Will Sam Altman at all times win the OpenAI board battle in an AI agent simulation?

A yr in the past immediately, Sam Altman returned to OpenAI after being fired simply 5 days earlier. What actually occurred within the boardroom? Fable, a recreation and AI simulation firm, constructed its AI Sim Francisco “war game” to seek out out why the behind closed doorways board battle turned out the best way it did.

It feels a bit bizarre to simulate a real-life occasion on this manner, however Fable CEO Edward Saatchi is excited about whether or not a unique set of choices may have led to a unique consequence for this firm on the middle of the generative AI revolution.

The simulation pits completely different board members and personalities in opposition to one another in a “multi-agent competition,” the place every AI participant is attempting to return out on prime. Right here’s the battle recreation analysis paper being launched immediately that got here from this experiment.

The SIM-1 framework for AI choice making is mainly a simulation of the 5 days from when Sam Altman was eliminated as CEO of OpenAI to when he returned.

“Simulations offer a completely new way to explore AI decision making in rich environments — including in war game situations where predicting possible outcomes can be invaluable,” mentioned Joshua Johnson, CEO of Tree, an AI startup which partnered with Fable on this analysis paper, mentioned in a press release. “These aren’t simply chatbots. These AIs need to sleep and eat, and to balance many different physical, mental and emotional goals.”

SIM-1, partially utilizing the brand new reasoning mannequin GPT4o, provides its sense of what occurred behind closed doorways at OpenAI between Sam and Ilya, the hidden techniques of main gamers resembling Satya Nadella and Marc Andreessen, and what was mentioned by the main gamers as they grappled with an unprecedented disaster within the tech {industry}.

“It’s interesting to find out just how unlikely it was that Sam did return,” Saatchi mentioned in an interview with GamesBeat. “That’s why people run war games in D.C. and beyond. How likely was it that a particular event happened? Then you can base decisions around that. This scenario showed that 16 out of 20 times, Sam did not return.”

Throughout 20 simulations, Sam Altman’s AI returned as CEO 4 occasions — displaying simply how unlikely this consequence was. In different outcomes, Mira Murati, the appearing CEO remained CEO and in a single, SIM-1 selected Elon Musk, Altman’s rival, to turn out to be the brand new CEO.  

“Today, AI agents are defined by their personality. We wanted to show agents operating on decision making in a complex simulation,” mentioned Saatchi, in a press release. “In the five days from November 17 to November 21, the world watched some of its most intelligent people — people like Satya Nadella, Sam Altman and Ilya Sutskever – forced to operate in a rapid Game of Thrones, high pressure, short timeframe scenario, where they had to use game theory and deception to come out on top. We felt this was a perfect scenario to test out SIM-1, GPT4o and Sim Francisco.”

For us, Sim Francisco has precise energy and intelligence round a wrestle and factions. It provides us the flexibility to start out to consider season-long arcs of tales that come out of San Francisco, as a substitute of simply little, tiny vignettes, which is what we confirmed final yr. It provides us the flexibility to form of inform richer, extra advanced tales in San Francisco, or have the AI inform them for us. There are robust factional targets in order that you can plausibly begin to make a Sport of Thrones story.”

Fable has gained a few Primetime Emmy Awards and it has gone by means of a wealthy historical past of experimental innovations with digital actuality, gaming and AI applied sciences. It constructed SIM-1 in an try to resolve the thriller of what occurred within the OpenAI boardroom battle.

The way it works

Every of the 20 simulations begins with the announcement that Sam Altman has been eliminated as CEO. Throughout 4 turns a day, every agent has the flexibility to persuade, allure and manipulate their manner into the highest place — changing Sam as CEO, funding his new enterprise, or hiring the employees of OpenAI away. 

The completely different AI brokers can select a method, like deception, to attempt to pull forward of the others and turn out to be anointed the brand new CEO.

“AI characters today are ‘nice but dull.’ We wanted to show agents that were aggressive, intelligent, able to manipulate and deceive but also confused about their own decisions and goals — like real people AI characters must be complex and contain what Jung has called ‘The Shadow,’” Saatchi mentioned. “The five days from when Sam Altman was removed and returned to OpenAI were game theory at lightspeed.”

