TRANSKRIPTIO 🔗
1
00:00:00,473 --> 00:00:02,490
(upbeat music)
2
00:00:02,490 --> 00:00:05,910
- Welcome to another episode
of "Supply Chain Frontiers,"
3
00:00:05,910 --> 00:00:09,690
the MIT CTL podcast where
we explore the trends,
4
00:00:09,690 --> 00:00:12,720
technologies, and innovations
shaping the future
5
00:00:12,720 --> 00:00:14,310
of supply chain management.
6
00:00:14,310 --> 00:00:17,337
Today's episode will be hosted
by Dr. Matthias Winkenbach,
7
00:00:17,337 --> 00:00:19,860
director of research at the MIT Center
8
00:00:19,860 --> 00:00:21,750
for Transportation and Logistics.
9
00:00:21,750 --> 00:00:24,480
Today we're diving into the use of AI
10
00:00:24,480 --> 00:00:27,360
and supply chain research
and applications.
11
00:00:27,360 --> 00:00:30,990
Matthias will interview
three MIT CTL researchers:
12
00:00:30,990 --> 00:00:33,150
Willem Guter, a research engineer
13
00:00:33,150 --> 00:00:36,030
at the MIT Intelligent
Logistics Systems Lab
14
00:00:36,030 --> 00:00:39,930
and the MIT Computational
and Visual Education Lab;
15
00:00:39,930 --> 00:00:43,350
Dr. Elenna Dugundji, a research scientist
16
00:00:43,350 --> 00:00:45,300
who leads the MIT Deep Knowledge
17
00:00:45,300 --> 00:00:47,400
for Supply Chain and Logistics Lab;
18
00:00:47,400 --> 00:00:50,490
and Dr. Bryan Reimer,
founder and co-director
19
00:00:50,490 --> 00:00:55,170
of MIT AgeLab's Advanced
Vehicle Technology Consortium.
20
00:00:55,170 --> 00:00:57,570
They will cover AI and
machine learning research
21
00:00:57,570 --> 00:01:00,930
in the context of warehouse
automation and robotics,
22
00:01:00,930 --> 00:01:04,500
demand forecasting, machine
learning and AI in sourcing,
23
00:01:04,500 --> 00:01:06,300
and the procurement function,
24
00:01:06,300 --> 00:01:09,123
and the use of AI in
advanced vehicle technology.
25
00:01:10,920 --> 00:01:12,240
- Thanks for joining us, Willem.
26
00:01:12,240 --> 00:01:13,650
Let's start with something simple.
27
00:01:13,650 --> 00:01:15,090
What's your current research focus?
28
00:01:15,090 --> 00:01:16,230
So what are you working on?
29
00:01:16,230 --> 00:01:19,050
- So in general, I would
say my research focuses
30
00:01:19,050 --> 00:01:22,380
on the intersection of machine learning
31
00:01:22,380 --> 00:01:24,900
and traditional optimization methods
32
00:01:24,900 --> 00:01:27,060
trying to solve business questions.
33
00:01:27,060 --> 00:01:30,360
So that's questions like
how do we route a truck,
34
00:01:30,360 --> 00:01:32,400
what does our fleet look like?
35
00:01:32,400 --> 00:01:34,650
And then also how do
we optimize a warehouse
36
00:01:34,650 --> 00:01:35,640
using these methods.
37
00:01:35,640 --> 00:01:37,370
- So how does AI come
into the picture here?
38
00:01:37,370 --> 00:01:39,367
How are you currently
using machine learning
39
00:01:39,367 --> 00:01:42,000
and AI methods and these
types of questions?
40
00:01:42,000 --> 00:01:44,340
- AI can fit in in lots
of different ways here.
41
00:01:44,340 --> 00:01:46,170
One of the ways I'm most excited about
42
00:01:46,170 --> 00:01:50,520
is sort of being able to
optimize almost in real time
43
00:01:50,520 --> 00:01:51,960
things we couldn't before.
44
00:01:51,960 --> 00:01:53,220
Lemme give an example of that.
45
00:01:53,220 --> 00:01:55,800
We're looking at telling AMRs,
46
00:01:55,800 --> 00:01:58,680
so that's robots in your warehouse,
47
00:01:58,680 --> 00:02:01,050
when to charge in real time.
48
00:02:01,050 --> 00:02:03,870
So before you had to optimize a policy
49
00:02:03,870 --> 00:02:05,280
and they would follow that policy,
50
00:02:05,280 --> 00:02:07,860
which was the perfect
average, you could say.
51
00:02:07,860 --> 00:02:09,840
It's the best most of the time.
52
00:02:09,840 --> 00:02:12,210
Now with machine learning you can optimize
53
00:02:12,210 --> 00:02:15,330
minute by minute and get
the best at every time.
54
00:02:15,330 --> 00:02:16,163
- Okay.
55
00:02:16,163 --> 00:02:18,300
So in a way you're not just
trying to make good decisions
56
00:02:18,300 --> 00:02:20,749
for the entire ensemble of these robots,
57
00:02:20,749 --> 00:02:22,890
but you're basically
trying to make each of them
58
00:02:22,890 --> 00:02:25,050
make smarter decisions individually.
59
00:02:25,050 --> 00:02:25,890
- Exactly.
60
00:02:25,890 --> 00:02:26,723
- Interesting.
61
00:02:26,723 --> 00:02:28,973
So why weren't we able to do this before?
62
00:02:28,973 --> 00:02:32,160
Why are we using machine
learning or AI for this
63
00:02:32,160 --> 00:02:35,760
and why did more traditional
methods fail at that?
64
00:02:35,760 --> 00:02:37,890
- Well, there's a combination of reasons.
65
00:02:37,890 --> 00:02:39,960
So one of those is speed.
66
00:02:39,960 --> 00:02:42,840
You might have heard
about how much compute
67
00:02:42,840 --> 00:02:44,790
it takes to train an AI model,
68
00:02:44,790 --> 00:02:46,230
but what you don't usually hear about
69
00:02:46,230 --> 00:02:47,640
is that after it's trained,
70
00:02:47,640 --> 00:02:49,710
it actually runs extremely quickly.
71
00:02:49,710 --> 00:02:50,820
So traditional methods,
72
00:02:50,820 --> 00:02:53,070
there's not this big upfront load,
73
00:02:53,070 --> 00:02:56,730
but there is a larger time
each time you have to run it
74
00:02:56,730 --> 00:02:59,370
that makes it difficult to do
in this real time scenario.
75
00:02:59,370 --> 00:03:02,340
The other piece is information.
76
00:03:02,340 --> 00:03:06,660
So we have more real-time
information than ever
77
00:03:06,660 --> 00:03:10,950
and AI lets us use all of
that in a way that is useful,
78
00:03:10,950 --> 00:03:13,320
whereas earlier you'd
really have to look at
79
00:03:13,320 --> 00:03:15,300
fine tuning information by hand
80
00:03:15,300 --> 00:03:17,370
and finding what's useful and what's not
81
00:03:17,370 --> 00:03:18,810
and constantly sorting through that.
82
00:03:18,810 --> 00:03:21,180
- And for the models that you
are working with right now,
83
Load full transcript
00:03:21,180 --> 00:03:24,510
how do they first of all
identify what data is out there
84
00:03:24,510 --> 00:03:26,250
and what could be useful?
85
00:03:26,250 --> 00:03:28,020
I mean, how do they distinguish between
86
00:03:28,020 --> 00:03:30,990
what's actually of value versus
what might just be noise?
87
00:03:30,990 --> 00:03:32,250
- So that's a great question
88
00:03:32,250 --> 00:03:34,620
and it's one that a lot of
people would love to answer.
89
00:03:34,620 --> 00:03:36,720
The models I'm working with right now
90
00:03:36,720 --> 00:03:39,270
tend to be what are called
deep learning models,
91
00:03:39,270 --> 00:03:40,980
which are something the mechanics
92
00:03:40,980 --> 00:03:43,620
behind how they learn
aren't fully understood.
93
00:03:43,620 --> 00:03:45,750
We can dive into the math behind
94
00:03:45,750 --> 00:03:47,940
what's called stochastic gradient descent,
95
00:03:47,940 --> 00:03:50,250
but the important piece
to get away is that
96
00:03:50,250 --> 00:03:52,590
they do sort of learn
to ignore this noise,
97
00:03:52,590 --> 00:03:55,680
but they can learn to tune
back into it over time
98
00:03:55,680 --> 00:03:56,880
if you continue training
99
00:03:56,880 --> 00:03:58,650
and if there's valuable information there.
100
00:03:58,650 --> 00:03:59,483
- I see.
101
00:03:59,483 --> 00:04:00,630
You briefly mentioned the use case
102
00:04:00,630 --> 00:04:01,860
that you're currently working on
103
00:04:01,860 --> 00:04:03,750
with those autonomous mobile robots.
104
00:04:03,750 --> 00:04:06,390
So in short, AMRs, can you provide
105
00:04:06,390 --> 00:04:07,740
a little bit more detail on that?
106
00:04:07,740 --> 00:04:09,000
Who are you working with there
107
00:04:09,000 --> 00:04:11,160
and what's the specific
industry application
108
00:04:11,160 --> 00:04:13,020
of what you're currently working on?
109
00:04:13,020 --> 00:04:15,090
- So we're working with
our partner Mecalux.
110
00:04:15,090 --> 00:04:19,860
They build and deploy
these AMRs to their clients
111
00:04:19,860 --> 00:04:22,200
and they're looking at both shrinking
112
00:04:22,200 --> 00:04:24,180
the required fleet size,
113
00:04:24,180 --> 00:04:27,210
so can we do more with fewer AMRs?
114
00:04:27,210 --> 00:04:28,560
And then they're also looking
115
00:04:28,560 --> 00:04:31,080
at optimizing existing fleets.
116
00:04:31,080 --> 00:04:32,940
So can I grow my warehouse
117
00:04:32,940 --> 00:04:34,680
or have higher throughput in my warehouse
118
00:04:34,680 --> 00:04:36,630
with my existing AMR fleet?
119
00:04:36,630 --> 00:04:39,390
- And what do you think is the key metric
120
00:04:39,390 --> 00:04:41,190
that they're actually looking at to assess
121
00:04:41,190 --> 00:04:43,800
whether all the investment into building
122
00:04:43,800 --> 00:04:46,900
a new AI-based approach to this pays off
123
00:04:46,900 --> 00:04:48,364
or differently put,
124
00:04:48,364 --> 00:04:50,370
how do we know whether it's worth it?
125
00:04:50,370 --> 00:04:51,720
- Yeah, that's an interesting question.
