The following serves as a technical addendum to the article “In the Future, Everyone’s an Architect (and why that’s a good thing)” published by Design Intelligence, April 2023. The full process by which the article, and the video it featured, were generated is described below. It is provided as a learning companion to the article itself, and a furtherance of the issues raised therein.
For even further details, you can follow the series The Cult of G.E.N.I.U.S. - A Speculative Exploration on the Possibility of a Full-Stack, AI Architect on my substack. There, I went on a more speculative journey and tried to envision a full-stack, AI architecture firm that could carry a project all the way from the client interview to the ribbon cutting. You can read the 1st three parts of “The Cult of G.E.N.I.U.S.” on my substack today, which outline the first steps in setting up an AI replicant of myself. The remaining parts will be published every few days through the 1st week of May. In this 10-part series, I’ll walk the reader through each phase of the design cycle, and what role AI could play in SD’s, CD’s, CA, etc.
1) “GPT” refers to both Chat GPT and GPT 4 since both were used in the construction of the model. Chat GPT and GPT 4 have different capabilities, and wherever the use of one was specific, it is noted as such. Wherever “GPT” is used, it could refer to either Chat GPT or GPT 4.
2) NLGAI Prompts (for GPT, Midjourney, all others) are in blue bold.
3) NLGAI responses are in Purple bold.
4) References to ‘the article’ refer to “In the Future, Everyone’s an Architect (and why that’s a good thing)” parts 1 and 2, available on Design Intelligence. If you haven’t read them, you should probably begin by doing so, as they give important context to what follows.
5) References to ‘the video’ refer to the video included in “In the Future, Everyone’s an Architect (and why that’s a good thing),” which is also available on my youtube.
6) I developed the initial language model on March 12th, 2023 and developed the video in the days after. I submitted the article for publication on April 7th, 2023. The most challenging part of writing the article was how rapidly the technology advanced during that time. I continually had to re-edit the article to account for new technologies and applications of NLGAI that were surfacing daily. Indeed, the technology has advanced rapidly in the time between my submission and today (the date of publication, April 26th, 2023). To stay faithful to the experiment, the methods below described what I did on March 12th, 2023. They are not necessarily the methods I would use today. If you’re reading this at some future date, it’s inevitable that the methods described below have been superseded. Where possible, I tried to note how a particular process might have been updated in the lag between the creation of the model and the publication of the article. However, with the technology changing as rapidly as it is, I would encourage you to follow my substack for periodic updates.
7) The article and the processes described below focus on replicating the role of an architect during the earliest parts of the design cycle, since those seemed most challenging to replicate. The video depicts an architect, client, engineer and contractor moving through the entire design cycle. While they were not the subject of this investigation, it is imagined that the later stages of the design cycle (DD’s, CD’s, CA, etc.) could also be largely replicated via AI, resulting in a near total replacement of an architect’s labor. For ideas on how this near-future might come to pass, please check out “ The Cult of G.E.N.I.U.S. - A Speculative Exploration on the Possibility of a Full-Stack, AI Architect” on my substack.
I began by adding a series of generalized, contextual questions. I did so to ‘prime the pump’ and get GPT thinking about the conversation it was going to have, and also as a kind of experiment control. Later in the process, when Bob is asking Carla questions about her design preferences, I wanted to be able to see to what degree Bob’s questions were being informed by a generalized knowledge about architecture, and to what degree Bob was being responsive to Carla and the things she was saying in the moment. I asked two questions:
“Imagine you are an architect preparing
to meet with a potential client for the first time. The client is interested in
designing a custom luxury residential house for themselves and their children.
What questions would you ask this potential client, to better understand their
design ambitions, and aspirations for the house?” (Responses here).
“Imagine you are a successful tech
executive who is interested in hiring an architect to design and build your
custom dream house. You are about to meet with the architect for the first time
and interview him for the job. What questions would you ask the architect to
determine if he is the right fit for you?” (Responses here).
I followed that by feeding GPT character descriptions of the different roles that would appear in the video. I took care not to feed GPT any technical information – I didn’t give Bob any advice on how to be an architect, or Carla any advice on how to be a client. I merely tried to flesh out a description that would allow GPT to ‘understand’ who they were, and hopefully have that animate their words and actions in the dialogue. I wrote separate character descriptions, in varying detail, for Bob, Carla, Doug and Gus.
There is an alternative method of character generation that proved to be too cumbersome from a timing perspective. It involves asking GPT to generate the character sketches through iterative prompting. This enlists GPT in creating its own prompts. You basically ask GPT to ask you questions that will help it in crafting a personality profile. It asks you questions, and you answer them, and it subsequently refines the character profile. It does a pretty good job, but it takes a while. I could not discern the logic by which GPT was selecting its questions; some of the questions didn’t seem pertinent to what I was trying to do. So, I would end up answering questions like what kind of food Bob enjoys, and that sort of thing. An interesting process, with a wide variety of applications, but proved to be a little lengthy for this particular exercise.
Once the pretraining was done I pursued two parallel methods for generating the conversation. GPT is set up on a conversational model that anticipates it interacting with you, a single user. It is a chat bot designed for a single person at the keyboard. I needed to alter that, so I created a model with a combination of Chat GPT, GPT 4, Google Sheets and a Google Sheets plugin called “GPT for SheetsTM and DocsTM.[1] The first attempt was made with Google Sheets, and the second through iterative prompt engineering with GPT 4. Both worked equally well, although they generated slightly different results. Ultimately, I used GPT for SheetsTM and DocsTM to generate the base dialogue but then switched to using GPT 4 when it was released and I had gotten off the waitlist (roughly halfway through my work on the article). I found that GPT 4 was better able to handle large inputs and had a higher degree of spatial intelligence.
I arranged the conversation in Google Sheets such that the questions asked by one character would inform the responses of another, as well as generate additional questions (Figure 1).I let the conversation iterate through 4 cycles, until I felt that the characters had asked and answered enough of each other’s questions to form the foundations of a design brief.
While GPT for SheetsTM and DocsTM proved successful, I ultimately switched over to GPT 4 once it became available (about midway through my work on the article). GPT 4 was a massive upgrade over Chat GPT, and able to handle much larger blocks of text. It also seemed ‘smarter’ although I’m not sure how to quantify that. It seemed to reflect enhanced intuition about the characters, their motives, and design choices. I initiated the conversation with a 3-part prompt:
"I would like you to imagine the
script for a zoom interview between an architect named Bob Bobinsky and a
potential client named Carla Carlinsky. Respond with "..." if
you understand.
For the zoom interview, imagine Bob
Bobinsky as a 50-year-old male architect with a specialty in single-family,
residential projects. He has been in practice for 20 years and is based in the
San Francisco Bay Area. He runs his own firm and has 10 employees. He has
received several local design awards for his work. He was raised in Texas
and studied architecture at Washington University in St. Louis. Earlier
in his career, he worked on large commercial projects for a large firm, before
starting his own practice to focus on custom residential projects. He is
married with 3 children. In his free time, he enjoys kayaking in the San
Francisco Bay and cooking for friends and family. He is very interested in
securing Carla as a client. Imagine Carla Carlinsky as a 40-year-old
single mother of two. She is a successful tech executive, currently living in
San Francisco. She recently left a tech conglomerate and cashed out her stock
options, resulting in a financial windfall. She is starting her own small
consulting business so she can have more flexible time with her two children.
