Skip to main content

Xiaochang Li

Communication scholar

Sep 14, 2026

15:06
Yeah.
15:06
Right? The underlying mechanisms, right? That maybe there are certain domains in which the inability to understand the why is actually a problem for us.
15:18
Um, and it's also, like, when we think about these kinds of models, one of the issues that does emerge is that you can't evaluate the model based on how well it represents any kind of underlying phenomenon, right? So in order to evaluate the model, you merely have to test its future predictions against past occurrences, right? That you set aside a b- a bit of your data a- as test data and see how well it can reproduce those past occurrences, right? Um, how well its predictions work.
15:52
And that's a bit of an issue when it then gets put into systems, um, in which you can't, in which, like, you- you're using this information then to make sort of preemptive interventions.
16:08
When we think about things like predictive policing, for instance, right? Um, that we're now using the same kind of thing where we can't know except for how well it predicts, but we use that information to intervene on the thing it's meant to predict.

19 MINS LATER

35:43
Mm
35:43
... um, uh, at AT&T, right? That it's a moment in which these kind of large, um, industrial laboratories, um, in the, starting in, like, the 1920s at GE, at AT&T Bell Laboratories, are, really start to sh- reshape the, the directions of certain scientific fields, um, because there's so much money coming from these areas.
36:11
And we see that reflected really in our present moment with the kind of re-emergence of this kind of large scale industrial research model, right? Be- beyond the kind of commercial manufacturing laboratories, that this is, like, industrial research model in which, um, the goal is to really shape and redefine a field of research that is, you know, relevant or helpful to the corporation's long range goals as opposed to a kind of Uh, research that would be aimed at more immediate commercial demands, right? That you see corporations in- getting in on what's known as, like, you know, basic research, right? Um-
19:40
Or were those kind of taken more as like just sort of inevitable tools to achieve this?
19:45
Yeah, I mean, so with the shift to large-scale data sets, that comes much later.
19:52
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of, well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures all the kinds of knowledge that that would be relevant in coming to a decision.
20:35
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
20:48
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, Using the computer.
21:02
And so in that context, speech recognition was really a kind of proving ground.
21:06
It was a test bed, but it wasn't necessarily like there wasn't really a set of.

7 MINS LATER

28:29
So what are the kind of key steps that you see kind of going back and investigating this history between that mid 90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
19:48
Or were those kind of taken more as like just sort of inevitable tools to achieve this? Yeah, I mean,
19:53
so with the shift to large-scale data sets, that comes much later.
20:00
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of Well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures all the kinds of knowledge that would be relevant to in coming to a decision.
20:42
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
20:55
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, using the computer.
21:09
And so in that context, speech recognition was really a kind of proving ground.
21:13
It was a test bed, but it wasn't necessarily like, there wasn't really a set of clear applications that could be envisioned.

7 MINS LATER

28:36
So what are the kind of key steps that you see kind of going back and investigating this history between that mid 90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
19:40
Or were those kind of taken more as like just sort of inevitable tools to achieve this?
19:45
Yeah, I mean, so with the shift to large-scale data sets, that comes much later.
19:52
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of Well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures all the kinds of knowledge that would be relevant to in coming to a decision.
20:35
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
20:48
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, using the computer.
21:02
And so in that context, speech recognition was really a kind of proving ground.
21:06
It was a test bed, but it wasn't necessarily like, there wasn't really a set of clear applications that could be envisioned.

7 MINS LATER

28:29
So what are the kind of key steps that you see kind of going back and investigating this history between that mid 90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
19:42
Or were those kind of taken more as like just sort of inevitable tools to achieve this?
19:47
Yeah, I mean, so with the shift to large-scale data sets, that comes much later.
19:54
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of, well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures, all the kinds of knowledge that would be relevant to in coming to a decision.
20:37
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
20:50
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, Using the computer.
21:04
And so in that context, speech recognition was really a kind of proving ground.
21:08
It was a test bed, but it wasn't necessarily like there wasn't really a set of clear applications there.

