DeepMind's AI Takes An IQ Test! 🤖

The paper "Measuring abstract reasoning in neural networks" is available here:

We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
313V, Andrew Melnychuk, Angelos Evripiotis, Brian Gilman, Christian Ahlin, Christoph Jadanowski, Dennis Abts, Emmanuel, Eric Haddad, Esa Turkulainen, Geronimo Moralez, Kjartan Olason, Lorin Atzberger, Marten Rauschenberg, Michael Albrecht, Michael Jensen, Morten Punnerud Engelstad, Nader Shakerin, Owen Skarpness, Rafael Harutyuynyan, Raul Araújo da Silva, Rob Rowe, Robin Graham, Ryan Monsurate, Shawn Azman, Steef, Steve Messina, Sunil Kim, Thomas Krcmar, Torsten Reil, Zach Boldyga.

Crypto and PayPal links are available below. Thank you very much for your generous support!
Bitcoin: 13hhmJnLEzwXgmgJN7RB6bWVdT7WkrFAHh
Ethereum: 0x002BB163DfE89B7aD0712846F1a1E53ba6136b5A
LTC: LM8AUh5bGcNgzq6HaV1jeaJrFvmKxxgiXg

Károly Zsolnai-Fehér's links:

[coma-body-stilllife]

Will the Ai constantly tell people it's in MENSA when it reaches IQ132 like people do? That's the true test.

[MIC_Enjoyer]

Maybe someday ai will be able to fully grasp the complexities of rick and morty

[MobyMotion]

Wow. I swear with every DeepMind paper I feel like they're on the cusp of cracking true AGI.

[stasyszy]

terminator was not an action movie, it was a documentary from the future

[DamianReloaded]

I don't think it means the same to perform these tests on humans that on ANNs. These kind of intelligence tests let you predict other capacities of the human making them. In the case of an ANN you can only say that it is good at solving that test. What would be interesting is to see the same ANN without retraining solving the same problem on a completely different domain, like humans do. That'd be the real deal! ^_^



(A general abstraction for all things in the world)

[M]

With the third question there's actually additional pattern as every column contains black, white and gray shapes. I used this pattern and the AND pattern to get to A, meaning that you don't necessarily need the first pattern.

[PaulBrunt]

"Perhaps even in the next few months": A statement I never thought I'd take seriously :-)

[karlkastor]

I don't see why training on IQ test problems, then testing on the same kinds of problems, is any different from any other kind of classification.

[NZOypHI]

Reasoning is not just discovering existing things, but also nonexisting and related things. It's critical for making strategies.

[AntsanParcher]

I find these kinds of tests to be extremely infuriating. To me they seem more like guesswork and dealing with visual noise than actual reasoning – that is, there is a bit of reasoning in them, but most of them being hard comes from the visual noise and I've seen multiple tests of these kinds where there exist alternative, equally reasonable answers.
Not to speak of the general problem of tasks of the kind
"1, 1, 2, 3, 5, 8, …
What is the next number in the sequence?"
Well, there is the obvious answer, but that's not how sequences work and it could be any series starting with these numbers. It's just looking at it and guessing "Well, everyone knows the Fibonacci sequence and so that's probably what they're talking about", but being a math student this kind of answer is really unsatisfactory and giving that answer I can't shake a feeling of "Maybe there's another pattern there that gives the same answers up to this point but different ones later on?". Like, I could reasonably go on with a "4" here – and that's not even a particularly creative answer!

[MargaretSpintz]

Regarding WReN, I spoke to the authors about this. The true labels (problem type) were shown during training. They'd suggested that the reason for the boost in performance could be that it was being trained more effectively since it was shown these direct reasons, which is kind of extra info. So the performance boost isn't as magical as it seems...

[npip]

1:30 "Only panel A has 4 circles in it" What? How is that a legitimate deduction. 234/345/234 is such a weak pattern, someone is supposed to click the answer without having any idea why the pips are in the precise locations that they're in? How would anyone be able to be that confident, even if they're told the pattern

[theisegeberg]

If they remove the distracting line and give it training sets, they're not training it towards general intelligence, they're just teaching it a rather obscure kind of image recognition. No matter how you shuffle, color and shape the objects, it will be working with the white background. It's going to be looking at the background more rather than the objects. (I work with machine learning).

[jamietherooster]

I love how all the naysayers always comment on these videos saying "but it cant do that". With the same sort of mindset would we'd still be riding horses. There are some incredibly smart people collaborating and spending shed loads of cash to further this technology, like never before. Every step of progress in this domain brings us closer to AGI, don't be one of those guys who looks back and wonders how it happened so fast.

[diamondguy]

You mentioned that the AI has to justify its answer to the problem, what do you mean by this? What format is the "explanation" given in?

[PandoraMakesGames]

Freaking cool video! You inspire me to make my own AI videos!

[Joshuaxiong]

1:23 I never thought of this before, nice, learned.

[DistortedV]

Thanks for the great content! Seems a bit naive to come to AGI conclusions from this as some of the commenters have done. Hopefully it can inspire more research about why some of the top networks (Resnet) perform so well.

[civimaxx]

I thought it would have been B for the second question, because each row has one panel with seven arrows, and the other two have the same amount (1, 1, 7 in the first; 7, 5, 5, in the second and 5, 7, 5 for the third).

[dellorttoguoy]

1:35 omfg i over analyzed this question so much you can't even comprehend. i was thinking about this for 1 hour... just thinking of hundreds of theories that i would test out in my brain without pen and paper... i was over analyzing and thinking of such over complicated things such as string theory... and when i found out the answer was so easy i was so pissed off -.-