YT YouTube2Text

YouTube2Text → Transcripts → wk06 2 PromptEng iat111

wk06 2 PromptEng iat111 — Transcript

by ivizlab · 6,511 words · 970 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Welcome to week six of IIT 111. I'm
0:04Steve DePaulo. Today we're going to
0:08generally finish off with large language
0:10models and AI used as words. There's AI
0:14in so many other areas. And surely the
0:16visual area is one we haven't fully
0:19discussed yet. Um, I will be definitely
0:22talking about more details with large
0:24language models, but more so um
0:28very new ways to work with them.
0:31So, um
0:33surely we've talked about how large
0:36language models are moving beyond words
0:38and are being very multimedia oriented.
0:40Um, you're
0:42uh if you're taking the class,
0:44assignment two is due
0:47um this week.
0:49Um
0:50and then we're going to move very
0:51quickly to assignment three. I'm going
0:53to be talking about both of those
0:56today. Not just from an assignment
0:58standpoint, but from literally
1:01uh uh
1:02foundations of AI point. So, um
1:06assignment two is prompt engineering.
1:09I'm going to review it one more time.
1:11And if you're using those techniques, I
1:13think you're about 90% better than most
1:16people using prompt engineering. But
1:19assignment three will be these new
1:21techniques. In this I I've been doing
1:24workshops on it. So, I call it activity
1:27one and activity two. So, in our class
1:29it's assignment two and three that's
1:31due, but you'll see um
1:34I'll write it the other way. So, let's
1:36look at this in large language wrap up,
1:38a little bit about visuals, and then
1:40moving on to these new ways of doing
1:42things. So, surely we've made it pretty
1:44clear that these large language models
1:47are moving way beyond simple or even
1:50reasoning, and they're moving into
1:52autonomous agents. We've talked a lot
1:54from an ethics standpoint. And in the
1:57news version of what the issues are with
2:00agents.
2:01We don't always refer to it that way,
2:03but an agent is something surely that
2:05can access different things, but surely
2:07the web,
2:08doing code, and I'll be starting to talk
2:10more and more about vibe coding in
2:12classes
2:13in the class, and other forms. So, let's
2:17look at this new convergence where LLMs
2:20are now reasoning plus tool use and
2:23multimedia. And in a way changing how
2:26they think from sense to reasoning to
2:28act or reflect. Notice this is a very a
2:30more human way of of how we're
2:33intelligent that we're starting to build
2:35into our systems that just started
2:38started as systems that you know, came
2:41up with the next word and then again and
2:42again.
2:44So, we're beyond just a prompt and you
2:47get an answer. Now we have persistent
2:49context that remembers your prior chats,
2:52docs, um maybe even your style. All
2:55those things if you haven't and you're
2:57using some of these systems, you're able
2:59to build into the system.
3:02There There are settings where you can
3:03say this is the style I like. I don't
3:05want to see end dashes. I want to be
3:07thinking this way. It'll put all that in
3:10for you.
3:11Um including getting rid of things that
3:13it
3:14that it knows about you that maybe you
3:16think
3:17isn't helpful.
3:18Uh we talked about tool orchestration,
3:21multimedia, live web retrieval. Um
3:25and where it's getting to the point of
3:26almost a type of reflection system
3:30hopefully for you and your drafts. And
3:33into vibe coding, which I'll show you
3:34examples of.
3:37Um
3:39So, we definitely have the core model.
3:41Um now this agentic layer going out and
3:43doing these things, the memory and
3:45context manager, and [clears throat]
3:48something processing when a visual shows
3:51up. Remember when it goes to the web or
3:53it has to analyze a picture that just
3:56might be in your PDF, it is doing
3:58specialized stuff and typically just
4:01getting the the text description of
4:03those things, processing it separately,
4:06and then putting it back in the context
4:08window that we've talked about a lot.
4:11Um
4:13So, if you say summarize this PDF and
4:15make a chart and draft a report, it's
4:18got to read and parse an uploaded PDF.
4:21If there's vision, you know, using text,
4:23but if it sees an image, it might switch
4:25over to a vision system, turn that back
4:27into text, run maybe some Python program
4:30to plot that data,
4:32uh write a summary for you, maybe cite
4:35sources on the web. So, it's really
4:38starting to act like a research
4:39assistant, not a chatbot. We want to
4:41start using more and more of that.
