Why AI doesn't follow length instructions, the best $40 you can spend, and more


This week's episode covers:
- Generative AI's paywall problem
- Anthropic release new Claude models that beat GPT
- Google has a bad week
- Why generative AI doesn't follow length instructions (and what you can do about it)
- The best $40 you can spend on generative AI
- More Useful Things releases some interesting AI resources
- Chain of thought versus few-shot prompting
--- AI generated description ---
Welcome to AI Goes to College, where we navigate the ever-changing world of generative AI in higher education. In this thought-provoking episode, I, your host, Dr. Craig Van Slyke, delve into the latest developments in the realm of generative AI, from the paywall problem to Anthropic's groundbreaking Claude models that outperform GPT. This episode sheds light on the ethical considerations and challenges facing academic researchers when working with biased training data and the potential limitations in reflecting findings from behind-the-paywall academic journals.
But it's not all about the challenges. I also uncover the exceptional potential of Anthropic's new Claude models and the significance of competition in driving innovation and performance in the AI landscape. You'll be immersed in the intriguing discussion about Google's stumbling block in implementing ethical guardrails for generative AI, a pivotal reminder that human oversight remains crucial in the current stage of AI utilization.
And let's not forget about practical tips. I share game-changing insights on prompting generative AI, covering the nuances between few shot and chain of thought prompting, and reveal the best $40 investment for enhancing productivity in your AI endeavors.
The conversation doesn't end there. I invite you to explore the transformative applications of generative AI in education through a fascinating interview with an industry expert. This episode promises to reshape your perspective on the potential and challenges of generative AI in higher education and leave you equipped with valuable knowledge and practical strategies for navigating this dynamic landscape.
Join us as we uncover the profound impact of generative AI on academic research, and gain invaluable insights that will shape your approach to utilizing AI effectively for success in the educational sphere. If you find this episode insightful, don't miss the chance to subscribe to the AI Goes to College newsletter for further invaluable resources and updates. Let's embark on the journey to embracing and leveraging generative AI's potential in higher education.
00:00:14,035 --> 00:00:17,870
Welcome to AI Goes to College, the podcast that helps higher education
2
00:00:17,930 --> 00:00:21,550
professionals navigate the changes brought on by generative AI.
3
00:00:22,010 --> 00:00:24,725
I'm your host, doctor Craig Van Slyke.
4
00:00:25,664 --> 00:00:29,285
The podcast is a companion to the AI Goes to College newsletter.
5
00:00:29,824 --> 00:00:33,510
You can sign up for the newsletter at ai goes to college dot com
6
00:00:33,969 --> 00:00:37,649
slash newsletter. Welcome to episode 4 of
7
00:00:37,649 --> 00:00:41,465
AI Goes to College. In this episode, I
8
00:00:41,465 --> 00:00:44,845
talk about generative AI's paywall problem,
9
00:00:45,305 --> 00:00:49,020
Anthropic's release of some excellent new Claude models that
10
00:00:49,260 --> 00:00:52,940
actually beat GPT. Talk about Google's
11
00:00:52,940 --> 00:00:56,620
bad week. And why generative AI doesn't
12
00:00:56,620 --> 00:00:57,520
follow linked
13
00:01:06,040 --> 00:01:09,640
versus few shot prompting and the best $40 you can
14
00:01:09,640 --> 00:01:12,940
spend on generative AI. It is not what you expect.
15
00:01:13,755 --> 00:01:17,595
I also wanted to give a shout out to Rob Crossler. If you haven't checked
16
00:01:17,595 --> 00:01:21,115
out his interview on AI Goes to College, you ought to. It's very
17
00:01:21,115 --> 00:01:24,780
interesting. Rob is a smart guy. Before we
18
00:01:24,780 --> 00:01:28,460
get into the main content though, I wanna thank Grammar Nut for catching
19
00:01:28,460 --> 00:01:31,920
a small typo on the AI Goes to College website.
20
00:01:32,185 --> 00:01:35,885
It's really great that some of the readers and listeners are so sharp eyed.
21
00:01:36,025 --> 00:01:39,625
I very much appreciate it. So here's my
22
00:01:39,625 --> 00:01:43,440
rant of the week. Large language models, as some of
23
00:01:43,440 --> 00:01:47,280
you probably know, are trained on data, and data
24
00:01:47,280 --> 00:01:50,785
that aren't included in the training data aren't reflected in the output.
25
00:01:51,585 --> 00:01:55,425
This means that biased data on the input side is reflected in biased
26
00:01:55,425 --> 00:01:59,125
output on the output side. So none of this is groundbreaking.
27
00:02:00,065 --> 00:02:03,490
We've talked about this before. You probably talked about this before with
28
00:02:03,490 --> 00:02:07,250
others. Demographic based bias
29
00:02:07,250 --> 00:02:10,965
is well known and it's a serious ethical issue related to generative
30
00:02:11,025 --> 00:02:14,865
AI. But when I was playing around over the last couple of
31
00:02:14,865 --> 00:02:18,620
weeks, it occurred to me that biased training data results in a different
32
00:02:18,620 --> 00:02:22,459
problem for academic researchers. I'm speculating here
33
00:02:22,459 --> 00:02:26,135
because I really don't know exactly what data the various large language
34
00:02:26,135 --> 00:02:29,975
models are trained on, but it seems to me that the training
35
00:02:29,975 --> 00:02:33,655
data may underrepresent articles from top academic journals,
36
00:02:33,655 --> 00:02:37,420
which is a huge problem. And here's why.
