Good morning! This is The Neuron. AI's moving fast but we move faster. Let's go.
Today in AI:
- We've Got More Microsoft/OpenAI Deal Details
- New Research: DeepMind Plays Minecraft, Meta Reads Lips
- Around the Horn
- Leo Sends His Regards
We've Got More Microsoft/OpenAI Deal Details
In hindsight, that obviously wasn't the full picture.
On Monday, we thought Microsoft was investing $10 billion, taking 75% of profit until it was paid back, then owning 49% of OpenAI afterwards.
We've got more deal details. Microsoft still puts in $10 billion. Then, it gets complicated, with 4 stages:
- [First close partners payback] A set of investors (called "first close partners" or FCP) get 100% of profits until their investments are paid back
- [Microsoft payback + first close partners profit] Then, Microsoft starts taking 75% of profits. The other 25% goes to employees and FCPs' "profit cap". This lasts until Microsoft gets its investment paid back.
- [Microsoft profit + employees] Then, 49% of profits go to Microsoft for its "profit cap", 41% goes to employees, 8% goes to FCPs' "profit cap", 2% goes to the OpenAI nonprofit.
- [OpenAI takes over] Once that's all done (everyone has maxed out their profit cap), the OpenAI nonprofit gets 100% moving forward.
Whew. That's a doozy.
Wait, hold on...
- "Profit cap"? OpenAI used to be a nonprofit. Then in 2019, it decided it was down to make some money, but *only* up to a 100x return.
- "First close partners"? It's not clear what this means, but the for-profit version of OpenAI took $1 billion from investors when it started. We think FCP refers to those investors.
So what's the final math? All in all, if we make it to Stage 4, Microsoft walks with $92 billion in profit and other investors with $150 billion.
Isn't that a lot? Yeah. About that. OpenAI is basically betting it's gonna be as big as Microsoft itself ($72 billion in 2022 profit). You only get to Stage 4 by spitting that kind of profit.
As complicated as it is, this deal structure makes more sense for OpenAI than before. Giving away 49% to Microsoft forever smelled fishy. Now, OpenAI's putting their money where their mouth is on AI being the next revolution.
New Research: DeepMind Plays Minecraft, Meta Reads Lips
A double header on new models.
#1: Google subsidiary DeepMind loves games. Its AI pushed the #1 ranked player in Go to retire, and now it's solving challenges in Minecraft.
Minecraft is one big random sandbox. The world goes on forever, with no "right" thing to do.
So mining a diamond can be a challenge even for human players. You have to scope out a randomly generated map, collect the right materials to build the right tools, then find and mine diamond. And you can't die.
DeepMind developed an algorithm called DreamerV3 that figured out how to do just that, with no hints or any idea of what Minecraft even was.
That last part is key. Other teams have completed the task before but they had to give their algorithms a few hints to show it the ropes. DreamerV3 got it "out of the box".
Why it's important: Games may seem like no big deal, but some act as "dumbed-down" versions of the real world. Especially when they involve things like randomness (like Minecraft) or negotiation (like Diplomacy, which Meta created a model for).
#2: Speaking of, Meta released ReVISE, a model that can read your lips.
Even better? It can do it when there's a ton of background noise, bad audio quality or people talking over each other.
Take a listen for yourself on their project page.
Around the Horn
- OpenAI and Georgetown published an 80-page report on ways this wave of AI models could be used for disinformation and how to mitigate the risk
- Take a pic of trash, and EcoSnap will tell you how to recycle it.
- Seek.ai (ask questions about your data) raises $7.5 million.
- Create vector illustrations using text with Illustroke.
- This Chrome extension reveals hidden features in ChatGPT.
- Vivid uses AI to generate front-end code.
- A collection of 400+ AI tools you can explore: SaaS AI Tools.
- [Job] Anthropic is looking for a Director of Communications.
- [Job] Students, a travel media company is looking for an AI intern to upgrade how they work using AI. Apply here.
- A 3,000-employee gaming company in China has named an AI as CEO.
- Law school dean calls for ChatGPT in the classroom.
- ChatGPT listed as an author in a research paper.
DM me links on Twitter: @nonmayorpete
PSA: If you see people online claiming that GPT-4 will have 100 trillion parameters vs. GPT-3's 175 billion, take it with a fat grain of salt. Rumors are that GPT-4 is powerful, but the numbers are speculative and clickbait-y.
Leo Sends His Regards