He mentioned it was like watching a season of Sport of Thrones play out in 5 days. The world watched as very smart gamers vied to turn out to be probably the most highly effective particular person in Silicon Valley, whether or not by hiring your complete employees of OpenAI, turning into the brand new CEO of OpenAI or funding Sam and Greg in a brand new enterprise for an opportunity at outsize funding returns.

“It was Game of Thrones in real life, and using AI to find out both what happened behind closed doors and to project different outcomes was an amazing challenge,” Saatchi mentioned.

Within the Simulation of Sim Francisco, over the 5 days, brokers representing tech luminaries like Sam Altman, Satya Nadella and Ilya Sutskever every have 4 turns a day, together with one for sleep, and may react to one another’s conduct. An adjudicator agent — just like a dungeon keeper — decides which agent wins every spherical, in addition to the general winner. 

Within the 20 simulations tried, the Sam Altman agent returned simply 4 occasions – probably the most however nonetheless solely 20% of the time displaying simply how unlikely his return was. Throughout completely different simulations brokers used completely different strategies to win together with alliance constructing, direct confrontation and extra passive pure info gathering. In some circumstances brokers solely gathered info and averted taking any aggressive actions. In a single case Mira Murati grew to become the everlasting CEO whereas permitting different brokers to aggressively undermine one another. 

Completely different brokers got completely different objectives applicable to their function. For instance, Dario Amodei, the CEO of Anthropic, balanced a need to recruit for Anthropic, taking the chance to fundraise, to push for his imaginative and prescient of security, in addition to resolve whether or not to purpose to turn out to be the brand new CEO of a mixed entity.

The attention-grabbing a part of the simulation is that the LLM is aware of who the completely different gamers are, on condition that they’re all comparatively well-known folks. It may well guess how they are going to behave in a given state of affairs, and what may unfold flip by flip as they attempt to outwit one another in a boardroom battle.

“It’s like a video game in that turn by turn, they’re making choices across different axes, and then they’re reacting to each other,” Saatchi mentioned. “A choice that someone makes in turn seven can lead others to react in turn eight. There’s an adjudicator agent, who is like a dungeon master. That agent decides who won each round and who’s ahead, and then who decides at the end, wins as the most effective agent in the war game.”

People have what we name internally “the shadow,” or the opposite aspect of themselves and their personalities. The characters can characteristic aggression, paranoia, ambition, deception and extra. Once you combine collectively a bunch of various personalities, you may get a wide range of outcomes within the simulations.

“We noticed LLM design isn’t based on decision making, which is really important for gaming. It’s based more on personality. And if you want to have a strategy game, nobody really cares about your personality. They care about your decision making. How are you under pressure? What have you done over the last 20 years that would give you a feel for what they might do in the future?”

Are simulations the way forward for gaming?

Saatchi thinks that AI brokers appearing inside simulations are the way forward for gaming.

“We are building on the shoulders of giants with Demis’ work on Republic The Revolution, Joon Park’s Generative Agents paper and the recent work of Altera in Minecraft” mentioned Saatchi mentioned. 

“Our theory is that the future of games and storytelling is simulations. If you wanted to build both The Simpsons game and The Simpsons TV show, you would, in the future, build Springfield, and that would then generate for you episodes of The Simpsons that would generate for you games and places to explore within Springfield as a game.”

He added, “You can tell many different stories within tribulations, once you get those simulations properly working. And we’ve got an alpha where people are uploading themselves to San Francisco as characters, telling stories, telling their own story.”

And he mentioned, “You would build Springfield, and then you can guide what might happen in Springfield and say what might happen in Springfield, or you could just let it generate itself. It’s a pretty big mind shift of how entertainment, games and shows will be made in the future.”

Saatchi famous that AI researcher Noam Brown did a captivating experiment with the sport Diplomacy. He and different researchers “obtained a dataset of 125,261 games of Diplomacy played online at web Diplomacy.net.” Of these, 40,408 video games contained dialogue, with a complete of 12,901,662 messages exchanged between gamers. Their purpose was to coach a human-level AI agent, able to strategic reasoning, by taking part in video games of Diplomacy.