126
00:04:51,720 --> 00:04:53,490
I think there are a few metrics
127
00:04:53,490 --> 00:04:56,430
and it does sort of depend
what kind of warehouse
128
00:04:56,430 --> 00:04:58,320
you're running and what you're looking at.
129
00:04:58,320 --> 00:05:01,110
Like you can talk about are
you picking alongside humans,
130
00:05:01,110 --> 00:05:03,180
that's gonna change
your metrics very much,
131
00:05:03,180 --> 00:05:05,340
but I think the main thing
that they're looking at here
132
00:05:05,340 --> 00:05:07,110
is warehouse throughput.
133
00:05:07,110 --> 00:05:11,490
So how much demand can I fulfill
from an existing warehouse
134
00:05:11,490 --> 00:05:13,560
and how much faster can I do that
135
00:05:13,560 --> 00:05:15,210
using these new techniques.
136
00:05:15,210 --> 00:05:18,240
- As you're trying to make
those robots smarter basically.
137
00:05:18,240 --> 00:05:21,210
You mentioned already
decisions around charging
138
00:05:21,210 --> 00:05:22,043
and the like.
139
00:05:22,043 --> 00:05:25,140
What other decisions are
you trying to either replace
140
00:05:25,140 --> 00:05:28,950
or augment or maybe both
with AI-based methods?
141
00:05:28,950 --> 00:05:31,530
- So one of the fun things
about AI-based methods,
142
00:05:31,530 --> 00:05:33,240
and I mentioned this a little bit earlier
143
00:05:33,240 --> 00:05:34,560
when I talked about the speed,
144
00:05:34,560 --> 00:05:37,620
is there is just a
whole range of decisions
145
00:05:37,620 --> 00:05:38,850
that we could improve here.
146
00:05:38,850 --> 00:05:41,700
So some examples of what
we're looking at right now
147
00:05:41,700 --> 00:05:44,700
are things like where
to park, how to drive,
148
00:05:44,700 --> 00:05:46,980
so how to get between point A and point B
149
00:05:46,980 --> 00:05:48,660
as efficiently as possible.
150
00:05:48,660 --> 00:05:52,380
When to drive, so when
to wait, when to go,
151
00:05:52,380 --> 00:05:53,910
when to send another AMR,
152
00:05:53,910 --> 00:05:55,470
even if you're a little further away
153
00:05:55,470 --> 00:05:57,660
to reduce traffic or whatever the reason.
154
00:05:57,660 --> 00:06:00,090
I think in the future we
could dive even further,
155
00:06:00,090 --> 00:06:03,360
whether that's looking at
the actual picking process
156
00:06:03,360 --> 00:06:06,540
or maybe the optimization of the placing
157
00:06:06,540 --> 00:06:09,300
to be put on a truck and
shipped outta the warehouse.
158
00:06:09,300 --> 00:06:11,550
There's really a whole
range of possibilities here.
159
00:06:11,550 --> 00:06:14,253
- And it sounds like what
you're working on with Mecalux
160
00:06:14,253 --> 00:06:16,650
and this particular project
161
00:06:16,650 --> 00:06:20,340
is relatively close to
real world application.
162
00:06:20,340 --> 00:06:22,710
It doesn't sound like this
is gonna become reality
163
00:06:22,710 --> 00:06:23,670
in five, six years from now,
164
00:06:23,670 --> 00:06:25,650
but potentially much sooner than that.
165
00:06:25,650 --> 00:06:28,140
Thinking about a little bit further ahead,
166
00:06:28,140 --> 00:06:29,580
what kind of breakthroughs
167
00:06:29,580 --> 00:06:32,850
do you think AI might make possible
168
00:06:32,850 --> 00:06:35,160
or what kind of
breakthroughs are you hoping
169
00:06:35,160 --> 00:06:36,870
it'll make possible at some point?
170
00:06:36,870 --> 00:06:38,880
- To me it all comes back to that
171
00:06:38,880 --> 00:06:40,710
every decision is thought out.
172
00:06:40,710 --> 00:06:43,440
So I've been talking about
that for the warehouse,
173
00:06:43,440 --> 00:06:45,090
but you can try to imagine
174
00:06:45,090 --> 00:06:46,770
what that would look like in the rest
175
00:06:46,770 --> 00:06:49,500
of your whole business operations,
176
00:06:49,500 --> 00:06:54,090
if every single action was
taken not to an optimal policy,
177
00:06:54,090 --> 00:06:59,090
but to a very close AI
approximation of what is optimal.
178
00:06:59,820 --> 00:07:02,910
We could see the whole
supply chain sped up
179
00:07:02,910 --> 00:07:05,280
to I think a surprising degree.
180
00:07:05,280 --> 00:07:08,100
- And do you see the
potential of more and more
181
00:07:08,100 --> 00:07:10,620
of these currently
somewhat isolated systems
182
00:07:10,620 --> 00:07:11,730
to become integrated?
183
00:07:11,730 --> 00:07:13,590
So for instance, right
now you're focusing on
184
00:07:13,590 --> 00:07:15,840
how to make warehouse robots smarter.
185
00:07:15,840 --> 00:07:17,940
I know that a few other
folks are looking into
186
00:07:17,940 --> 00:07:20,760
how to use AI for things
like inventory management
187
00:07:20,760 --> 00:07:21,593
and the like.
188
00:07:21,593 --> 00:07:23,220
Do you think we'll ever get to a point
189
00:07:23,220 --> 00:07:27,810
where those models will
become one big nexus of models
190
00:07:27,810 --> 00:07:29,040
that work together?
191
00:07:29,040 --> 00:07:30,000
- Absolutely.
192
00:07:30,000 --> 00:07:31,650
I've already seen in my research
193
00:07:31,650 --> 00:07:33,300
that some models I developed
194
00:07:33,300 --> 00:07:37,950
for one piece of warehouse
optimization, say when to drive,
195
00:07:37,950 --> 00:07:41,090
are surprisingly applicable
to other parts, say parking.
196
00:07:41,090 --> 00:07:45,720
And I think that sort of
cross model working together
197
00:07:45,720 --> 00:07:48,240
is something that we're gonna
see expand more and more
198
00:07:48,240 --> 00:07:50,220
as well as, like I mentioned earlier,
199
00:07:50,220 --> 00:07:52,260
the information they can integrate
200
00:07:52,260 --> 00:07:54,210
is going to expand as well.
201
00:07:54,210 --> 00:07:57,240
And so you're gonna be
seeing not only information
202
00:07:57,240 --> 00:07:58,413
about your supply chain,
203
00:07:58,413 --> 00:08:01,110
but your models are gonna
be integrating information
204
00:08:01,110 --> 00:08:03,810
about your
suppliers, your customers,
205
00:08:03,810 --> 00:08:06,657
and using all of that to make
the best decision for today.
206
00:08:06,657 --> 00:08:09,630
- And so the work that
you're currently focused on
207
00:08:09,630 --> 00:08:11,040
is obviously targeted
208
00:08:11,040 --> 00:08:14,100
at a relatively narrowly defined use case.
209
00:08:14,100 --> 00:08:17,970
How transferable is the knowledge,
210
00:08:17,970 --> 00:08:20,610
but also the technology that
you and the team are developing
211
00:08:20,610 --> 00:08:22,710
there to other contexts
212
00:08:22,710 --> 00:08:25,230
that might even fall
outside of the four walls
213
00:08:25,230 --> 00:08:26,063
of a warehouse?
214
00:08:26,063 --> 00:08:29,160
So basically what do we learn
from these types of projects
215
00:08:29,160 --> 00:08:31,710
that can be generalized to
other types of problems?
216
00:08:31,710 --> 00:08:33,990
- Yeah, some of it absolutely is specific
217
00:08:33,990 --> 00:08:35,040
to within the four walls.
218
00:08:35,040 --> 00:08:37,710
So when you're thinking
about things like parking,
219
00:08:37,710 --> 00:08:40,140
that is probably a warehouse problem,
220
00:08:40,140 --> 00:08:41,760
but at the same time you can think of,
221
00:08:41,760 --> 00:08:44,550
well, I'm parking my
trucks outside my warehouse
222
00:08:44,550 --> 00:08:46,140
at a certain level of abstraction,
223
00:08:46,140 --> 00:08:48,420
I'm parking my goods in different places.
224
00:08:48,420 --> 00:08:51,510
So I think a lot of
these overarching models
225
00:08:51,510 --> 00:08:54,090
really can transfer well
to different things.
226
00:08:54,090 --> 00:08:55,860
- And so far we've talked a lot about
227
00:08:55,860 --> 00:08:58,920
the positive outlook on
AI and machine learning,
228
00:08:58,920 --> 00:09:00,660
the great potential that it bears,
229
00:09:00,660 --> 00:09:02,430
but I'm sure there must be,
230
00:09:02,430 --> 00:09:04,650
at least currently there
must be a lot of limitations
231
00:09:04,650 --> 00:09:06,570
as well that keep you from doing
232
00:09:06,570 --> 00:09:08,550
what you would probably love to do.
233
00:09:08,550 --> 00:09:12,214
So for a non-technical audience,
234
00:09:12,214 --> 00:09:15,390
how would you describe the
most critical limitations
235
00:09:15,390 --> 00:09:16,857
of that technology today
236
00:09:16,857 --> 00:09:19,290
and what are you guys
doing to overcome those?
237
00:09:19,290 --> 00:09:22,110
- So to me there are two main limitations
238
00:09:22,110 --> 00:09:23,880
when you're working
with this AI technology.
239
00:09:23,880 --> 00:09:26,100
The first one is the easiest to overcome
240
00:09:26,100 --> 00:09:27,540
and that's data availability.
241
00:09:27,540 --> 00:09:30,600
So you typically need
huge amounts of data.
242
00:09:30,600 --> 00:09:33,630
We're looking at months
if not years of data
243
00:09:33,630 --> 00:09:35,070
to train these systems.
244
00:09:35,070 --> 00:09:37,440
However, there are some
cutting edge techniques
245
00:09:37,440 --> 00:09:39,330
using things like transfer learning.
246
00:09:39,330 --> 00:09:41,100
So where you train it on one set of data
247
00:09:41,100 --> 00:09:43,380
and then you only need a
little bit to transfer it.
248
00:09:43,380 --> 00:09:45,390
Or looking at things like simulation
249
00:09:45,390 --> 00:09:48,600
where maybe you only have
a very small amount of data
250
00:09:48,600 --> 00:09:52,168
but you can simulate whole
months or years of data virtually
251
00:09:52,168 --> 00:09:53,940
and gather it that way.
252
00:09:53,940 --> 00:09:56,520
What I foresee is really the bigger issue
253
00:09:56,520 --> 00:09:58,710
is the hallucination problem.