37
00:02:38,120 --> 00:02:41,740
A lot of top journals are behind paywalls of various sorts.
38
00:02:42,255 --> 00:02:45,555
For example, if I want to access a recent article
39
00:02:45,855 --> 00:02:49,235
from MIS Quarterly, I either have to get it through
40
00:02:49,700 --> 00:02:53,300
Association For Information Systems or maybe through my
41
00:02:53,300 --> 00:02:57,060
library. And a lot of top journals are like
42
00:02:57,060 --> 00:03:00,834
that. I'm not sure in other fields, but that's certainly the case in
43
00:03:00,834 --> 00:03:04,594
my field and I suspect it's the case in most fields. You know,
44
00:03:04,594 --> 00:03:08,280
sometimes we can get lucky lucky and an article's available through Google Scholar
45
00:03:08,340 --> 00:03:12,019
or some other non paywalled repository like
46
00:03:12,019 --> 00:03:15,845
ResearchGate or something like that. And eventually many
47
00:03:15,845 --> 00:03:19,364
of these articles make their way around the paywalls to become more freely
48
00:03:19,364 --> 00:03:23,180
available. But if those articles weren't available as part
49
00:03:23,180 --> 00:03:26,700
of the training data, then their
50
00:03:26,700 --> 00:03:30,305
findings may not be reflected when you interact with that
51
00:03:30,305 --> 00:03:32,965
large language model for your research.
52
00:03:34,225 --> 00:03:37,365
Now the training data may include abstracts or citations
53
00:03:37,665 --> 00:03:41,159
from articles that are included, but
54
00:03:41,379 --> 00:03:45,219
those are limited. The abstract only contains so much
55
00:03:45,219 --> 00:03:49,065
information and so you're really not going to get the full
56
00:03:49,065 --> 00:03:52,745
representation of the article. And since many of our top
57
00:03:52,745 --> 00:03:56,440
journals are behind paywalls, you may miss out
58
00:03:56,440 --> 00:04:00,280
on articles that are in top journals. That's kind of been
59
00:04:00,280 --> 00:04:03,965
my experience so far with perplexity. I think this
60
00:04:03,965 --> 00:04:07,645
problem is even bigger for disciplines that rely more on books than journal
61
00:04:07,645 --> 00:04:11,485
articles. Now, I'm not sure about the extent of the problem, but
62
00:04:11,485 --> 00:04:15,030
it's something to be aware of. For now,
63
00:04:15,650 --> 00:04:19,410
I advise caution. Look, generative AI
64
00:04:19,410 --> 00:04:23,235
is great. I'm a big believer in it, but it really
65
00:04:23,235 --> 00:04:26,294
isn't a shortcut for the hard work of scholarship.
66
00:04:27,074 --> 00:04:29,495
You still have to put in human thought and effort.
67
00:04:31,860 --> 00:04:35,699
So the big news item over the last couple of weeks is Anthropic releasing
68
00:04:35,699 --> 00:04:39,435
some new Claude models. This is a little bit
69
00:04:39,435 --> 00:04:43,035
complicated, so I'll refer you to the newsletter which is available
70
00:04:43,035 --> 00:04:44,575
at ai goes to college.com.
71
00:04:46,650 --> 00:04:50,270
But here's the low down. Anthropic,
72
00:04:51,370 --> 00:04:54,685
who's a competitor to OpenAI and produces
73
00:04:54,824 --> 00:04:58,185
competitors to JetGPT, released 3 new
74
00:04:58,185 --> 00:05:02,025
models: Claude 3 Haiku, Claude 3 Sonet,
75
00:05:02,025 --> 00:05:05,610
and Claude 3 Opus. I don't think Haiku is available
76
00:05:05,610 --> 00:05:08,670
yet, but Sonnet, Sonnet, and Opus are.