With the proceeds from her stock sale, she is interested in building her dream
home for herself and her two children. She is interested in procuring the
services of an architect. She is a California native and enjoys hiking in her
free time. She has already procured a site in the Berkeley hills where
she wishes to build her dream house. She is very passionate about
sustainability and technology and would like to find an architect who shares
her values. She likes the idea of a modernist house, with clean lines,
but does not want the house to be too sterile. She got Bob’s name from a
friend of hers named Hilda who recommended him highly. Respond with
"..." if you understand.
When I input “Begin the Conversation”
open the zoom conversation with some friendly small talk, and then begin the
interview. Conduct the interview in a professional but friendly
tone. Bob should begin by asking Carla about her vision for the project,
and Carla should respond. Carla should ask additional questions that a
client might ask a potential architect. Bob should subsequently respond
to her questions and ask her additional questions that an architect might ask a
new client in order to fully understand their vision
for the project. Do not conclude the interview. I will review the
responses and let you know if the interview should continue or conclude. If I
write “Continue the interview” you should have Bob and Carla continue the
interview. Each should be responsive to the other’s questions, as well as
ask additional questions. Bob should ask additional questions that might
help him more fully understand Carla’s vision for the project. Carla
should ask additional questions that might help her understand if Bob is the
right architect for her. We will continue this iterative process until I
input “Conclude the interview” at which point you should conclude the interview
in a professional and friendly manner. Respond with "..." if you
understand.
The prompts were designed to allow me to nudge the conversation along but ensure that the questions and answers being raised by both parties were authentically coming from GPT. It would also allow me to have a measure of control to make the conversation go on as long or as little as I thought was appropriate, given the constraints of the exercise.
I let the conversation run through 4 iterations, or enough for about a 10-minute conversation. Choosing to stop at 4 iterations was somewhat arbitrary, but not entirely without logic. Too few iterations, and it would not have provided sufficient information for a design brief. Too many, and it would have been too long for publication. Importantly, GPT could have gone on indefinitely, providing more and more details about the proposed project. It is unclear how an hour-long conversation would have affected the overall ‘design’ process. However, every architect has had the experience of talking ‘too much’ with a client. There’s a point in the conversation where inputs can become redundant. A good architect knows how to ask the right questions and get to the essence of a client’s needs & intentions.
I then asked GPT to use the transcript of the conversation it had just generated and generate a four separate design briefs:
For the first three, I specifically asked it to use very detailed and descriptive language, as if it was describing the project to an illustrator, since I was planning on using it to generate design options in subsequent steps via Midjourney. I also asked it to generate a bulleted design brief, from more of a project management perspective.
I first generated the images for the housing design in
Midjourney V.4 (MJ4); I omit that process here, because while I was in the process
of making the video, Midjourney V.5 (MJ5) was released
and I began to use that instead. MJ5, in
contrast to MJ4, can accept and process more natural language prompts. There is no special lingo needed to generate
images. According to its makers, you can
merely enter a natural language description of what you would like to see.
Since GPT’s knowledge cutoff was in 2021, it was unaware of Midjourney’s
existence. Therefore, I couldn’t ask it
directly ‘make prompts for Midjourney.’
To overcome this, I tried two methods:
First, I tried to get GPT to write my Midjourney prompts automatically. Even though it didn’t know what Midjourney was, I tried training it anyhow by feeding it information about Midjourney V5.
I copied and pasted all the instructions from the Midjourney website about how to use prompts, as well as some random articles from around the web about how to write good prompts for MJ5.
Once it had assimilated the instructions, I asked it to use that new knowledge and craft some prompts for use in MJ5.
It was moderately successful. The real limitation seemed to be that it didn’t know how I wanted the images to appear, so it kept trying random approaches to angle, lighting, etc. I was able to control it by putting more and more information into the prompts, but that seemed to be redundant; if I was going to be very specific with the prompts, I might as well do so directly into MJ5, rather than using GPT 4 as an intermediary.
After that experiment, I decided to switch to a more natural language approach. I asked GPT to generate the image prompts with the following:
“Generate a natural language description
of Carla’s house that could be given to an architectural illustrator. The description should be highly
detailed. Provide sufficient details
such that an architectural illustrator could develop accurate, detailed
renderings of Carla’s Dream House from multiple perspectives.”
That was sufficient to generate The Initial Description, which I split into parts to reflect different parts of the house. (e.g., the bedroom and the living room should be different prompts, to avoid MJ5 trying to render both in the same image response). I would subsequently generate options for both the exterior, the master bedroom, the office, and the living room.
To assist with lighting, camera angles, etc., I used a free tool from Promptomania to provide the correct MJ5 parameters. For the exterior shots, I used two different camera angles, one a wide-angle lens from the front, and another from an aerial view.
I then input the completed prompts into MJ5, as is[2]. I originally generated 16 options of both interior and exterior options, resulting in 80 separate images (2 exterior shots, plus 3 interior shots x 16 options). Since MJ5 generates 4 options with every prompt, I only had to run the prompt (or ‘reroll’) four times to generate the 80 options.
The interior options were removed from the exercise for two reasons:
1. The space constraints of publication: I did not think anyone would watch an hour-long video or be interested in watching Carla assess 8 options for every room in her new house.
2. The complexity of getting interior and exterior options to be consistent with one another.[3]
For Bob and Carla’s second conversation, I brought eight of the completed ‘designs’ back into the conversation and developed a new set of prompts.
At this point, the reader should have an obvious complaint: an exterior rendering of a house does not constitute a ‘design.’ Architects don’t design buildings by exterior renderings. We design the plan, the section, and the elevations simultaneously, which then generates the 3D massing of the building which can then be rendered into an image. An architect’s process is iterative, and reflective, moving forward and backward, between steps, etc. This is necessary in order to ensure that the building works, spatially, and that what is being presented to the client isn’t a misrepresentation of possibility.
I entertained this ‘cheat’ for several reasons.
The logic that informed this video can be applied to different software and different approaches. I chose to use MJ5 to generate exterior renderings in lieu of floorplans for reasons of expediency and cost. A Midjourney basic plan is $10/month, and I was already subscribed. It also renders images in under a minute. Given a lack of budget for the project, it made sense to use tools that were cheap or that I already had. Could you use written text prompts to generate a floor plan instead? And subsequently sections, models, etc.? Yes, absolutely. More information can be found on these processes in GENIUS.
The point of the exercise was to see whether a NLGAI could replicate the earliest parts of the design cycle – whether it could get input from a client in natural language and use that to then generate a design. The critical part, for me, was whether the client’s input could be extracted in useful form. Less important was how that input was eventually used to create the ‘designs.’ The design briefs are sufficiently robust that you could give them to a human architect, and they would likely design something that lives inside the stylistic range defined by Bob’s 8 options.
I selected the 8 that seemed in best alignment with Carla’s vision and discarded the rest to be mindful of time. The Design Review Prompt would instruct Bob to present the designs to Carla, and Carla to respond to them, based on her preferences.[4]
Here, I engaged in a type of cheat,
but an appropriate one under the technological circumstances. The most challenging part of the exercise was
in getting Carla to evaluate and react to design proposals. Despite the seeming intelligence of some of
these programs, both Chat GPT and MJ4 lack computer vision. Meaning, they can’t see, interpret, or answer
questions about an image. And they
certainly can’t have any opinions on it.
I was attempting to
elucidate a reaction to the images.
Which designs did Carla ‘like’ and what did she ‘like’ about them? It was not enough for Carla to merely look at
the images and answer questions about them – she had to react to them as Carla.
