High Signal Podcasts Evidence ledger
Method
Browse
How the evidence works

Public evidence record

Tim Scarfe

Host · Machine Learning Street Talk

Claims
168
Episodes
35
Shows
1
Named items
4

Books, apps, and tools

The evidenced stack.

Browse the grouped index →

tool / uses

MATLAB

“during my PhD, I used Mathematica and MATLAB”

Machine Learning Street Talk · 3 Mar 2026

Evidence receipt · Source ↗

tool / uses

Mathematica

“during my PhD, I used Mathematica and MATLAB”

Machine Learning Street Talk · 3 Mar 2026

Evidence receipt · Source ↗

person / likes

Douglas Hofstadter

“Yes. I I love Douglas Hofstadter. So so there's this kind of self modeling and then second order self modeling and third order self modeling which could be applied to other agents and of course, know, in in the real world, we are computationally bounded.”

Machine Learning Street Talk · 21 Oct 2025

Evidence receipt · Source ↗

person / likes

Andy

“I'm I'm huge fan of Andy's, big hero of mine. You probably didn't think that they would be so relevant later on in your career.”

Machine Learning Street Talk · 18 Oct 2025

Evidence receipt · Source ↗

Claim ledger

What Tim said.

56 transcript-backed records

02 / evaluation

A a lot of people are. And the these agents have an incredible amount of intelligence and flexibility, which means we don't precisely specify what they do.

“A a lot of people are. And the these agents have an incredible amount of intelligence and flexibility, which means we don't precisely specify what they do.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

03 / evaluation

Whereas in the in this sense, we're we're now abstracting them because there's also this orthogonality thesis as well, which is this intelligence of final goals are are, you know, disconnected from each other.

“Whereas in the in this sense, we're we're now abstracting them because there's also this orthogonality thesis as well, which is this intelligence of final goals are are, you know, disconnected from each other.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

04 / evaluation

Maybe we can RL train it to refactor it, and we just have to kind of and it's very dangerous keeping going because now we're messing all of our code bases up and we're we're creating all of this slop everywhere.

“Maybe we can RL train it to refactor it, and we just have to kind of and it's very dangerous keeping going because now we're messing all of our code bases up and we're we're creating all of this slop everywhere.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

05 / evaluation

similar context around them. And this is very pertinent because you've got a paper out basically saying that we should predict in the latent space, not the token space.

“similar context around them. And this is very pertinent because you've got a paper out basically saying that we should predict in the latent space, not the token space.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

06 / evaluation

I guess the interesting thing for me is that when we think of reinforcement learning algorithms like AlphaGo Zero, it makes sense that they are reward seeking because there is this structured inference process.

“I guess the interesting thing for me is that when we think of reinforcement learning algorithms like AlphaGo Zero, it makes sense that they are reward seeking because there is this structured inference process.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

07 / evaluation

Right? And it's a serious problem, because it's all about epistemic subjectivity, which is that you generate things that you don't understand, and it convinces you that it's correct, and you can't see the glitches.

“Right? And it's a serious problem, because it's all about epistemic subjectivity, which is that you generate things that you don't understand, and it convinces you that it's correct, and you can't see the glitches.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

08 / evaluation

Even that, I think you can argue, maybe not entirely the story, but it's definitely not the story for proteins because that's the hardest problem is the large scale structure.

“Even that, I think you can argue, maybe not entirely the story, but it's definitely not the story for proteins because that's the hardest problem is the large scale structure.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

09 / evaluation

Physicists, for example, they they work very, low level and they talk about, you know, the the the dynamics of particle systems and and whatnot. And what I'm really fascinated in, I mean, you come at it from an economics perspective, which is traditionally dominated by this agential lens, and you talk about equilibria and incentives and and so on.

“Physicists, for example, they they work very, low level and they talk about, you know, the the the dynamics of particle systems and and whatnot. And what I'm really fascinated in, I mean, you come at it from an economics perspective, which is traditionally dominated by this agential lens, and you talk about equilibria and incentives and and so on.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

10 / evaluation

Well, suppose another thing that doesn't help is that these these systems are like soup, and there's even a field called mechanistic interpretability that tries to kind of dig into the soup and and it's almost like they're they're searching for UFOs.