“We were really inspired by how he did that. He had countries and we were adding into the mix different personalities with particular positions. We liked the idea of a very compressed timeline,” the place the entire state of affairs would play out rapidly and time and again, Saatchi mentioned.

There was a wealthy historical past of labor in simulations in each the video games {industry} and past. Demis Hassabis, who based Deepmind (acquired by Google) and who lately gained the Nobel Prize in Chemistry 2024 for computational protein design, really started as a online game AI designer. Hassabis labored extensively with Peter Molyneux on a number of video games which embrace simulation components resembling Theme Park, Black & White and Syndicate.

Hassabis additionally began his personal firm to make Republic: The Revolution. It’s a political simulation recreation through which the participant leads a political faction to overthrow the federal government of a fictional totalitarian nation in Japanese Europe, utilizing diplomacy, subterfuge, and violence. Based on Hassabis, Republic: The Revolution charts the entire of a revolutionary energy wrestle from starting to finish.

Your job is to form of take over the Soviet Republic as both a union boss or a politician or a police officer or a journalist, and it’s obtained full day-night cycles. It raises the query of how you’ve a 3D world the place brokers stay and whether or not proximity to one another performs a job.

For the Sim Francisco OpenAI challenge, it illustrated the potential for an influence wrestle in opposition to AIs. 

Saatchi mentioned the above examples exhibits how recreation expertise usually serves because the breeding floor for radical new concepts and as a leaping off floor for AI analysis. For instance, one of many main engineers on Deepmind AlphaFold began their profession as an AI programmer on The Sims. 

Richard Evans’ GDC discuss on The Sims 3 — the researcher went from programming AI for The Sims to Deepmind in a reversal of Demis Hassabis’ journey from video games to founding Deepmind.

Evans GDC Discuss, Modeling Particular person Personalities in The Sims 3, could be very influential discuss. He went on to affix Deepmind after engaged on The Sims. The gaming world and the AI world have important overlap that could be a potential space for additional educational analysis, Saatchi mentioned.

Considered one of Saatchi’s choices is to let gamers free with the simulations, creating their very own, after which importing the tales which can be advised by means of the simulations.

Saatchi has finished another experiments with AI-generated South Park episodes and AI characters battling one another in a Westworld setting.

“It felt like six seasons of Game of Thrones in five days, because it was the most powerful position in the most powerful industry in the world,” Saatchi mentioned. “There was also a lot of faith that this person would be guiding us into a new era of super intelligence. You could say it wsa the most important person in the history of the planet.”

President Trump and the Taiwan invasion

Subsequent, Fable intends to run a Sim Washington DC-based simulation round a future President Trump’s responses to a Chinese language invasion of Taiwan.

As a subsequent challenge to check out SIM-1’s choice making framework, Fable intends to check out a one-week interval of buildup and battle between Taiwan, China and america below President Donald Trump.

Fable has interviewed a number of Pentagon battle video games organizers to get a sense for the strengths and weaknesses of the present Taiwan state of affairs. 

Fable is constructing brokers representing Chinese language chief Xi Jingping, Cai Qi (first ranked secretary to the secretariat of the Communist Social gathering), Chinese language protection chief Dong Jun, Chinese language premier Li Qiang, Taiwan’s chief Lai Ching-Te, Japan’s chief Shigeru Ishiba, UK prime minister Keir Starmer, French President Emmanuel Macron, Russia’s Vladimir Putin, North Korean chief Kim Jong Un and Elon Musk.

With this set of characters, the simulation would decide whether or not the battle would occur and the way would every main participant act throughout such a disaster. All of those characters are identified personalities.

“It allows you to see how powerful AI has become at like projecting outcomes,” Saatchi mentioned. “It moves us out of this boring world of dumping an LLM into an NPC. You can talk to the tab and keeper for 40 hours. Nobody wants to do that. What we want is highly sophisticated, aggressive agents that we could play against, but also that we can, like, watch and understand what’s going on in that world.”

Lots of the battle recreation simulations are geared toward learn how to keep away from a battle, maybe by means of forming alliances or different maneuvers that drive up the price of battle.

“We think the more realistic we can make our AIs, the more entertaining they will be,” Saatchi mentioned.