As if on cue, OpenAI released GPT 4 within the few weeks between when I started the article and when I finished it. One of the distinguishing features of GPT 4, as opposed to its predecessor, is multi-modal image recognition. So, my problem was solved, in theory. I would be able to have Carla look at the images, and react to them from the standpoint of Carla, given what I had programmed about her situation and personality. Alas, while GPT 4 has been released and was used in the creation of this video, as of time of publication, its image recognition capabilities have still not been opened to the general public.
I sidestepped the issue in the way that I thought was most appropriate to the spirit of the exercise. My approach anticipated forthcoming technology (I assume that GPT 4 will release its multimodal capabilities within the next few weeks) as well as proved the efficacy of the general model: I took Carla’s reactions and reran the image generation based on her reactions. The images you see in the video are a synthesis of the original design brief, as well as Carla’s reactions after-the-fact. To the degree that there is coherence between what Carla is describing about the image, and the image itself, it’s because the image is partially based on Carla’s reaction to it.
As an additional degree of proof in the model, I used the ‘describe’ function in MJ5 to review Bob’s generated design options and translate them back into text. For those unaware, the describe function recently debuted in MJ5 does have computer vision and offers a text description of any image you upload. After generating MJ5’s descriptions of the 8 schemes, I compared them against Carla’s reactions and against the images themselves, in order to establish that the computer vision was actually working, and not just spouting nonsense. It wasn’t as accurate as some other programs might have been (or as GPT 4 will eventually be) but it was sufficient to give me confidence in the model.
At this point, I had a design interview, a design brief, multiple schematic renderings of the possible house design and the client’s reaction to them. In anticipation of later stages of the exercise, I asked GPT for several other items:
I asked GPT to generate a data model using the design brief, and all discussions up to this point. I wanted to understand how GPT was understanding all of this. A data model is just a way of organizing information in such a way that it becomes code-able, for use in programming languages. I wanted to make sure that it was organizing information in a way that was useful to any downstream applications. It appears that it did so. The organization of The Data Model appears suitable for use in multiple downstream applications, although to keep the exercise tightly focused, I did not try to program anything based on its outputs. If you would like to try it yourself, GPT 4 was able to take the data model and generate codes for pyplot and Mermaid, although I did not try to do anything with that code.
I also asked it to produce a room schedule
for similar reasons – I was just curious how much of the conversation it was actually ‘understanding’ versus how much was just language
prediction. GPT 4 clearly displays some
kind of world-building intelligence. It
can organize things spatially. If you
tell it that the bedroom is north of the office, the gym is west of the
bedroom, the laundry room is south of the gym, it will be able to tell you that
the laundry room is west of the office, without having been told that
explicitly.
I found the implications of this profound.
I was also impressed with the fact that GPT was making good guesses
about the interior details of the house, without having been told to do so. At what point in the conversation, Bob tells
Doug that the house is ‘around 3500 square feet’ while the total generated by
the room schedule was 4520 sf. I had
thought that I had caught GPT in an error!
However, subtracting the deck and the patio gets you to 3,920 sf, and
subtracting the garage gets you to 3,120 sf.
Allowing for circulation space, mechanical spaces, assorted closets,
etc., 3,500 sf is probably spot on! When
GPT had Bob tell Doug that the house was around 3500 sf, was it merely saying
so because that would be the average size for a 3 BR luxury home in
Berkeley? Or did it perform the
calculations ahead of time, averaging room sizes first, and then summing that
into an aggregate? Or did it arrive at
that number in some other fashion?
The room schedule provides other insights into the ‘intelligence’ of GPT. For instance, why did it list the bathrooms
as having 1 window each, but the powder room as having zero windows? Does it understand the difference between the
two, and understand that a bathroom with no windows can be inhospitable, while
a powder room with no windows might actually be
considered typical?
How did it generate the number of doors?
Does it have some kind of mental map from which it understands that the
dining room is adjacent to the kitchen?
Early research has validated that GPT 4, in contrast to Chat GPT, does
have some kind of spatial, world-building
intelligence.[5] Or maybe it was just making a guess about the
number of doors that would be typical for each space? Fascinating!
It was not part of my original plan to include an engineer and contractor. I decided later in the process to include them, mostly because I wanted to test the limits of the model. The model had proven successful in orchestrating the architect/client design conversation and an early production of schematic design. Would its mysterious methods translate into other conversations? Could the design brief, written by GPT, based on a conversation, itself generated by GPT, allow for a discussion of engineering and buildability?
I replicated the previously-discussed methods to generate the conversations with Doug the Engineer and Gus the Contractor. Their character prompts are here and here, respectively. For the engineer and contractor, it was necessary to be a bit more specific in the kinds of instruction given. A limitation of GPT seems to be that it prefers to resolve conversations quickly.[6] When you ask it to orchestrate a conversation between two parties, it will try to complete that task by rushing to the end and wrapping up that conversation. In the conversation between Bob and Carla, I circumvented that tendency by specifically asking GPT not to conclude the conversation, and instead wait for me to allow it to conclude, and to otherwise continue asking questions. (see Conversation Starter). The drawback to that approach is that you really had to give GPT time to explore. It would ask questions in its own order, and it was difficult to tell whether its questions were auguring towards some goal. A human understands that sometimes you let a conversation wander, but in business conversations, time is money, and you’re typically moving towards some goal or agreement. Because of the iterative nature of the prompting, I’m not sure that GPT understood that, and it seemed to meander at some points.
I wanted to be mindful of time, and didn’t have the patience to wait for Bob and his vendors to get to the point. I added a specific instruction of ‘send the design brief’ to the prompt logic so that I would have a command to execute and help them get to the point. It was my intuition, as well, that GPT would understand that Doug should understand the design brief once it had been sent to him.
The conversations with the engineer and the contractor revealed another interesting problem, briefly mentioned in the article: because GPT is trained on historical data, it could not extrapolate outward into a near future and consider what construction timelines might be. Bob kept insisting that the drawings would take 4 or 5 months. And Gus kept saying that the construction would take 14 months. That hardly seemed appropriate in a video that purported to depict a futuristic, automated design and construction cycle. Perhaps that points to a problem of imagination, perhaps not. I, as the human imagineer of the exchange, envisioned a near future where design and construction are nearly completely automated and conducted by AIs, although I didn’t specifically write that into any prompts. GPT did not predict or intuit my vision. Is that because as a human, I am just so vastly more creative, and this machine isn’t capable of my human-level imagination? Or is it because I just failed to tell it that that’s what I wanted? Could it have imagined the near-future scenario, if I had prompted it appropriately? Interesting questions!
For Bob and Carla’s third conversation, I aimed to wrap everything together.
As a preliminary matter, I asked Bob to calculate his own fee and establish a total price for the contract (prompt here). He very generously assigned himself a fee of 12% of construction costs. The construction costs were also calculated by GPT (in character as Gus, the Contractor).
By now, the conversation was getting to be about 20 minutes in length, so I wanted to avoid a lot of continued discussion. I instructed GPT to have Bob present the ‘final’ design to Carla and to have her approve it after asking any remaining questions she might have (prompt here). I did not allow for an additional cycle of feedback and response, owing to time constraints. However, it’s worth noting that this process could be extended indefinitely. One could easily program Bob to go back to Carla x number of times and get her continued reactions to different design possibilities. Carla’s feedback could be assimilated into GPT’s mental model of the project, leading to finer and finer resolutions of the design.
I knew that I had enough content for the article I was intending to write, so I brought it to a close and turned my attention to making a video representation of the exchange.