77
00:05:09,450 --> 00:05:11,755
I kind of like the names, since they're loosely
78
00:05:20,439 --> 00:05:24,120
the one that's the most capable. Both SONNET and
79
00:05:24,120 --> 00:05:27,800
Opus have 200,000 token context
80
00:05:27,800 --> 00:05:31,505
windows. So the context window is how much data the model
81
00:05:31,505 --> 00:05:35,345
can consider when it's creating its output. You can think of it as
82
00:05:35,345 --> 00:05:39,070
the model's memory capacity. Roughly, this is very
83
00:05:39,070 --> 00:05:42,910
roughly, a 200 k context window should be
84
00:05:42,910 --> 00:05:46,705
able to handle 3 to 400 pages of text. Now there are a lot of
85
00:05:46,705 --> 00:05:50,545
variables there. Opus can even
86
00:05:50,545 --> 00:05:54,370
go further for some uses. According to Anthropic, it
87
00:05:54,370 --> 00:05:58,210
can have a context window of up to 1,000,000 tokens,
88
00:05:58,210 --> 00:06:01,845
which is huge, but right now that's reserved for special
89
00:06:01,845 --> 00:06:05,525
cases. You can get the 200,000 context
90
00:06:05,525 --> 00:06:08,664
window for Claude 3 directly from Anthropic,
91
00:06:09,530 --> 00:06:12,570
or you can get it through Poe, and you know I'm a big fan of
92
00:06:12,570 --> 00:06:16,270
poe.com. But in Poe at least, Claude
93
00:06:16,705 --> 00:06:20,544
3 OPUS and Claude 3 OPUS 200 k are different
94
00:06:20,544 --> 00:06:24,220
models. Claude 3 OPUS is a smaller
95
00:06:24,220 --> 00:06:27,740
model than Claude 3 Opus 200 ks. Yeah. It gets
96
00:06:27,740 --> 00:06:31,099
confusing, but if you're trying to deal with large
97
00:06:31,099 --> 00:06:34,485
documents, either just default to the 200 k version
98
00:06:35,105 --> 00:06:38,545
or try the non 200 k version and see how things go. If they don't
99
00:06:38,545 --> 00:06:42,180
go well, then try the 200 k version. There's some
100
00:06:42,180 --> 00:06:45,860
limits through Poe of how many interactions you can have with Claude in a
101
00:06:45,860 --> 00:06:49,620
24 hour period, but they're pretty liberal, so I don't
102
00:06:49,620 --> 00:06:53,415
think it's going to be a big big deal. I don't think you're
103
00:06:53,415 --> 00:06:57,195
going to run into the usage limits for most use cases.
104
00:06:58,215 --> 00:07:01,760
What is a big deal is this ever enlarging context
105
00:07:01,820 --> 00:07:05,280
window. That just opens up a lot of interesting
106
00:07:05,420 --> 00:07:09,120
possibilities. For example, club 3
107
00:07:09,945 --> 00:07:13,625
should be able to summarize and synthesize across multiple
108
00:07:13,625 --> 00:07:17,385
journal articles in a single session. I haven't tested this out
109
00:07:17,385 --> 00:07:20,360
yet, but I'm going to soon and I will certainly let you know how it
110
00:07:20,360 --> 00:07:24,199
goes. The other big thing about Claude 3 is that
111
00:07:24,199 --> 00:07:27,715
according to Anthropic, Claude 3 outperforms
112
00:07:27,935 --> 00:07:31,775
GPT 4 across the board. Now if you go to the
113
00:07:31,775 --> 00:07:35,410
newsletter there's a nice little table, it's
114
00:07:35,410 --> 00:07:39,190
from Anthropic though, that shows a bunch of popular
115
00:07:39,330 --> 00:07:43,030
benchmarks and how well Claude
116
00:07:43,515 --> 00:07:47,275
3 and the various models and GPT 4 and
117
00:07:47,275 --> 00:07:51,020
Gemini Ultra and Pro did. And
118
00:07:51,020 --> 00:07:54,700
in every single instance, Claude 3 was
119
00:07:54,700 --> 00:07:58,425
better. Now who knows? You know, is Anthropic
120
00:07:58,585 --> 00:08:02,345
cherry picking here? Maybe, but even if they are, the
121
00:08:02,345 --> 00:08:06,160
performance is quite intriguing. And I think
122
00:08:06,160 --> 00:08:09,699
it bodes well for the future of these models.
123
00:08:10,240 --> 00:08:13,860
Competition is gonna push everybody. Google and Anthropic and
124
00:08:13,974 --> 00:08:17,495
OpenAI, they're all going to push each other and whenever Apple gets into the
125
00:08:17,495 --> 00:08:20,854
game and Meta and on and on and on. So I think the
126
00:08:20,854 --> 00:08:24,389
competition is good. It's gonna help us push the boundaries of what's
127
00:08:24,389 --> 00:08:28,229
possible with these models. Okay. Let's switch
128
00:08:28,229 --> 00:08:31,210
to another little bit of kind of amusing but
129
00:08:31,935 --> 00:08:35,315
insightful news. Google had a bad week.
130
00:08:35,615 --> 00:08:38,755
This is a couple of weeks ago by the time you listen to this.