254
00:09:58,710 --> 00:10:01,680
If you spend enough
time talking to ChatGPT,
255
00:10:01,680 --> 00:10:04,530
I'm sure at some point or
another it's hallucinated to you.
256
00:10:04,530 --> 00:10:06,360
It's made up a fact that's not true
257
00:10:06,360 --> 00:10:08,760
or said something that you never told it
258
00:10:08,760 --> 00:10:10,470
and doesn't check out in the real world,
259
00:10:10,470 --> 00:10:13,200
but this is a problem
that can find its way
260
00:10:13,200 --> 00:10:15,270
into other AI systems.
261
00:10:15,270 --> 00:10:17,760
So when they're making what they think
262
00:10:17,760 --> 00:10:20,340
is the perfect decision for the moment,
263
00:10:20,340 --> 00:10:22,110
sometimes it's really not
264
00:10:22,110 --> 00:10:25,980
and it's very hard to verify
that they will always do that.
265
00:10:25,980 --> 00:10:28,410
So I think that's probably
the biggest roadblock
266
00:10:28,410 --> 00:10:29,790
to deploying these in the real world
267
00:10:29,790 --> 00:10:30,975
is this verification
268
00:10:30,975 --> 00:10:33,840
and checking that they're
doing things that make sense.
269
00:10:33,840 --> 00:10:36,300
- That's probably also a major limitation
270
00:10:36,300 --> 00:10:38,130
that at least partially might explain
271
00:10:38,130 --> 00:10:40,560
why we see much more autonomous systems
272
00:10:40,560 --> 00:10:42,510
already being deployed within
273
00:10:42,510 --> 00:10:44,190
let's say the four walls of a warehouse
274
00:10:44,190 --> 00:10:46,170
or a factory than outside of it.
275
00:10:46,170 --> 00:10:47,760
Even though the public discussion
276
00:10:47,760 --> 00:10:50,010
has been very much about
autonomous driving,
277
00:10:50,010 --> 00:10:51,780
delivery robots, delivery drones,
278
00:10:51,780 --> 00:10:54,060
but none of that or very little of that
279
00:10:54,060 --> 00:10:55,440
has become a reality yet
280
00:10:55,440 --> 00:10:56,700
while the types of robots
281
00:10:56,700 --> 00:10:57,533
that you're currently working on,
282
00:10:57,533 --> 00:11:00,900
they exist and they work on
a daily basis already, right?
283
00:11:00,900 --> 00:11:01,890
- Yeah, absolutely.
284
00:11:01,890 --> 00:11:05,070
Inside a warehouse it's
relatively easy to say,
285
00:11:05,070 --> 00:11:06,600
hey, if you get in a situation
286
00:11:06,600 --> 00:11:08,250
where things look wrong
287
00:11:08,250 --> 00:11:10,620
or you're trying to drive
too fast, just stop.
288
00:11:10,620 --> 00:11:12,150
If you try to do that on the road,
289
00:11:12,150 --> 00:11:13,590
people tend to get a little angry.
290
00:11:13,590 --> 00:11:15,270
- Yeah, understandably,
291
00:11:15,270 --> 00:11:16,260
this has been super interesting.
292
00:11:16,260 --> 00:11:17,100
Is there anything else
293
00:11:17,100 --> 00:11:19,530
that you think the audience should know
294
00:11:19,530 --> 00:11:21,300
about the great work
that you guys are doing
295
00:11:21,300 --> 00:11:22,350
before we closing?
296
00:11:22,350 --> 00:11:23,700
- I think you should just follow it
297
00:11:23,700 --> 00:11:24,540
and see where we're going.
298
00:11:24,540 --> 00:11:25,860
You know, there's exciting stuff.
299
00:11:25,860 --> 00:11:27,540
Like I said, we're
looking at all these sorts
300
00:11:27,540 --> 00:11:29,160
of decisions that right now
301
00:11:29,160 --> 00:11:30,510
you're not even thinking about
302
00:11:30,510 --> 00:11:32,340
as something you can explore.
303
00:11:32,340 --> 00:11:34,350
- Alright, that's a positive note
304
00:11:34,350 --> 00:11:35,460
to end this interview on.
305
00:11:35,460 --> 00:11:36,293
Thank you Willem.
306
00:11:36,293 --> 00:11:38,055
This has been very insightful.
307
00:11:38,055 --> 00:11:40,020
(upbeat music)
308
00:11:40,020 --> 00:11:43,020
- This episode of "Supply Chain
Frontiers" is brought to you
309
00:11:43,020 --> 00:11:46,110
by the MIT Intelligent
Logistics Systems Lab,
310
00:11:46,110 --> 00:11:47,550
an MIT CTL lab,
311
00:11:47,550 --> 00:11:49,710
which is a research initiative designed
312
00:11:49,710 --> 00:11:52,140
to revolutionize logistics operations
313
00:11:52,140 --> 00:11:53,700
through cutting edge research
314
00:11:53,700 --> 00:11:55,980
at the intersection of
operations research,
315
00:11:55,980 --> 00:11:59,250
artificial intelligence and
machine learning technologies.
316
00:11:59,250 --> 00:12:00,660
This lab is made possible
317
00:12:00,660 --> 00:12:02,610
by foundational support and collaboration
318
00:12:02,610 --> 00:12:06,180
with Mecalux, a global
leader in intralogistics.
319
00:12:06,180 --> 00:12:09,813
Learn more at intelligent.mit.edu.
320
00:12:10,801 --> 00:12:14,010
(upbeat music continues)
321
00:12:14,010 --> 00:12:15,600
- Thanks for joining us, Elenna.
322
00:12:15,600 --> 00:12:16,890
It's a pleasure to have you.
323
00:12:16,890 --> 00:12:19,080
Let's start with a simple question.
324
00:12:19,080 --> 00:12:21,150
Can you describe your
current research focus?
325
00:12:21,150 --> 00:12:23,100
What are you and your team
working on these days?
326
00:12:23,100 --> 00:12:24,090
- Oh, thank you Matthias.
327
00:12:24,090 --> 00:12:25,380
I'm delighted to be here.
328
00:12:25,380 --> 00:12:27,600
My work with the Deep Knowledge Lab
329
00:12:27,600 --> 00:12:29,280
for Supply Chain and Logistics,
330
00:12:29,280 --> 00:12:30,600
it focuses primarily
331
00:12:30,600 --> 00:12:34,680
on deep learning
applications in global trade,
332
00:12:34,680 --> 00:12:38,850
and the global trade has
two main pillars of work.
333
00:12:38,850 --> 00:12:41,520
One is the actual movement of goods
334
00:12:41,520 --> 00:12:45,480
through airports, for
example, special air cargo
335
00:12:45,480 --> 00:12:48,570
and the other is through
maritime terminals
336
00:12:48,570 --> 00:12:51,048
and the maritime terminal logistics,
337
00:12:51,048 --> 00:12:53,340
yard logistics, et cetera.
338
00:12:53,340 --> 00:12:56,850
Also ocean congestion
arriving at the terminals.
339
00:12:56,850 --> 00:12:58,260
That's one pillar.
340
00:12:58,260 --> 00:12:59,700
The other pillar of work
341
00:12:59,700 --> 00:13:02,790
is related to sourcing, strategic sourcing
342
00:13:02,790 --> 00:13:06,210
and the procurement function
and spend analytics.
343
00:13:06,210 --> 00:13:08,187
- So how are you currently using AI
344
00:13:08,187 --> 00:13:10,650
and machine learning in
that particular space?
345
00:13:10,650 --> 00:13:12,480
- So AI and machine learning,
346
00:13:12,480 --> 00:13:14,490
we started this trajectory
347
00:13:14,490 --> 00:13:17,700
in the procurement function
in spend analytics.
348
00:13:17,700 --> 00:13:19,650
That was an interesting challenge
349
00:13:19,650 --> 00:13:21,090
presented to us
350
00:13:21,090 --> 00:13:25,230
where the company had a very large volume
351
00:13:25,230 --> 00:13:28,894
of uncategorized spend every quarter
352
00:13:28,894 --> 00:13:32,970
because of new vendors onboarding.
353
00:13:32,970 --> 00:13:36,120
In a company where you have
maybe 50,000 suppliers,
354
00:13:36,120 --> 00:13:37,380
you can imagine
355
00:13:37,380 --> 00:13:41,670
that there are just routinely
a large number of SKUs,
356
00:13:41,670 --> 00:13:45,150
stock keeping units, each quarter
357
00:13:45,150 --> 00:13:47,010
which need to be categorized.
358
00:13:47,010 --> 00:13:49,560
And if you are not categorizing them,
359
00:13:49,560 --> 00:13:52,260
then you don't have a good
overview on your spend.
360
00:13:52,260 --> 00:13:55,320
So this work had been previously done
361
00:13:55,320 --> 00:13:57,990
by the company, by a team of people
362
00:13:57,990 --> 00:14:00,120
going through all these line items
363
00:14:00,120 --> 00:14:02,580
and manually categorizing them.
364
00:14:02,580 --> 00:14:04,530
Then this thought of,
365
00:14:04,530 --> 00:14:06,450
well there should be a better way.
366
00:14:06,450 --> 00:14:08,250
They had outsourced it
367
00:14:08,250 --> 00:14:10,920
to a consultancy company
368
00:14:10,920 --> 00:14:13,920
that made a rule-based model,
369
00:14:13,920 --> 00:14:16,590
but then that rule-based
model quickly broke
370
00:14:16,590 --> 00:14:19,440
after two or three quarters.
371
00:14:19,440 --> 00:14:20,700
So they were looking for something
372
00:14:20,700 --> 00:14:22,260
that could naturally learn.
373
00:14:22,260 --> 00:14:26,220
So in this case, we applied
a machine learning model.
374
00:14:26,220 --> 00:14:28,350
We tested several different functions
375
00:14:28,350 --> 00:14:30,720
that could do classification
376
00:14:30,720 --> 00:14:32,730
and continuously learn
377
00:14:32,730 --> 00:14:36,060
so that every time there
becomes a new description
378
00:14:36,060 --> 00:14:38,730
you can properly put it
in the right category.
379
00:14:38,730 --> 00:14:40,770
So this was a huge win
380
00:14:40,770 --> 00:14:42,780
because now the people
381
00:14:42,780 --> 00:14:45,780
who were previously doing
this very tedious work
382
00:14:45,780 --> 00:14:49,140
were freed up to do more
important strategic work
383
00:14:49,140 --> 00:14:50,520
for the company
384
00:14:50,520 --> 00:14:54,180
and also made their life a lot easier.
385
00:14:54,180 --> 00:14:56,190
That work was work we started
386
00:14:56,190 --> 00:15:00,540
before generative AI really
exploded in the world.