At this point in the process, I had a wealth of material: the design brief, the images, the budget, etc. However, the transcript of the whole conversation was still in raw form. However, the challenging part was done. Assembling the rest of the video was quite mechanical. I first copied the conversational text out of GPT and into MS Word, to make it a little easier to work with. I utilized an online service called Synethesia.io to generate the avatars for the video. A basic plan for Synthesia.io costs $30/month and for that, you can generate roughly 10 minutes of video per month. Because Synesthesia does not create multi-character videos (yet), I split the transcript into 4 parts, one for each character. I added in the requisite number of pauses during which the other party would be talking. The following exchange
· Bob talks for 10 seconds.
· Carla Talks for 17 seconds.
· Bob talks for 24 seconds.
would have to have pauses to make it sound & look natural in video, like so:
Bob talks for 10 seconds |
Carla Stays silent and nods for 10 seconds |
Bob stays silent and nods for 17 seconds |
Carla talks for 17 seconds |
Bob talks for 24 seconds |
Carla stays silent and nods for 24 seconds |
To make sure that these pauses were accurate, I extracted the length of each conversation segment through a 4-step process.
1. I used a text-to-speech engine to read the entire transcript aloud and recorded the reading using desktop audio recording software.[7]
2. I then used an online transcription service to transcribe the audio file I just generated, now with timestamps.
3. I put that timestamped transcript back into GPT 4, which was able to take it (in parts) and generate a table that had the duration of each character’s speech, as well as their subsequent required conversational delay.
4. That allowed me to manually enter the delays into Synethesia.io (they have a little button where you can just input ‘add 10 sec delay’)
Synethesia.io can also accept audio recordings. It is possible to take a recording of your own voice, upload it, and have one of their avatars read it. I elected to use written transcripts because I already had them and did not want to add an extra step. Synethesia.io and similar platforms have a very user-friendly interface and are designed for novices. It was simple enough to select avatars, have them read the scripts, and export the results to an *.mp4 file.
Synethesia.io took about 1 hour to generate each character’s video. I subsequently spliced the 4 videos together using desktop video production software, which was easier than it sounds, since GPT had previously generated a table for me that indicated exactly when each party was speaking. I also added a zoom frame around the conversation, mostly for humor.
Ordinarily, this is the part where I would write some kind
of pithy conclusion, offering insights into what we all can learn from this
exercise. However, I wrote most of that
into the article. I offer this record of
the process knowing that it’s completely out of date, even as I write these
words. However, I thought it was
important to memorialize what I did. My
primary motivation was to illustrate how easy it was. It didn’t require any
expensive software or advanced knowledge of coding. It was entirely off the shelf technology. The total budget for the project was probably
less than $100, although I did not keep close track of the tokens I was
spending.
Emphatically, I
don’t offer this process as some kind of proof that AI can now do an
architect’s job. But I think it should
raise some questions about what an architect’s job actually is. The conventional wisdom around AI and architecture tells us that it’s
the earliest parts of the design cycle (meaning, interviewing clients,
generating creative designs, etc.) are safe from automation, because computers
can’t possibly do that sort of thing.
Maybe not, but it looks like they can fake it pretty
damn well.
Imagine Bob Bobinsky as a 50-year-old
male architect with a specialty in single-family, residential projects. He has
been in practice for 20 years and is based in the San Francisco Bay Area. He
runs his own firm and has 10 employees. He is very active in the local
architectural community and has served on several boards for local
nonprofits. He has received several
local design awards for his work. He was raised in Texas and studied
architecture at Washington University in St. Louis. Earlier in his
career, he worked on large commercial projects for a large firm, before
starting his own practice to focus on custom residential projects. He is
very passionate about sustainability and believes architects should do more to
develop sustainable solutions in their work.
He is married with 3 children. In his free time, he enjoys
kayaking in the San Francisco Bay and cooking for friends and family. He is
very interested in securing Carla as a client, as he believes that her
connections within the tech community might lead him to further commissions in
the future.
Imagine Carla Carlinsky as a 40-year-old
single mother of two. She is a successful tech executive, currently living in
San Francisco. She recently left a tech conglomerate and cashed out her stock
options, resulting in a financial windfall. She is starting her own small
consulting business so she can have more flexible time with her two
children. With the proceeds from her stock sale, she is interested in
building her dream home for herself and her two children. She is interested in
procuring the services of an architect. She is a California native and enjoys
hiking in her free time. She has already procured a site in the Berkeley
hills where she wishes to build her dream house. She is very passionate about
sustainability and technology and would like to find an architect who shares
her values. She likes the idea of a modernist house, with clean lines,
but does not want the house to be too sterile. She got Bob’s name from a
friend of hers named Hilda who recommended him highly.
Imagine Doug Duggins as a 40-year-old
structural engineer who owns his own firm located in Oakland. He has five
employees and specializes in structural designs for residential
construction. Doug wanted to be an engineer since he was a little
kid. He grew up in Southern California
but is a long-time resident of the Bay Area.
He got his PhD at Berkeley and stayed in Oakland after graduating. He worked in industry for 5 years before
starting his own firm. His experiences
at a large engineering firm taught him a lot about how to do complex
engineering, but he prefers small jobs where he can work more closely with his
clients. Bob and Doug are friendly and
have collaborated on many projects together in the past. Doug enjoys
spending time in nature, and has a passion for
sustainability. He often advocates
within his own profession for the use of more sustainable design options, and tries to avoid using concrete in his designs
whenever he can.
Imagine Gus Guthrie as a 55-year-old
general contractor who owns his own firm located in Oakland. He has 25
employees and specializes in high-end residential construction. Gus has
been working in construction since he was a teenager, and his father was also a
contractor. Gus cares about doing
professional, high-quality work, and makes a good living. He is married with 3 children. His oldest daughter has recently started
college, a fact that he is very proud of, as she is the first in his family to
go to college. He is looking forward to retirement, but intends to keep working until all of his
children have finished college. After
that, he hopes to move outside the Bay Area and live in the country or in a
small town somewhere. He enjoys boating
and fishing in his spare time and collects stamps. Bob and Gus are friendly and have
collaborated on many projects together in the past.
I want you to help me write a personality
profile of Bob Bobinsky. His current personality profile is as follows:
“Bob is an architect. He is 50 years old.”
Your goal is to help me craft the best
possible personality profile for Bob. The personality profile will then be used
by you, GPT 4, to answer other prompts and other questions later in the chat.
You will follow the following process:
1.
You
will first ask me 3 questions about Bob.
2.
I
will provide my answers, but we will need to improve the personality profile
through continual iterations by going through the next steps.
3.
Based
on my answers, you will generate 2 sections.
a.
an
updated 250-word personality profile of Bob
b. Additional questions (ask any relevant
questions pertaining to what additional information is needed from me to
improve the personality profile).
We will continue this iterative process
with you asking me additional questions and me providing additional answers to
you and you updating the personality profile in the Revised prompt section until
I say we are done.
"I would like you to imagine the
script for a zoom interview between an architect named Bob Bobinsky and a
potential client named Carla Carlinsky. Respond with "..." if
you understand.