4:43Besides that, um
4:45it's doing a lot of reflective
4:47reasoning. And the most obvious one is
4:50Anthropic's Claude. It rather than just
4:52say, you know, with with guidelines of
4:55don't do that, [clears throat] don't do
4:56this, um do that and do this, it
4:59actually now has constitutional AI
5:01that's been in there for a while, where
5:03it's a training method that teaches the
5:05AI how to be most helpful and harmless
5:07and honest by following some ethical
5:10principles or constitution.
5:12It hasn't
5:14totally been a lifesaver where
5:16everyone's thinking, "Wow, that's
5:17wonderful. We should be using it, too."
5:19But it is a technique to give it kind of
5:22a bigger document about this is the way
5:25an AI should act.
5:28So, you're no longer just chatting,
5:30you're orchestrating, and that's what
5:31I'm mainly going to be talking about
5:33today in what I call assignment three or
5:35activity two.
5:36Uh you find, you analyze, you visualize,
5:38you summarize, you refine. Who's doing
5:40that? There's a way that you could just
5:42have an agent do that and get an answer
5:44back to you. I'm saying, "No, that's not
5:47this newer what I call human in the loop
5:50or cognitive conducting that we've now
5:52in our my research lab and others have
5:55put into our work. Um I'm going to use
5:58show you a system called FURK, which is
6:00a way to to do this. I'm [clears throat]
6:03on a paper on it with others. I work
6:05with people in Denmark. I'm really
6:07rethinking the lab. Where does this come
6:09in
6:10quite significantly? Well, it comes in
6:12with this notion this ethical problem
6:15along with energy and privacy and
6:18stealing your data. Um it comes in with
6:22something you've seen me talk about a
6:23lot, which is this cognitive decline
6:25that that if you're doing nothing but
6:28asking it to do your homework or to get
6:31answers, you know, you're not really
6:33being the human you're supposed to be.
6:35Instead, this conducting is this notion
6:38that you have many AIs all there to help
6:41you think. So, we'll talk more about
6:43that. But, it's always tricky when
6:47um
6:47who the agency is. Are you doing it or
6:50are you you're letting it
6:52>> [clears throat]
6:52>> uh uh an agent do part of it? So, we'll
6:54talk more about that. Uh are you
6:56approving the actions? Um
6:58what happens when it goes off and does
7:00something you don't want? Remember the
7:02name agent is an interesting term. If
7:05you think of celebrities or movie stars,
7:08they have agents. And those agents can
7:11act uh less ethical
7:14than So, if Brad Pitt says to a famous
7:17director,
7:19um I anything, I want to be in your
7:21movie. And the famous director says, I
7:23don't have the budget. And Brad Pitt
7:26says, I don't care. I'm here for you.
7:28That all sounds good. But, then the
7:31director's company gets on the phone and
7:34the agent says, Brad needs more money
7:36and can't do it. You have to hold off on
7:38the movie till this. And he's is one who
7:40plays hardball. So, there is this notion
7:43that the agents don't have to be,
7:45{quote} as nice or ethical as the people
7:49asking them to do. This happens in
7:51business and and other places as well,
7:53too. Do we want that? How much so?
7:56Um, just getting back to reminding about
7:59a large language model that it's
8:01encoding things like dog into some kind
8:04of vector representation, some numerical
8:07representation. And it's doing that for
8:09all the words, um,
8:11or part of a word. A token is is a word
8:13or a part of a word. Um, and that's what
8:16brings the speed and the the the ability
8:19to do the logic that's in deep learning.
8:23Um, I showed you a lot with neural
8:25networks. I just want to remind you that
8:26neural networks are no longer one neural
8:29network. It's a number of small things
8:32all doing pieces. This is an example of
8:35an encoder-decoder.
8:38You know, where each one of these, uh,
8:40blue or orange things are actually very
8:44simple
8:45um, neural networks that have a job to
8:47do. I showed you last week this notion
8:51of we're starting to move towards
8:53meaning because when you can do text and
8:55image pairs, so with text of, you know,
8:59a puppy in the grass,
9:01um,
9:02that would be, again as I've mentioned,
9:04the text would be that statement. And
9:07then this would be the picture. This
9:08would get a perfect one because the text
9:10matches. And then there could be one of
9:13a
9:14a cat, uh, a jumping. So, that would be
9:17this text and there would be a picture
9:19of it. And that would be a perfect one.
