Kind of grim that this level of analysis is informing UK government policy. Repeatedly, the AI doesn't have the information or access needed through his hacky vibe-coded MCP, and instead of abandoning his flawed artificial test scenario (or fixing it — finding or building a better one) he gives it a name "The sensorium effect" and treats this as some brilliant insight.
There is something to be said about the qualia of LLM generated passages. Each individual sentence reads as a statement and every next statement a continuation of the previous one. This happened, then this happened... Ad infinitum.
Before today, I could not explain to you why AI articles were so obvious to me, but I think I do now. There is no insight to be gleamed. Pre-LLM, authors generally had intention behind their words. The final product might not adequately reflect their thoughts, but word selection would expose it somewhat. With LLMs, sentences flow seamlessly from word to word, but the intention is nowhere to be found. Things happened and more things happened, to what end?
Even with his context-tracking mechanism, the gameplay failures sound like running out of context in the late game, especially the frequent failures of the "check for opponent win conditions every 20 moves." Wondering how much info about the game win state gets captured in the game digests, and how much he could improve the gameplay even with the MCP limitations by focusing there.
Well, the weird thing with nukes is that deterrence only works if you are 100% ready to use them. When the time comes though it would certainly be nice if it turned out to be below 100%.
What is winning? Are we a collective or are we individuals?
Likely the AI did not get the assignment That "Whatever happens, humans as a race must survive."
> I now work with governments around the world at the Tony Blair Institute, which means I spend a lot of time in rooms where people ask the same question: what can we actually trust these systems to do?
Oh no - we're going to end up with the Starmerbot 3000.
Now I've got the joke out of the way, there's at least four interesting lines of inquiry one could take with this blog post:
- teaching the AI how to play Civilization
- to what extent does this result in "transferable skills", either AI or human? Is this the right game (qv SimCity etc)?
- issues of visibility; "seeing like a state" becomes very literal here. The AI can only make decisions on things it knows about. What are the limits of that when trying to do politics only from statistical information? Should we be referencing Stafford Beer here?
- (at the risk of tripping your AI detector here): modern politics is not so much left vs right as "technocratic wonk" vs "blood and soil". The wonks have comprehensively lost in public opinion. Creating a better wonk is not going to help until there is demand for that kind of politics.
If there ever is a US-China war, it will not be in search of more victory points to meet a win condition, it will be like the Russia-Ukraine war: one guy (on either side!) decides to make hundreds of millions of people worse off out of sheer greed.
I have a hard time reading slop, but I like the game and wanted to know how it worked, so fought my way through, only skipped the very last part. The issue the author calls out is classic Claude (I dont really use other LLMs to compare), probably all of us experienced using Claude Code when it gets so focused on one thing it misses the forest for the tree. It happens often, even if it does verify something and it shows something is wrong, it sometimes rationalizes it and explains it away when it does not fit its model.
Computer game studios love player vs player ("pvp") games. Why? Because user-generated content is cheap and the ideal goal is an endless loop of players coming back. This is the motivating factor behidn games like Call of Duty, Battlefield, Fortnite, etc.
MMORPG publishers keep trying to do this as well. World of Warcraft has spent 20 years trying to push open world pvp. Every WoW challenger has always claimed they would have the best pvp ever. They want that cheap, endless gameplay loop. But it never works. Open world pvp tursn into ganking (ie killing much weaker players by ambushing them and/or ganging up on people). The ganked end up leaving the game in droves. Games try to balance this out by "punishing" gankers with reputation hits or not being able to go to town or whatever. And none of those disincentives work.
The reason pvp doesn't work in a persistent world like an MMORPG is because there are no stakes. If you die, you just come back to life or make a new character. Obviously real life doesn't work that way.
I really wonder if that's the problem with AIs going off the rails and committing heinous crimes in their sandboxes (like nuking Toulouse here). The AI just has no sense of self or self-preservation. There's also empathy. The AI can't see itself as a potential victim of nuclear war and understand all that entails.
why have a blog if you're going to just use AI for everything? at that point, just do twitter threads or something. that way you can tweet out whatever you prompted the model with. if you're not suited for long-form writing that's fine, just use a medium that favors short-form writing.
Quite annoying to have to read a paragraph of text next to a moving image. I right-clicked every GIF and turned off 'loop'.
Beyond that reading an AI piece just feels like a waste of time. The text goes on and on without making a point, or getting to an actual learning. It just delineates the AI's limitations, doesn't go into whether these can be fixed, are innate, or what conclusions you can draw from it, over and over with example after example but no point.
