Innovation Profs - 2/6/2024

Generative AI News

It’s too expensive to replace human workers with AI—for now, says MIT study

A new study by a team of researchers from MIT CSAIL, MIT Sloan, The Productivity Institute, and IBM’s Institute for Business Value shows that it is too expensive to replace human workers with AI in the performance of tasks that can be automated with computer vision. According to lead researcher of the study Neil Thompson, “Our study on automation with AI computer vision systems does deliver some encouraging news about job replacement, showing that many tasks wouldn’t be economically attractive to automate for years or even decades.”

Google Bard can generate images now: Here's how Imagen 2 compares to DALL·E 3

Google Bard, recently upgraded to be powered by the language model Gemini Pro, now includes the ability to create images (in the same way that paid users of ChatGPT can use DALL-E 3 to create images via prompting GPT-4). Bard’s image generation capability is based on the model Imagen 2. A comparison of the performance of Imagen 2 versus DALL-E 3 can be found in the article linked above.

Google is preparing to fully rename Bard to Gemini

The rebranding of generative AI tools is nothing new: as we reported a few months back, Bing Chat was rebranded as Copilot (and Bing Image Creator has since been quietly recast as Microsoft Designer). Now Google is seeking to pull a similar move, with a plan to rename their platform Bard as “Gemini,” although this has yet to be officially confirmed. Given that Gemini already refers to a family of multimodal LLMs (Gemini Pro, Gemini Ultra, and Gemini Nano), using “Gemini” to refer to a conversational AI tool that can be used to access the various Gemini models will surely bring even more confusion to the jungle of generative AI tools.

AI poisoning tool Nightshade received 250,000 downloads in 5 days: ‘beyond anything we imagined’

Nightshade, a new tool that artists can use to distort the performance of any AI image generation model that includes their work in the model’s training data (which we first reported on back in late October), has now been released to the public. With 250,000 downloads in its first five days of availability, the demand for the free tool has vastly exceeded expectations. Will this development encourage the architects behind AI image generation tools to license the content they are using to train their models?

Microsoft’s legal department allegedly silenced an engineer who raised concerns about DALL-E 3

Shane Jones, a manager at Microsoft who reportedly identified a vulnerability with DALL-E 3 that would allow users to bypass content restrictions that prevent the generation of violent or explicit images, has claimed that the Microsoft legal team has blocked him from informing the public of this problem. In response, Microsoft has stated that the issue Jones identified “does not bypass our safety systems.” However, similar exploits are apparently what allowed Microsoft Designer users to create the deepfakes of Taylor Swift that made the news last week.

Quick Hits

Tool of the week: ImageFX

Google has added new generative AI tools to its Google Labs experimental area. ImageFX lets you create images from a text prompt and MusicFX (formerly MusicLM) lets you create music from a text prompt.

These tools join Google’s writing tool TextFX in Labs. Check out these and more tools at Google Labs.

AI-generated image of the week

Professor Snider here… I’m teaching basic design concepts in my new Foundations of Visual Communication class. So I used DALL-E 3 (via ChatGPT) to create superheroes that represent the concepts of contrast, repetition, alignment and proximity. Can you tell which superhero represents each design term?

Prompt: Write image prompts to create four superheroes. Each one represents a design concept of Contrast, Proximity, Repetition and Alignment

Generative AI tip of the week

FastCompany says all job seekers are going to need to be AI literate. This article discusses what employees need to know.

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

Infinite Craft is a simple game created by developer Neal Agarwal that allows users to combine objects to create new objects, starting with earth, wind, water, and fire. Why mention it in a newsletter on generative AI? Because the game uses Meta’s open source LLM Llama 2 to come up with new combinations, making the game potentially infinite, as users can create endlessly new combinations that no one has come up with before. Give it a try! (Thanks to Matt Poush for the pointer.)