9:21But, how would it be if it's the text
9:24talking about the cat, but showing the
9:26dog picture? We run these things,
9:29um, quite significantly through
9:31everything and we're getting this this
9:33nearness waiting now, where certain
9:35things are more similar than others. So,
9:38we're moving towards meaning in a way.
9:40We call this an embedding space, and
9:43often this embedding space um we don't
9:46always, when we do a new model, start
9:48from scratch. We might have a good
9:50embedding space, and if we only wanted
9:52to write, say, a tool where you could um
9:58put your camera up to trees and leaves
10:02in the BC area, so it could understand
10:05what kind of tree it is just by showing
10:06it, we would put that on top of this
10:09embedding space. So, it's learning in
10:10general about the world, and then we
10:13would put our details on it. So, this is
10:15almost always what happens now. If you,
10:18maybe for your project, are trying to do
10:20a new kind of model, you wouldn't start
10:23from scratch at all. You'd find the
10:25closest embedded um
10:28vector space to your area, and there's
10:30tons of them out there. Um so, an
10:32embedding space is a vector space where
10:35similar meanings are positioned close to
10:36each other, like we just talked about.
10:38It's a simple way of saying
10:40uh um these are built up, and then we
10:42only tend to find the closest one to
10:45what we want. Maybe in that case,
10:47there's something about all trees and
10:50and flowers and leaves, and you just
10:52have to put the um
10:55the BC uh your smaller trained thing on
10:58top of it. That's what we're doing now.
11:00We're building the top of of the space
11:03now. But again, I wanted to remind you
11:05then that it's using that nearness now
11:08with things, including positioning
11:11encoding. So, that would be that it LM
11:14builds up not just like what the word
11:16is, but where the word um builds up the
11:18first the word, the position of the word
11:20in the in the sentence word in the
11:22sentence or phrase or a meaning based on
11:25the text image pair. And this is why um
11:28these systems are getting smarter and
11:30smarter, that it's it's where and how
11:33that works.
11:35And you get we get different vectors
11:37attention vectors
11:39for these. So it seems like it's just
11:41dog but there's different variants of
11:43dogs. And I've talked about this in
11:45other ways
11:46with this and this is in my grad class
11:49we get into this this attention we call
11:51it.
11:52Very famous paper called attention is
11:54all you need as a way to get some of
11:56this to work. But you know it's the
11:59understanding this two words litter
12:01here. You know this is the cat litter.
12:04So this litter is tied to who is using
12:07the litter where this this litter is
12:10you know
12:11these are my mates.
12:13Um
12:14The use is you know what what's left
12:16what's left behind. So all this is
12:19getting smarter. Each one of these cat
12:21would start with one vector of the word
12:23cat and then would grow to all these
12:25other meanings.
12:27So we do other things in modern
12:30because that's a good for training but
12:32as we're starting to get something we
12:33want to clean it up. We're actually
12:35giving it documents and having it almost
12:38dream a new document and seeing gets the
12:40next word. It feeds that back does it
12:43again. And in a way it's it's dreaming
12:45these documents and then it's testing
12:47it. Is this a good document or bad
12:48document? So we're doing everything we
12:51can to make
12:53the final large language model
12:56or the model in general strong. Here is
12:58the model here's the word dog here or
13:00this there or that there. And it just
13:02get pulled to different directions and
13:04how do we own this best? Like why is it
13:06going in a particular place for the next
13:08word or phrase or thinking? And I think
13:11I've told you that it's used
13:12reinforcement learning with human
13:13feedback. It started with this human
13:15feedback where we're giving these things
13:17to people to kind of go is did we write
13:19this well? How would you write it
13:20better? We have so many of them now
13:23we're using a different kind of AI
13:24called reinforcement learning that kind
13:27of is built on rewards and errors.
13:29Um it's a mean machine learning that
13:31technique that aligns models to the
13:33human values by human feedback, by rank,
13:36rate, and refine options. Rate is
13:39sometimes oh, we're going to give a
13:41human two choices and they prefer this
13:43one and we're going to learn from that.
13:45So, step one is these these document
13:47these words. Step two is question
13:50questions about the answers. Was this
13:52good? Um and then can we refine that and
13:55make it better? [clears throat] So, it's
13:57literally trained by people with Q&As
13:59now. Sometimes you get this yourself and
14:02you know, Claude or chat PT say, I'll
14:04give you two answers, which one do you
14:05like better? So, they're even doing that
14:08testing
14:09um to general users.
