What's been pinging my AI scanner

Things I'm tracking, late November 2023 edition

Greetings, readers. It’s time for another scanner post, where I share what I’ve recently found in my research into generative AI and the future of education.

I have a lot to share now, too much for a single post, and so will send one or two more of these in a sequence.

TECHNOLOGY

My favorite recent project might be MonadGPT, a chatbot trained on early modern English, French, and Latin texts. It uses the OpenHermes 2 - Mistral 7B open source AI tool and the results read a lot like they were written around 1600. I find them to be delightful. For example:

Setting aside my literary bemusement, this is an example of the narrow-function generative AI I’ve been forecasting. Again, imagine someone grabbing OpenHermes and training it on, say, a vast corpus of one religion’s primary texts to create a religious counselor. Or doing this with a lot of Chinese Communist texts to yield a Maobot.

Microsoft is considering atomic power to power its AI services. Apparently (paywalled) the goal is to switch away massive electricity draws from CO2-emitting sources. One interesting angle involves using generative AI in the enormous permitting process:

For the past six months, a team of Microsoft employees have been training a large language model with U.S. nuclear regulatory and licensing documents, hoping to expedite the paperwork required for such approvals, which can take years and cost hundreds of millions of dollars.

Integrating AI into non-AI apps and services More examples of this process keep appearing. For example, Snapchat launched an image generator, which users control through text prompts. Meanwhile, Instagram has added an AI-backed image generator, specifically for backgrounds:

When users will tap on the background editor icon on an image they will get ready prompts like “On a red carpet,” “Being chased by dinosaurs,” and “Surrounded by puppies.” Users can write their own prompts to change the background as well.

Once a user posts a Story with the newly generated background, others will see a “Try it” sticker with the prompt so they can also play with the image generation tool.

Here’s AI in Mailchimp, when I worked on my other newsletter earlier this month:

That’s text. Examples occur from the images world as well. Shutterstock launched an AI image editor. Canva added an image generator.

Why this matters: these stories indicate the spread of AI into the digital world beyond AI-focused applications. They are also more support for my hypothesis that those AI-specific apps will decline or vanish over time, as most users become more likely to use the tech in software aimed at other, more commonplace purposes.

IBM launched a governance tool through Watson for AI. They call it governance, but it sounds more like a management dashboard. This might be a growing service area to watch and work in.

Meta disbanded its responsible AI unit. Staff were reassigned to other functions, which look more production-oriented. It’s hard to assess this, given a lack of transparency, but it looks like a signal that Meta is less interested in that kind of internal guidance or critique:

RAI was created to identify problems with its AI training approaches, including whether the company’s models are trained with adequately diverse information, with an eye toward preventing things like moderation issues on its platforms.

I don’t know how far we can generalize from this story - i.e., will other businesses cut their AI ethics units. Popular skepticism about AI is growing and criticism continues to appear, so it should be a good move for a firm to at least make a formal show of responding to it.

A big open source release Open Source AI Red Pajama posted a gigantic training dataset. RedPajama-V2 includes 100 billion texts or 30 trillion tokens. I’m not sure how much memory it takes up; a HackerNews thread spoke of around 100TB. (I haven’t downloaded it myself because I haven’t wrangled a big PC for my AI work yet.)

Building a computer game entirely by AI One Javi Lopez created a computer game, ann Angry Birds knockoff, by using two AI tools. First, he prompted ChatGPT into building the code. Second, he used MidJourney to create visual assets. It took him 10 hours.

Why does this matter? First, I’m going to use it as a prompt in at least one class. More to the point, it represents an extension of the generative AI generating code and potentially generating coder unemployment. Now it’s functions which combine code with media - gaming, here - which may face un- or underemployment. Tom’s Hardware observes accurately that “Games development is an expensive affair.” So we might see AI eat up gaming from above (investors/managers automating design) and below (individual creators using AI to make games).

AI for weather forecasting. A new DeepMind tool conducted decent weather work, more quickly than traditional digital methods.

---

There is a lot more in my hopper, but this is enough for now, especially given the utter madness of my schedule. Stay tuned for the next scanner post.