making the joke that I inadvertently used the same "there are two types of people in the world" metaphor from my namesake's movie.
 
You must log in or register to reply here.