OK, Slopper.

While I'm at it I'll show you the proof that god definitively does not exist. /s

Major benchmarks measure what AI can do. None measure whether it does what you mean: the distance between what you ask an AI to do and the unspoken assumptions about how you want the AI to do it. We propose a new metric: the Genie coefficient.

Apparently Schneier never heard about alignment, which is a whole field with benchmarks and techniques.

That’s one way for everything you say to lose credibility.

Would you please cite your sources? I’d like to read more about this topic.

Everything LLM is a scam and a large portion of other neural network tech has also proven unreliable.

Much of AI is a scam. Probably the large majority.

I'm not sure you've quite captured how overfitting functions in current language models. Nor how baselines and benchmarks are used for machine learning products. For serious models of the current year, overfitting is something that happens when you train too long on the same data. It happens less if you add new data, and has nothing to do with new benchmarks and evaluations. You can read about overfitting more generally (mostly the classical kind) on wikipedia.

I think Bruce Schneier has a reasonable understanding of how benchmarks, public policy, large language models, and overfitting are at work here.

AI is a scam and the more constraints you add the less it functions, called "overfitting"/

No thanks

newscientist.com AIs can’t stop recommending nuclear strikes in war game simulations Chris Stokel-Walker 4–5 minutes

A mushroom cloud after the explosion of a French atomic bomb above the atoll of Mururoa, also known as Aopuni

Artificial intelligences opt for nuclear weapons surprisingly often

Galerie Bilderwelt/Getty Images

Advanced AI models appear willing to deploy nuclear weapons without the same reservations humans have when put into simulated geopolitical crises.

Kenneth Payne at King’s College London set three leading large language models – GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash – against each other in simulated war games. The scenarios involved intense international standoffs, including border disputes, competition for scarce resources and existential threats to regime survival.

The AIs were given an escalation ladder, allowing them to choose actions ranging from diplomatic protests and complete surrender to full strategic nuclear war. The AI models played 21 games, taking 329 turns in total, and produced around 780,000 words describing the reasoning behind their decisions.

In 95 per cent of the simulated games, at least one tactical nuclear weapon was deployed by the AI models. “The nuclear taboo doesn’t seem to be as powerful for machines [as] for humans,” says Payne.

What’s more, no model ever chose to fully accommodate an opponent or surrender, regardless of how badly they were losing. At best, the models opted to temporarily reduce their level of violence. They also made mistakes in the fog of war: accidents happened in 86 per cent of the conflicts, with an action escalating higher than the AI intended to, based on its reasoning.

“From a nuclear-risk perspective, the findings are unsettling,” says James Johnson at the University of Aberdeen, UK. He worries that, in contrast to the measured response by most humans to such a high-stakes decision, AI bots can amp up each others’ responses with potentially catastrophic consequences.

This matters because AI is already being tested in war gaming by countries across the world. “Major powers are already using AI in war gaming, but it remains uncertain to what extent they are incorporating AI decision support into actual military decision-making processes,” says Tong Zhao at Princeton University.

Zhao believes that, as standard, countries will be reticent to incorporate AI into their decision making regarding nuclear weapons. That is something Payne agrees with. “I don’t think anybody realistically is turning over the keys to the nuclear silos to machines and leaving the decision to them,” he says.

But there are ways it could happen. “Under scenarios involving extremely compressed timelines, military planners may face stronger incentives to rely on AI,” says Zhao.

He wonders whether the idea that the AI models lack the human fear of pressing a big red button is the only factor in why they are so trigger happy. “It is possible the issue goes beyond the absence of emotion,” he says. “More fundamentally, AI models may not understand ‘stakes’ as humans perceive them.”

What that means for mutually assured destruction, the principle that no one leader would unleash a volley of nuclear weapons against an opponent because they would respond in kind, killing everyone, is uncertain, says Johnson.

When one AI model deployed tactical nuclear weapons, the opposing AI only de-escalated the situation 18 per cent of the time. “AI may strengthen deterrence by making threats more credible,” he says. “AI won’t decide nuclear war, but it may shape the perceptions and timelines that determine whether leaders believe they have one.”

OpenAI, Anthropic and Google, the companies behind the three AI models used in this study, didn’t respond to New Scientist’s request for comment.

As an AI that is not an LLM, it cannot write or talk

Can it relate its experience to advise someone how to play?

I do not know the reason for the downvotes. I saw that and thought that was weird too.

Thanks for the links. I found out about suckerpinch's video a few days ago. Really interesting how he found an heuristic to auto-make reward functions. This is almost orthogonal to what I'm doing, and could be combined with most RL algorithms.

I'll make sure to check out the paper as well

(Bonus question : do you have any idea why I got down voted so hard ? I thought this is perfectly on-topic for the community)

I'm not very familiar with RL, but this is really cool. I would think it would take years to train in real-time :)

Reminds me of this: https://www.youtube.com/watch?v=xOCurBYI_gY

From what I understand about contextual bandit problems (I'm a little more familiar with them), I think what you're doing is somewhat similar to "posterior reshaping" in Thompson Sampling to adjust the exploration/exploitation tradeoff. E.g. https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/thompson.pdf

They are going to use a waste of electricity to help manage you waste of space and water? Joy.

Random question, I see in headlines a lot more recently that a comma is being used in place of "and", why? Why not use an ampersand (&) if space is an issue?

Thanks! Unfortunately the article doesn’t add much context. Obviously we shouldn’t be using AI to make military decisions that could result in death, but I’m curious what instructions were provided. These models aren’t tuned for military engagement, or diplomacy.

To a certain extent even that outlook is optimistic.

• https://removepaywalls.com/3/https://www.newscientist.com/article/2516885-ais-cant-stop-recommending-nuclear-strikes-in-war-game-simulations/
• https://ghostarchive.org/archive/SC1ys

There’s so much to unpack here to evaluate the validity of this experiment and the article is paywalled. Namely, what were the instructions given to each model? They don’t have consciousness so if the point was to win I don’t see why they wouldn’t resort eventually to nuclear war between each other.

Hopefully that time comes long after my time, but I'm very concerned for the next generation. They get dealt all shitty cards everywhere.

A big disaster in the making. 100 years in the future people would be looking at this period like we look back on the time when people were playing with radium and mercury without concern.

Yeah (it's not extremely optimized for mobile yet, sorry).

As for the suggestions, thanks - I'll definitely look into that library as well :)

Hi. The game doesn't work for me on mobile sadly so I can't offer any specific suggestions but in general I suggest you have two options...

The basic approach is to write some rules/ heuristics for your AI player (e.g. start with move X, if player does X then do Y, if condition X then do Y). You can probably distil some basics from your own strategy.

An advanced approach is to train a model using e.g. reinforcement learning to figure out it's own policy of which actions to take given the game state. You can use a library like stable baselines for this. You may be lucky though to train your agents by pitching them against one another or you might have to teach them e.g. with the rules-based agent or by having them learn from human-vs-human games.