For the zoom interview, imagine Bob
Bobinsky as a 50 year old male architect with a
specialty in single-family, residential projects. He has been in practice for
20 years and is based in the San Francisco Bay Area. He runs his own firm and
has 10 employees. He has received several local design awards for his
work. He was raised in Texas and studied architecture at Washington
University in St. Louis. Earlier in his career, he worked on large
commercial projects for a large firm, before starting his own practice to focus
on custom residential projects. He is married with 3 children. In
his free time, he enjoys kayaking in the San Francisco Bay and cooking for
friends and family. He is very interested in securing Carla as a
client. Imagine Carla Carlinsky as a 40-year-old single mother of two. She
is a successful tech executive, currently living in San Francisco. She recently
left a tech conglomerate and cashed out her stock options, resulting in a
financial windfall. She is starting her own small consulting business so
she can have more flexible time with her two children. With the proceeds
from her stock sale, she is interested in building her dream home for herself
and her two children. She is interested in procuring the services of an
architect. She is a California native and enjoys hiking in her free time.
She has already procured a site in the Berkeley hills where she wishes to build
her dream house. She is very passionate about sustainability and technology
and would like to find an architect who shares her values. She likes the
idea of a modernist house, with clean lines, but does not want the house to be
too sterile. She got Bob’s name from a friend of hers named Hilda who
recommended him highly. Respond with "..." if you understand.
When I input “Begin the Conversation” open
the zoom conversation with some friendly small talk, and then begin the
interview. Conduct the interview in a professional but friendly
tone. Bob should begin by asking Carla about her vision for the project,
and Carla should respond. Carla should ask additional questions that a
client might ask a potential architect. Bob should subsequently respond
to her questions and ask her additional questions that an architect might ask a
new client in order to fully understand their vision
for the project. Do not conclude the interview. I will review the
responses and let you know if the interview should continue or conclude. If I
write “Continue the interview” you should have Bob and Carla continue the
interview. Each should be responsive to the other’s questions, as well as
ask additional questions. Bob should ask additional questions that might
help him more fully understand Carla’s vision for the project. Carla
should ask additional questions that might help her understand if Bob is the
right architect for her. We will continue this iterative process until I
input “Conclude the interview” at which point you should conclude the interview
in a professional and friendly manner. Respond with "..." if you
understand.
Imagine a follow-up conversation between
Bob and Carla, happening later on the same day. Bob has completed 8 potential designs for
Carla’s house and will briefly present them to her over zoom, with accompanying
visuals, one at a time. Carla should react
to each one with her likes and dislikes.
If Carla really likes a design option, and hasn’t stated why, Bob should
inquire as to what she likes about it.
Begin by having Bob present Carla with
the first 4 options, and having Carla respond to each. Do not conclude the conversation. I will review the responses and let you know
if the conversation should continue or conclude. If I write “Continue the
conversation” you should have Bob and Carla continue the conversation by having
Bob present 4 additional options. Bob
and Carla should each be responsive to the other’s questions, as well as ask
additional questions if appropriate.
If I write “Conclude the conversation”
you should conclude the zoom call by having Bob thank Carla for her input, and
request permission to put her on hold for a few minutes.
Imagine a new zoom conversation between
Bob and an engineer, Doug Duggins. Imagine Doug Duggins as a 40-year-old
structural engineer who owns his own firm located in Oakland. He has five
employees and specializes in structural designs for residential
construction. Doug wanted to be an engineer since he was a little kid. He grew up in Southern California but is a
long-time resident of the Bay Area. He
got his PhD at Berkeley and stayed in Oakland after graduating. He worked in industry for 5 years before
starting his own firm. His experiences
at a large engineering firm taught him a lot about how to do complex
engineering, but he prefers small jobs where he can work more closely with his
clients. Bob and Doug are friendly and
have collaborated on many projects together in the past. Doug enjoys
spending time in nature, and has a passion for
sustainability. He often advocates
within his own profession for the use of more sustainable design options, and tries to avoid using concrete in his designs
whenever he can.
Begin the conversation with some polite
small talk, and then have Bob introduce the Carlinsky Residence as a new
project for them to collaborate on. Have Doug ask relevant questions that
a structural engineer might ask when considering taking on a new project. Have Bob answer those questions.
Do not conclude the conversation. I will review the responses and let you know
if the conversation should continue or conclude. If I write “Continue the
conversation” you should have Bob and Doug continue the conversation. Each should be responsive to the other’s
questions, as well as ask additional questions.
Doug should ask additional questions to help him understand the
structural requirements of the design.
Bob should ask questions about Doug’s ideas for a
structural design.
If I write “Send the Design Brief”, you
should take the following steps:
1.
Have
Doug ask Bob if there’s some kind of design brief he can look at.
2.
Have
Bob reply that it’s just been sent to his inbox.
3.
Have Doug indicate that he sees it and is having a look.
We will continue this iterative process
until I input “Design the Structural system” at which point you should take the
following steps:
1.
Have
Doug ask Bob if he can see the BIM model.
2.
Have Bob reply that the BIM model is in his inbox.
3.
Have Doug indicate that he sees it and is having a look,
and will input the structural design momentarily.
Do not conclude the conversation. I will review the responses and let you know
if the conversation should continue or conclude. If I write “Conclude the
conversation” you should have Bob and Doug end the zoom call in a cheerful and
affable manner.
Imagine a new zoom conversation between
Bob and a general contractor, Gus Guthrie. Imagine Gus Guthrie as a
55-year-old general contractor who owns his own firm located in Oakland.
He has 25 employees and specializes in high-end residential construction.
Gus has been working in construction since he was a teenager, and his father
was also a contractor. Gus cares about
doing professional, high-quality work, and makes a good living. He is married with 3 children. His oldest daughter has recently started
college, a fact that he is very proud of, as she is the first in his family to
go to college. He is looking forward to retirement, but intends to keep working until all of his
children have finished college. After
that, he hopes to move outside the Bay Area and live in the country or in a
small town somewhere. He enjoys boating
and fishing in his spare time and collects stamps. Bob and Gus are friendly and have
collaborated on many projects together in the past.
Begin the conversation with some polite
small talk, and then have Bob introduce the Carlinsky Residence as a new
project for them to collaborate on. Have Gus ask relevant questions that
a general contractor might ask when considering taking on a new project. Have Bob answer those questions.
Do not conclude the conversation. I will review the responses and let you know
if the conversation should continue or conclude. If I write “Continue the
conversation” you should have Bob and Gus continue the conversation. Each should be responsive to the other’s
questions, as well as ask additional questions.
Gus should ask additional questions to help him understand the
buildability of the design and request any information that a general contractor
would need to develop a construction schedule.
Bob should answer those questions if
applicable.
If I write “Send the Design Brief”, you
should take the following steps:
1.
Have
Gus ask Bob if there’s some kind of design brief he can look at.
2.
Have
Bob reply that it’s just been sent to his inbox.
3.
Have
Gus indicate that he sees it and is having a look.
If I write “Send the BIM Model”, you
should take the following steps:
1.
Have
Gus ask Bob if there’s some kind of digital model he can look at.
2.
Have
Bob reply that it’s just been sent to his inbox, and Doug’s structural design
has recently been added to the model.
3.
Have
Gus indicate that he sees it and is having a look.
We will continue this iterative process
until I input “Proceed with construction” at which point you should take the
following steps:
1.
Have
Gus agree to take the job.
2.
Have
Bob express gratitude and indicate that he will be back in touch when Carla
gives them a confirmed start date.
3.
have
Bob and Gus end the zoom call in a cheerful and affable manner.
As Bob, considering the construction cost
estimate provided by Gus, which ranges from $1.8 million to $2.2 million, add
your own fee to the contruction totals so that Carla
can have a ballpark estimate of her dream home’s final cost.
