The Generative Workplace: How AI Is Already Disrupting Familiar Careers
Version 1.1
00:00 The Learning With AI initiative
Introduced by Jon Ippolito.
01:54 Study: 9 out of 10 businesses want ChatGPT skills
05:40 Simplifying an interface
Will AI be the death of web forms?
10:20 Learning new technologies
Will generative AI put coders out of a job?
AI may put quality assurance (QA) jobs at risk, but coders who can think abstractly will gain superpowers.
13:29 Translating code between programming languages
ChatGPT lies confidently.
15:42 Struggles with tabular data
ChatGPT can solve a word problem more easily than it can average numbers in a table.
Prompt engineering will be in demand.
18:00 Inspiration from OpenAI, Twitter, Copilot, and sample prompts
23:05 How should students be adapting to the future landscape
24:21 Generative AI is really derivative AI
In generative art, creators set the rules, but in generative AI a machine reverse-engineers rules based on outputs.
The industry is not predicting and steering around AI's failures but mitigating damage after encountering them.
"ChatGPT is essentially an interface without an application behind it."
28:02 Better code through dialogue
LLMs are good at making necessary components, less so at the plumbing required to fit them together.
LLMs can help with paperwork and overcoming "blank page" anxiety, but are not so good for literature review.
Large language models don't inherit human context.
33:18 Derivative AI as meme generator
34:41 Aesthetics of convenience
Overuse of the Ken Burns Effect made it cliché.
35:33 Text personalized for specific audiences
36:12 Skepticism about extracting structured data from notes
37:13 Are smaller models better?
Best uses may be specific, especially visual, cases since imagery is less lossy than language.
AI can produce 3d landscape from photos or a 3d avatar from a 2D video.
We may soon use text prompts to add assets and textures to virtual worlds (Geppetto's Workshop).
41:00 Parsing notes fields into structured data may get better
Misuse in healthcare and journalism remains a real danger.
Reducing a complicated text for a specific audience may be a better use.
44:29 Do LLMs create a logical model of the world? (Mike Hanley)
45:45 Improvement from new models. (Scott Arndt)
46:27 Text customization as danger
Tailoring for audiences might be helpful but risks reinforcing echo chambers and privacy concerns.
49:58 How will salaries be affected by generative AI? (Fox Gleason)
Rote jobs may be replaced, but not the ability to apply more abstract concepts.
52:28 History and social context matters
Before large language models, artists wielded more control over their models.
Precedents include rule-based generative art (Harold Cohen) and Generative Adversarial Networks (Anna Ridler and Daniel Hanley)
58:13 Today's indiscriminate data sets
By contrast, LLMs are based on "diffused" databases--they span almost every topic and image style.
Artists' contributions to the data set are like elections: "Your individual vote doesn't count, but without individual votes, there is no election."
60:10 The models can reproduce the original images in their training data
Professional creatives don't want to use someone else's work, both for legal cover and because they value originality.
61:44 "These tools are stereotype engines"
62:56 Who is responsible for context-less racist and sexist imagery?
"There are Nazis in the data set associated with the word hero."
Google performed a system-wide intervention to prevent stereotypic results for searches like "black girls," but AI systems generally do not.
"For our legal system to say these images are produced by AI is like saying the legal rights to a photograph belong to the camera."
70:17 Rights over AI-generated art and music
What will it do to the music industry when anyone can generate a song with the voice of another artist? (Emily Brule)
Automation already plays a major role in producing Pop music. (Eryk Salvaggio)
When you use a "song extension" tool like OpenAI's Jukebox, are you a producer or just an active consumer?
74:23 Should intellectual property treat AI-generated images, languages, and code differently? (Vijayanta Jain)
"AI is not one thing." (Eryk Salvaggio)
Recent court cases have hinged on the difference between generic function and expressive code. (John Bell)
79:40 Can/are individual websites be disallowed from the dataset? (Aiden Fuller)
Guardrails are by definition put on after the fact.
"Prompt injection": DALLE-2 injects unseen modifiers to force its model to return more diverse images than, say, just white male doctors. (Eryk Salvaggio)
82:41 New standards for labeling web content (John Bell)
New HTML headers could help structure content, eg labeling AI-generated or diverse sources, to prevent self-reinforcing datasets.
84:23 Future solutions: outlaw, regulate, specialize, or reset the systems?
Some researchers are exploring "machine unlearning," looking for ways to remove specific sources from an AI data set. (Eryk Salvaggio)
We could go back to pre-statistical AI "expert systems," and use ChatGPT as their interface. (John Bell)
89:46 Become a better overall writer (John Bell)
"Prompt engineering" syntax will change as the tools evolve, but natural language may become the standard for interacting with all digital systems.
92:57 Broaden your perspective (Eryk Salvaggio)
Rather than the details of today's tools, learn their purposes as well as social and environmental context.
This presentation was recorded by the University of Maine's New Media program. For more information, contact ude.eniam@otiloppij.
Timecodes are in minutes seconds
Generative AI is expected to transform any career that depends on writing, coding, or audiovisual media--in other words, just about all of them. Who better to forecast how workers will adapt to these disruptions than New Media alumni already incorporating AI into their own planning and practice?
This New Media webinar offers reports from working professionals on ways that ChatGPT and Midjourney are having an impact today on business, software, and art.
Presenters include artist Eryk Salvaggio, whose research focuses on the impact of generative AI on visual art and design; software engineer and accessibility expert Pattie Reaves, who's worked for Variety and Rolling Stone magazine; John Bell, who directs augmented/virtual reality and digital humanities programs at Dartmouth; and CourseStorm co-founder Matt James, who's designed applications that hundreds of thousands of people across the country use to connect to education.
This is part of a series of free webinars on cutting-edge technologies offered by the University of Maine's New Media program, which teaches animation, digital storytelling, gaming, music, physical computing, video, and web and app development. These are not Powerpoint lectures but guided demonstrations that students can follow at school or at home on their laptops. Learn more about these webinars or UMaine's Learning With AI initiative.
Watch the entire video or choose an excerpt from the menu on this page.