387
00:15:00,540 --> 00:15:02,040
That was work we had done
388
00:15:02,040 --> 00:15:04,290
with a classical machine learning
389
00:15:04,290 --> 00:15:05,790
with natural language processing
390
00:15:05,790 --> 00:15:08,070
and tokenization of the text.
391
00:15:08,070 --> 00:15:11,220
And then we subsequently
explored with a company
392
00:15:11,220 --> 00:15:14,670
what you could do using generative AI
393
00:15:14,670 --> 00:15:16,830
and reading text.
394
00:15:16,830 --> 00:15:19,890
And we did a couple of cases
395
00:15:19,890 --> 00:15:23,040
and you might have heard
of something called RAG,
396
00:15:23,040 --> 00:15:25,950
Retrieval-Augmented Generation.
397
00:15:25,950 --> 00:15:30,950
This RAG is a way to use
a large language model
398
00:15:31,170 --> 00:15:33,990
to convert your language
399
00:15:33,990 --> 00:15:37,920
into specific queries on your own data.
400
00:15:37,920 --> 00:15:41,250
So you are using a trained model
401
00:15:41,250 --> 00:15:43,140
that understands human language,
402
00:15:43,140 --> 00:15:46,710
but you are doing the
query on your own data.
403
00:15:46,710 --> 00:15:48,390
So this allows you
404
00:15:48,390 --> 00:15:53,100
to not expose your queries
to the outside world.
405
00:15:53,100 --> 00:15:55,650
If you're using OpenAI for example,
406
00:15:55,650 --> 00:15:59,310
or any of a number of other
large language models,
407
00:15:59,310 --> 00:16:03,360
you can run this behind
your own company firewall
408
00:16:03,360 --> 00:16:04,560
in a secure way.
409
00:16:04,560 --> 00:16:09,300
And we developed over
the past couple years
410
00:16:09,300 --> 00:16:11,670
a number of different examples
411
00:16:11,670 --> 00:16:16,050
where in one case we use RAG for text
412
00:16:16,050 --> 00:16:20,449
and another case we use
RAG for SQL queries.
413
00:16:20,449 --> 00:16:24,990
An example of SQL queries is,
414
00:16:24,990 --> 00:16:28,830
for example, you want to know how
much
415
00:16:28,830 --> 00:16:32,520
if you're preparing for a
negotiation with a vendor
416
00:16:32,520 --> 00:16:35,760
and instead of having to
click through dashboards,
417
00:16:35,760 --> 00:16:37,920
you can simply ask the question
418
00:16:37,920 --> 00:16:40,170
and get the answer right back.
419
00:16:40,170 --> 00:16:42,870
Now dashboards are already very helpful,
420
00:16:42,870 --> 00:16:45,450
but if you have a complicated query,
421
00:16:45,450 --> 00:16:47,970
sometimes that dashboard is not giving
422
00:16:47,970 --> 00:16:49,620
exactly what you want
423
00:16:49,620 --> 00:16:51,300
and it can be very frustrating
424
00:16:51,300 --> 00:16:54,360
and particularly if you're
under high time pressure
425
00:16:54,360 --> 00:16:57,600
to negotiate, it's a huge win for people
426
00:16:57,600 --> 00:17:00,394
whose main task is negotiating
427
00:17:00,394 --> 00:17:03,000
not to have to get bogged down
428
00:17:03,000 --> 00:17:04,890
with clicking through dashboards.
429
00:17:04,890 --> 00:17:05,723
- I'm just curious
430
00:17:05,723 --> 00:17:07,620
for not so technically versed audience
431
00:17:07,620 --> 00:17:09,660
if you had to put it in simple terms.
432
00:17:09,660 --> 00:17:11,550
You mentioned earlier
that existing methods,
433
00:17:11,550 --> 00:17:13,200
like for instance, rule-based methods
434
00:17:13,200 --> 00:17:17,190
for the SKU classification
task quickly broke.
435
00:17:17,190 --> 00:17:19,530
So I think you've already
answered my next question,
436
00:17:19,530 --> 00:17:21,690
which is why are you
using machine learning
437
00:17:21,690 --> 00:17:22,980
and AI in the first place
438
00:17:22,980 --> 00:17:25,890
and kind of what makes
these types of methods
439
00:17:25,890 --> 00:17:28,890
more robust or more
capable for these tasks?
440
00:17:28,890 --> 00:17:31,590
Simply speaking, why don't
they break so quickly?
441
00:17:31,590 --> 00:17:33,270
- This is an excellent question.
442
00:17:33,270 --> 00:17:36,090
So this has to do with the fact
443
00:17:36,090 --> 00:17:39,543
that machine learning is
inherently stochastic.
444
00:17:40,590 --> 00:17:42,240
In the previous example,
445
00:17:42,240 --> 00:17:43,710
what had been done at the company
446
00:17:43,710 --> 00:17:46,500
was a deterministic rule-based system.
447
00:17:46,500 --> 00:17:50,130
So it said if this, then
that, if this, then that.
448
00:17:50,130 --> 00:17:53,266
And there was no leeway
449
00:17:53,266 --> 00:17:57,600
to have something specifically different.
450
00:17:57,600 --> 00:18:00,480
So if you followed if this then that
451
00:18:00,480 --> 00:18:03,584
and you ended up somewhere
it wasn't exactly
452
00:18:03,584 --> 00:18:08,104
what you wanted, it
could either just abort
453
00:18:08,104 --> 00:18:13,104
and say don't know, or it could
give you something spurious.
454
00:18:13,710 --> 00:18:18,180
So having a stochastic model allows you
455
00:18:18,180 --> 00:18:22,470
to have more flexibility to make splits
456
00:18:22,470 --> 00:18:25,620
and also allows you to learn
457
00:18:25,620 --> 00:18:29,580
as you're having new data every quarter.
458
00:18:29,580 --> 00:18:31,440
So that's a very important part
459
00:18:31,440 --> 00:18:33,540
of the machine learning itself.
460
00:18:33,540 --> 00:18:37,200
Now for the generative problem,
461
00:18:37,200 --> 00:18:40,050
this is also a very interesting
462
00:18:40,050 --> 00:18:45,050
that the generative processing
of data also allows you
463
00:18:45,930 --> 00:18:50,820
to ask if, for example, in
the past we did have chatbots,
464
00:18:50,820 --> 00:18:53,490
but I don't know how many people out there
465
00:18:53,490 --> 00:18:55,770
might have had the frustrating experience
466
00:18:55,770 --> 00:18:57,420
where you're asking a question
467
00:18:57,420 --> 00:18:59,190
and the chatbot doesn't answer
468
00:18:59,190 --> 00:19:02,970
and you try to ask it again
and you try to ask it again
469
00:19:02,970 --> 00:19:05,430
and you ask it in a different
way and a different way
470
00:19:05,430 --> 00:19:07,590
and the chatbot is just not understanding
471
00:19:07,590 --> 00:19:08,610
what you're trying to say
472
00:19:08,610 --> 00:19:11,490
and finally you get frustrated and say,
473
00:19:11,490 --> 00:19:14,400
oh, this chatbot is stupid, or you hang up
474
00:19:14,400 --> 00:19:15,660
if you're in a call.
475
00:19:15,660 --> 00:19:19,500
But again, the stochasticity
in the generative approach
476
00:19:19,500 --> 00:19:23,460
allows you to custom tailor questions.
477
00:19:23,460 --> 00:19:27,750
So you don't need to
have specifically seen
478
00:19:27,750 --> 00:19:30,270
the exact question before.
479
00:19:30,270 --> 00:19:33,150
The model understands,
if you've seen this,
480
00:19:33,150 --> 00:19:37,080
it's probably a 90% or 95% probability,
481
00:19:37,080 --> 00:19:40,560
this is what you mean even if
you haven't seen it before.
482
00:19:40,560 --> 00:19:44,970
So that same principle in the
classical machine learning
483
00:19:44,970 --> 00:19:46,740
is applying in this chatbot
484
00:19:46,740 --> 00:19:49,470
that you can flexibly answer questions
485
00:19:49,470 --> 00:19:52,320
that the chatbot has not seen before
486
00:19:52,320 --> 00:19:54,390
because of the stochasticity.
487
00:19:54,390 --> 00:19:56,160
It's extremely powerful.
488
00:19:56,160 --> 00:19:59,190
- And that's already a good
segue to what's next, right?
489
00:19:59,190 --> 00:20:01,020
Because I mean you've
talked about what you guys
490
00:20:01,020 --> 00:20:04,200
are currently working on with
a bunch of research partners,
491
00:20:04,200 --> 00:20:06,180
but if you look a few years down the line,
492
00:20:06,180 --> 00:20:08,010
what do you think or what do you hope
493
00:20:08,010 --> 00:20:10,947
are the breakthroughs
that AI will make possible
494
00:20:10,947 --> 00:20:13,410
in the let's say medium term future?
495
00:20:13,410 --> 00:20:15,870
- Oh, it's developing so fast.
496
00:20:15,870 --> 00:20:17,670
Difficult question to answer.
497
00:20:17,670 --> 00:20:21,360
So you asked what breakthroughs
AI makes possible.
498
00:20:21,360 --> 00:20:23,730
So in the generative case,
499
00:20:23,730 --> 00:20:27,900
this makes possible to
be able to ask questions
500
00:20:27,900 --> 00:20:32,280
on the fly and being able
to get answers on the fly.
501
00:20:32,280 --> 00:20:35,280
Now we talked about an example
502
00:20:35,280 --> 00:20:40,280
of using asking a question
where you are getting an answer,
503
00:20:40,650 --> 00:20:44,460
your question is
translated into a SQL query
504
00:20:44,460 --> 00:20:47,550
and that SQL query acts on a database.
505
00:20:47,550 --> 00:20:51,120
Now this is something we did
a couple years ago already,
506
00:20:51,120 --> 00:20:54,690
but you can extend this in a
number of interesting ways.
507
00:20:54,690 --> 00:20:59,430
One is instead of asking a
question to a SQL database,
508
00:20:59,430 --> 00:21:02,460
you can ask a question
to a graph database.
509
00:21:02,460 --> 00:21:05,580
And this is something I find very exciting
510
00:21:05,580 --> 00:21:08,160
because this allows you to connect
511
00:21:08,160 --> 00:21:12,510
very disparate systems in
the procurement function.
512
00:21:12,510 --> 00:21:16,530
So for example, you have
a case in resilience,
513
00:21:16,530 --> 00:21:20,520
you have a case where a
supplier is not available,
514
00:21:20,520 --> 00:21:23,700
that supplier might have
gotten hit by a flood
515
00:21:23,700 --> 00:21:26,760
or a hurricane or a fire
516
00:21:26,760 --> 00:21:30,660
or there's some geopolitical
risk or a bankruptcy
517
00:21:30,660 --> 00:21:33,330
or any number of reasons a supplier
518
00:21:33,330 --> 00:21:34,980
if you have 50,000 of them
519
00:21:34,980 --> 00:21:38,370
or even if you have
only a handful of them,
520
00:21:38,370 --> 00:21:41,610
if that supplier is not
available, it's a problem.