“Well, suppose another thing that doesn't help is that these these systems are like soup, and there's even a field called mechanistic interpretability that tries to kind of dig into the soup and and it's almost like they're they're searching for UFOs.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

11 / evaluation

You know, and you make this series of decisions, and then you've got people using your application, and then you can't really wind that back. It doesn't matter if you've got the magical automation machine because you can't easily roll that back because there's lots of complexities, you know, do you see ICD testing?

“You know, and you make this series of decisions, and then you've got people using your application, and then you can't really wind that back. It doesn't matter if you've got the magical automation machine because you can't easily roll that back because there's lots of complexities, you know, do you see ICD testing?”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

13 / evaluation

Because the whole reason we created agile software development as a methodology is because it's inconceivable, It's outside our cognitive horizon.

“Because the whole reason we created agile software development as a methodology is because it's inconceivable, It's outside our cognitive horizon.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

14 / evaluation

These are problems that have been solved before in in part or in whole, which means when you look at the epistemic tree, many of the building blocks for solving them are very high up in the tree.

“These are problems that have been solved before in in part or in whole, which means when you look at the epistemic tree, many of the building blocks for solving them are very high up in the tree.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

15 / evaluation

Right? So we're using neural networks because they're incredibly flexible and they understand a lot of things about the world, but they don't have the kind of constraints that we want.

“Right? So we're using neural networks because they're incredibly flexible and they understand a lot of things about the world, but they don't have the kind of constraints that we want.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

18 / evaluation

I think the counterfactual thing is an important feature here because we could take something which was conscious or something which had agency, and we could just take a trace of the actual path which was found.

“I think the counterfactual thing is an important feature here because we could take something which was conscious or something which had agency, and we could just take a trace of the actual path which was found.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

20 / evaluation

We agents because you have consistent beliefs and ideas and you're not just an impulse response machine that's your being actions aren't determined entirely by this situation, you're a person.

“We agents because you have consistent beliefs and ideas and you're not just an impulse response machine that's your being actions aren't determined entirely by this situation, you're a person.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

21 / evaluation

I think if I remember correctly, at 1 point you drew an imaginary kind of graph where you said on 1 axis we have science realism, which is where our scientific theories actually represent things in the world, then we have empiricism, is the idea that facts we receive tell us something about the world.

“I think if I remember correctly, at 1 point you drew an imaginary kind of graph where you said on 1 axis we have science realism, which is where our scientific theories actually represent things in the world, then we have empiricism, is the idea that facts we receive tell us something about the world.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

23 / evaluation

Of course Pavlov and the dogs, folks at home will know about that. And Newton is still around, we still use that, but we don't use reflex theory anymore.

“Of course Pavlov and the dogs, folks at home will know about that. And Newton is still around, we still use that, but we don't use reflex theory anymore.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

25 / evaluation

I interviewed her recently. And she said that 1 of the most pervasive myths in neuroscience is that we use these leaky abstractions and idealizations to talk about cognition.

“I interviewed her recently. And she said that 1 of the most pervasive myths in neuroscience is that we use these leaky abstractions and idealizations to talk about cognition.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

28 / evaluation

Yeah. It's really interesting what you said because the way I read that is things like chat GBT and language models, they are entropy smuggling or agency smuggling.

“Yeah. It's really interesting what you said because the way I read that is things like chat GBT and language models, they are entropy smuggling or agency smuggling.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

29 / evaluation

I I think part of it is it's a kind of acquiescence. So I think you're sequestering your agency when you externalize too much of your cognition, particularly if it's parts of your cognition that are useful in the sense that it has core knowledge which would generalize and help you acquire new knowledge or, it's just the the the proto ability of, you know, discovering knowledge.

