WEBVTT

00:00.000 --> 00:04.540
of the real human data, or if that captures different answers.

00:04.540 --> 00:07.880
For example, if I say, instead of a sovereign German woman,

00:07.880 --> 00:11.820
you're a city English boy, I know,

00:11.820 --> 00:16.020
if you have explored those kind of things.

00:16.020 --> 00:20.920
Western politics of Hawaii

00:20.920 --> 00:25.480
kind of look at that as those democratic variables really

00:25.480 --> 00:28.880
matter in terms of that.

00:28.880 --> 00:34.020
Yeah, yeah, like the LLM was prompted

00:34.020 --> 00:37.060
to answer as a German woman, but then your sample

00:37.060 --> 00:40.600
is not all the German women, I guess, like the real data.

00:40.600 --> 00:43.380
So how do you account for those demographic variations

00:43.380 --> 00:46.380
when you prompt them all to answer just as one persona?

00:46.380 --> 00:46.880
You know?

00:46.880 --> 00:50.380
Yeah, yeah, yeah, that's what makes it so clear.

00:50.380 --> 00:55.260
So basically, he took like 200 persons for the survey,

00:55.260 --> 00:57.260
and all of those kind of .

00:58.880 --> 01:05.060
And did you explore like the difference between the personas

01:05.060 --> 01:05.720
that you made?

01:05.720 --> 01:10.700
Like did the LLM show some variability

01:10.700 --> 01:14.980
according to these different demographics?

01:14.980 --> 01:19.600
Yeah, yes, I had a lot of .

01:19.600 --> 01:22.900
I did run this analysis with the latest data.

01:22.900 --> 01:26.780
Previously, it was like, you could really kind of see that they kind of cast this kind of .

01:26.780 --> 01:27.060
They were like, I don't know, I'm not sure.

01:27.060 --> 01:27.320
I'm not sure.

01:27.320 --> 01:27.700
I'm not sure.

01:27.700 --> 01:27.860
I'm not sure.

01:27.860 --> 01:28.020
I'm not sure.

01:28.020 --> 01:28.280
I'm not sure.

01:28.280 --> 01:28.460
I'm not sure.

01:28.460 --> 01:28.860
I'm not sure.

01:29.400 --> 01:31.700
I don't keep them in a background.

01:31.700 --> 01:40.400
But they are like missing or some of that

01:40.400 --> 01:42.740
in contrast .

01:42.740 --> 01:44.900
Yes, there's some .

01:44.900 --> 01:46.200
There's some .

01:46.200 --> 01:48.260
Right. Well, thank you very much.

01:48.700 --> 01:54.540
The next presentation is bunun

01:54.540 --> 01:58.860
of theömni intensivpamc towards .

01:58.860 --> 02:02.360
in order to general intelligence with a causal temper approach

02:02.360 --> 02:04.800
to the cognitive execution substrate.

02:19.800 --> 02:20.280
Yes.

02:26.640 --> 02:28.400
Okay, thank you for this opportunity.

02:28.860 --> 02:37.000
Like my two main teams I'm interested in is how is cognitive science,

02:37.100 --> 02:43.080
is it possible for cognitive science to help to build AI systems?

02:45.300 --> 02:48.080
Or is it just an engineering task?

02:48.320 --> 02:53.840
And the other one is on the next slide,

02:53.840 --> 02:58.220
but I'm going to tractability.

02:58.860 --> 03:03.860
And then some questions related to these issues.

03:04.740 --> 03:07.860
And I finalize with the hypothesis that is kind of checklist

03:08.500 --> 03:14.860
that what we should know up to this point, maybe also later.

03:16.200 --> 03:27.620
So, my method is like all the participatory design, meaning that, okay, I set a goal that,

03:28.160 --> 03:28.740
like, unfiltered.

03:28.860 --> 03:29.620
I set a feasible goal.

03:30.520 --> 03:37.420
Let's build a robot which has AGI capabilities, meaning artificial general capabilities.

03:37.580 --> 03:39.620
Obviously, that's kind of unfeasible.

03:39.800 --> 03:49.820
But the purpose is that in the process we find problems and criteria that are, you know, on the way.

03:51.180 --> 03:58.040
So, and the other one good thing is that this kind of moves from the classical symbolic paradigm,

03:58.240 --> 03:58.740
symbolic paradigm to the AI.

