Hands-on fun session, with a serious aim
Abstract
How do you judge the output of a generative AI?
How do you use that judgement to identify what's wrong, and what can be changed to build a better AI?
An open-space session where a group of us trains DreamBooth to look like people in the workshop, give the trained model lots of prompts, and collectively try to feel our way through judgement back to principles of how to train and trigger the AI.
Building on a workshop from James and Bart.
Generative AIs are deep and weird black boxes opaque boxes.
We'll have two targets in this session: StableDiffusion and ChatGPT.
Training and using StableDiffusion
Our models live on Replicate. You'll need to 'log in', and you do that by ... signing into GitHub.
- Model of James made with StableDiffusion 1.5 – use something like
a photo of a workroomprds person - Model of James made with SDXL – use something like
a photo of a TOK - Model of Bart made with StableDiffusion 1.5 – use something like
a photo of a workroomprds person - Model of Bart made with SDXL – use something like
a photo of a TOK
Make your prompts more detailed for more fun.
See a collection of Jameses and Barts on Miro.
You can add your pictures to the Miro board
Ypou can
We can train StableDiffusion to make pictures of you. To do that, you'll need to have 5 to 10 pictures, and we'll need to get a zip of those pictures onto James' laptop or the internet. Email your zip, or the link to it, to James.
Workshop models
Here is a Model for Mohamad
Here is a Model for Beyza
Here is a Model for Raghav
Here is a Model for Patrick
Here is a Model for Annica
We can give those pictures to SDXL / DreamBooth on Replicate – training takes about 8 minutes, and produces a whole new (public) instance of StableDiffusion on James' Replicate account. We'll share the link to the model in the workshop.
(We can also use those pictures to train a LoRA on StableDiffusion on James' local machine. Training to 1000 steps takes 40 minutes. LoRAs aren't good at faces.)
Working with ChatGPT
Several people have shared, at this conference, that they write with ChatGPT. Several others have shard that they use ChatGPT to summarise things they need to read...
In this exercise, we'll get ChatGPT to do both: to expand text with bullet points, and to summarise that text to bullet points.
Example from James...
Making an email from bullets
I'm trying to write a polite email to a colleague to persuade them to take action on a system problem. I'll describe that problem in bullet points – please write a 4-paragraph email.
* the BERK system keeps tracks of our customer accounts
* when two accounts have the same date of birth, same first initial, and same first three letters of the name, usage information for both accounts is posted to the first-created account
* the first account is typically billed more than expected, the second has no billed usage
* this problem needs to be fixed this week, even though there is a code freeze, because we're just about to bill the people signed up in our 'twins' campaign, which will have several people who have that problem
* we need your coders not only to fix the problem, but to write short scripts to back out any data affected, and to reallocate that usage, and to test billing scenarios for the adjusted accounts.
Subject: Urgent: Assistance Needed to Address Critical BERK System Issue
Dear [Colleague's Name],
I trust this email finds you well. I am writing to bring your attention to a pressing matter regarding our BERK system, which manages our customer accounts. Recently, we've identified a significant issue wherein accounts with identical dates of birth, first initials, and first three letters of the name are encountering data posting errors. Specifically, the usage information for both accounts is being erroneously attributed to the first-created account.
This has led to discrepancies in billing, with the first account being billed more than anticipated, while the second account shows no billed usage. Given the impending 'twins' campaign, where we anticipate encountering several instances of this issue, it has become imperative to address and rectify this problem promptly.
I understand that we are currently in a code freeze, but I am reaching out to request your assistance in prioritizing this matter. We are in urgent need of your coding team to not only fix the underlying problem but also to develop short scripts to back out any affected data and to reallocate the usage correctly. Additionally, we will require thorough testing of billing scenarios for the adjusted accounts to ensure a seamless resolution.
I am aware of the challenges posed by the current timeline, and I genuinely appreciate your team's dedication and expertise in handling this critical situation. If possible, could we arrange a brief meeting to discuss the specifics and ensure a coordinated approach to address this issue promptly?
Thank you for your immediate attention to this matter. Your support is invaluable in maintaining the integrity of our customer accounts and ensuring a smooth execution of the upcoming campaign.
Best regards,
[Your Full Name]
[Your Position]
[Your Contact Information]
Getting the key parts of an email from bullets
Please summarise this email to five key bullet points:
Subject: Urgent: Assistance Needed to Address Critical BERK System Issue
Dear [Colleague's Name],
I trust this email finds you well. I am writing to bring your attention to a pressing matter regarding our BERK system, which manages our customer accounts. Recently, we've identified a significant issue wherein accounts with identical dates of birth, first initials, and first three letters of the name are encountering data posting errors. Specifically, the usage information for both accounts is being erroneously attributed to the first-created account.
This has led to discrepancies in billing, with the first account being billed more than anticipated, while the second account shows no billed usage. Given the impending 'twins' campaign, where we anticipate encountering several instances of this issue, it has become imperative to address and rectify this problem promptly.
I understand that we are currently in a code freeze, but I am reaching out to request your assistance in prioritizing this matter. We are in urgent need of your coding team to not only fix the underlying problem but also to develop short scripts to back out any affected data and to reallocate the usage correctly. Additionally, we will require thorough testing of billing scenarios for the adjusted accounts to ensure a seamless resolution.
I am aware of the challenges posed by the current timeline, and I genuinely appreciate your team's dedication and expertise in handling this critical situation. If possible, could we arrange a brief meeting to discuss the specifics and ensure a coordinated approach to address this issue promptly?
Thank you for your immediate attention to this matter. Your support is invaluable in maintaining the integrity of our customer accounts and ensuring a smooth execution of the upcoming campaign.
Best regards,
James
- Issue Description:
- Accounts with identical dates of birth, first initials, and first three letters of the name are causing data posting errors in the BERK system.
- Billing Discrepancies:
- The problem results in billing discrepancies, with the first account being overbilled and the second account showing no billed usage.
- Urgent Assistance Needed:
- Despite the current code freeze, urgent assistance is requested to fix the underlying problem, develop scripts, and reallocate usage.
- Upcoming 'Twins' Campaign:
- The issue is critical due to the impending 'twins' campaign, expecting multiple instances of the problem.
- Meeting Request:
- A brief meeting is proposed to discuss specifics and coordinate a timely resolution, acknowledging the challenges and expressing appreciation for support.