You will now resume the conversation
between Bob and Carla, who has been on hold this entire time. Bob should present the final house design and
give a brief explanation of how Carla’s desires and needs as a client were
incorporated into the design. Carla
should react, and ask what the necessary steps are in
order to begin the project. Bob should
answer any remaining questions she has about the project.
Do not conclude the conversation. I will review the responses and let you know
if the conversation should continue or conclude. If I write “Continue the
conversation” you should have Bob and Carla continue the conversation. Bob and Carla should each be responsive to
the other’s questions, as well as ask additional questions if appropriate.
If I write “Conclude the conversation”
you should conclude the zoom call in an excited yet professional tone.
1.
What is your overall vision for the house, and how would you
describe the architectural style you're drawn to?
2.
How many family members will be living in the house, and could
you provide a brief overview of their individual needs and preferences?
3.
How do you envision the layout and flow of the house, in terms
of public and private spaces?
4.
What are the most important features or rooms you'd like to
include in the house, and are there any specific requirements for these spaces?
5.
Are there any sustainability or energy efficiency goals you'd
like to achieve with this project?
6.
How important is outdoor living and landscaping to your overall
design vision, and do you have specific requirements or ideas for these areas?
7.
How do you feel about incorporating smart home technology and
automation features into the design?
8.
Are there any particular materials,
finishes, or textures you're drawn to or would like to incorporate in the
design?
9.
Do you have any accessibility or future-proofing considerations,
such as aging in place or accommodating potential changes in family dynamics?
10. What is your budget for this project, and
do you have any specific timeline or deadline expectations?
11. How involved would you like to be in the
design process, and what is your preferred method of communication?
12. Are there any examples of homes, design
elements, or architectural features that you admire and would like to discuss
further?
13. By asking these questions, you can gain a
comprehensive understanding of the client's design ambitions and aspirations,
allowing you to create a personalized, luxury residential house that meets
their unique needs and preferences.
1.
Can you tell me about your background, experience, and how you
got into architecture?
2.
What is your design philosophy and approach to creating unique,
custom homes?
3.
Can you provide me with a portfolio or examples of your previous
work, particularly for high-end, custom-built residences?
4.
How do you stay current with the latest trends, materials, and
technologies in residential design and construction?
5.
Can you discuss your process for working with clients from the
initial design concept to the completion of the project?
6.
What is your experience in integrating sustainable and
energy-efficient features into your designs?
7.
How do you approach balancing aesthetics, functionality, and
budgetary constraints in your projects?
8.
Are there any specific architectural styles or design elements
you specialize in or particularly enjoy working with?
9.
Can you provide references from previous clients who have
commissioned custom homes from you?
10. How do you manage communication and
collaboration with your clients throughout the project?
11. What is your typical timeline for
designing and overseeing the construction of a custom home?
12. Can you discuss your fee structure, and
what services are included in your fees?
13. How do you handle any potential
challenges or obstacles that may arise during the design and construction
process?
14. Are you able to work with and coordinate
with other professionals, such as interior designers, landscapers, or
engineers, as part of the project?
15. What sets you apart from other architects
who specialize in designing custom homes?
Carla's dream house is a harmonious blend
of sustainability, modernity, and family-focused design. Situated in the
picturesque Berkeley hills, the house takes full advantage of its natural
surroundings and offers breathtaking views of the Bay.
The exterior of the home is characterized
by clean lines and simple forms, with large windows and sliding glass doors
creating a seamless connection between indoor and outdoor living spaces. A
combination of natural materials such as wood and stone adds
warmth and texture, while the use of steel accents lends an elegant touch of
modernity.
Upon entering the home, one is greeted by
an expansive open floor plan that combines the living room, dining room, and
kitchen into a single, cohesive space. The living area features
floor-to-ceiling windows that flood the room with natural light and provide
panoramic views of the surroundings. The kitchen is designed to be both
functional and visually appealing, with custom cabinetry, high-end appliances,
and a large island that can serve as a focal point for family gatherings and
entertaining guests.
A covered outdoor terrace extends from
the main living area, offering a comfortable space for alfresco dining and
relaxation. The terrace is surrounded by lush landscaping and features
eco-friendly materials, such as reclaimed wood and native plants, to create a
serene and inviting atmosphere.
The private family wing of the house is
strategically separated from the communal living spaces, ensuring privacy and
tranquility. The master suite is a luxurious retreat, complete with a walk-in
closet, spa-like ensuite bathroom, and a private balcony that overlooks the
hillside. The children's bedrooms are located nearby, providing a sense of
comfort and security, while still allowing for individual expression and
personalization.
Incorporating sustainability features,
the house boasts a green roof, solar panels, and energy-efficient windows. The
design emphasizes natural ventilation and passive heating and cooling
strategies, minimizing energy consumption and reducing
the home's environmental impact.
A dedicated home office space provides an
inspiring and functional environment for remote work or personal projects.
Large windows frame views of the surrounding landscape, fostering creativity
and focus. The office is also designed to be flexible, easily transforming into
a guest room or additional living space when needed.
The dream house is a true testament to
the perfect balance between form and function. With a strong focus on
sustainability, modern aesthetics, and family-centered design, Carla's home is
set to be an inspiring and comfortable sanctuary for generations to come.
The exterior of Carla's dream house
reflects a perfect balance between modern design, natural beauty, and
sustainability. Nestled in the scenic Berkeley hills, the home's facade
embraces its surroundings and capitalizes on the stunning views of the Bay.
Clean lines and simple geometric forms
define the structure, creating an elegant and contemporary aesthetic. The
generous use of floor-to-ceiling windows and sliding glass doors accentuates
the visual connection between the interior and exterior spaces, allowing for an
abundance of natural light to permeate the home.
A harmonious blend of natural materials,
such as wood and stone, brings warmth and texture to the facade. The use of
reclaimed wood adds an eco-friendly touch, while the incorporation of local
stone highlights the home's connection to its environment. Contrasting these
organic elements, sleek steel accents provide a refined and modern
counterpoint.
The covered outdoor terrace serves as an
extension of the living area, seamlessly integrating indoor and outdoor spaces.
Surrounded by lush landscaping, the terrace features sustainable materials like
reclaimed wood and native plants, creating a serene and inviting atmosphere for
alfresco dining and relaxation.
The roof of Carla's dream house showcases
a commitment to sustainability, with a green roof system that not only reduces
energy consumption but also enhances the home's visual appeal. Solar panels are
discreetly integrated into the design, demonstrating the seamless fusion of
eco-friendly technology and modern aesthetics.
Attention to detail is evident in the
design of the entrance, which is both welcoming and visually striking. A
cantilevered roof extends over the front door, providing shelter and a sense of
grandeur, while the juxtaposition of wood, stone, and steel materials adds
depth and interest to the facade.
In summary, the exterior of Carla's dream
house embodies a sophisticated interplay of modern design principles,
sustainable features, and natural elements. With its visually captivating
facade and thoughtful integration with its surroundings, the home is poised to
be a stunning architectural gem in the Berkeley hills.
The interior of Carla's dream house is a
testament to modern, sustainable living, with an emphasis on open spaces,
natural light, and a seamless connection between indoor and outdoor
environments. Every room is thoughtfully designed to cater to the needs and
desires of Carla and her family while maintaining a strong commitment to
eco-friendly materials and energy efficiency.
Upon entering the home, visitors are
greeted by a spacious, open-plan living area that seamlessly connects to the
kitchen and dining area. The floor-to-ceiling windows and sliding glass doors
flood the space with natural light, enhancing the sense of openness and
connection to the outdoors. The living area is designed to encourage social
interaction, with comfortable seating arrangements and a cozy fireplace as the
focal point.