131
00:08:39,520 --> 00:08:43,280
So the basic problem was that Gemini was creating some, shall we
132
00:08:43,280 --> 00:08:46,800
say, interesting images. America's founding fathers as
133
00:08:46,800 --> 00:08:50,395
African Americans, the pope is a woman, and there were some
134
00:08:50,395 --> 00:08:54,015
others. Of course, the world being what the world is,
135
00:08:54,075 --> 00:08:57,839
there was a bunch of outrage over this. Although I
136
00:08:57,839 --> 00:09:01,680
kind of chuckled and I think a lot of other people did. So
137
00:09:01,680 --> 00:09:05,495
according to Google, the problem came when
138
00:09:05,495 --> 00:09:09,175
they were fine tuning Gemini's image generation model, which is
139
00:09:09,175 --> 00:09:12,720
called Imagen 2. So the fine tuning was
140
00:09:12,720 --> 00:09:15,540
intended to prevent the tool from, and I'm quoting here,
141
00:09:16,079 --> 00:09:19,920
creating violent or sexually explicit images or depictions
142
00:09:19,920 --> 00:09:23,745
of real people. Engineers were also trying to
143
00:09:23,745 --> 00:09:27,505
ensure gender and ethnic diversity, but Google
144
00:09:27,505 --> 00:09:31,260
seems to have overcorrected, which resulted in some of these curious
145
00:09:31,320 --> 00:09:35,080
and historically inaccurate images. I think we need to get
146
00:09:35,080 --> 00:09:38,605
used to this. Widespread use of generative AI is
147
00:09:38,605 --> 00:09:42,045
still very new and we're still trying to figure out how to
148
00:09:42,045 --> 00:09:45,800
implement appropriate guardrails. In fact, there's not even
149
00:09:45,800 --> 00:09:49,480
widespread agreement on what those guardrails ought to be. So
150
00:09:49,480 --> 00:09:53,260
we're gonna continue to see these sorts of problems. There'll be a problem,
151
00:09:53,825 --> 00:09:57,585
There'll be an over correction. These are going to go back and forth swinging
152
00:09:57,585 --> 00:10:01,345
like a pendulum until we eventually find the equilibrium and the right
153
00:10:01,345 --> 00:10:04,570
balance between freedom of use and freedom from harm.
154
00:10:05,190 --> 00:10:08,710
I'm pretty confident that we'll get there eventually, but it's
155
00:10:08,710 --> 00:10:12,404
gonna take a while. So when you see these kinds of problems, it's
156
00:10:12,404 --> 00:10:16,005
good to be aware of them, but don't get unduly upset that
157
00:10:16,005 --> 00:10:19,820
there's some right wing or left wing conspiracy going on. I
158
00:10:19,820 --> 00:10:23,520
think most of it is just honest engineers trying to find the right balance
159
00:10:23,580 --> 00:10:26,960
between freedom of use and freedom from harm.
160
00:10:27,340 --> 00:10:30,965
So in the meantime, I think one of the big
161
00:10:30,965 --> 00:10:34,265
messages that I wanna take away and I want you to take away from this
162
00:10:34,805 --> 00:10:38,165
is be careful of relying on generative AI for anything important or
163
00:10:38,165 --> 00:10:41,720
anything or anything that might be seen by the public
164
00:10:42,339 --> 00:10:46,100
unless there's human review. The human in the loop is
165
00:10:46,100 --> 00:10:49,635
critical especially at this stage of generative AI.
166
00:10:50,175 --> 00:10:53,875
So make a human check part of the process whenever you use generative
167
00:10:54,015 --> 00:10:57,660
AI for anything important. Maybe it wasn't practical
168
00:10:57,660 --> 00:11:01,500
for Google, but for most of us, it will be. If you
169
00:11:01,500 --> 00:11:05,120
want some more details, there are a couple of links to articles about this whole
170
00:11:05,355 --> 00:11:09,115
brouhaha in the newsletter. Again, the newsletter's
171
00:11:09,115 --> 00:11:12,415
available at aighostocolllege.com. You really should subscribe.
172
00:11:14,000 --> 00:11:17,839
Okay. So here's my first tip of the week. And this
173
00:11:17,839 --> 00:11:21,360
comes from listener Ralph Estep, who's also a friend of
174
00:11:21,360 --> 00:11:25,074
mine, who sent me an email asking me why generative AI
175
00:11:25,074 --> 00:11:28,295
is so bad at following instructions about length.
176
00:11:28,915 --> 00:11:32,580
By the way, Ralph has a really good daily podcast that focuses
177
00:11:32,580 --> 00:11:35,700
on financial health. You ought to check it out. It's,
178
00:11:36,100 --> 00:11:39,940
available at askralphpodcast.com. It really
179
00:11:39,940 --> 00:11:43,365
is very good, and we all need to keep an eye out on our financial
180
00:11:43,365 --> 00:11:47,045
health, especially given inflation and some of the uncertainties in the
181
00:11:47,045 --> 00:11:50,600
world. So you have probably experienced the length problem of
182
00:11:50,600 --> 00:11:54,280
generative AI. You tell Chad GPT or Gemini or Claude or whomever to
183
00:11:54,280 --> 00:11:57,895
produce an output of 500 words, and there's no telling how
184
00:11:57,895 --> 00:12:01,595
long the output will be, but I'll bet it's not 500 words.