7 MINS LATER

28:31
So what are the kind of key steps that you see kind of going back and investigating this history between that mid 90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
19:53
mean,
19:54
so with the shift to large-scale data sets, that comes much later.
20:00
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of...
20:14
Well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures all the kinds of knowledge that that would be relevant in coming to a decision.
20:43
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
20:55
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, Using the computer.
21:09
And so in that context, speech recognition was really a kind of proving ground.

7 MINS LATER

28:37
So what are the kind of key steps that you see kind of going back and investigating this history between that mid-90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
19:48
Or were those kind of taken more as like just sort of inevitable tools to achieve this?
19:53
Yeah, I mean, so with the shift to large-scale data sets, that comes much later.
20:00
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of, well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures, all the kinds of knowledge that would be relevant to in coming to a decision.
20:43
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
20:56
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, Using the computer.
21:10
And so in that context, speech recognition was really a kind of proving ground.
21:14
It was a test bed, but it wasn't necessarily like there wasn't really a set of clear applications there.

7 MINS LATER

28:37
So what are the kind of key steps that you see kind of going back and investigating this history between that mid 90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
20:47
Or were those kind of taken more as like just sort of inevitable tools to achieve this?
20:52
Yeah, I mean, so with the shift to large-scale data sets, that comes much later.
20:59
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of Well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures all the kinds of knowledge needed.
21:38
that would be relevant in coming to a decision.
21:42
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
21:55
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, Using the computer.
22:09
And so in that context, speech recognition was really a kind of proving ground.

7 MINS LATER

29:36
So what are the kind of key steps that you see kind of going back and investigating this history between that mid 90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
20:40
Or were those kind of taken more as like just sort of inevitable tools to achieve this? Yeah, I mean, so
20:47
with the shift to large-scale data sets, that comes much later.
20:52
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of...
21:06
Well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures all the kinds of knowledge that would be relevant to in coming to a decision.
21:35
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
21:47
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, Using the computer.
22:01
And so in that context, speech recognition was really a kind of proving ground.

7 MINS LATER

29:29
So what are the kind of key steps that you see kind of going back and investigating this history between that mid 90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
16:48
Or were those kind of taken more as like just sort of inevitable tools to achieve this?
16:53
Yeah, I mean, so with the shift to large-scale data sets, that comes much later.
17:00
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of...
17:14
Well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures all the kinds of knowledge that would be relevant to in coming to a decision.
17:42
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
17:55
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, using the computer.
18:09
And so in that context, speech recognition was really a kind of proving ground.

7 MINS LATER

25:36
So what are the kind of key steps that you see kind of going back and investigating this history between that mid-90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.
20:48
Or were those kind of taken more as like just sort of inevitable tools to achieve this?
20:53
Yeah, I mean, so with the shift to large-scale data sets, that comes much later.
21:00
And so for quite a while, for several decades, all through the 60s and well into the 70s, automatic speech recognition research was still following a kind of AI paradigm in the sense of Well, not what we think of AI to be now, but in what it was through this period, which was first kind of modeling of symbolic reasoning, followed by modeling of knowledge and expertise, right? These kinds of rules-based expert systems in which you would try to formalize into a set of rules and procedures all the kinds of knowledge that would be relevant to in coming to a decision.
21:43
And so that was really the dominant mode through much of the 60s and 70s, largely because this is coming out of AI research and all of this AI funding.
21:56
And so the goal of that was to work on questions of artificial intelligence, questions of simulating human faculties, Using the computer.
22:10
And so in that context, speech recognition was really a kind of proving ground.
22:14
It was a test bed, but it wasn't necessarily like there wasn't really a set of clear applications there.

7 MINS LATER

29:37
So what are the kind of key steps that you see kind of going back and investigating this history between that mid 90s moment and more where we're at today with big data and also sort of black box systems too? Yeah.

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.