521
00:21:41,610 --> 00:21:45,900
And if you have a large number of SKUs,
522
00:21:45,900 --> 00:21:48,270
then you might want to understand
523
00:21:48,270 --> 00:21:51,349
how a supplier is affecting
524
00:21:51,349 --> 00:21:53,820
how many different finished goods.
525
00:21:53,820 --> 00:21:57,150
Now typically the way
this data is structured,
526
00:21:57,150 --> 00:21:59,280
you're having a finished good
527
00:21:59,280 --> 00:22:02,220
and on your bill of materials you see
528
00:22:02,220 --> 00:22:04,920
this finished good is
made of this component,
529
00:22:04,920 --> 00:22:07,290
this component which is
made of this raw material
530
00:22:07,290 --> 00:22:08,760
and that raw material
531
00:22:08,760 --> 00:22:12,450
and you can trace it from
the finished good down
532
00:22:12,450 --> 00:22:16,230
to different tiers of suppliers.
533
00:22:16,230 --> 00:22:18,540
If you're not creating a graph,
534
00:22:18,540 --> 00:22:22,650
it's very difficult to
understand the full impact
535
00:22:22,650 --> 00:22:24,870
of one supplier going out,
536
00:22:24,870 --> 00:22:27,390
how it's impacting maybe two or three
537
00:22:27,390 --> 00:22:29,040
different finished goods.
538
00:22:29,040 --> 00:22:32,760
So by making this graph you
can see that in an instant
539
00:22:32,760 --> 00:22:36,390
and you can also query
it using the same methods
540
00:22:36,390 --> 00:22:37,920
I was just describing before.
541
00:22:37,920 --> 00:22:41,340
You can also connect it not
only your bill of materials
542
00:22:41,340 --> 00:22:42,750
to your supplier database,
543
00:22:42,750 --> 00:22:45,600
you can also connect it
to your spend database
544
00:22:45,600 --> 00:22:49,020
so you can immediately see
this supplier goes out,
545
00:22:49,020 --> 00:22:51,660
these are the finished
goods that are affected
546
00:22:51,660 --> 00:22:55,380
and this is how much
total spend is affected.
547
00:22:55,380 --> 00:22:59,370
And if you are managing a
large complex supply chain,
548
00:22:59,370 --> 00:23:03,381
that is really a huge
win because right now
549
00:23:03,381 --> 00:23:06,240
it just by looking only
at the bill of materials,
550
00:23:06,240 --> 00:23:09,960
it's a lot of work to go manually
551
00:23:09,960 --> 00:23:13,080
and have to check everything
your supplier database
552
00:23:13,080 --> 00:23:14,670
and your spend database.
553
00:23:14,670 --> 00:23:17,730
When you can connect all this
together in a large graph
554
00:23:17,730 --> 00:23:20,730
and query the graph,
it's extremely powerful.
555
00:23:20,730 --> 00:23:21,870
- That sounds fascinating.
556
00:23:21,870 --> 00:23:23,970
I imagine having a system like this
557
00:23:23,970 --> 00:23:24,990
would've been very useful
558
00:23:24,990 --> 00:23:28,830
for instance during COVID when
things suddenly went missing
559
00:23:28,830 --> 00:23:30,417
and nobody really knew why
560
00:23:30,417 --> 00:23:32,850
and the network effects that
you were just describing
561
00:23:32,850 --> 00:23:34,200
were probably the reason why
562
00:23:34,200 --> 00:23:36,240
because there were no existing methods
563
00:23:36,240 --> 00:23:38,280
that were able to capture
these dependencies.
564
00:23:38,280 --> 00:23:39,270
I suppose?
565
00:23:39,270 --> 00:23:40,290
- That's correct,
566
00:23:40,290 --> 00:23:44,130
but what you're saying now
actually touches on another
567
00:23:44,130 --> 00:23:46,530
part of my research in the global trade,
568
00:23:46,530 --> 00:23:50,730
which has to do with the
movements of air cargo
569
00:23:50,730 --> 00:23:54,810
and ocean freight and how
that picture connects.
570
00:23:54,810 --> 00:23:56,250
It's a little bit of a segue,
571
00:23:56,250 --> 00:24:00,660
but we are also using machine
learning in that area as well.
572
00:24:00,660 --> 00:24:05,660
In particular, once a purchase
order is issued in say Asia,
573
00:24:06,210 --> 00:24:11,210
and that product is then
shipped on an ocean liner
574
00:24:11,264 --> 00:24:14,010
and then it comes at the ocean liner
575
00:24:14,010 --> 00:24:16,290
and then you have a terminal logistics
576
00:24:16,290 --> 00:24:17,970
that need to be processed
577
00:24:17,970 --> 00:24:21,360
and then from there it
maybe goes onwards by road
578
00:24:21,360 --> 00:24:22,530
or by rail.
579
00:24:22,530 --> 00:24:27,360
Now in all steps of this
process, you also require,
580
00:24:27,360 --> 00:24:32,070
if it's going onward by road,
you need chassis available.
581
00:24:32,070 --> 00:24:35,190
And if you're going onward by rail,
582
00:24:35,190 --> 00:24:37,500
you need intermodal equipment available,
583
00:24:37,500 --> 00:24:39,450
you need rail cars available.
584
00:24:39,450 --> 00:24:42,360
So this is related to
another stream of work,
585
00:24:42,360 --> 00:24:43,560
which we're doing,
586
00:24:43,560 --> 00:24:47,490
is understanding the
complete supply chain,
587
00:24:47,490 --> 00:24:50,490
starting from the beneficiary cargo owner
588
00:24:50,490 --> 00:24:53,970
to the ocean carrier to the terminal,
589
00:24:53,970 --> 00:24:58,710
to the equipment providers
and motor carriers
590
00:24:58,710 --> 00:25:02,820
and rail carriers and
also warehousing space.
591
00:25:02,820 --> 00:25:07,200
In COVID, what largely happened
was people are wondering
592
00:25:07,200 --> 00:25:10,710
why is a particular good
that I want not on the shelf.
593
00:25:10,710 --> 00:25:14,700
It's a whole breakdown if you have ships
594
00:25:14,700 --> 00:25:18,510
that can't get to shore,
if you have chassis
595
00:25:18,510 --> 00:25:19,920
not where they need to be.
596
00:25:19,920 --> 00:25:23,340
If you have rail cars not
where they need to be.
597
00:25:23,340 --> 00:25:28,340
If you have a rail yard
being used as a warehouse
598
00:25:28,770 --> 00:25:31,230
because the warehousing
space in the region
599
00:25:31,230 --> 00:25:33,690
is not available, the
whole system broke down.
600
00:25:33,690 --> 00:25:36,450
So that's a very unique situation,
601
00:25:36,450 --> 00:25:39,150
but we can use machine learning here also
602
00:25:39,150 --> 00:25:40,590
in a number of different ways.
603
00:25:40,590 --> 00:25:44,220
We talked a little bit
about generative models
604
00:25:44,220 --> 00:25:47,490
and we talked a bit about
classification models.
605
00:25:47,490 --> 00:25:50,280
There's also other kinds
of models you can use
606
00:25:50,280 --> 00:25:53,850
for machine learning
which are using clustering
607
00:25:53,850 --> 00:25:57,390
and also making predictions
for forecasting.
608
00:25:57,390 --> 00:26:02,390
And in our work, which we do
for the import global trade,
609
00:26:03,150 --> 00:26:07,890
we are using clustering in order
to understand ship signals.
610
00:26:07,890 --> 00:26:10,620
So we use something called DBSCAN
611
00:26:10,620 --> 00:26:14,520
in order to identify from AIS pings
612
00:26:14,520 --> 00:26:16,620
where congestion is happening
613
00:26:16,620 --> 00:26:18,810
and you can look at that real time,
614
00:26:18,810 --> 00:26:20,730
you can also look at it historically
615
00:26:20,730 --> 00:26:23,190
and also build a predictive model
616
00:26:23,190 --> 00:26:26,910
to understand if
something is congested now
617
00:26:26,910 --> 00:26:28,980
historically how long did it take
618
00:26:28,980 --> 00:26:30,300
to break that congestion
619
00:26:30,300 --> 00:26:32,880
or is this something that
we've not seen before
620
00:26:32,880 --> 00:26:36,450
like the Baltimore Bridge and
suddenly there's congestion.
621
00:26:36,450 --> 00:26:41,137
So this is an interesting
interplay of clustering
622
00:26:42,510 --> 00:26:47,370
to detect something both
in present and in the past
623
00:26:47,370 --> 00:26:50,850
and then putting it
into a prediction model
624
00:26:50,850 --> 00:26:55,850
in order to make a forecast,
how long is this going to take?
625
00:26:56,640 --> 00:26:58,520
And in fact, we teach this
626
00:26:58,520 --> 00:27:00,990
in our supply chain masters program.
627
00:27:00,990 --> 00:27:04,770
You can make a single
time series predictions,
628
00:27:04,770 --> 00:27:08,220
but you can also make a graph prediction
629
00:27:08,220 --> 00:27:11,850
and that's very important
in the maritime sector
630
00:27:11,850 --> 00:27:16,850
because if you are congested in New York
631
00:27:17,130 --> 00:27:21,720
and then ocean carriers are
making a particular circle route
632
00:27:21,720 --> 00:27:25,920
and the next route they
hit is Norfolk, Virginia,
633
00:27:25,920 --> 00:27:28,410
well then Norfolk, Virginia is likely
634
00:27:28,410 --> 00:27:32,370
also going to be delayed if
you're delayed in New York.
635
00:27:32,370 --> 00:27:37,260
So understanding this graph
of how the ports work together
636
00:27:37,260 --> 00:27:40,950
for commodities is also
a very nice application
637
00:27:40,950 --> 00:27:45,000
of different kinds of machine
learning models that we use.
638
00:27:45,000 --> 00:27:46,860
- Very briefly before we close,
639
00:27:46,860 --> 00:27:48,300
if you look at your current work,
640
00:27:48,300 --> 00:27:52,120
what are maybe the one or
two biggest limitations,
641
00:27:52,120 --> 00:27:54,701
if you wish, of these types of methods
642
00:27:54,701 --> 00:27:56,580
that keep you from doing
643
00:27:56,580 --> 00:27:58,350
what you would love to do right now?
