AI Roundup 009: Italy, industry, and image detection

Apr 07, 2023

Last weekend, Italy banned ChatGPT. Though it happened in the wake of last week's AI safety letter, the ban was due to privacy concerns.

What to watch:

Rapid AI progress is forcing the hand of governments worldwide, and we don't yet know where the chips will fall.
Some, like Germany, are looking to follow in Italy's footsteps. Whereas India has said it does not plan to regulate AI.
In the US, the FTC has written multiple warnings about AI. And while President Biden recently weighed in on AI safety, there’s not much motion from Congress.

Yes, but:

After insults and injuries, OpenAI has responded with its approach to AI safety.
The write-up covers internal safeguards, factual accuracy, and, of course, protecting the children.

Stanford published a gargantuan 386-page report on the state of AI. If you don't have that kind of time, here's the high-level summary.

The big picture:

Academics used to lead the way in AI innovation; now it's firmly in the hands of industry.
More and more businesses and industries are deploying AI. Meanwhile, AI tools like Copilot are tangibly helping workers.
Generative AI has stolen the spotlight, but other breakthroughs are happening behind the scenes.

What's coming:

Github projects offer a peek at the next big thing: GPT agents.
Microsoft is working on a system where ChatGPT plans tasks and has other specialized models execute them.
Likewise, amateur projects like babyagi and Auto-GPT aim to give GPT-4 a goal and let it loose. Frankly, this is slightly terrifying.

This week Meta released its Segment Anything Model (SAM) which accurately IDs different objects in photos or videos.

Why it matters:

SAM seems to fundamentally "understand" object placement. How many humans could correctly outline the coral reef example?
AI has a concept of "foundation models," like GPT-3 and Stable Diffusion. When released, foundation models trigger a chain reaction of downstream apps and innovation.
SAM aims to become a computer vision foundation model. The more foundation models we have, the greater the diversity and capability of AI.

Between the lines:

Meta has been less visible in the AI headlines recently, but they aren't sitting on their hands. They're an AI research powerhouse, so assume we haven't seen all their cards yet.
We do know they're planning to bring generative AI to ad creation later this year.
In direct contrast to OpenAI, Meta open-sourced SAM’s code and training data. Meta's LLaMa model has also been open-sourced, mostly against their will.