>AI beat humans at chess but it will never be do so at Go!
>AI may beat humans at Go but it will never solve protein folding it's too complex!
>AI may have solved protein folding but it will never create art!
>AI may have created art but it will never be able to explain a joke
>ok it can do that too but it will never be able to do blablabla
What is wrong with those people? Do they not get tired of getting blown the fuck out? Seriously.
Alphafold is very far from 'solving protein folding'. It's merely better than the shitshow that existed before, but many much simpler, fully end-to-end models exist that perform much better than alphafold already, and beside this, it's insanely far from 'solved' (the reason I mention the simpler models is that it shows the problem was the use of DL at all, not alphafold itself -- the sota was so ass that a random forest would have improved it).
isnt it like 90% accurate when stuff before was like in the 40% range?
Yes, but not really. In the first place, that's not how folding works at all anyway. For example, proteins don't actually fold into a specific 3d structure, but rather wiggle around constantly, at different rates depending on the substructure. It is far more accurate to represent a protein folding as an ensemble of solutions, not a single solution. "Folding" really means "solving the one configuration that was elucidated by crystalography/NMR in this specific preparation context". In addition, you can't talk about 'accuracy' in this context, what is produced is the position of Calpha atoms, and the measure is the angstrom distance between the ground truth and prediction (e.g. in pic related, root mean square error thereof). As you can see, alphafold is about "3x better" than other methods, but that doesn't mean much without a reference frame, and once again only applies to a very structurally specific subset of structures and sequences (because there are plenty of proteins we can't crystalize at all for instance). Specifically, in context, you really want an error rate of 0.3 or lower RMSD, so alphafold is far too poor yet to consider the problem 'solved' even in this very limited subset.
You may be interested in
http://occamstypewriter.org/scurry/2020/12/02/no-deepmind-has-not-solved-protein-folding/
or
https://www.chemistryworld.com/opinion/behind-the-screens-of-alphafold/4012867.article
does this happen to be your field? seeing an informed post on BOT is like seeing a shooting star
Almost. I work in proteomics, but not on structure. Obviously I end up with some decent knowledge of adjacent areas as a result.
what is a 150+ IQ person doing on this godforsaken board for mental retards
lolwut
What would you like me to clarify?
Don't assisted folding pathways make a computational solution of all protein structures based solely on the AA sequence impossible as well?
Except for the blog not emphasizing the caveats much (i.e. the author is on the "very optimistic" camp, I'm always more on the skeptical side), it's pretty accurate. The author also outlines the same kinds of caveats I listed, for example, despite quickly moving on to hopeful stuff.
What I would like to add is that, exactly as the opening of the article points out, it is the "protein folding people" that declare this solved, but the chemistry people (he didn't name them specifically, but implied it, and in my experience that's accurate) that say it's far from solved.
Regarding
>the science of protein folding is solved and the rest is just engineering?
I both agree and disagree.
I greatly disagree with the notion that the *science* of protein folding is solved, because this is still just using a statistical black box based on pattern matching, it tells us nothing about how proteins fold at all. In addition, the caveats haven't been addressed at all.
However, I do agree that engineering is all we need from there to get <0.2A accuracy on C_alpha atom positions in the static structure case, and that AF2, modern/simpler models that do the same thing, and such are already incredibly useful for many tasks.
I don't now enough to pronounce myself on this but I will say the following:
With the ensemble view, it is possible to find the solution set. The problem is how to model it, and how to observe it, as well as how to select a solution from the ensemble post-hoc after knowing the conditions of capture (the last step being required to actually verify the predictions).
A good article on the topic of solution ensembles and why they matter includes:
https://www.pnas.org/doi/10.1073/pnas.1411798111
This article also references papers explaining this ensemble view in more details for anyone who wants to dig into this.
Well Anon they don't understand the quick pace of technological progress.
Some things I'd love to see AI do, that it has actually failed to deliver on so far:
>self driving cars
>beating the top human at StarCraft (laugh all you will, this turns out to be harder than Go)
>programming
cant it just micro two zerglings frame perfectly and win on 6-pool?
I remember it was doing perfect stalker micro while looking at the whole map at one point and fucking everyone up, but the meatbags complained about it being unfair.
Now, what exactly is fair is debatable, but looking at the whole map and having uncapped apm probably isnt.
I think that's ridiculous. No one is capping the Go/Chess bots for being "unfair", what makes Starcraft so different?
The difference is that go and chess are turn-based. Imagine if the chess bot could move its pieces anytime it wanted, regardless of whose turn it is. That's what the starcraft bot is basically doing. Also, in starcraft, you don't normally see the map, there's a fog of war and you have to explore to locate your enemy at all, and then keep scouting to figure out what the enemy is doing and how to counter that (the info is not automatically updated just because you reached some point on the map once, you have to keep a unit within vision range to have visual updates).
The AI bypassed all that, seeing the whole map at all time, and also having unlimited action control (so it can bypass all tactic by just abusing animation delays in the enemy units because it can click a billion times a second).
I see. that does sound broken, the AI should only be provided the same information as a human player. but the whole action control seems debatable to me - I mean, you are already competing against a non-human, limiting its action speed just seems like kidding ourselves.
It's because starcraft is not an action game, it's a tactics and strategy game. A player cannot simultaneously manage build queues and juggling individual units to mess with animations, it's basically one or the other (a limited version is possible though). This is important for game-play purposes, because it means the player has to make the tactical decision between attending one task or the other. Logically, the same should be true of the AI for it to be fair (see analogy with AI playing chesspieces at any time). Limiting the AI to, say, 200 apm would be reasonable and force the AI to actually place the 200 apms tactically as well instead of having free reign.