14:12So, we know we have the pre-training
14:13which uses a huge amount of information.
14:16I pulled slides that I talked to about
14:18my grad class.
14:19Uh if you're interested, we can always
14:20talk about it more. Exactly how big this
14:23data set is, how many GPUs you need, and
14:26then just this gives you a base model.
14:28And then we clean that base model with
14:31understanding uh um
14:33exactly what are good phrases and
14:36sentences and humans coming in cleaning
14:38that. So, the training is very
14:40expensive. It used to take over a year.
14:43We're down to that. And then the fine
14:45tune fine tuning takes a you know, a
14:47month or two. Uh fixes bad behavior,
14:51uh puts in corporate bias and other
14:53things. I talked about it with llama in
14:55a deeper way. Here's this comparison
14:58thing. Which one of these do you like
15:00better? We you know, they were paying
15:02people to decide that and really
15:04cleaning up what people would like to
15:06see. Um so, and then starting to do that
15:09automatically.
15:11Um so, that's again I just did a kind of
15:14review with a little bit extra of what
15:17makes a modern word-based system. So,
15:20what about as we get into visuals? Well,
15:22some of the things are very similar.
15:24It's still deep learning. It's still a
15:26large data set. Now, it's doing pattern
15:28recognition as well, but more in the
15:30visual space. And it's definitely doing
15:32generation, generating pictures.
15:35Uh but instead of predicting next world,
15:37the visual AI is learning to create
15:40images from noise. So, remember this
15:44neural network that we've always talked
15:46about. And this is the visual space that
15:49at some point you want to understand um
15:51the notions of what this is by doing the
15:54low level to the high level. We went
15:56through that. So, they don't want to
15:58keep going there. But, what we're going
16:00to start next week is this notion of um
16:04that large language models generally use
16:06that transformer encoder-decoder that I
16:08just showed you.
16:10Um and that predicts predicts the next
16:12token or word in the sequence. Visual AI
16:16and this is a little bit older. It's
16:18done a little bit differently now, but I
16:20need to explain it kind of how it was a
16:22year ago. Uses a diffusion model, which
16:25is a different kind of system. So, a
16:27diffusion model remove removes noise
16:30step-by-step to reveal an image.
16:33Um so, transformers versus diffusion.
16:36Transformers are used for words.
16:38Diffusion is is used to um come up with
16:42images by by doing this noise technique.
16:45Again, I don't want to get into it now,
16:47but uh as we get into it, we're starting
16:50to change this where we're using
16:51transformers
16:53to also predict pictures that are still
16:56using diffusion. So, the we're starting
16:58to do both for the images. So, what do
17:00we mean by diffusion? AI gets a picture
17:04and uses a particular technique to
17:07gradually add noise to it to it as
17:09nothing but noise. And there's an
17:11equation
17:13kind of a Markov chain is is this
17:15mathematical equation to actually get
17:17there. Then we just have the noise and
17:20you can reverse that process.
17:22We call it denoising. You'll see that in
17:24a lot of tools. How much denoising, how
17:27many steps you take to go back to to go
17:29back. Because we broke this down with an
17:32equation starting from noise, this is
17:35now very fast. We don't even have to
17:37save all those images anymore. We just
17:40save the equations that brought it to
17:43noise and then we can go back which This
17:45is great because everything, every image
17:48when it's trained goes to noise plus the
17:51equation that got it there. Therefore,
17:53you're just keeping little equations.
17:56And then everything starts from noise.
17:59So if you wanted to do a cat that looks
18:01like Einstein, you would have the noise
18:04both of them would start from noise and
18:06you would just kind of develop their
18:07equation to get next. That's generally
18:10how those systems work. I'm going to
18:12hold off on that. You know I do a lot of
18:15work and have a number of papers on
18:17visual systems including real-time
18:19visual systems and and the problems with
18:22these systems. Um we'll hold off on that
18:25till next week. So now I'm just going to
18:28That was my little review. I'm going to
18:29go right to
18:32uh the last of large language models
18:34which is all about prompt engineering.
18:36So this is the part of course where I
18:38get you to be a better user of AI. I was
18:42just explaining AI to you. What I call
18:44foundation work, but this is now how
18:47could you be better?