644
00:27:58,350 --> 00:28:00,180
- There's nothing that keeps us actually
645
00:28:00,180 --> 00:28:01,680
from doing what we love to do.
646
00:28:01,680 --> 00:28:04,740
It's important to have good tracking data
647
00:28:04,740 --> 00:28:06,600
in order to be able to find
648
00:28:06,600 --> 00:28:08,520
where this congestion is happening.
649
00:28:08,520 --> 00:28:12,120
But this is what is
advancing so rapidly now
650
00:28:12,120 --> 00:28:15,870
and it's so exciting to see that this data
651
00:28:15,870 --> 00:28:18,690
is becoming available
which feeds these models.
652
00:28:18,690 --> 00:28:21,300
So I see a bright future ahead
653
00:28:21,300 --> 00:28:24,138
for the work
of tracking cargo,
654
00:28:24,138 --> 00:28:27,660
whether that's by air, by
ocean and global trade.
655
00:28:27,660 --> 00:28:30,570
Now the other question
related to the pillar
656
00:28:30,570 --> 00:28:32,640
in sourcing and procurement,
657
00:28:32,640 --> 00:28:35,970
you have very heterogeneous
systems of data
658
00:28:35,970 --> 00:28:37,950
that need to be put together.
659
00:28:37,950 --> 00:28:41,010
So here there's also a world to win
660
00:28:41,010 --> 00:28:44,700
in the procurement function
of different kinds of data
661
00:28:44,700 --> 00:28:48,570
that suddenly we have visibility into
662
00:28:48,570 --> 00:28:52,650
in terms of being able
to read and classify
663
00:28:52,650 --> 00:28:56,550
heterogeneous text and
also mixed with tables too.
664
00:28:56,550 --> 00:28:58,200
For example, if you have a contract,
665
00:28:58,200 --> 00:28:59,850
you have text and tables,
666
00:28:59,850 --> 00:29:03,540
you want to extract that
and also organizing well
667
00:29:03,540 --> 00:29:06,840
so that you can query
appropriately text for terms
668
00:29:06,840 --> 00:29:08,880
and tables for maybe a price rate.
669
00:29:08,880 --> 00:29:13,880
So here there's also very
rapid advancements being made.
670
00:29:14,160 --> 00:29:16,800
So it's really exciting to see
671
00:29:16,800 --> 00:29:18,870
that the things that you would like, oh,
672
00:29:18,870 --> 00:29:22,080
I'd love to be able to
read this legal contract
673
00:29:22,080 --> 00:29:23,970
and separate the text from the tables.
674
00:29:23,970 --> 00:29:27,120
It's possible, you can
put that into a database
675
00:29:27,120 --> 00:29:29,160
and you can query it
and you can connect it
676
00:29:29,160 --> 00:29:30,720
to other databases.
677
00:29:30,720 --> 00:29:34,170
So it's all about organizing,
678
00:29:34,170 --> 00:29:37,710
having your data to be able
to run these models on.
679
00:29:37,710 --> 00:29:42,710
And right now I just see
advancements going very quickly,
680
00:29:42,780 --> 00:29:46,020
but there is one very important part,
681
00:29:46,020 --> 00:29:48,780
particularly in the procurement function
682
00:29:48,780 --> 00:29:53,780
that the category managers,
they need to be on board.
683
00:29:53,788 --> 00:29:56,700
In the example which we
tested at the company
684
00:29:56,700 --> 00:30:00,270
of upgrading their existing system,
685
00:30:00,270 --> 00:30:02,220
which was rule-based with a thing,
686
00:30:02,220 --> 00:30:04,050
this was very seamless for them
687
00:30:04,050 --> 00:30:06,000
because they had something
that wasn't working
688
00:30:06,000 --> 00:30:07,650
and now suddenly it is working.
689
00:30:07,650 --> 00:30:10,830
They can just ask a question
and they get the answer
690
00:30:10,830 --> 00:30:12,810
instead of having to
click through dashboards
691
00:30:12,810 --> 00:30:14,820
and try to piece things together.
692
00:30:14,820 --> 00:30:18,870
In all of these cases, it
makes the lives of the people
693
00:30:18,870
--> 00:30:23,010
who are working more efficient
and saves them tedious work,
694
00:30:23,010 --> 00:30:26,190
which is not what their main function is.
695
00:30:26,190 --> 00:30:28,440
Our category manager should be procuring,
696
00:30:28,440 --> 00:30:29,580
they should be negotiating,
697
00:30:29,580 --> 00:30:31,140
they don't need to get dragged down
698
00:30:31,140 --> 00:30:32,910
in clicking through dashboards,
699
00:30:32,910 --> 00:30:37,638
but the tools we provide
should be embedded
700
00:30:37,638 --> 00:30:39,330
in their work process.
701
00:30:39,330 --> 00:30:40,320
That's essential.
702
00:30:40,320 --> 00:30:42,780
If you are just providing an extra tool
703
00:30:42,780 --> 00:30:45,660
which is not embedded in their
work process, it's useless.
704
00:30:45,660 --> 00:30:48,180
That's part of our success also,
705
00:30:48,180 --> 00:30:50,640
that when we're developing something,
706
00:30:50,640 --> 00:30:52,860
we're meeting with the people whose lives
707
00:30:52,860 --> 00:30:55,500
we like to make easier, right from day one
708
00:30:55,500 --> 00:30:57,750
to understand exactly what are their needs
709
00:30:57,750 --> 00:31:02,010
and what is their workflow and
how this can facilitate it.
710
00:31:02,010 --> 00:31:03,330
- That's a strong message to end this on.
711
00:31:03,330 --> 00:31:05,190
Thank you so much for joining us, Elenna,
712
00:31:05,190 --> 00:31:07,982
look forward to seeing
more of your research.
713
00:31:07,982 --> 00:31:10,649
(upbeat music)
714
00:31:12,210 --> 00:31:13,800
Thank you so much for joining us Bryan,
715
00:31:13,800 --> 00:31:15,060
it's a pleasure to have you.
716
00:31:15,060 --> 00:31:17,640
So let's start this off easily.
717
00:31:17,640 --> 00:31:18,840
Can you explain to me,
718
00:31:18,840 --> 00:31:20,550
but more importantly to our audience
719
00:31:20,550 --> 00:31:22,590
what your current research focus is?
720
00:31:22,590 --> 00:31:25,320
- One of my big research's
focus is really centered
721
00:31:25,320 --> 00:31:27,697
around a book that I just published,
722
00:31:27,697 --> 00:31:29,730
"How to Make AI Useful,"
723
00:31:29,730 --> 00:31:32,640
very much focused around the human's role
724
00:31:32,640 --> 00:31:35,820
with complex social
technical systems like AI.
725
00:31:35,820 --> 00:31:38,580
So we often think about AI
726
00:31:38,580 --> 00:31:40,560
in the context of its capabilities
727
00:31:40,560 --> 00:31:42,180
to do things for us, automation,
728
00:31:42,180 --> 00:31:46,500
but the real value of AI to
me and the work that I'm doing
729
00:31:46,500 --> 00:31:50,160
is its ability to amplify human expertise.
730
00:31:50,160 --> 00:31:53,370
So improving the decision
systems around us
731
00:31:53,370 --> 00:31:55,200
so that we can operate better
732
00:31:55,200 --> 00:31:57,546
where impartial
information presents itself
733
00:31:57,546 --> 00:31:59,250
in the real world.
734
00:31:59,250 --> 00:32:02,520
Automation and machine
intelligence is really good
735
00:32:02,520 --> 00:32:04,260
at black and white decisions,
736
00:32:04,260 --> 00:32:06,210
but when disruption happens,
737
00:32:06,210 --> 00:32:09,480
human expertise merged with
that advanced decision making
738
00:32:09,480 --> 00:32:12,000
is really the area of focus.
739
00:32:12,000 --> 00:32:12,833
- Okay.
740
00:32:12,833 --> 00:32:15,480
Can you give us an example
of how you're actually
741
00:32:15,480 --> 00:32:18,180
using AI in that particular
way in your work,
742
00:32:18,180 --> 00:32:20,370
in your research at MIT and elsewhere?
743
00:32:20,370 --> 00:32:23,280
- I am increasingly
using more and more GenAI
744
00:32:23,280 --> 00:32:27,810
as a part of my research
focus, especially in writing.
745
00:32:27,810 --> 00:32:29,550
A lot of folks, especially
in the academics
746
00:32:29,550 --> 00:32:30,930
who are setting all those things well,
747
00:32:30,930 --> 00:32:32,340
it's gotta be all human written
748
00:32:32,340 --> 00:32:35,640
and GenAI tools are a
tool that doesn't sleep.
749
00:32:35,640 --> 00:32:36,900
So two o'clock in the morning
750
00:32:36,900 --> 00:32:39,810
instead of calling a
colleague to help revise
751
00:32:39,810 --> 00:32:41,130
and edit something more importantly,
752
00:32:41,130 --> 00:32:43,980
often take a hundred words
out of a thousand word essay.
753
00:32:43,980 --> 00:32:47,250
GenAI tools are incredibly useful
754
00:32:47,250 --> 00:32:50,640
at helping provide suggestions
755
00:32:50,640 --> 00:32:53,280
on how to merge and change texts,
756
00:32:53,280 --> 00:32:55,440
how to summarize information.
757
00:32:55,440 --> 00:32:58,800
But what's really key is
that all AI is capable
758
00:32:58,800 --> 00:33:02,370
of doing today is providing
me decision support,
759
00:33:02,370 --> 00:33:05,670
providing me insight and
direction that I can use
760
00:33:05,670 --> 00:33:09,210
and honing either
messages that I'm writing
761
00:33:09,210 --> 00:33:10,980
or how I'm thinking about framing things
762
00:33:10,980 --> 00:33:12,990
in front of an executive audience.
763
00:33:12,990 --> 00:33:15,510
We're also using AI increasingly
764
00:33:15,510 --> 00:33:18,210
to augment data flows through our systems,
765
00:33:18,210 --> 00:33:20,640
both more traditional
machine learning approaches
766
00:33:20,640 --> 00:33:23,910
and increasingly to synthesize information
767
00:33:23,910 --> 00:33:26,970
that once required an
enormous amount of human labor
768
00:33:26,970 --> 00:33:27,803
to go through.
769
00:33:27,803 --> 00:33:29,190
- That's a very interesting take,
770
00:33:29,190 --> 00:33:31,380
especially I was intrigued
by the first point
771
00:33:31,380 --> 00:33:32,213
that you were making
772
00:33:32,213 --> 00:33:34,980
that you're using GenAI increasingly
773
00:33:34,980 --> 00:33:36,810
as your writing assistant
774
00:33:36,810 --> 00:33:39,270
or as someone who can provide you opinions
775
00:33:39,270 --> 00:33:41,640
that you as the
human
decision maker then can weigh.