185
00:12:02,455 --> 00:12:06,280
When I try this, I get 200 words. I might get 750 words. And this
186
00:12:06,280 --> 00:12:09,800
can be kind of frustrating. So I wanted to understand what
187
00:12:09,800 --> 00:12:13,595
this is all about. So I asked Gemini, why is this a
188
00:12:13,595 --> 00:12:17,355
persistent problem with generative AI tools? I
189
00:12:17,355 --> 00:12:20,795
actually kind of liked Gemini's response, so I put it
190
00:12:20,795 --> 00:12:24,590
verbatim in the newsletter. I even put a link to the conversation in
191
00:12:24,590 --> 00:12:28,430
the newsletter, so you ought to check it out. But here's the bottom
192
00:12:28,430 --> 00:12:32,154
line. What you need to do is give it a range, not a
193
00:12:32,154 --> 00:12:35,675
target. So don't say 500 words, say between 3
194
00:12:35,675 --> 00:12:39,420
56 100 words, or something like that. You can provide
195
00:12:39,420 --> 00:12:43,100
examples of writing that fits the length that you want. These can
196
00:12:43,100 --> 00:12:46,845
be templates that AI can follow. Another good
197
00:12:46,845 --> 00:12:50,685
approach is to start small and then build up. Ask for a
198
00:12:50,685 --> 00:12:54,490
short summary first and then ask for more detail on specific
199
00:12:54,630 --> 00:12:57,990
points. This gives you more control. And
200
00:12:57,990 --> 00:13:01,670
so, how you phrase your request might also make a
201
00:13:01,670 --> 00:13:05,355
difference. And this is according to Gemini. If you
202
00:13:05,355 --> 00:13:09,035
say summarize this topic in about 400 words, it might
203
00:13:09,035 --> 00:13:12,860
work better than write 400 words on this topic. It's
204
00:13:12,860 --> 00:13:16,080
gonna take practice, so I
205
00:13:16,460 --> 00:13:20,220
just wouldn't rely on it ever to give me a specific number of words. But
206
00:13:20,220 --> 00:13:23,975
as you practice, you can find a way to get it closer
207
00:13:23,975 --> 00:13:27,735
and closer to the link that you want. This is kind of a good
208
00:13:27,735 --> 00:13:31,170
thing for those of us who are instructors because it's gonna
209
00:13:31,170 --> 00:13:34,770
make students actually work a little bit instead of just spitting
210
00:13:34,770 --> 00:13:38,290
out 500 word papers. Okay. Here's
211
00:13:38,290 --> 00:13:42,115
a very useful tool. There's a little bit
212
00:13:42,115 --> 00:13:45,555
of a pun there. More useful
213
00:13:45,555 --> 00:13:49,400
things, which comes from the creator of
214
00:13:49,400 --> 00:13:52,460
one useful thing, which is a newsletter I really like,
215
00:13:53,240 --> 00:13:56,665
Ethan Moloch and Lillek. I think it's
216
00:13:56,665 --> 00:14:00,504
Lillek, Moloch. They've got a new website called More Useful
217
00:14:00,504 --> 00:14:04,310
Things. You can go to more useful things dot com and find it
218
00:14:04,310 --> 00:14:07,910
there. It includes an AI resources page
219
00:14:07,910 --> 00:14:11,589
that, no surprise, includes some pretty useful AI
220
00:14:11,589 --> 00:14:15,415
resources. There are 3 sections. 1 is a
221
00:14:15,415 --> 00:14:19,255
pre order section for Ethan's book, Co Intelligence Living and
222
00:14:19,255 --> 00:14:22,880
Working with AI. I pre ordered it, and I think it's probably gonna be pretty
223
00:14:22,880 --> 00:14:26,720
good. There's other resources that has some stuff
224
00:14:26,720 --> 00:14:30,535
like an AI video, not an AI video, a video
225
00:14:30,535 --> 00:14:34,154
on AI and links to some of their research
226
00:14:34,375 --> 00:14:37,894
on AI. But what I really want to talk to you about is their
227
00:14:37,894 --> 00:14:41,460
prompt library. Their prompt library
228
00:14:41,520 --> 00:14:45,200
includes instructor aids, student exercises, and some
229
00:14:45,200 --> 00:14:48,995
other stuff. The instructor aid prompts are really pretty
230
00:14:48,995 --> 00:14:52,755
good, but they are long and they they are
231
00:14:52,755 --> 00:14:56,120
complex. For example, they've got one that will create a
232
00:14:56,120 --> 00:14:59,720
simulation and that prompt is over 600 words
233
00:14:59,720 --> 00:15:02,885
long. Look. There's nothing wrong with this. In fact,
234
00:15:03,365 --> 00:15:07,205
complicated prompts are often very effective, especially for
235
00:15:07,205 --> 00:15:10,805
more complicated tasks. But I want to be careful
236
00:15:10,805 --> 00:15:14,550
here. I don't want you to look at the complexity of these prompts and
237
00:15:14,550 --> 00:15:17,850
go, oh good Lord. I'm never going to be able to learn this.
238
00:15:18,445 --> 00:15:21,825
Don't have to know how to write those prompts, especially not at first.
239
00:15:22,285 --> 00:15:25,905
You can accomplish a lot with pretty simple
240
00:15:26,069 --> 00:15:29,430
prompts. So both simple and complex prompts have their
241
00:15:29,430 --> 00:15:33,129
places. You can start off simple, everything will be fine.
