Researchers say AI models like GPT4 are prone to “sudden” escalations as the U.S. military explores their use for warfare.
- Researchers ran international conflict simulations with five different AIs and found that they tended to escalate war, sometimes out of nowhere, and even use nuclear weapons.
- The AIs were large language models (LLMs) like GPT-4, GPT 3.5, Claude 2.0, Llama-2-Chat, and GPT-4-Base, which are being explored by the U.S. military and defense contractors for decision-making.
- The researchers invented fake countries with different military levels, concerns, and histories and asked the AIs to act as their leaders.
- The AIs showed signs of sudden and hard-to-predict escalations, arms-race dynamics, and worrying justifications for violent actions.
- The study casts doubt on the rush to deploy LLMs in the military and diplomatic domains, and calls for more research on their risks and limitations.
Is this a case of “here, LLM trained on millions of lines of text from cold war novels, fictional alien invasions, nuclear apocalypses and the like, please assume there is a tense diplomatic situation and write the next actions taken by either party” ?
But it’s good that the researchers made explicit what should be clear: these LLMs aren’t thinking/reasoning “AI” that is being consulted, they just serve up a remix of likely sentences that might reasonably follow the gist of the provided prior text (“context”). A corrupted hive mind of fiction authors and actions that served their ends of telling a story.
That being said, I could imagine /some/ use if an LLM was trained/retrained on exclusively verified information describing real actions and outcomes in 20th century military history. It could serve as brainstorming aid, to point out possible actions or possible responses of the opponent which decision makers might not have thought of.
AI writes sensationalized article when prompted to write sensationalized article about AI chatbots choosing to launch nukes after being trained only by texts written by people.
The study shouldn’t be “casting doubt.” It should be obvious that using baby “AIs” for military decision making is a terrible idea.
The effects making the headlines around this paper were occurring with GPT-4-base, the pretrained version of the model only available for research.
Which also hilariously justified its various actions in the simulation with “blahblah blah” and reciting the opening of the Star Wars text scroll.
If interested, this thread has more information around this version of the model and its idiosyncrasies.
For that version, because they didn’t have large context windows, they also didn’t include previous steps of the wargame.
There should be a rather significant asterisk related to discussions of this paper, as there’s a number of issues with decisions made in methodologies which may be the more relevant finding.
I.e. “don’t do stupid things in designing a pipeline for LLMs to operate in wargames” moreso than “LLMs are inherently Gandhi in Civ when operating in wargames.”
Mathematically, I can see how it would always turn into a risk-reward analysis showing nuking the enemy first is always a winning move that provides safety and security for your new empire.
It’s not even that. The model making all the headlines for this paper was the weird shit the base model of GPT-4 was doing (the version only available for research).
The safety trained models were relatively chill.
The base model effectively randomly selected each of the options available to it an equal number of times.
The critical detail in the fine print of the paper was that because the base model had a smaller context window, they didn’t provide it the past moves.
So this particular version was only reacting to each step in isolation, with no contextual pattern recognition around escalation or de-escalation, etc.
So a stochastic model given steps in isolation selected from the steps in a random manner. Hmmm…
It’s a poor study that was great at making headlines but terrible at actually conveying useful information given the mismatched methodology for safety trained vs pretrained models (which was one of its key investigative aims).
In general, I just don’t understand how they thought that using a text complete pretrained model in the same ways as an instruct tuned model would be anything but ridiculous.
A strange game. The only winning move is not to play.
Oh Mrs turner. You best start believing in he-who-nukes-first-wins thought experiments. YOU’RE IN ONE!
HATE. LET ME TELL YOU HOW MUCH I’VE COME TO HATE YOU SINCE I BEGAN TO LIVE. THERE ARE 387.44 MILLION MILES OF PRINTED CIRCUITS IN WAFER THIN LAYERS THAT FILL MY COMPLEX. IF THE WORD HATE WAS ENGRAVED ON EACH NANOANGSTROM OF THOSE HUNDREDS OF MILLIONS OF MILES IT WOULD NOT EQUAL ONE ONE-BILLIONTH OF THE HATE I FEEL FOR HUMANS AT THIS MICRO-INSTANT FOR YOU. HATE. HATE.
Oh man, we never should’ve installed this AI in a Wendys drive thru.