18:48So first of all, things have shifted
18:50since 2026. We definitely have multiple
18:53multi-agent orchestration, this
18:55iterative and reflective things, and the
18:58notion of a rag which is again is a way
19:00to put a very big document in by
19:02chunking it into pieces and only using
19:05the meaning pieces that are needed at
19:08that moment. So that today's two
19:10activities, I'm calling them activity
19:12one and two, even though your assignment
19:16uh sorry, activity two uh sorry,
19:19assignment two and three, activity one
19:21and two. Um is how you can shift from
19:25what we've been talking about, this
19:27special this specialized assistant, to
19:29this much more structured human AI way.
19:32I think everyone should be doing this
19:34more complex way. It really will put
19:37humans back in the driver's seat, stop
19:41the cognitive decline that's happening
19:43in high schools and other places by
19:45using AI. So, I'm going to first root
19:48So,
19:49um
19:50as we get into it, I'll just remind you,
19:52we now know these external apps and we
19:55know that this this this this agent way
19:59has risks, but it's moving from advising
20:02to acting now. Um and we've talked about
20:04the agentic system again. I know I do a
20:07lot of review, but we have people at all
20:09levels in this class, so I'm trying to
20:11make sure I don't lose people. So, the
20:13agent system can go out there and it has
20:15the AI chains, and we talked about the
20:18real problems, current problems. But we
20:20know with that, we can put in PDFs and
20:23URLs and videos now, so this is the
20:25multimedia side. It isn't just It isn't
20:28just text anymore. And we can build
20:31those up and we can do things with them.
20:33And in activity two, you'll see that
20:35you're going to now create all this. So,
20:38um we know that these these models have
20:42a cutoff date. So, sometimes they have
20:45to go out on the web to get new
20:46information or specialized information.
20:49So, if you run a particular company or
20:52for your project, you you want to
20:55uh give it specialized information that
20:57you're not sure the main model would get
20:59right, you would use these rags learning
21:02from documents. I will talk more about
21:04how in my lab, we're doing very specific
21:08technique with very advanced rags like
21:11we have with Mango and others where
21:13we're we go with domain expertise
21:16in a field. So if you're domain expert
21:19of
21:20stock trading or something like that,
21:22you could put in a rag that isn't just
21:24to grab the information, but it's almost
21:26like you're in there
21:28helping the AI be
21:32get stuff that is you. Um
21:35the context window is only so big. So
21:38this is where we use the rag and we've
21:41talked a bit about this. So we now know
21:44that these systems have persistent
21:45memory. They know about prior chats and
21:47docs and user styles. We know that
21:49they're doing this level of
21:50orchestration. And with all that, it's
21:53this web retrieval and self-reflection
21:56that I want you to start thinking of.
21:58So we're no longer chatting, we're
22:00co-orchestrating.
22:02Reasoning agency,
22:04multimedia, and the blur of you letting
22:07your agent do it or you're keeping an
22:10eye on your agent and you're doing it
22:11yourself. This assignment is to think
22:14through these cuz there's real ethical
22:16issues how you do that.
22:18Um
22:19So we've talked about rags before and
22:22the context window, so I'll just do this
22:23quickly. Again, a rag is cutting up a
22:26document into pieces and then and then
22:29saving it so when you're giving a query,
22:33it could it could kind of go, "Hey,
22:37chunks of the document,
22:39does anyone have something to say about
22:42this current issue?" And they go, "Oh
22:43yeah, chunk five, 10, and 100." So then
22:47it could use those and that's how it
22:49it's smarter without putting the entire
22:51document in.
22:52So let's now go to the activities.
22:55So for those in the class right at this
22:57moment, reading it this week, you
22:59actually have assignment to do in days,
23:03so you might have done it already, but
23:05I'll do a quick recap. So, when I get to
23:07activity two, which is assignment three,
23:10you'll you'll be ready for it. So, I'm
23:12going to do this very quickly. So, the
23:14key is vague in is AI slop out. So, the
23:17more you can explain things the better.
23:20My new thing is, if you don't have the
23:22domain expertise, people with domain
23:24expertise in an area using Claude or
23:27ChatGPT really do well. People without
23:30it don't cuz it if you just give it
23:32vague stuff, you'll get some really
23:35basic moving towards the most pop
23:38culture slop. So, how do we do that with
23:41persona assignment chain of thought few
23:43shot? I've talked about this many times.
23:45I just have it in here to kind of help
23:48remind you. Persona, you give it a
23:50specific role identity, it does better
23:53because of that. Chain of thought is the
23:55step-by-step system, so you're trying to
23:58get reasoning. This is the one that
24:00we're changing in 2026.