242
00:15:33,415 --> 00:15:37,255
But you really ought to check out more useful things. Even if it just
243
00:15:37,255 --> 00:15:41,015
gives you some ideas about how generative AI can be used, it's
244
00:15:41,015 --> 00:15:44,720
worthwhile checking out if for no other reason than that.
245
00:15:45,180 --> 00:15:48,400
So check it out. Okay. On to the next topic.
246
00:15:48,940 --> 00:15:52,685
Recently I was listening to an episode of Dan Shipper's How You
247
00:15:52,765 --> 00:15:56,365
Use Chat GPT. It's really good. I think it's on YouTube. I
248
00:15:56,365 --> 00:16:00,205
listen to it as a podcast. Dan was interviewing Nathan
249
00:16:00,205 --> 00:16:04,050
Lebens on how he uses chat gpt as a copilot for
250
00:16:04,050 --> 00:16:07,670
learning. Episode was very interesting. Check it out.
251
00:16:07,889 --> 00:16:11,675
But what caught my attention was a discussion of something called chain
252
00:16:11,675 --> 00:16:15,355
of thought versus few shot prompting. This is a
253
00:16:15,355 --> 00:16:19,010
little advanced, so I want you to stay with me
254
00:16:19,010 --> 00:16:22,290
here. But if it gets to be too much, just move on to the next
255
00:16:22,290 --> 00:16:26,070
segment. Few shot prompting is pretty easy to understand.
256
00:16:26,745 --> 00:16:29,885
You just follow your task description with a few examples.
257
00:16:30,745 --> 00:16:34,505
So let's say that you wanna create some open
258
00:16:34,505 --> 00:16:38,270
ended exam questions. Do you give chat GPT
259
00:16:38,330 --> 00:16:42,010
or Gemini or whomever, whomever, whatever? Is it
260
00:16:42,010 --> 00:16:45,635
whomever or whatever? Oh, that's scary. You give the
261
00:16:45,635 --> 00:16:49,154
tool of choice your prompt, say, create some open
262
00:16:49,154 --> 00:16:52,915
ended exam questions on this topic
263
00:16:52,915 --> 00:16:56,540
and give it some parameters, and then you give it 2 or 3 examples.
264
00:16:57,560 --> 00:17:01,399
Now what the AI tool will do I'm gonna say chat gpt here
265
00:17:01,399 --> 00:17:05,175
to just make it easy. What chat gpt will do is it will look at
266
00:17:05,175 --> 00:17:09,015
and analyze your examples and try to create questions
267
00:17:09,015 --> 00:17:12,839
that are similar to your examples. Sometimes just giving
268
00:17:12,839 --> 00:17:16,539
a single example is really useful. They call that one shot prompting.
269
00:17:17,240 --> 00:17:20,140
Chain of thought prompts are a lot more complicated.
270
00:17:20,954 --> 00:17:24,414
The main idea is that you ask chat gpt
271
00:17:25,275 --> 00:17:29,115
to think aloud. So I'm gonna give you an example of a chain of
272
00:17:29,115 --> 00:17:32,140
thought prompt, and this one was produced by Chatt GPT.
273
00:17:33,720 --> 00:17:37,480
Explain the concept of chain of thought prompting using the chain of thought
274
00:17:37,480 --> 00:17:41,205
approach. I thought that was clever. Begin by defining what chain
275
00:17:41,205 --> 00:17:44,885
of thought prompting is. Next, break down the process into its
276
00:17:44,885 --> 00:17:47,785
key components explaining each one step by step.
277
00:17:48,660 --> 00:17:52,500
Then illustrate how these components work together to guide an AI in
278
00:17:52,500 --> 00:17:55,560
processing and responding to complex tasks.
279
00:17:56,875 --> 00:18:00,555
Finally, conclude by summarizing the advantages of using chain of thought
280
00:18:00,555 --> 00:18:04,335
prompting in AI interactions, especially in educational
281
00:18:04,715 --> 00:18:08,510
contexts. And then the result was pretty long. I'm
282
00:18:08,510 --> 00:18:12,190
gonna have to send you the newsletter send you to the newsletter to check that
283
00:18:12,190 --> 00:18:15,565
out, but this can be
284
00:18:15,945 --> 00:18:19,705
a good way to really get chat
285
00:18:19,705 --> 00:18:22,960
gpt to do more complicated
286
00:18:23,179 --> 00:18:26,539
things. I don't use chain of thought prompting very
287
00:18:26,539 --> 00:18:30,059
much. I think few shot prompting works really
288
00:18:30,059 --> 00:18:33,345
well, but few shot prompts
289
00:18:33,485 --> 00:18:36,625
require knowing what good output will look like.
290
00:18:37,325 --> 00:18:40,840
If you're not sure what you want, you might consider chain of thought
291
00:18:40,840 --> 00:18:44,380
prompting. But if you're a beginner, stick with
292
00:18:44,440 --> 00:18:47,420
few shot prompting. Even one shot prompts
293
00:18:48,414 --> 00:18:51,554
really are quite useful. I use that quite a bit actually.