They limited the AI to basically 1/4 of the apm of a starcraft pro and they still were shitting their pants saying "the AI uses clicks perfectly meanwhile moronclicking is about clicking many times in useless spaces!" so the team just dropped the project
Starcraft community just couldn't handle it, not even DOTA community was as butthurt as starcraft about the AI
Without that limitation the original AIs made twenty years ago would have easily beaten the best players, using strategies like what this anon said:
It's like making a bot in an FPS game, you don't need neural networks, it's faster than people are so you can beat people with regular programming, it can headshot people before they notice there is an enemy on screen, again with no neural networks involved.
the bot should be able to beat you through strategic calculation too, not just quicker response optimizations on normal play
its cope but its also boring and doesnt solve any interesting problems if the difference is just "be able to constantly perceive and react to new information at 1/10000th of a second for 20 minutes straight"
If you wanted it to be TRULY fair then you would give the AI a robot arm to control the mouse and keyboard, which would probably be roughly as fast as a quick fingered human elite.
I don't think it was actually seeing past fog of war, that shouldn't be possible or it would be trivial for hackers to exploit.
It was just looking at the entire map at once, fog of war and all, which meant that it could have its attention on all units at once without having to move its screen like normal players have to.
When I saw the AI wasn't watching the pixels like a person but had access to game data I knew it was all bullshit.
Of course it's harder, the StarCraft state space is unquantifiably large, there have probably never been two games identical past the first couple of seconds.
They said the same about Go, it has an insane state space. But AI will need to be able to grapple with such complexity to deal with the real world.
Yeah, but at least Go can reasonably be expressed as a decision tree.
Also this
Man imagine the seething when an AI can code better and faster than shitstain BOTentoomen
Current deep learning AI is incapable of abstract thought and will 1) always require human handholding to achieve anything and 2) can only handle simple, straightforward tasks.
what happened with starcraft?
>beat some top players with cheats
>played randoms without cheats and got to a very high level but not the best
>then abandoned
They moved to other projects like Alphafold. Kind of disappointing that they were unable to truly conquer Starcraft, but their goal wasn't to focus solely on making good videogame AI so it makes sense. I hope someone in the future works more on this though, some games (like the entire strategy game genre) could benefit from having decent AI.
NTA, but what are your thoughts on one of the first articles from the time of the original Alphafold
https://moalquraishi.wordpress.com/2020/12/08/alphafold2-casp14-it-feels-like-ones-child-has-left-home/
and more specifically, the claim that the science of protein folding is solved and the rest is just engineering?
AI beat the fuck out of humans in Dota 2
and that's more complex than starcraft
there were thousands of games and humans only won once through incredible rat style gameplay with sniper whittling down their towers for literal 3 hours (because AI were absolutely unbeatable and perfect in an actual teamfight)
nah, ai had no idea about strategy, the only advantage was it was clickclickclick, which we already had literal decade ago without aiaiaiai in starcraft for example
They beat Starcraft anon, it just starcraft community seethe'd so hard they dropped the project, gaymers started denouncing cheats and changing the rules mid way and then it became an exercise in sophistry
> starcraft community seethe'd so hard they dropped the project
holy based.
why can artists and coders be this based?
idk if people say that. but if they do say that they're retarded.
reality is cause and effect with some randomness thrown in. it's just a fact that we'll be able to simulate human brains, problem solving, intelligence, and super intelligence using 1s and 0s. it's just a question of when.
Quantum processes that may exist in the brain fundamentally cannot be emulated in binary
Yes, but that assumes quantum processes have macroscopic implications at the level of neural processing, which is very unlikely. Even then, QC is happening, too.
ai will never be able to make me happy
gene editing will
my ai chatbot is making me happy
That;s a you problem, you dont want to be happy. Happiness always has been a choice.
AI will delete pepe and wojak
I'll kneel if so
Ludd was right and the AI will be the cause of the collapse.
When's the fucking singularity going to happen, tired of going to work everyday
>>AI may have created art but it will never be able to explain a joke
It can't though. Have you seen an AI trying to explain the sneeds feed and seed joke?
Yes.
AI may have solved programming but it will never hold you, because you're too poor
>ai beat a crippled stockfish no one can replicate because trust me bro
>the only people care about Go is muh ai snakeoilsellers so they have a cheap target to gloat over. How many Go players you know besides the one dead moron they shove in your face? That is your answer
>ai solves protein folding in headlines, same goes for matrix multiplication
>ai is capable of composing e-boi images from other e-boi images
thing is, they use these non achievements to prep up the garbage they want to sell like self driving and npcs will nod, but the cold harsh reality is that ai $T$ scam has a net negative impact
i like ai because it devalues other people's achievements and brings them down to my level
nobody cares about your "achievements", narcissistic gay.
>solved games
>pre trained datasets
>statistical models can do things given a big enough sample set
big fucking revelation
Why are brainlet zoomzoommorons like you so prevalent on nu/g/?
>AI beat humans at chess but it will never be do so at Go!
Not really.
If you ever study the problem, you'll see the tree of posible games of chess is limited. It was just a matter of time. (Moore's law.). Go is more complex, but still limited.
>art
depends on your definition of art. I do think it is doing "art", but is nothing new. A kid doodling is making art, caveman did art, and applying a filter to modify colors is a computer doing art IMO.
Anyway the issue is not what can or can't do. but Who owns it.