“I I think part of it is it's a kind of acquiescence. So I think you're sequestering your agency when you externalize too much of your cognition, particularly if it's parts of your cognition that are useful in the sense that it has core knowledge which would generalize and help you acquire new knowledge or, it's just the the the proto ability of, you know, discovering knowledge.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

30 / evaluation

So for example, someone might come up to me and say, blue swirly thing is over there. And I'll say, well, I don't know what you mean about the blue swirly thing because I've I've never seen 1 before.

“So for example, someone might come up to me and say, blue swirly thing is over there. And I'll say, well, I don't know what you mean about the blue swirly thing because I've I've never seen 1 before.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

31 / evaluation

I mean, because the way I read it in active inference literature, it's a very principled definition of an agent. And there's still a bit of a gap because I think Friston would argue, in the natural world, because of the the laws of physics and particles and and whatnot you get the emergence of things and things become agents when they have a certain you know depth of planning shall we say.

“I mean, because the way I read it in active inference literature, it's a very principled definition of an agent. And there's still a bit of a gap because I think Friston would argue, in the natural world, because of the the laws of physics and particles and and whatnot you get the emergence of things and things become agents when they have a certain you know depth of planning shall we say.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

32 / evaluation

We we know from experience, right, that when we have experience, we get better at things. And we have this this weird, I don't know whether it's an illusion that all we need to do is just write down our understanding into into a wiki document.

“We we know from experience, right, that when we have experience, we get better at things. And we have this this weird, I don't know whether it's an illusion that all we need to do is just write down our understanding into into a wiki document.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

33 / evaluation

You you can't seriously tell me that if you simulate fire using a computer program that that the computer would get hot or that you would get hot. It it just doesn't work like that.

“You you can't seriously tell me that if you simulate fire using a computer program that that the computer would get hot or that you would get hot. It it just doesn't work like that.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

34 / evaluation

Knowledge decays very, very quickly. And that's actually a good thing because it's the way that we can adapt our strategies because things that don't work die off.

“Knowledge decays very, very quickly. And that's actually a good thing because it's the way that we can adapt our strategies because things that don't work die off.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

35 / evaluation

And we'll move to the next level of our evolution. And in a sense, I think, even though I don't agree with it, because I think even with Neuralink, there's bandwidth problems.

“And we'll move to the next level of our evolution. And in a sense, I think, even though I don't agree with it, because I think even with Neuralink, there's bandwidth problems.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

36 / evaluation

Yes, and on that point, I think maybe the most exciting thing about your paper is, you know, we were talking about path dependence and having this understanding which is built step by step, this process of complexification.

“Yes, and on that point, I think maybe the most exciting thing about your paper is, you know, we were talking about path dependence and having this understanding which is built step by step, this process of complexification.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

37 / evaluation

I think he would, because he he has a bunch of criteria, but 1 of them is is a fundamental coarse graining and reorganization of the micro substrate such that the new phenomena can be described with with a, you know, with within, you know, simple new variable.

“I think he would, because he he has a bunch of criteria, but 1 of them is is a fundamental coarse graining and reorganization of the micro substrate such that the new phenomena can be described with with a, you know, with within, you know, simple new variable.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

38 / evaluation

You know, when when something is, you know, trivially easy to mechanize, no 1 actually thought it was intelligent. But I think the Turing test is bad because we know that in my opinion, language models aren't actually that intelligent, yet we've passed it with flying colors.

“You know, when when something is, you know, trivially easy to mechanize, no 1 actually thought it was intelligent. But I think the Turing test is bad because we know that in my opinion, language models aren't actually that intelligent, yet we've passed it with flying colors.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

40 / evaluation

The reason for that, as we discuss in today's show, is that current AI does not understand the world in a grounded way. It doesn't have a deep abstract understanding of the world, which is why the only way that we can make AI work effectively is by grounding the generation and supervising the training of AI models with human data.

“The reason for that, as we discuss in today's show, is that current AI does not understand the world in a grounded way. It doesn't have a deep abstract understanding of the world, which is why the only way that we can make AI work effectively is by grounding the generation and supervising the training of AI models with human data.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

42 / evaluation

He's talking about things at a level of abstraction which, you know, can refer to anything, but it's beyond most people's cognitive horizon. But there is something to be said for that when when you can respect the history deep down into the epistemic tree, the creative stepping stones you take, because they respect the history, they actually have more evolvability.