294
00:18:52,095 --> 00:18:55,820
Okay. So there are a couple of messages here. First, keep
295
00:18:55,820 --> 00:18:59,260
things simple when you can. Simple is often very
296
00:18:59,260 --> 00:19:02,940
effective. 2nd, don't be afraid to experiment with different
297
00:19:02,940 --> 00:19:06,735
approaches and even to blend approaches. So
298
00:19:06,735 --> 00:19:10,275
you can decide. Your prompts can be simple or they can be complicated.
299
00:19:11,530 --> 00:19:14,910
All right. Here's my favorite little part of this episode.
300
00:19:15,370 --> 00:19:19,070
This is, I think, the best $40 you can spend, and it's $40
301
00:19:19,210 --> 00:19:22,785
a year, on your generative AI use and productivity.
302
00:19:24,045 --> 00:19:27,760
Look. I know monthly software subscriptions are totally out of
303
00:19:27,760 --> 00:19:31,360
hand. I don't even wanna know how much I'm spending at 5,
304
00:19:31,360 --> 00:19:35,015
10, 15, $20 a pop every month. But
305
00:19:35,015 --> 00:19:38,795
despite this, I recently added a $40 per year subscription,
306
00:19:39,655 --> 00:19:43,280
and it's already proven to be one of my best software investments ever.
307
00:19:43,840 --> 00:19:46,660
Alright. So what is this great investment?
308
00:19:47,440 --> 00:19:50,880
Well, it's for a text expander, and I wanna give a shout out here to
309
00:19:50,880 --> 00:19:54,485
Dave Jackson of the school of podcasting who talked about this on one of his
310
00:19:54,485 --> 00:19:58,005
episodes. And so a text expander just
311
00:19:58,005 --> 00:20:00,825
replaces a short bit of text with a longer bit.
312
00:20:01,870 --> 00:20:05,470
For example, if I want to type the web address for AI Goes TO College,
313
00:20:05,470 --> 00:20:09,275
I type semicolon uaig. U is
314
00:20:09,275 --> 00:20:12,875
short for URL. And the text expander, it just gives the full
315
00:20:12,875 --> 00:20:16,555
address. The semicolon here is used to indicate that
316
00:20:16,555 --> 00:20:19,940
what follows is gonna be or could be an abbreviation for a text
317
00:20:19,940 --> 00:20:23,780
snippet. The semicolon works well because it's usually followed by
318
00:20:23,780 --> 00:20:27,165
a space rather than characters, but you could really use anything you
319
00:20:27,165 --> 00:20:30,625
wanted. Now this doesn't save me a lot of time,
320
00:20:31,005 --> 00:20:34,685
but it saves me 5 or 6 seconds every time I wanna
321
00:20:34,685 --> 00:20:37,820
type the AI goes to college website.
322
00:20:39,320 --> 00:20:43,160
And it's long. You know, it's got the HTTPS, colon, etcetera,
323
00:20:43,160 --> 00:20:46,135
etcetera, etcetera. It just takes a while.
324
00:20:46,595 --> 00:20:49,975
I have to give people my biography
325
00:20:51,235 --> 00:20:54,880
periodically. Matter of fact, once a month or so somebody asks for
326
00:20:54,880 --> 00:20:58,560
it. So normally I go find the Word file and attach it to an
327
00:20:58,560 --> 00:21:02,355
email or copy and paste it into the message. Now I just type semi colon
328
00:21:02,355 --> 00:21:06,195
bio and my bio pops up. And I
329
00:21:06,195 --> 00:21:09,955
use this for student advising. I use this for grading.
330
00:21:09,955 --> 00:21:13,610
If you're a professor, you have to grade a lot
331
00:21:13,610 --> 00:21:17,390
of kind of projects, that sort of thing. A text expander
332
00:21:17,530 --> 00:21:21,054
will change your life when you have to grade. My spring
333
00:21:21,054 --> 00:21:24,835
class has 85 students in it, and I'll grade 85
334
00:21:24,975 --> 00:21:28,355
projects for twice. What is that, A 170
335
00:21:28,575 --> 00:21:32,120
projects? And a lot of the feedback will be
336
00:21:32,120 --> 00:21:35,960
exactly the same. You know, they forget to use the 1,000 separator,
337
00:21:35,960 --> 00:21:39,795
the comma, or their their spreadsheets aren't formatted
338
00:21:39,795 --> 00:21:43,095
well, that sort of thing. Well, now I can just in a few characters
339
00:21:43,955 --> 00:21:47,790
pop in the feedback and the number of points I'm taking off for that.
