ROUNDUP: Meta AI Wins at Diplomacy
Plus GPU shipments fall, and new Dive Computer App for Apple Watch Ultra
Facebook started its AI department in 2013 as Facebook Artificial Intelligence Research, or FAIR. When Facebook changed its parent company name to Meta, the department became Meta AI. Earlier this month it made news for developing the Evolutionary Scale Modeling algorithm that significantly accelerated the ability to discover new ways proteins can fold.
And now near the end of the month it has announced a new model called CICERO, that made it into the top 10% of people playing the board game Diplomacy at WebDiplomacy.net. This is the latest in a string of milestones for machines being able to play games as well or better than humans. IBM's Deep Blue famously defeated Gary Kasparov in 1997. DeepMind's AlphaGo bested Lee Sedol in 2016. DeepMind's AlphaStar got good enough to beat almost everybody at Starcraft in 2019. And now Meta's Cicero can beat most people at Diplomacy.
Each one of these milestones is step up in complexity. Diplomacy is a board game similar to Risk, but without dice, requiring players to make strategic alliances with other players in order to win. For a machine to do well at this, it must be good at both the strategy and the negotiations.
CICERO combined a strategic reasoning engine, similar to AlphaGo, with natural language models, much like GPT-3. The language processing used a 2.7 billion parameter model pre-trained on text from the internet, then fine tuned during more than 40,000 games against humans on WebDiplomacy.net. Filters were put in place to insure intelligibility and relevance. It then trained the model to accurately predict human interactions and to choose effective policies to implement. While the model isn't perfect, it has the advantage of not playing emotionally. So it never makes mistakes because it wants revenge, nor does it annoy other players with unnecessary ruthlessness.
The ability to pair a strategic model with natural language could be useful in the gaming world for things like making non-player characters- aka NPCs- more dynamic when providing quests and instructions to human players. It could also be used by intelligent assistants to better anticipate what you need during a conversation. This could be a boon for the Siris, Echos and Google Assistants of the world. And of course the usual, "somebody will try using this for bad things" warning applies. If you want a closer look at the system, MetaAI published the code on GitHub under an MIT OpenSource license.
Here are a few more things I wrote about last week.
Apple announced the Oceanic+ app is available for the Apple Watch Ultra.
Amazon and Russo Brothers to Make TV Show About FTX
GPU Shipments Fall. Chip Shortage May Have Reached Bottom
3D TVs Poised to Try a Glasses-Free Comeback
The Physics of the New World Cup Soccer Ball
Netflix Looking for AAA Video Game Director
Paid subscribers can scroll down to read more.
Keep reading with a 7-day free trial
Subscribe to Tom Merritt Tech Newsletter to keep reading this post and get 7 days of free access to the full post archives.