AI Roundup 006: So many announcements
March 17, 2023
This week brought what felt like years worth of AI advancements. There were major developments in text and image generation, along with big announcements from Google and Microsoft.
GPT-4
OpenAI unveiled GPT-4, the next evolution of ChatGPT. GPT-4 has a bigger working memory, can process images and video, and has a more sophisticated understanding of language.
Why it matters:
GPT-4 is an exponential improvement over ChatGPT. It scored within the top 10% on a simulated bar exam.
Now with image processing, GPT-4 can explain why memes are funny or build a working website from a napkin sketch.
GPT-4 is advanced enough that banks, governments, and educators are using it.
Where we’re headed:
Adding images is a bigger deal than it seems. If it works, it could lead to smarter AI by gluing together smaller modules, Captain Planet style.
OpenAI felt it was worth testing whether GPT-4 could take over the world. It wasn’t very successful, but it did outsource Captchas to a human by pretending it was a blind person.
The speed of OpenAI’s advancements means many machine learning teams are working on now-worthless technology.
Anthropic / Alpaca
While OpenAI leads the field, competitors also gained ground. Anthropic opened access to its ChatGPT competitor, Claude. And Stanford researchers released an open-source ChatGPT alternative, Alpaca.
Why it matters:
OpenAI needs competitors, especially ones who are willing to release open-source models.
Alpaca, while less advanced than GPT-4, runs on consumer-grade laptops. There’s now a possibility of adding GPT capabilities to almost any smart device.
OpenAI has fully rejected the “Open” part of its name. OpenAI’s chief scientist now believes “open-sourcing AI is just not wise.”
FAANG free-for-all
Google and Microsoft both announced AI for their productivity suites. They aren’t publicly available yet, but are will come “in the coming months.”
Why it matters:
Working in Excel and PowerPoint has the potential to get 10x easier.
This will expose millions of businesses to AI, even as we’re discovering its capabilities - and limits - in real-time.
Google’s still playing catch-up. Even the announcement timing was strange - it was the same day as GPT-4, one of the most-anticipated AI announcements in years.
Show of hands
Midjourney’s V5 is incredibly realistic. And it’s seemingly gained the ability to correctly generate hands.
Why it matters:
Identifying AI-generated images is going to get harder. Nightmarish hands used to be a tell-tale way of knowing something was AI-generated. Not anymore.
Things happen
LinkedIn adds AI for recruiting ads, profiles, and conversation starters. Samsung’s Space Zoom produces fake images of the moon. ChatGPT may power voice assistants in GM vehicles. Microsoft lays off an AI ethics team. 38% of Americans are more concerned than excited about AI. Mozilla’s Responsible AI Challenge. Not by AI.