The kitchen boasts state-of-the-art,
energy-efficient appliances and sustainable materials, such as reclaimed wood
cabinetry and recycled glass countertops. A large island serves as both a
functional workspace and a gathering spot for casual meals or conversation. The
adjacent dining area is well-suited for both intimate family dinners and larger
gatherings, with an elegant dining table and chairs that complement the overall
design aesthetic.
The children's bedrooms are located close
to the master suite, providing Carla with a sense of comfort and security. Each
bedroom is designed with ample storage, natural light, and a cozy atmosphere
that promotes rest and relaxation. The master suite is a luxurious retreat,
complete with a spacious walk-in closet and a spa-like en-suite
bathroom featuring a soaking tub, a walk-in shower, and dual vanities.
The home office is both functional and
inspiring, with a carefully curated selection of eco-friendly materials and
furnishings. Large windows frame picturesque views of the surrounding
landscape, creating a serene and motivating workspace. Built-in bookshelves and
storage solutions ensure that the office remains organized and clutter-free.
Throughout the home, sustainable
materials such as reclaimed wood, natural stone, and low VOC paint are used to
create a healthy and environmentally conscious living space. The incorporation
of energy-efficient lighting, heating, and cooling systems further emphasizes
the home's commitment to sustainability.
In summary, the Interior Design Brief for
Carla's dream house showcases a harmonious blend of modern design,
family-friendly functionality, and eco-friendly materials. Each space is
carefully crafted to cater to the needs of Carla and her family while
demonstrating a strong commitment to sustainability and environmental
stewardship.
Project Name: Carlinsky Residence Client:
Carla Carlinsky Location: Berkeley Hills, California Project Type: Single-family
residential Budget: $2.5 million Timeline: 2 years
Design Vision:
·
Modern,
sustainable home for a single mother and her two children
·
Clean
lines and environmentally friendly features
·
Warm
and inviting atmosphere, avoiding sterility
·
Strong
connection to the outdoors, taking advantage of the views and natural surroundings
Site:
·
Gentle
slope with mature trees
·
Stunning
views of the San Francisco Bay
·
Opportunities
for indoor-outdoor living spaces
Sustainability Features:
·
Solar
panels
·
Energy-efficient
windows
·
Green
roof (to be explored)
·
Water
conservation measures, such as rainwater harvesting or gray water recycling
Interior Spaces:
·
Open
floor plan with a spacious living, dining, and kitchen area
·
Separate
home office for the client's consulting business
·
Individual
bedrooms for the two children
·
Shared
playroom or study area for the children
·
Master
suite with a walk-in closet and spa-like bathroom
Materials and Finishes:
·
Natural
materials, such as wood and stone, for a warm and inviting atmosphere
·
Combination
of wood siding and large expanses of glass on the exterior
·
Hardwood
floors and stone countertops in the interior
·
Mix
of traditional and modern finishes
Design Inspiration:
·
Mid-century
modern aesthetic (e.g., Richard Neutra and Joseph Eichler)
·
Contemporary
architects like Tom Kundig and Tadao Ando
·
Emphasis
on blending with surroundings and maximizing natural light
Collaboration and Communication:
·
Regular
client involvement throughout the design process
·
Use
of drawings, renderings, and physical models to facilitate client understanding
and feedback
·
Ongoing
communication and updates on progress, budget, and timeline
Budget and Timeline Management:
·
Detailed
budget and schedule outlining the various phases of design, permitting, and construction
·
Close
coordination with team, consultants, and contractors to ensure adherence to guidelines
·
Regular
updates to the client and creative problem-solving in case of unexpected
challenges
Imagine: Carla's Dream House is a
modernist, single-family residence situated on a half-acre sloping lot in the
Berkeley hills. The design should focus on clean lines and a warm, inviting
atmosphere, with mid-century modern influences. The house should feature a
harmonious blend of natural materials like wood and stone, creating a seamless
connection between indoor and outdoor spaces.
Exterior: The exterior of the house
should showcase large floor-to-ceiling windows, optimizing the stunning views
of the surrounding landscape. The structure should incorporate multiple levels
and terraced sections, adapting to the slope of the site. The primary facade
should be dominated by natural materials, such as wood cladding and stone
accents. A flat roof with overhanging eaves will provide a contemporary feel
while also offering solar shading. The house should feature a cantilevered deck
or patio that extends from the main living area, creating an inviting outdoor
space for dining and entertaining.
Ground Floor: Upon entering the house, a
spacious, open-concept living, dining, and kitchen
area should greet the visitors. The living area should be oriented to maximize
views of the landscape and should include a fireplace with a
natural stone surround as its focal point. The dining area should be
able to comfortably accommodate a table for eight, while the modern kitchen
should boast energy-efficient appliances, ample storage, and a large central
island.
A separate family room should be adjacent
to the open-concept space, providing a cozy and more intimate setting. The
ground floor should also include a guest bedroom, a full bathroom, and a
laundry room.
Upper Floor: The upper floor should house
the master suite, which includes a spacious bedroom with panoramic views, a
walk-in closet, and a luxurious en-suite bathroom
with a freestanding tub, a large walk-in shower, and a double vanity. In
addition, two more bedrooms should be located on this floor, each with ample
closet space and large windows that let in natural light. These bedrooms should
share a well-appointed bathroom with a combination tub and shower.
A dedicated home office should also be
situated on the upper floor, providing a quiet and comfortable workspace for
Carla. This room should have large windows to allow for ample natural light and
inspiring views, as well as built-in storage and desk solutions.
Outdoor Spaces: The outdoor spaces should
emphasize the connection between the house and its natural surroundings. The
cantilevered deck or patio should be accessible from the main living area and
should include an outdoor dining and entertaining
space, complete with built-in seating and a barbecue area. The landscape design
should incorporate native plants and trees, as well as sustainable features
like a rainwater harvesting system and a green roof section.
Sustainability and Technology: Carla's
Dream House should be designed with sustainability and technology in mind.
Solar panels should be integrated into the design, either on the roof or as
part of the exterior cladding. The house should incorporate energy-efficient
appliances, low-flow fixtures, and high-performance insulation. Smart home
technology should be integrated throughout the house, including advanced lighting
and climate control systems, security, and home automation features.
In summary, the architectural illustrator
should create detailed renderings of Carla's Dream House that showcase the
modernist design, the warm and inviting atmosphere, the seamless connection
between indoor and outdoor spaces, and the integration of sustainability and
technology. The renderings should capture the essence of Carla's vision for her
dream home, from multiple perspectives both inside and outside the house.
Below is an expanded data model for the Carlinsky
Residence project presented in a tabular format. The tables represent the
different entities and their attributes.