“He's talking about things at a level of abstraction which, you know, can refer to anything, but it's beyond most people's cognitive horizon. But there is something to be said for that when when you can respect the history deep down into the epistemic tree, the creative stepping stones you take, because they respect the history, they actually have more evolvability.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

43 / evaluation

I've never been able to completely pin you down, professor Friston, because there have been so many interpretations of the free energy principle that that lean internalist and externalist and even, the the hybrid version, which Maxwell also wrote a paper about.

“I've never been able to completely pin you down, professor Friston, because there have been so many interpretations of the free energy principle that that lean internalist and externalist and even, the the hybrid version, which Maxwell also wrote a paper about.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

44 / evaluation

We'll just turn up the temperature and we'll just sample tokens from the tail. And and you you really get garbage there because you're kind of, you know, you're you're a little bit out of distribution now.

“We'll just turn up the temperature and we'll just sample tokens from the tail. And and you you really get garbage there because you're kind of, you know, you're you're a little bit out of distribution now.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

45 / evaluation

There is just 1 potential objection, which is that you know what a lot of theories of consciousness do is is they kind of they they brush it to 1 side and they treat it as something which is epiphenomenal, which means that it's not like causally embedded in in the system.

“There is just 1 potential objection, which is that you know what a lot of theories of consciousness do is is they kind of they they brush it to 1 side and they treat it as something which is epiphenomenal, which means that it's not like causally embedded in in the system.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

46 / evaluation

So as as you've said, we use analogies like, electricity. And I think that's quite a good 1, because as as AI becomes enmeshed in into society, imagine how hard it would be to shut down a power station.

“So as as you've said, we use analogies like, electricity. And I think that's quite a good 1, because as as AI becomes enmeshed in into society, imagine how hard it would be to shut down a power station.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

47 / evaluation

But in Genie 2, there was an ST transformer, so a special temporal transformer, which was conceptually quite similar to like a VIT. And there was a latent action model, which means even from non interactive data, you could infer some low cardinality action space.

“But in Genie 2, there was an ST transformer, so a special temporal transformer, which was conceptually quite similar to like a VIT. And there was a latent action model, which means even from non interactive data, you could infer some low cardinality action space.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

49 / evaluation

We are here at Google DeepMind in London, and you guys have just demoed to me something which I think I'm more impressed with this, I think, than anything I've seen probably ever before.

“We are here at Google DeepMind in London, and you guys have just demoed to me something which I think I'm more impressed with this, I think, than anything I've seen probably ever before.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

50 / evaluation

Perhaps in the future, we might have an outer loop, which makes the system more open ended. But right now, my opinion, Genie 3, like all AI, gives you exactly what you asked for in the prompts and isn't creative on its own.

“Perhaps in the future, we might have an outer loop, which makes the system more open ended. But right now, my opinion, Genie 3, like all AI, gives you exactly what you asked for in the prompts and isn't creative on its own.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

52 / evaluation

I've also noticed that this model is is even better than v o 3 at things like text. It's because you you would think that they would they would dumb down the model to make it interactive and to make it this sophisticated.

“I've also noticed that this model is is even better than v o 3 at things like text. It's because you you would think that they would they would dumb down the model to make it interactive and to make it this sophisticated.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

53 / evaluation

You know, it it it feels to me that the missing link is having the correct level of abstraction and being able to do this iterative open ended search. They can't do that because they simply don't have the abstractions.

“You know, it it it feels to me that the missing link is having the correct level of abstraction and being able to do this iterative open ended search. They can't do that because they simply don't have the abstractions.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk

54 / evaluation

These are continuous functions because most of the time in a neural network, like, if you if you give it a test sample, which is outside of the training support, it you're in no man's land.

“These are continuous functions because most of the time in a neural network, like, if you if you give it a test sample, which is outside of the training support, it you're in no man's land.”
Speaker
Tim Scarfe
Publisher
Machine Learning Street Talk
Search evidence