340
00:21:48,890 --> 00:21:51,935
So what does this have to do with generative AI? Well, as you start
341
00:22:00,575 --> 00:22:04,380
I Now I just type semicolon w d y
342
00:22:04,380 --> 00:22:08,000
t. I have one for please critique
343
00:22:08,060 --> 00:22:11,515
this text. I have one for this is really
344
00:22:11,515 --> 00:22:15,195
lazy. I have one for thank you. And you'll
345
00:22:15,195 --> 00:22:18,860
find more and more and more uses for a text expander once you get
346
00:22:18,860 --> 00:22:22,380
into it. So I use generative AI for a lot of
347
00:22:22,380 --> 00:22:26,140
tasks, so this helps a lot. But it also helps when I
348
00:22:26,140 --> 00:22:29,945
need to provide some context. So for example, I try
349
00:22:29,945 --> 00:22:33,705
to tell generative AI whether I'm working on something related to my teaching or
350
00:22:33,705 --> 00:22:37,539
one of my podcasts or this newsletter. When I'm working
351
00:22:37,539 --> 00:22:41,299
on the newsletter, I have a whole blurb. It's, I
352
00:22:41,299 --> 00:22:45,115
don't know, 5 or 6 sentences long. Now, I just type
353
00:22:45,115 --> 00:22:48,735
in an abbreviation. That's it. So it's
354
00:22:48,875 --> 00:22:52,549
semicolondaig, which just means
355
00:22:53,570 --> 00:22:56,870
description of AI goes to college and this whole
356
00:22:57,570 --> 00:23:01,404
bit of context pops up. So if you're not using a
357
00:23:01,404 --> 00:23:05,245
text expander, you really ought to consider it. I use one called you're
358
00:23:05,245 --> 00:23:08,860
gonna love this text expander. I use it. That's the one that
359
00:23:08,860 --> 00:23:12,000
costs about $40 a year. I use it because it's cross platform.
360
00:23:12,539 --> 00:23:15,260
It'll work on a Mac. It'll work on a PC. I think it may even
361
00:23:15,260 --> 00:23:18,215
work on smartphones, although I haven't tried that yet.
362
00:23:19,315 --> 00:23:22,995
So I really would encourage you to consider making that
363
00:23:22,995 --> 00:23:26,700
investment, not just for generative AI, but for your
364
00:23:26,700 --> 00:23:30,460
use in general. Okay. All right. Last thing I
365
00:23:30,460 --> 00:23:32,559
wanna talk about is the interview I had
366
00:23:42,809 --> 00:23:45,389
AI Goes to College podcast, which is surprisingly
367
00:23:46,250 --> 00:23:49,929
available at ai goes to college.com. We talked a
368
00:23:49,929 --> 00:23:53,034
lot about a lot of different things. Rob is a really smart guy. He's got
369
00:23:53,034 --> 00:23:56,815
a lot of experience. He's a department chair. He's a fantastic
370
00:23:56,955 --> 00:24:00,475
researcher. So he's used generative AI in a lot of
371
00:24:00,475 --> 00:24:04,220
different contexts. We talked about how he's using
372
00:24:04,220 --> 00:24:07,279
AI to create more creative assignments and to generate questions,
373
00:24:08,140 --> 00:24:11,935
how he's helping students learn to use AI tools tools to
374
00:24:11,935 --> 00:24:15,695
explore and understand important concepts. We talked about the
375
00:24:15,695 --> 00:24:19,535
importance of being willing to experiment and fail with
376
00:24:19,535 --> 00:24:23,270
generative AI. We discussed
377
00:24:23,490 --> 00:24:27,270
why it's important to help students become confident but critical
378
00:24:27,330 --> 00:24:31,010
users of these fantastic new tools, and we talked about a lot of other
379
00:24:31,010 --> 00:24:34,455
things. So go to ai goes to college.com/rob,
380
00:24:36,035 --> 00:24:39,635
r o b, and you can check out the entire interview. I'd
381
00:24:39,635 --> 00:24:43,380
love to hear what you think about it. You can email me at
382
00:24:43,380 --> 00:24:47,140
craig@aighostocollege.com. Let me know if you've got
383
00:24:47,140 --> 00:24:50,945
any ideas for future episodes or if there's something you wanna see in
384
00:24:50,945 --> 00:24:54,625
the newsletter, and you might get featured. Alright.
385
00:24:54,625 --> 00:24:58,300
That's it for this time. Thank you. Thanks for listening
386
00:24:58,300 --> 00:25:02,140
to AI Goes to College. If you found this episode useful, you'll love
387
00:25:02,140 --> 00:25:05,815
the AI Goes to College newsletter. Each edition brings you
388
00:25:05,815 --> 00:25:09,575
useful tips, news, and insights that you can use to help you figure out what
389
00:25:09,575 --> 00:25:13,335
in the world is going on with generative AI and how it's affecting higher
390
00:25:13,335 --> 00:25:16,960
ed. Just go to ai goes to college.com to sign
391
00:25:16,960 --> 00:25:20,799
up. I won't try to sell you anything, and I won't spam you or share
392
00:25:20,799 --> 00:25:24,345
your information with anybody else. As an incentive for
393
00:25:24,345 --> 00:25:28,025
subscribing, I'll send you the getting started with generative AI
394
00:25:28,025 --> 00:25:31,830
guide. Even if you're an expert with AI, you'll find the guide
395
00:25:31,830 --> 00:25:34,330
useful for helping your less knowledgeable colleagues.