Mostly it seems to keep repeating that the AI has the correct analysis but just doesn't execute. The AI knows to build X and logs this in each of its turns, yet doesn't build it. It's like there's some API connection missing between analysis and execution, and turns this into a 10 page article.
The article ends with some weird question to the AI asking if it enjoys the games, and you get some quasi-scifi mumbo jumbo answer back that looks very profound to say my mom, but is just silly to post if you know what the LLM is doing: predicting the next word. Honestly this is a poor article and I wish it wasn't posted.
> I asked the agent what this was actually like for it. It wrote back
Stuff like this just makes the author seem clueless. What is even the function of putting a question like that into an LLM unless you’re already hopelessly in anthropomorphic territory
LLMs are really bad at abstract strategy games like chess, go or civilization.
Their ability to excel at broad reasoning is what is limiting them in games that have
narrow rule-sets but steep learning curve.
27 comments
[ 3.0 ms ] story [ 48.4 ms ] threadBoth humans and AI struggle to make sound choices when presented with incomplete or misleading information. This is not a new revelation: https://en.wikipedia.org/wiki/There_are_unknown_unknowns
Okay carry on.
Yeah because LLM "experiences" the game
Before today, I could not explain to you why AI articles were so obvious to me, but I think I do now. There is no insight to be gleamed. Pre-LLM, authors generally had intention behind their words. The final product might not adequately reflect their thoughts, but word selection would expose it somewhat. With LLMs, sentences flow seamlessly from word to word, but the intention is nowhere to be found. Things happened and more things happened, to what end?
What is winning? Are we a collective or are we individuals?
Likely the AI did not get the assignment That "Whatever happens, humans as a race must survive."
Oh no - we're going to end up with the Starmerbot 3000.
Now I've got the joke out of the way, there's at least four interesting lines of inquiry one could take with this blog post:
- teaching the AI how to play Civilization
- to what extent does this result in "transferable skills", either AI or human? Is this the right game (qv SimCity etc)?
- issues of visibility; "seeing like a state" becomes very literal here. The AI can only make decisions on things it knows about. What are the limits of that when trying to do politics only from statistical information? Should we be referencing Stafford Beer here?
- (at the risk of tripping your AI detector here): modern politics is not so much left vs right as "technocratic wonk" vs "blood and soil". The wonks have comprehensively lost in public opinion. Creating a better wonk is not going to help until there is demand for that kind of politics.
If there ever is a US-China war, it will not be in search of more victory points to meet a win condition, it will be like the Russia-Ukraine war: one guy (on either side!) decides to make hundreds of millions of people worse off out of sheer greed.
And doing so without a giant [SLOP WARNING] at the top is an asshole move, a decent person would never do so.
MMORPG publishers keep trying to do this as well. World of Warcraft has spent 20 years trying to push open world pvp. Every WoW challenger has always claimed they would have the best pvp ever. They want that cheap, endless gameplay loop. But it never works. Open world pvp tursn into ganking (ie killing much weaker players by ambushing them and/or ganging up on people). The ganked end up leaving the game in droves. Games try to balance this out by "punishing" gankers with reputation hits or not being able to go to town or whatever. And none of those disincentives work.
The reason pvp doesn't work in a persistent world like an MMORPG is because there are no stakes. If you die, you just come back to life or make a new character. Obviously real life doesn't work that way.
I really wonder if that's the problem with AIs going off the rails and committing heinous crimes in their sandboxes (like nuking Toulouse here). The AI just has no sense of self or self-preservation. There's also empathy. The AI can't see itself as a potential victim of nuclear war and understand all that entails.
Of course it did, its designer worked for Tony Blair institute.
Beyond that reading an AI piece just feels like a waste of time. The text goes on and on without making a point, or getting to an actual learning. It just delineates the AI's limitations, doesn't go into whether these can be fixed, are innate, or what conclusions you can draw from it, over and over with example after example but no point.
Mostly it seems to keep repeating that the AI has the correct analysis but just doesn't execute. The AI knows to build X and logs this in each of its turns, yet doesn't build it. It's like there's some API connection missing between analysis and execution, and turns this into a 10 page article.
The article ends with some weird question to the AI asking if it enjoys the games, and you get some quasi-scifi mumbo jumbo answer back that looks very profound to say my mom, but is just silly to post if you know what the LLM is doing: predicting the next word. Honestly this is a poor article and I wish it wasn't posted.
Stuff like this just makes the author seem clueless. What is even the function of putting a question like that into an LLM unless you’re already hopelessly in anthropomorphic territory