03:58.740 --> 03:58.820
So, I'm going to talk about the AI.

03:58.820 --> 03:58.840
So, I'm going to talk about the AI.

03:59.000 --> 04:06.340
And it's a kind of manipulation paradigm to the embodied, embedded, and enacted and extended cognition,

04:06.920 --> 04:12.740
because it's a robot and it functions in a physical world.

04:13.400 --> 04:16.860
And this causal chamber is a device from physics.

04:17.640 --> 04:28.100
It's a box where you can measure and do, measure real physical phenomena and do experiments in that way,

04:28.220 --> 04:28.620
in real physical phenomena.

04:28.620 --> 04:37.920
real world instead of in simulation world so I have this thinking tool or

04:37.920 --> 04:48.600
concept of cognitive executive substrate meaning that everything that our head is

04:48.600 --> 04:54.700
the executive organ that interprets every language data just like in

04:54.700 --> 05:02.980
computer we have processor the processor executes certain kinds of programs so

05:02.980 --> 05:11.020
like we humans we execute philosophy we understand philosophy through this to

05:11.020 --> 05:16.660
our brain to the system all the knowledge we have and all the skills we

05:16.660 --> 05:24.480
have we need that in order to interpret that so and and the point is that

05:24.480 --> 05:24.680
that

05:24.680 --> 05:32.100
we should think always what is the system we are kinda what is the system

05:32.100 --> 05:38.760
that kinda executes the theories or anything or problems that we investigate

05:38.760 --> 05:42.800
instead of like in many times in philosophy you focus in the discourse

05:42.800 --> 05:49.460
what people say and what happens in this like conceptual realm

05:49.460 --> 05:54.460
it's like same thing like in you can have extended mathematics but

05:54.460 --> 06:01.240
only part of that is relevant for physics same thing here we should kind

06:01.240 --> 06:12.640
of ground our our theories in in cognitive executive substrate that's the

06:12.640 --> 06:16.720
conceptual tool so what is artificial general intelligence there are many

06:16.720 --> 06:24.700
definitions but but I like this one it's quite practical there's similarly

06:24.700 --> 06:30.100
like self-driving cars there's this levels of autonomy and then there is

06:30.100 --> 06:44.680
this is it the general system or is it very narrow narrow system and yeah and

06:44.680 --> 06:46.520
this image shows that

06:46.720 --> 06:58.660
resistance is futile that the crafts are going that direction where the in

06:58.660 --> 07:05.340
different tasks AI is able to surpass human performance so that's why I'm

07:05.340 --> 07:12.960
interested in in knowing why how would computer science help in going that

07:12.960 --> 07:15.220
direction

07:16.720 --> 07:23.880
but there's a problem there's this practical condition thesis that states

07:23.880 --> 07:32.020
that the AGI is not computable in practice

07:32.020 --> 07:37.540
5 minutes I have 5 minutes what left

07:37.540 --> 07:46.540
ok so but there's a problem

07:46.720 --> 07:54.280
they don't define that they define the problem problem space to broadly without

07:54.280 --> 08:01.480
any inductive biases and they think AI system as a monolithic agent instead of

08:01.480 --> 08:06.400
they should think it as a modular agent

08:06.400 --> 08:11.400
and

08:11.400 --> 08:14.400
yeah

08:16.720 --> 08:46.720


08:46.720 --> 09:16.720


09:16.720 --> 09:46.720


09:46.720 --> 10:16.720


10:16.720 --> 10:28.500
, that by its own perspective like what it like to cognize perspective without

10:28.500 --> 10:35.020
without minimal or suitable inductive biases that we throw into them how would

10:35.020 --> 10:46.680
that system kind of be able to create AGI so instead of having a thinking a

10:46.720 --> 10:54.580
model monolithic system we should have that set of modules where each module

10:54.580 --> 11:03.420
represents a cognitive capacity its module is tractable as an algorithm and

11:03.420 --> 11:13.560
also in this into input space the modules are organized in an architecture

11:13.560 --> 11:16.660
where each module is independent

11:16.720 --> 11:26.720
there's no cycles, and they are best effort broadcasters, meaning that they just function

11:26.720 --> 11:33.060
as their own and get modulated by others.

11:33.060 --> 11:43.240
Then these models should be, they can mimic sensitive periods and we can learn from, start

11:43.240 --> 11:56.420
from most digital levels like sensory levels and build on those and the whole system has

11:56.420 --> 12:01.800
different time loops where the lowest, the most digital levels are the most fastest and

12:01.800 --> 12:03.580
the highest levels are more fastest.