776
00:33:41,640 --> 00:33:43,800
Don't you see a risk there though
777
00:33:43,800 --> 00:33:47,820
that the incentive is
high to give more and more
778
00:33:47,820 --> 00:33:52,170
of the actual decision-making
authority to the GenAI?
779
00:33:52,170 --> 00:33:55,290
Don't you see a risk that
people take it too lightly
780
00:33:55,290 --> 00:33:57,510
and let a GenAI tool then, for instance,
781
00:33:57,510 --> 00:33:59,820
write an entire paper or a book
782
00:33:59,820 --> 00:34:02,640
and without the necessary
checks and balances?
783
00:34:02,640 --> 00:34:06,510
- I think that is a huge
fear and ultimate reality.
784
00:34:06,510 --> 00:34:09,810
Automation complacency,
human's ability to get lazy
785
00:34:09,810 --> 00:34:12,690
in the advent of
automation is a real thing.
786
00:34:12,690 --> 00:34:15,030
We hear increasingly in the business world
787
00:34:15,030 --> 00:34:17,400
around the context of
work slop in essence,
788
00:34:17,400 --> 00:34:19,470
I can generate 10 pages of slop
789
00:34:19,470 --> 00:34:21,390
that satisfy a deliverable.
790
00:34:21,390 --> 00:34:25,380
We desperately need to be
thinking about creating
791
00:34:25,380 --> 00:34:28,980
within organizations an intolerable policy
792
00:34:28,980 --> 00:34:30,300
against work slop.
793
00:34:30,300 --> 00:34:33,611
Unless we set the social
standard here high
794
00:34:33,611 --> 00:34:37,410
where AI and especially
GenAI today is used
795
00:34:37,410 --> 00:34:39,450
as a copilot, not as an autopilot,
796
00:34:39,450 --> 00:34:44,190
a good portion of the
populace is going to get lazy
797
00:34:44,190 --> 00:34:45,180
with these systems.
798
00:34:45,180 --> 00:34:48,780
I think we're going to
unfortunately see that occur,
799
00:34:48,780 --> 00:34:50,940
and in that world what we will see
800
00:34:50,940 --> 00:34:54,510
is the best and brightest learn to use AI
801
00:34:54,510 --> 00:34:56,580
as an accelerator of their capabilities.
802
00:34:56,580 --> 00:35:00,870
I think that you're going
to see a number of people
803
00:35:00,870 --> 00:35:05,100
using this to accelerate,
to move faster, to hone,
804
00:35:05,100 --> 00:35:07,230
to do impressive things
that were not capable
805
00:35:07,230 --> 00:35:09,060
in the time that they had available.
806
00:35:09,060 --> 00:35:10,740
But I think there is a good segment
807
00:35:10,740 --> 00:35:11,940
of the working population
808
00:35:11,940 --> 00:35:13,950
that's going to get
awful lazy, awful quick,
809
00:35:13,950 --> 00:35:16,680
and that needs to be
intolerable in the workplace,
810
00:35:16,680 --> 00:35:18,750
intolerable in the educational setting.
811
00:35:18,750 --> 00:35:22,020
I want to know how I can use
these tools to be better.
812
00:35:22,020 --> 00:35:24,000
You know, I had a project
the other day I was doing
813
00:35:24,000 --> 00:35:26,250
and that I knew there
was a document out there
814
00:35:26,250 --> 00:35:28,860
and I knew what
the
document generally said,
815
00:35:28,860 --> 00:35:31,170
but I went to Google to
try to find that document
816
00:35:31,170 --> 00:35:32,430
and the document was around some,
817
00:35:32,430 --> 00:35:35,070
some federal guidance
on automated vehicles
818
00:35:35,070 --> 00:35:37,200
from about eight years ago.
819
00:35:37,200 --> 00:35:40,620
So I spent about an
hour failing with Google
820
00:35:40,620 --> 00:35:41,820
to find this document.
821
00:35:41,820 --> 00:35:43,380
So I said, oh wait a second,
822
00:35:43,380 --> 00:35:46,200
why the heck didn't I go
over to ChatGPT and try?
823
00:35:46,200 --> 00:35:48,120
In the matter of two queries
824
00:35:48,120 --> 00:35:49,500
in less than three minutes
825
00:35:49,500 --> 00:35:51,450
I had a document I needed at hand
826
00:35:51,450 --> 00:35:54,630
because how I was framing this question
827
00:35:54,630 --> 00:35:58,170
was slightly different than
how the document was written
828
00:35:58,170 --> 00:36:00,930
and ChatGPT is capable
of making that leap.
829
00:36:00,930 --> 00:36:04,500
So the ability to harness these tools
830
00:36:04,500 --> 00:36:07,560
in the middle of the
night in a moments' notice
831
00:36:07,560 --> 00:36:10,380
reshapes what work looks like
832
00:36:10,380 --> 00:36:12,150
and it allows people
to become more creative
833
00:36:12,150 --> 00:36:13,800
and allows the folks that we would use
834
00:36:13,800 --> 00:36:15,330
in a daily basis to be helping us
835
00:36:15,330 --> 00:36:16,710
to do other creative things.
836
00:36:16,710 --> 00:36:19,530
- Obviously a lot of our
research here at the center,
837
00:36:19,530 --> 00:36:22,590
but also like a lot of
our audience is faced
838
00:36:22,590 --> 00:36:23,940
more or less on a daily basis
839
00:36:23,940 --> 00:36:26,580
with very numbers driven
decision problems,
840
00:36:26,580 --> 00:36:27,960
very analytical work.
841
00:36:27,960 --> 00:36:30,600
Now let's say I had a
large supply chain dataset
842
00:36:30,600 --> 00:36:32,880
and I knew nothing about it actually.
843
00:36:32,880 --> 00:36:35,040
I just know what the dataset contains,
844
00:36:35,040 --> 00:36:36,990
but I don't know whether
there's anything interesting
845
00:36:36,990 --> 00:36:37,823
in there or not.
846
00:36:37,823 --> 00:36:41,610
Would you go out and
actually still use a tool
847
00:36:41,610 --> 00:36:43,890
like ChatGPT or others
848
00:36:43,890 --> 00:36:47,880
to extract the interesting
insights from such a dataset?
849
00:36:47,880 --> 00:36:49,230
So you would for instance,
850
00:36:49,230 --> 00:36:50,880
ask a research assistant today, hey,
851
00:36:50,880 --> 00:36:52,260
can you go through this data set,
852
00:36:52,260 --> 00:36:54,240
analyze it backwards and forwards,
853
00:36:54,240 --> 00:36:57,090
and tell me what's interesting
or unexpected about it?
854
00:36:57,090 --> 00:36:58,500
Are we already at a point
855
00:36:58,500 --> 00:37:01,950
where we can trust an AI model to do that
856
00:37:01,950 --> 00:37:04,380
or is that still one level of complexity
857
00:37:04,380 --> 00:37:08,880
and analytical capability
too far at least today?
858
00:37:08,880 --> 00:37:10,620
- That's a great question
859
00:37:10,620 --> 00:37:13,170
and I would absolutely want to see
860
00:37:13,170 --> 00:37:14,820
what the GenAI model tells me,
861
00:37:14,820 --> 00:37:17,220
but all it is telling me is an opinion
862
00:37:17,220 --> 00:37:19,650
that I have no ground truth for.
863
00:37:19,650 --> 00:37:22,470
So if I had the resources,
864
00:37:22,470 --> 00:37:25,230
I would be having my research assistant
865
00:37:25,230 --> 00:37:28,080
interrogate the data
right, left up and down
866
00:37:28,080 --> 00:37:30,600
the same way I would've a year or two ago.
867
00:37:30,600 --> 00:37:32,430
At the same time,
868
00:37:32,430 --> 00:37:36,720
I would be asking ChatGPT to
provide me an interpretation.
869
00:37:36,720 --> 00:37:39,420
Then given there is two humans in the loop
870
00:37:39,420 --> 00:37:43,380
and one machine who can't
defend itself quite as well yet,
871
00:37:43,380 --> 00:37:46,770
I would take those two
pieces of information.
872
00:37:46,770 --> 00:37:49,350
And much like we in the
social sciences world
873
00:37:49,350 --> 00:37:51,510
used to do something called
inter-coder reliability
874
00:37:51,510 --> 00:37:52,343
all the time,
875
00:37:52,343 --> 00:37:54,660
I had two different independent
codings of this data
876
00:37:54,660 --> 00:37:58,350
that I want to try to
extract a true meaning from.
877
00:37:58,350 --> 00:38:00,300
I would take those two interpretations
878
00:38:00,300 --> 00:38:03,870
and infuse my expert judgment
on what do I think is right
879
00:38:03,870 --> 00:38:05,100
and what do I think is wrong,
880
00:38:05,100 --> 00:38:07,650
because the machine is not perfect
881
00:38:07,650 --> 00:38:09,360
and the human is not perfect.
882
00:38:09,360 --> 00:38:13,500
We are both non-perfect systems
looking to provide answers
883
00:38:13,500 --> 00:38:15,000
based upon all the information
884
00:38:15,000 --> 00:38:16,440
that we've learned in the past.
885
00:38:16,440 --> 00:38:18,030
In a lot of cases today,
886
00:38:18,030 --> 00:38:20,460
the machine may have biases in other ways,
887
00:38:20,460 --> 00:38:21,570
any statistical model
888
00:38:21,570 --> 00:38:23,700
or is making assumptions with their data
889
00:38:23,700 --> 00:38:26,280
that as readers don't
fully appreciate either.
890
00:38:26,280 --> 00:38:30,210
So I think it is very
much using these opinions
891
00:38:30,210 --> 00:38:32,280
and fusing a model on top of that.
892
00:38:32,280 --> 00:38:35,640
Now given my trust in machine intelligence
893
00:38:35,640 --> 00:38:37,380
is not fully evolved yet,
894
00:38:37,380 --> 00:38:39,930
I might take my synthesis from the machine
895
00:38:39,930 --> 00:38:41,910
and go back to my analyst and say, hmm,
896
00:38:41,910 --> 00:38:44,310
do you see something similar in
the data?
897
00:38:44,310 --> 00:38:46,080
And the hope would be is yeah,
898
00:38:46,080 --> 00:38:48,390
that's an interesting point
that the machine generates.
899
00:38:48,390 --> 00:38:49,920
That's something I didn't pick up
900
00:38:49,920 --> 00:38:51,660
in the patterning of the data
901
00:38:51,660 --> 00:38:53,610
and that's exactly what we're hoping
902
00:38:53,610 --> 00:38:55,650
that that the combination
of machine intelligence
903
00:38:55,650 --> 00:38:57,540
and human intelligence
can bring us together.
