
Thore Graepel
Computer scientist
5
APPEARANCES
5
PODCASTS
012
DEC 30
JAN 6
JAN 13
JAN 20
JAN 27
FEB 3
FEB 10
FEB 17
FEB 24
MAR 3
MAR 10
MAR 17
MAR 24
MAR 31
APR 7
APR 14
APR 21
APR 28
MAY 5
MAY 12
MAY 19
MAY 26
JUN 2
JUN 9
JUN 16
JUN 23
JUN 30
JUL 7
JUL 14
JUL 21
JUL 28
AUG 4
AUG 11
AUG 18
AUG 25
SEP 1
SEP 8
SEP 15
SEP 22
SEP 29
OCT 6
OCT 13
OCT 20
OCT 27
NOV 3
NOV 10
NOV 17
NOV 24
DEC 1
DEC 8
DEC 15
DEC 22
DEC 29
JAN 5
JAN 12
JAN 19
JAN 26
FEB 2
FEB 9
FEB 16
FEB 23
MAR 2
MAR 9
MAR 16
MAR 23
MAR 30
APR 6
APR 13
APR 20
APR 27
MAY 4
MAY 11
MAY 18
MAY 25
JUN 1
JUN 8
JUN 15
JUN 22
JUN 29
JUL 6
JUL 13
JUL 20
JUL 27
AUG 3
AUG 10
AUG 17
AUG 24
AUG 31
SEP 7
SEP 14
SEP 21
SEP 28
OCT 5
OCT 12
OCT 19
OCT 26
NOV 2
NOV 9
NOV 16
Oct 28, 2026
Thore Graepel, CEO & Co-Founder at Metis Reasoning: AI needs to reason like a scientist before we can trust it
6:24
6:38
S
6:06Seb JohnsonHOST
Can you explain a bit about that, how it's different to modern LLMs and then I guess like where we went wrong potentially, like why did we move from AlphaGo in this amazing moment at Mood37 to the LLM architecture where we kind of lost reasoning?

Thore GraepelGUEST
So going back 10 years in time, so to speak, what does the AlphaGo architecture actually look like? It has these two sides to it.

Thore GraepelGUEST
The policy network decides what are promising moves in a given position, and the value network is able to judge if a position is good for black or for white.
10 MINS LATER
S
16:42Seb JohnsonHOST
How do we try and rebuild that for a much more general intelligence, do you think?
AlphaGo co-creator comes out of stealth with new startup
4:32
4:53
5:08
5:15
5:45
8:16

Mike ButcherHOST
Sure, but the token approach has produced some pretty spectacular results, hasn't it? And so are you saying that you're sort of going back to a kind of if this, then that approach or a kind of algorithmic, traditional machine learning algorithmic approach, but rather than a neural

Thore GraepelGUEST
network approach? So for what we do, we definitely need neural networks because they've developed to a point where they can now do semantic reasoning.

Thore GraepelGUEST
You can give them images and they can give you an understanding of the images.

Thore GraepelGUEST
The problem is if you delegate the whole reasoning process to large language models, because then inside of the model, you get these complex token manipulations, but you never really get the full picture of a system that knows what it knows, knows what it doesn't know, has this knowledge state and then advances that knowledge state.

Thore GraepelGUEST
but using large language models and other neural operators to interface with the world.

Mike ButcherHOST
So if you're producing an auditable system, do you think you'll be able to address the problems with agents and rogue agents that have flared up in the media in recent times? Agents breaking out of sandboxes, for instance?
Ex-DeepMind Scientist WARNS: Why We Still Can’t Trust Today’s AI
56:30
56:36
56:43
56:52
56:58
60:56

Thore GraepelGUEST
And, um, AlphaGo actually embodied both of these concepts and brought them together.

Thore GraepelGUEST
There were these two neural networks, the policy net and the value net, that look at the board and make judgment calls.

Thore GraepelGUEST
What are good moves in this position? How good is this position for me? How good it is for the opponent? That was an important part of the system.

Thore GraepelGUEST
But, um, the other half, the system two part of it, was the game tree search.

Thore GraepelGUEST
And the game tree search is this process by which you start with a, with a position, and then you look at possible moves you can make that take you to future positions, and you, then you think about what the opponent might answer, and you-- So you develop this, this big data structure of hypothetical moves that you could make, and, and the system reasons about these moves and tries to find the one that leads to the best outcome.

Julian IssaHOST
It's fascinating to think about these things, and we think about it in, uh, with self-driving cars as well, being far more safe than, than people driving.
Bonus audio. Versión Original Entrevista Pushmeet Kohli y Thorne Graepel (v.o. en inglés)
6:35
6:48
7:05
7:16

Thore GraepelGUEST
And so what that does is it makes the tree n- more narrower because these are all the moves, but now we're looking only at a small selection of moves at each stage.

Thore GraepelGUEST
The value net is very useful because, uh, ordinarily we'd have to always look at all the variations to the very end because only then would we know if black or white has won the game.

Thore GraepelGUEST
You can show it an incomplete board, maybe only after 10 moves or 15 moves, and it can already judge if it's more favorable for black or more favorable for white.

Thore GraepelGUEST
And this, of course, has also been trained on game data, um, by using machine learning.
16 MINS LATER
Is the Earth warming faster than we expected?
19:18
19:28
19:35
19:45

Tom WhippleHOST
The, the architecture of your program, AlphaGo, was very different, wasn't it? How, how did it, how did it know how to play Go?

Thore GraepelGUEST
So, um, Deep Blue falls into the category of what we might call good old-fashioned AI.

Thore GraepelGUEST
It, uh, pretty much did a brute force search of all the possibilities that it could within the time limit and with the computer available to find the best move.

Thore GraepelGUEST
In contrast, uh, AlphaGo has a much more human-like architecture, if you like.