12:03.580 --> 12:12.100
So I woke up this morning and I saw, I realized in a dream there was something more, but I

12:12.100 --> 12:13.100
forgot it.

12:13.100 --> 12:23.220
I woke up so, so I invite you to like tell me what is missing here or what I need to

12:23.220 --> 12:26.940
consider more.

12:26.940 --> 12:31.980
Yes, any questions?

12:31.980 --> 12:40.280
I think there's time for one or two more questions, but it's probably great and then it's time

12:40.280 --> 12:41.280
for the next one.

12:41.280 --> 12:42.280
Great.

12:42.280 --> 12:42.780
Sorry.

12:43.100 --> 12:44.100
Thank you.

12:44.100 --> 12:45.100
Hi, Janne.

12:45.100 --> 12:46.100
Excellent.

12:46.100 --> 12:47.100
Thank you.

12:47.100 --> 12:48.100
Always nice to have discussions with you.

12:48.100 --> 12:49.100
You are researching on a very broad basis here.

12:49.100 --> 12:50.100
I have one very general question for you.

12:50.100 --> 12:55.100
So sorry to read it out from paper and sorry for making you to listen to it.

12:55.100 --> 13:06.100
But even if AGI, which we have, would be better in every single task than total analysis,

13:06.100 --> 13:09.940
what would you say would be the best solution for the data?

13:09.940 --> 13:10.940
So, I'm sorry.

13:10.940 --> 13:11.940
I'm sorry.

13:11.940 --> 13:12.940
I'm sorry.

13:12.940 --> 13:15.940
I'm sorry when I say this one's not the best solution.

13:15.940 --> 13:18.940
What I actually need to say is, that perhaps, creators andudio найти a superior intelligence

13:18.940 --> 13:19.980
to strawberry plums that you describe not merely as man, but as an even Senhor and you

13:19.980 --> 13:20.980
know, not only as this man top human, it might not mean that it has superior intelligence.

13:20.980 --> 13:24.500
This is rest injection on the position that intelligence is made of different common tasks.

13:24.500 --> 13:29.940
There's always a possibility about an entirely new task and a need for what we call shave

13:29.940 --> 13:30.940
on creativity.

13:30.940 --> 13:33.740
The long of human intelligence.

13:33.740 --> 13:38.940
So just coming up with different cognitive tasks in a machine does not create anything

13:38.940 --> 13:39.900
more than the superhuman and

13:39.900 --> 13:40.940
all is clear.

13:40.940 --> 13:41.940
Know?

13:41.940 --> 13:49.220
more and there's some of those parts it doesn't create suddenly something this is a promise and

13:49.220 --> 13:56.500
hype of or has been of ai that suddenly there will be this emergence of the general intelligence and

13:57.060 --> 14:05.300
doesn't seem to happen we still have to teach it different tasks i don't know uh my answer would be

14:05.300 --> 14:13.220
like uh using the other mind problem and the monolithic agent problem that if we are able if

14:13.220 --> 14:22.420
if we are afraid to that ai is replacing humans then then it certainly has certain kinds of

14:22.420 --> 14:33.460
capacities and the kind of also like if the tractable cognizant thesis is true then we can

14:33.460 --> 14:35.140
also they have to say that

14:35.860 --> 14:42.740
because they are critical everybody's afraid of ai doom so i would i would ask them that how do you

14:42.740 --> 14:49.540
prove that the ai doom is not intractable i would i would kind of throw it out there that it's

14:49.540 --> 14:56.260
intractable if we are in that school but if if we are in my school then

14:59.460 --> 15:02.740
it's tractable but i don't believe

15:04.740 --> 15:05.140
they do

15:05.300 --> 15:12.740
thank you

15:12.740 --> 15:25.860
now we have a coffee break of half an hour and 2 30 is the last keynote speech thanks again to all

15:25.860 --> 15:27.860
the great presentations

15:27.860 --> 15:31.860
yes

15:31.860 --> 15:33.860
yes

15:33.860 --> 15:35.140
yes

15:35.300 --> 16:05.300


16:05.620 --> 16:07.860
for you

16:07.860 --> 16:13.860
i

16:13.860 --> 16:14.660
thank

16:14.660 --> 16:14.740
you