904
00:38:57,540 --> 00:38:58,920
- And you've done some amazing work
905
00:38:58,920 --> 00:39:01,560
over the last couple of
years in various fields
906
00:39:01,560 --> 00:39:02,940
including autonomous driving.
907
00:39:02,940 --> 00:39:05,640
And so your bar is
relatively high when it comes
908
00:39:05,640 --> 00:39:10,500
to expectations towards AI
capabilities of the future.
909
00:39:10,500 --> 00:39:12,510
But compared to what's possible today,
910
00:39:12,510 --> 00:39:15,150
what are the breakthroughs
that you're hoping
911
00:39:15,150 --> 00:39:18,030
are going to be made possible by AI
912
00:39:18,030 --> 00:39:19,590
in the next couple of years?
913
00:39:19,590 --> 00:39:21,360
- I think that AI and automation
914
00:39:21,360 --> 00:39:24,360
is gonna solve immense
things over the next century,
915
00:39:24,360 --> 00:39:25,680
but I don't think automation
916
00:39:25,680 --> 00:39:27,600
is going to have the big impact
917
00:39:27,600 --> 00:39:29,817
that we would all love
in the next few years
918
00:39:29,817 --> 00:39:33,660
and all of a sudden transformatively
shift the workforce.
919
00:39:33,660 --> 00:39:35,280
That's gonna take a little more time.
920
00:39:35,280 --> 00:39:38,100
And it's not that I fear
unemployment all over the place.
921
00:39:38,100 --> 00:39:40,499
I think we are gonna
reallocate human expertise
922
00:39:40,499 --> 00:39:42,060
throughout the system.
923
00:39:42,060 --> 00:39:44,670
I think we're gonna create many jobs
924
00:39:44,670 --> 00:39:47,280
as we re-shift the workforce around.
925
00:39:47,280 --> 00:39:50,580
If I look forward over the next few years,
926
00:39:50,580 --> 00:39:52,710
what I think AI enables
927
00:39:52,710 --> 00:39:55,620
is something that I was
working on 25 years ago
928
00:39:55,620 --> 00:39:58,590
in the early stages of my doctoral work,
929
00:39:58,590 --> 00:40:00,750
decision support at scale,
930
00:40:00,750 --> 00:40:03,150
and providing me an opinion
931
00:40:03,150 --> 00:40:06,060
that I may decide is worthy of taking
932
00:40:06,060 --> 00:40:09,720
or I may dismiss because
of some of my expertise,
933
00:40:09,720 --> 00:40:12,420
but at least I have some
decision support information
934
00:40:12,420 --> 00:40:13,860
coming along in real time
935
00:40:13,860 --> 00:40:16,680
that I wouldn't have
had 2, 3, 4, 5 years ago
936
00:40:16,680 --> 00:40:19,680
because of AI's capabilities
of summarizing data.
937
00:40:19,680 --> 00:40:21,720
Is that information perfect?
938
00:40:21,720 --> 00:40:22,950
Probably not.
939
00:40:22,950 --> 00:40:24,510
Is it good enough to act on?
940
00:40:24,510 --> 00:40:27,240
Hopefully more times than it isn't,
941
00:40:27,240 --> 00:40:30,210
but often I think it will
enhance our decisions
942
00:40:30,210 --> 00:40:31,560
in the real world,
943
00:40:31,560 --> 00:40:33,690
leaving most of the important decisions
944
00:40:33,690 --> 00:40:35,310
for the foreseeable future
945
00:40:35,310 --> 00:40:37,170
in the hands of human decision makers.
946
00:40:37,170 --> 00:40:38,003
- That makes sense.
947
00:40:38,003 --> 00:40:40,590
And if you look at the tools
that are available to us
948
00:40:40,590 --> 00:40:43,290
right now, the methods that
are currently being used,
949
00:40:43,290 --> 00:40:46,050
what do you see are
currently the biggest hurdles
950
00:40:46,050 --> 00:40:48,870
or limitations or maybe even risks
951
00:40:48,870 --> 00:40:50,670
that come with these methods?
952
00:40:50,670 --> 00:40:52,090
- Yeah, that's a good question.
953
00:40:52,090 --> 00:40:55,140
GenAI based upon large
language models in particular,
954
00:40:55,140 --> 00:40:57,120
has been a dominant point in the headlines
955
00:40:57,120 --> 00:40:58,110
for a few years now.
956
00:40:58,110 --> 00:41:01,260
I think that the future
over the next several years
957
00:41:01,260 --> 00:41:05,805
is really small language
models and domain specific AI
958
00:41:05,805 --> 00:41:08,100
tailoring the applications of AI
959
00:41:08,100 --> 00:41:11,658
to the specific needs of a smaller problem
960
00:41:11,658 --> 00:41:14,730
because in that context we
can be much more confident
961
00:41:14,730 --> 00:41:16,050
in the answers it's giving us.
962
00:41:16,050 --> 00:41:17,820
So, you know, there's a lot of critiques
963
00:41:17,820 --> 00:41:20,687
when OpenAI released its latest model of,
964
00:41:20,687 --> 00:41:24,427
of the model is not performing
as well as the last.
965
00:41:24,427 --> 00:41:26,670
And I think we're gonna
continue to see that
966
00:41:26,670 --> 00:41:29,940
every time we release a new
model, and part of that,
967
00:41:29,940 --> 00:41:32,610
it's because we are now
using an ensemble approach
968
00:41:32,610 --> 00:41:34,710
of dividing the different queries up
969
00:41:34,710 --> 00:41:37,800
and sending 'em to different
places based upon the context,
970
00:41:37,800 --> 00:41:41,547
trying to become more efficient
with our resources energy
971
00:41:41,547 --> 00:41:44,760
and using smaller models to solve problems
972
00:41:44,760 --> 00:41:47,880
that they become much more
strategically trained for.
973
00:41:47,880 --> 00:41:52,680
The whole area of domain
specific AI is exploding.
974
00:41:52,680 --> 00:41:54,870
This, you know, reduces the compute needs,
975
00:41:54,870 --> 00:41:56,640
allows you to tailor the models
976
00:41:56,640 --> 00:41:59,010
to specifically the types of questions
977
00:41:59,010 --> 00:42:00,690
you're looking to solve today.
978
00:42:00,690 --> 00:42:01,950
- So we should start building
979
00:42:01,950 --> 00:42:05,100
a supply chain foundation
model, if I hear you right?
980
00:42:05,100 --> 00:42:06,540
- Supply chain foundation model
981
00:42:06,540 --> 00:42:09,870
and maybe it looks at the
movement of energy differently
982
00:42:09,870 --> 00:42:11,370
than the movement of bananas
983
00:42:11,370 --> 00:42:14,400
because even the supply
chain may be just too big.
984
00:42:14,400 --> 00:42:16,920
That's where a lot of
organizations are focusing
985
00:42:16,920 --> 00:42:20,246
and that's where maybe all
of this background investment
986
00:42:20,246 --> 00:42:22,890
in large scale computing,
987
00:42:22,890 --> 00:42:27,660
maybe overshooting what our
near and midterm needs are.
988
00:42:27,660 --> 00:42:30,780
You see the big tech companies
being asked by Wall Street,
989
00:42:30,780 --> 00:42:32,490
what is the return on investment
990
00:42:32,490 --> 00:42:34,830
in continuing to build, build, build?
991
00:42:34,830 --> 00:42:36,360
And I think that's a good question
992
00:42:36,360 --> 00:42:40,590
because quite frankly we
love investing in technology
993
00:42:40,590 --> 00:42:41,760
for technology's sake,
994
00:42:41,760 --> 00:42:43,980
but is it providing us the decision tools
995
00:42:43,980 --> 00:42:46,140
that are needed for the
problems we have today?
996
00:42:46,140 --> 00:42:48,300
And I wonder if we're overshooting here
997
00:42:48,300 --> 00:42:50,550
about the promise of AI
998
00:42:50,550 --> 00:42:52,380
as opposed to grounding ourselves in,
999
00:42:52,380 --> 00:42:53,820
what is this going to do to help me
1000
00:42:53,820 --> 00:42:56,610
and how am I gonna drive revenue,
1001
00:42:56,610 --> 00:42:59,700
operational efficiencies
using these tools?
1002
00:42:59,700 --> 00:43:02,370
And I think we're gonna come
back to that really quickly.
1003
00:43:02,370 --> 00:43:03,203
- Thank you, Bryan.
1004
00:43:03,203 --> 00:43:04,530
This has been a fascinating conversation
1005
00:43:04,530 --> 00:43:06,060
and I'm sure we could go on forever,
1006
00:43:06,060 --> 00:43:08,460
but for now I'll just
look even more forward
1007
00:43:08,460 --> 00:43:09,510
to reading your book.
1008
00:43:10,440 --> 00:43:13,380
- That wraps up this episode
of "Supply Chain Frontiers."
1009
00:43:13,380 --> 00:43:16,350
A big thank you to Matthias
Winkenbach, Elenna Dugundji,
1010
00:43:16,350 --> 00:43:18,270
Willem Guter, and Bryan Reimer
1011
00:43:18,270 --> 00:43:20,340
for sharing their expertise and insights
1012
00:43:20,340 --> 00:43:21,750
into the current state of AI
1013
00:43:21,750 --> 00:43:23,917
and supply chain research
and applications.
1014
00:43:23,917 --> 00:43:26,970
"Supply Chain Frontiers" is
recorded on the MIT campus
1015
00:43:26,970 --> 00:43:28,830
in Cambridge, Massachusetts.
1016
00:43:28,830 --> 00:43:32,220
Our sound editors are Dave
Lashinsky and Danielle Simpson
1017
00:43:32,220 --> 00:43:33,840
at David Benjamin Sound.
1018
00:43:33,840 --> 00:43:36,360
And our audio engineer
today is Kurt Schneider
1019
00:43:36,360 --> 00:43:38,910
of MIT Audio Visual Services.
1020
00:43:38,910 --> 00:43:41,130
Our producer is myself, Mackenzie Berry.
1021
00:43:41,130 --> 00:43:43,050
Be sure to check out previous episodes
1022
00:43:43,050 --> 00:43:48,050
of "Supply Chain Frontiers"
at ctl.mit.edu/podcast
1023
00:43:48,150 --> 00:43:50,760
or search for us on your
preferred podcast platform.
1024
00:43:50,760 --> 00:43:52,770
I'm Mackenzie Berry, thanks for listening,
1025
00:43:52,770 --> 00:43:55,769
and we'll catch you next time
on "Supply Chain Frontiers."
1026
00:43:55,769 --> 00:43:58,436
(upbeat music)