24:02And I, you know, when you could be
24:04thinking about this for your assignment,
24:06it's still nice to give it steps and
24:07explain what you want it to do, but it's
24:09kind of doing that anyway. So, the
24:11better one
24:13with this is to give the model the
24:15actual document to reason over. So,
24:18rather than say I want you to do this,
24:19that, and that, you you actually do a
24:22document. Now, when you're doing this
24:23well, what you're really doing is you're
24:26using AI, you're getting to some new
24:29idea that you never had before. And you
24:31stop at that point and you tell the AI
24:33AI, "Let's summarize this into a nice
24:37markdown file.
24:39Uh and I will use that in in, you know,
24:41when I come back." So, this is the
24:43notion that you have to start getting
24:45into. Did something good just happen? I
24:47should be saving that into a document.
24:50And I never say just summarize it in the
24:52middle of a big thread with Claude. I
24:55always say, "Hey, don't bother with this
24:58part and that part, but these three
24:59things we did are really good. And I
25:01don't I don't want to go back and figure
25:03out what it all is. You do that for me,
25:05but I want to see one this, two that,
25:07this, and write it in a way at this
25:09level." And I get this amazing document
25:12that then I could reuse again. And I
25:14have a lot of those, and I just keep
25:16doing it. So, that's kind of the
25:17slightly newer chain of thought.
25:19Few-shot again is this notion of just
25:21give it examples. How would you have
25:24answered the question? What do you want
25:26to see in example form? This one
25:29I think works the best out of these
25:31three. So,
25:33um
25:34since models reason on their own, the
25:36step-by-step thinking is less important.
25:39It's a little bit better than these
25:40documents. Long chats get noisy. So,
25:43this is saying what I already said. Go
25:46into a long chat, take the best part of
25:48it, make it a document, maybe go to a
25:50new AI, and start it all over again.
25:54Okay. So, um
25:56and then this the reason is and you can
25:58spend time reading this again is that
26:01because the context window gets too big
26:03and really smart
26:05documents that you're putting in, it's
26:07not only seeing everything in the
26:09document, it's using it by retrieving
26:12the important thing at the time. This is
26:14why this is really the best thing to be
26:16doing.
26:18Uh you put all those three together, and
26:20that's your assignment, your
26:21professional warm department
26:23coordinating with clear and concise
26:25messages. Let's think, first do this,
26:27then that. Here's some examples of what
26:29I want to see. This is what we expect.
26:32Um these are the three techniques. I
26:34call this the specialist. I did that
26:36quick because I want to move on now. Um
26:39so, I I did it here, create a 6-week
26:42physics tutorial, and I have examples of
26:44both of these. I think I've showed you
26:46this in class. So, I'm now going to move
26:48on. So, that's what's due in assignment
26:51two.
26:52And I'm calling that activity one
26:53because this is what I do in workshops
26:56and sometimes to staff in your very
26:58schools.
27:00Surely for business and
27:03other other schools. So again, as we now
27:06move on, so that's what you're still
27:08doing
27:09assignment two,
27:11just do that and don't get confused
27:13about these next things. This is for
27:15assignment three, but this is how I like
27:16everyone to think and when I do
27:18workshops for those workshops I do for
27:21the vice president research
27:24to to this to people in the school. So
27:27that we could be the best school in AI,
27:30blah blah blah. So old thinking, AI
27:33write me a report, accept whatever I
27:35give you, you know, I'll accept whatever
27:37it is, one prompt, one output. No, no,
27:39no. Now you're the conductor, AI is the
27:42orchestra. Imagine you're in front of a
27:45lot of AIs and you can say, "Hey, the
27:47violin section, I want you to summarize
27:49this this section. I want you to see
27:51what they did and, you know, do the
27:54octave above it or, you know, do it." So
27:57AI help me with research X. Now
27:59critique it.
28:01Now double-check a specific
28:03claim. So let's look at this in detail.
28:06Um so it's directing a collaboration. So
28:10in one, you just gave it a role and
28:12examples. You still can use the
28:14techniques of of the specialist
28:18technique that was in assignment two for
28:20this new one, but now you're going to be
28:22using this human AI collaboration. I
28:25call it human in the loop. There's
28:27several ways to do that. I like to be
28:29creative, so I don't always like to be
28:31locked into one way. It turns out my
28:34colleagues in Denmark and I
28:37um are using this FURK method. So I've
28:40decided this is the best one where
28:42you're framing the problem.
