Leo's Lightbulbs

Hello everyone and welcome to Leo’s Lightbulbs! I’ll be covering the top news from the past week every Monday morning in your inbox. This newsletter was written completely by hand, without any AI assistance or even AI editing.

So, let’s jump right in! What happened in AI this week?

I think the most interesting news has been the “diversity gone wrong” in Google’s image generator, and the company’s response. First, let’s look at what happened.

Basically, in order to add diversity to prompts, many generators will add descriptors like “Black” or “Indian” to prompts, to counter the fact that image training datasets predominantly feature skinny young white people. It’s a kludge not a fix, since it doesn’t address the root cause.

How’d it go wrong? Rather than reading the context like engines such as Dall-E, Google’s generator just blindly applies races no matter what you request. Hence, Black nazis, Native American kings of England in the year 1300 etc.

What does Google’s response tell us? They pulled all generations of humans, temporarily, but once again this guardrail is a kludge since it doesn’t really change the underlying model or training data. This is the same as Google and Adobe’s kludge-y fix for copyright. They still use copyrighted materials to TRAIN their AIs, they just block using certain words in prompts, which doesn’t really address the issue.

💡Breaking news: Anthropic releases Claude 3 today!

💡Google’s Sundar Pichai offers rare public apology over “diverse nazis” and other weird images from Imagen2

💡Wendy’s faces outrage after announcing plans for demand-sensitive pricing using AI, pleasing nobody

💡Microsoft’s AI investments are spurring antitrust concerns. Microsoft pinky promises not to be anticompetitive.

💡Microsoft invests $16 billion in French AI startup Mistral, a competitor to ChatGPT/OpenAI

💡AI robotics startup Figure AI raises $675 million from private investors

💡Image generation startup Ideogram, which specializes in text, raises $80 million

💡Meta’s Zuck visits Samsung and LG in Korea for secret talks about competing with the Apple Vision Pro

💡Microsoft investigating harmful responses about suicide in its latest Copilot chatbot (based on ChatGPT4)

💡Not only does positive reinforcement and encouragement make AIs perform better, but telling them to pretend to be on Star Trek makes them better at math (really). The point of this sort of research is that due to the way AI was created (training rather than programming), it is a “black box” and we don’t understand how it really works.

💡Microsoft Windows 11 new update adds plugins to Windows Copilot

💡Coming soon, generative music from Adobe

Looking for more on AI? Check out Simplified AI by my friend Nayeem!

Have a wonderful week everybody! I check this email address ([email protected]) daily, so don’t hesitate to reply with any thoughts or questions for a chance to get featured next week!

Yours,

Leo💡