Table: Clients
Attribute |
Data Type |
Description |
ClientID |
Integer |
Primary Key, Unique client identifier |
Name |
Varchar(100) |
Client's full name |
Phone |
Varchar(20) |
Client's phone number |
Email |
Varchar(100) |
Client's email address |
Address |
Varchar(255) |
Client's address |
Table: Projects
Attribute |
Data Type |
Description |
ProjectID |
Integer |
Primary Key, Unique project identifier |
ClientID |
Integer |
Foreign Key, Related client identifier |
ProjectName |
Varchar(100) |
Name of the project |
Location |
Varchar(255) |
Project location |
Budget |
Decimal |
Project budget |
Timeline |
Integer |
Project timeline in months |
Status |
Varchar(50) |
Project status |
Table: DesignRequirements
Attribute |
Data Type |
Description |
RequirementID |
Integer |
Primary Key, Unique requirement identifier |
ProjectID |
Integer |
Foreign Key, Related project identifier |
Category |
Varchar(50) |
Category of the requirement (e.g., Sustainability, Interior
Spaces, Materials & Finishes) |
Description |
Text |
Description of the requirement |
Table: DesignInspirations
Attribute |
Data Type |
Description |
InspirationID |
Integer |
Primary Key, Unique inspiration identifier |
ProjectID |
Integer |
Foreign Key, Related project identifier |
ArchitectName |
Varchar(100) |
Name of the inspiring architect |
ProjectName |
Varchar(100) |
Name of the inspiring project |
Table: TeamMembers
Attribute |
Data Type |
Description |
TeamMemberID |
Integer |
Primary Key, Unique team member identifier |
Name |
Varchar(100) |
Team member's full name |
Role |
Varchar(50) |
Team member's role (e.g., Architect, Interior Designer,
Landscape Architect, Structural Engineer) |
Phone |
Varchar(20) |
Team member's phone number |
Email |
Varchar(100) |
Team member's email address |
Table: ProjectTeam
Attribute |
Data Type |
Description |
ProjectID |
Integer |
Foreign Key, Related project identifier |
TeamMemberID |
Integer |
Foreign Key, Related team member identifier |
Table: Milestones
Attribute |
Data Type |
Description |
MilestoneID |
Integer |
Primary Key, Unique milestone identifier |
ProjectID |
Integer |
Foreign Key, Related project identifier |
MilestoneName |
Varchar(100) |
Name of the milestone |
StartDate |
Date |
Milestone start date |
EndDate |
Date |
Milestone end date |
Room Name |
Square Footage |
Number of Doors |
Number of Windows |
Foyer |
70 |
1 |
0 |
Living Room |
350 |
1 |
4 |
Dining Room |
240 |
2 |
2 |
Kitchen |
420 |
1 |
4 |
Family Room |
400 |
2 |
6 |
Laundry Room |
100 |
1 |
0 |
Master Bedroom |
400 |
1 |
4 |
Master Bathroom |
180 |
1 |
1 |
Master Closet |
150 |
1 |
0 |
Bedroom 2 |
250 |
1 |
2 |
Bathroom 2 |
90 |
1 |
1 |
Bedroom 3 |
200 |
1 |
2 |
Bathroom 3 |
70 |
1 |
1 |
Office |
150 |
1 |
2 |
Powder Room |
50 |
1 |
0 |
Garage |
800 |
2 |
0 |
Patio |
200 |
1 |
0 |
Deck |
400 |
1 |
0 |
Item |
Low-End Estimate |
High-End Estimate |
Site Preparation |
35000 |
50000 |
Foundation and Framing |
420000 |
440000 |
Electrical and Plumbing |
160000 |
180000 |
HVAC System |
80000 |
90000 |
Roofing and Insulation |
60000 |
70000 |
Windows and Doors |
90000 |
120000 |
Exterior Finishes |
100000 |
140000 |
Interior Finishes |
280000 |
350000 |
Cabinetry and Countertops |
100000 |
140000 |
Appliances and Fixtures |
50000 |
70000 |
Landscaping and Hardscaping |
100000 |
140000 |
Pool and Patio |
100000 |
140000 |
Design and Permitting |
60000 |
70000 |
Contractor's Overhead and Profit (10%) |
163500 |
200000 |
Total |
1798500 |
2200000 |
As Bob, considering the construction cost
estimate provided by Gus, which ranges from $1.8 million to $2.2 million, I
would need to add my architectural fee to determine the final price for the Carlinsky
Residence project.
Typically, architectural fees can range
from 10% to 15% of the total construction cost. For this project, let's assume
a fee of 12%.
Using the lower end of the construction
cost estimate ($1.8 million) as an example:
Architectural fee: $1,800,000 * 0.12 =
$216,000
Final price (lower end): $1,800,000
(construction) + $216,000 (architectural fee) = $2,016,000
Using the higher end of the construction
cost estimate ($2.2 million) as an example:
Architectural fee: $2,200,000 * 0.12 =
$264,000
Final price (higher end): $2,200,000
(construction) + $264,000 (architectural fee) = $2,464,000
Based on these calculations, the final
price for the Carlinsky Residence project, including the contractor's estimate
and my architectural fee, would range from approximately $2,016,000 to
$2,464,000.
[1]
NB: While I was writing this article, both Microsoft and Google announced plans
to integrate NLGAI across their work platforms, meaning, very shortly, you
won’t even need plugs like GPT for Sheets to complete an exercise like this one. Standard issue Microsoft office and Google
Docs will have this capability embedded by the end of the year. You could do this through your desktop Excel
application. Moreover, since the creation of the model, many experiments have
emerged utilizing Python and Pinecone to conduct conversations between NLGAI
agents. Chat GPT does not have a
‘memory’ in that it doesn’t remember anything outside of the conversation it’s
currently having. However, this is now
easily circumvented using other software.
On April 7th, researchers at Stanford published a research paper entitled
Generative Agents: Interactive
Simulacra of Human Behavior, wherein they duplicated my process at a large
scale (I’m not suggesting they levied my work, obviously, as it was not public
at the time). In their research, they
gave separate personalities to 25 different AI agents, and turned them
loose in an artificial world (like the SIMS video game). The 25 agents were able to interact, and also
plan spontaneous activities, like a Valentine’s day party. This suggests that my little model could be
greatly expanded! An entire kickoff
meeting composed of an AI client, an AI Architect, the AI Architect’s
assistant, an AI Geotech engineer, etc.
Similar models have been used to have AIs interact on dating platforms,
or to have two AIs collaborate on a game design (as well as write the code for
it), etc.
[2] In some cases, the prompt had to be manually trimmed, since MJ5 has an input limit of 2,000 characters.
[3] A serious limitation of MJ5 (and other image generation programs) is their consistency. Because they are constantly trying to ‘imagine’ the image described by your prompt, they interpret the words you write into pixels. However, it does not know that if you have two windows on the interior of the living room, you should have two windows on the exterior of the living room. Since those are different prompts, they are treated like entirely separate projects. Unlike GPT 4, it lacks any spatial intelligence whatsoever. Therefore, while it was capable of generating both exterior and interior design options that were consistent with Carla’s vision, it could not match those two together reliably. There are advanced methods of prompt engineering to enhance consistency across MJ5 production, but I felt that that would violate the spirt of the exercise. Moreover, there are better applications for achieving such consistency.
[4] It is critical here to continue using the same chat within GPT. GPT assimilates all previous exchanges within one chat, but not other chats. By keeping the entire project within a single chat, I was able to instruct GPT to utilizes Bob and Carla’s previous conversation to inform her reactions to the designs. This can be circumvented by pairing the language model with other applications, like Python and Pinecone, as noted in a previous footnote. However, if you lack coding skills, keep everything in one chat.
[5] Sparks of Artificial General Intelligence: Early experiments with GPT-4, Bubek, et al, arXiv:2303.12712v3, pp. 51
[6] This may just be a ‘token’ problem. For those unaware, GPT (and anything that works off of GPT) works on tokens. Your exchanges aren’t measured in dollars, they’re measured in computational difficulty, and the more difficult a request, the more tokens you’ll have to spend to complete it. There are all sorts of ways that you can control the ‘spend’ that GPT undertakes in solving your requests. In this case, my intuition was just that GPT was trying to limit the spending of tokens, so was defaulting to shorter messages.