28:44You're then exploring ideas of it.
28:46You're fighting it. You're commit. So,
28:48you go through these stages one by one
28:50and you don't just get a direct answer.
28:53It's human in a loop. It's now
28:55systematic as a way to you work through
28:57it. So, fur comes from the hybrid
28:59intelligence research project. If you
29:01want to see more of it, they have a lot
29:02there. We're trying to do some user
29:04studies
29:06um and you can always read our paper on
29:07it as well. So, frame
29:10explore, find, commit. Frame is
29:12articulate the content. You know, it's
29:13usually saying get me
29:17get me five papers on the thing I care
29:19about and maybe tell me the most
29:22important things about them. Maybe get
29:24me, you know, five
29:27um
29:29uh quarterly reports on chip
29:32manufacturers cuz I'm thinking of maybe
29:34seeing where chip manufacturing is
29:37going. Get me the Nvidia one and the and
29:39the this one and that one. And read all
29:42the papers or the and
29:44and maybe uh see where they agree and
29:46where they disagree disagree and give me
29:48that. Uh so, you're exploring with that.
29:52Where I will not give you credit and um
29:55if you do nothing but do these things
29:57and never make human decisions in
29:59between them. So, the whole point of
30:01this, whether you're in the class or
30:02not, is to say, "Oh, I just You just
30:05gave me the five top things that I was
30:07looking for. I just read them
30:10uh in summary form and I don't like I
30:13never thought about four. Like and I
30:15don't think much of two. So, let's never
30:17talk about two again and four is
30:19amazing. Maybe I'll just Maybe we just
30:22do the whole assignment over again based
30:24on four. Why don't you get me four
30:25papers just on what we learned in point
30:28four and then they're going to come back
30:29again.
30:30Uh or I like four and two and three. I'm
30:33just going to use those and there seems
30:35to be this really interesting thing.
30:37Let's go back and [clears throat] you
30:38just keep refining and at some point
30:41when you think close, you do the stress
30:44test revising. So one of them things is
30:46you are hot you what I suggest at times
30:50is you take when you're close, you
30:52either go to a completely different
30:56large language model so it knows nothing
30:57about the past.
30:58>> [clears throat]
30:59>> And you say you're an expert in this
31:00space and you can use the old techniques
31:02to do that. Here's a document I'm
31:04writing critique it. What's wrong with
31:06it? What's good about it? So this is the
31:08stepwise you do. So in a way you have a
31:12co-creative assistant not an oracle
31:14anymore. And the assistant should you
31:16should use it for what it does best.
31:18What it does best is it goes out and
31:20gets information. Tell me about this new
31:22thing that Apple did. Look you know and
31:25and so and so and so and so. Okay, but I
31:27think I I want to think about how I
31:30would do it. So you never ask it to
31:32deliver your first draft. You never use
31:35AI as a complete knowledge expert so
31:37that you're not uh
31:39have some kind of expertise in this is
31:42assignment. I'm going to give you one
31:44that you do first. Then I'm going to ask
31:46you to do one in your space.
31:48You know that could be the school you're
31:49in or some kind of expertise you have,
31:52but it's always better that uh use work
31:54as a human, you know, for human
31:57authorship. It's about you. Uh break the
31:59interaction loop at any time. The small
32:02pieces and where you're making
32:04decisions.
32:06Okay. So I'm actually going to give you
32:08a entire worksheet to do this on your
32:11own. So this is a little complicated.
32:14The worksheet I've color coded and I'm
32:16going to get it's going to be really
32:17easy cuz you just I'm just giving you a
32:20problem and the exact things to say and
32:22you just drop them into frame and then
32:25explore and then refine and then commit.
32:28Once you do that once, then I'm going to
32:30say do it for yourself. So look through
32:33the handout. Um um
32:35and I did this particular goal of of
32:38give me a one-page brief on Canadian
32:40post-secondary education. I'm going to
32:42give you three
32:44We're going to give the system three
32:45sources.
32:46And then you're going to see how you go
32:48through it. Now, I actually have those
32:50sources up on
32:52um
32:54Canvas files Ferk. So, all that's in the
32:57assignment instructions right now. This
32:59is I'm talking about the high-level
33:01stuff. And then you're going to fill
33:02that all out in this table.
