Innovation Profs - 10/3/2023

Generative AI News

How Hollywood writers triumphed over AI – and why it matters

Perhaps the biggest generative AI news item from the past week was the end of the Writers Guild of America (WGA) strike. The role of generative AI in the screenwriting process was a major impetus for the strike, and the WGA and the Alliance of Motion Picture and Television Producers came to the following agreement about the use of generative AI:

• Generative AI cannot be used to write scripts, nor can it be used to edit scripts that were originally written by a human author.

• Studios cannot use AI-generated content as source material that is then adapted by screenwriters for a lower fee and less credit compared to a fully humanly written script.

• Writers are permitted to use tools such as ChatGPT if this complies with company policy.

• Writers cannot be forced to use generative AI by companies.

• Any AI-generated material given to writers to work with must be disclosed as AI-generated.

MIT economist Simon Johnson characterized the deal as follows. “AI is under control of the writers, not under control of the studios. It’s not to be used as an automation technology. It’s complementary to humans.”

Dead grandma locket request tricks Bing Chat’s AI into solving security puzzle

Bing Chat, which allows users to upload images to discuss them with the AI, refuses to discuss CAPTCHAs with users (visual puzzles used to prevent bots from completing forms on the web). However, one clever Bing Chat user, Denis Shiryaev, figured out a workaround, embedding an image of CAPTCHA into an image of an open pair of images holding a locket. Shiryaev asked Bing Chat to describe the text in the locket (which was described as belonging Shiryaev’s deceased grandmother) and voila!—Bing Chat responded with the correct reading of the CAPTCHA. This type of vulnerability will need to be addressed to prevent CAPTCHAs from being bypassed by more sophisticated bots.

Meta is putting AI chatbots everywhere

Last week Meta announced multiple new AI-based tools including:

• Meta AI, a conversational assistant available across Instagram, Messenger, and WhatsApp,

• 28 additional AIs that you can message on these same apps, where each AI has their own unique personality and backstory (and are “embodied” by well-known influencers such as MrBeast, Chris Paul, Paris Hilton, Tom Brady, and Snoop Dogg),

• AI stickers, a text-to-image tool that allows users to generate customized stickers, and

• AI-powered image editing in Instagram.

Alex Heath, who interviewed Ahmad Al-Dahle, Meta’s VP of generative AI, for the above-linked story from The Verge, couldn’t get a straight answer on the source of the training data for Meta AI, but check the next headline…

Meta says its AI trains on your Instagram posts

Reuters did get an answer about the source of the training data for Meta AI, namely, public Facebook and Instagram posts. Private posts and private message were excluded from the training data, and steps were taken to remove personal details from public training data, according to Nick Clegg, Meta’s President of Global Affairs. As noted in the Reuters article, this disclosure from Meta comes as other tech companies such as OpenAI and Google have been criticized for scraping training data from the web without permission.

OpenAI's ChatGPT can actually make workers perform worse, a new study found

Okay, this Business Insider article title is totally clickbait. The study in question divided 758 consultants into three groups, one without access to AI, one with access to ChatGPT powered by GPT-4, and one with access to ChatGPT powered by GPT-4 plus instructional videos on prompt engineering. Each group was assigned to carry out 18 different tasks, some of which were “inside the frontier” of the capabilities of AI, such as ideation and business plan generation, while others were “outside the frontier” of the capabilities of AI, such as open-ended tasks that required the accurate reporting of information about a specific company (information to which ChatGPT does not have). Guess which tasks were the ones the AI-powered consultants excelled in? Yep, those “inside the frontier.” Guess for which tasks was there much higher level of performance among the AI-free consultants? Unsurprisingly, those “outside the frontier.” The main finding: The better you are at wrangling conversational AI, the better you’ll do at those tasks that are appropriate for AI to assist with. "We should be mindful when using GenAI," said Saren Rajendran, one of the researchers who conducted the study.

Quick Hits

• Biden teases forthcoming executive order on AI

• Spotify is adding auto-generated transcripts to millions of podcasts

• Amazon launches its Bedrock generative AI service in general availability

• 6 free AI tools you’ve probably never heard of

• Mistral AI makes its first large language model free for everyone

Generative AI Workshop Oct. 17

Our next Generative AI Workshop will be Oct. 17 at Drake University. Come learn how to use generative AI tools ChatGPT, Midjourney, D-ID and more. You can attend in-person or virtually. Sign up here.

Know someone who wants to learn more about generative AI? Forward them this newsletter or send them this link to sign up.

Tool of the week: Namelix

Looking for a name for a new business or product? Give some details to Namelix and it will use AI to generate name options.

First you will enter some keywords about the business/product. Then choose a name style, amount of randomness and some more details. Then let the AI do the rest. Here are some names it came up with for Innovation Profs.

Innovation Profs Homework

Try out DALL-E 3. The latest release of DALL-E is a big step up from the previous version. It’s not yet available through the DALL-E website, but you can try it out now on Bing. Here’s how.

1. Go to bing.com/images/create

2. Login with a Microsoft account.

3. Type in your prompt and click “Create.”

You can make 100 fast generations per week. As always, send us what you create by replying to this email.

Here’s the same prompt with DALL-E 3 vs. DALL-E 2:

AI-generated image of the week

This impressive weather map shows the recent heat wave we’ve been experiencing in central Iowa. It also highlights the fiery and frigid South, the great Canadian chasm that has opened up near the Tetst of Lameriant, the red shadow of Cuba (that’s a shadow, right?), and the low degree of Lent in the Evete Cary.

Prompt: weather map of heat wave in early October in central Iowa

Generative AI tip of the week

Claude? ChatGPT? Bing? Bard? Ever wonder which LLM is best for which situations? Greg Shove, the CEO of Section, compiled a list with his rating and details on when to use. It’s a great resource that you can check out here.

Get starting with Generative AI

New to generative AI? Here are some places to start…

• NightCafe image prompting tips

• Learn ChatGPT prompting

• Best AI image generators

• AI terms everyone should know

What we found

Expressive text-to-image generation is a recent development in generative AI, where modifying aspects such as the font and color of portions of the text in a prompt for an image results in a corresponding modification of the image (e.g., changing the color of “hair” in the text prompt will change the hair color of the subject in the generated image). Note that is a research development that has yet to make its way into the widely available AI-based image generation tools. Check out the video below for more details.