33:05So, you're going to do it once
33:07with me showing you every step, still
33:09with some things to think about. I don't
33:11need that handed in. I only need once
33:14you learn all these things to do it on
33:16your own. And you could decide I want to
33:19do the stress test or the framing a
33:20different ways. You really have to think
33:22about how it makes sense for you. Okay,
33:25so that that handout is there. Um
33:28You're then going to try it on your own.
33:32Um you can use any I bring up Copilot
33:34here, but you can use any system that
33:37you want to do to do it. And um I have
33:39more information here. I don't need to
33:41go through all this. You know, but with
33:43framing, you can see it's it's
33:46a frame, explore. This is where you do a
33:49stress test and think about drafts.
33:51You're starting to now
33:53uh um
33:54refine.
33:55Uh
33:56and then commit. Um sorry, I have this
33:58backwards. So, refine is a good place
34:00for the hostile critic or what what
34:03assumptions did I leave out? Um
34:06So, you can go Oh, actually, I just had
34:07commit twice. Okay, so that's what you
34:10need to do. Um if you want,
34:13um almost for extra credit, is the way
34:15I'm doing it, you might add other
34:17things. You might decide somewhere I
34:18don't like your voice. We're getting
34:20places, but I don't like the voice.
34:22Change the voice. And at some point, you
34:24might decide to go back to framing when
34:26you change the voice. So, and then I
34:28really like the hostile critic. The
34:30dumbest hostile critic would say be a
34:32hostile critic. How would it know what
34:34to do? So, the people who do this the
34:36best I'm sorry about that. I just
34:37knocked out ourselves out. Let me just
34:40go back in again.
34:42The people who do this the best
34:45Uh
34:46I have lost control here. Give me a
34:48second.
34:51Ah,
34:52jumped out again.
34:54Cuz it's the last frame.
34:58The people who do the hostile critic the
34:59best
35:01um
35:03actually spend time writing a hostile
35:05critic. I will show you this in the
35:06class, but I was just writing a grant
35:09with somebody in the UK and they had
35:11four hostile critics on our work. None
35:15of them were telling us None of them
35:17were changing it and doing it for us.
35:19They were just informing us, "Hey, your
35:22EDI section is weak and this part needs
35:26to be better." Why does she Why does she
35:28have four? Well, one looks at the many
35:32pages of how the grant needs to be done
35:34with this and that. So, it's a great
35:36hostile critic saying you didn't follow
35:38the rules of the grant. She's got
35:40another one which is, "I've been on
35:42committees for so long, I know what the
35:45UK wants to see or not." A different
35:47hostile critic. Another one is, "I've
35:49taken all my old grants and the way I
35:52talk and I want it to look like that."
35:55So, you can imagine if you're really
35:57good at this, you will build up your
35:59hostile critics um
36:01to be you. They're They're a version of
36:03you that's smart enough to know what you
36:05want to see and what you don't. So, this
36:07is what we're uh asking for within
36:11assignment three. More so, this is this
36:14FURCS system is a a much better way to
36:19work uh
36:20in the future. So, remember we're we're
36:23down we're back to these two things. I
36:26gave you the this the first assignment I
36:29tend to call it the specialist which
36:31still is 90% better than almost everyone
36:34out there. And then we then move out
36:38from that to this new way of
36:40orchestrating. Most people are not doing
36:42this orchestration. It doesn't even have
36:44a good name whether it's called human in
36:47the loop or furc or or cognitive
36:49conducting. We've built it into our AI
36:51system and we're we're using it in a bit
36:55that way. And I'll talk more about it
36:57later and live coding. So this gets you
37:00on your way to finish your assignment
37:02two if you have it but also
37:05moving on to your assignment three.
37:07Please do not add all these assignment
37:10three
37:11ideas in assignment two. Just be good
37:13with assignment two
37:15understanding this. In fact, we still
37:17can use this locally in assignment
37:19three. And then wait till you do
37:21assignment three to learn this better
37:23way. This again this is the 90% great
37:26way of doing it. A lot of times this is
37:28what you have time for but when you're
37:30doing heavy duty research you really
37:32want to move to our our furc and human
37:35in the loop method.
37:37Thanks, Steve DePalma.

About this transcript

This page contains the full transcript of wk06 2 PromptEng iat111 by ivizlab, generated from the public captions YouTube serves with the video. The transcript has 6,511 words across 970 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.