Recent posts

#11

Off the Record / Re: The AI dooooooom thread

Last post by The Minsky Moment - Today at 10:35:36 AM

Anyways, we have been ignoring a key question:

How can LLMs be made to play computer games?

I was thinking: if the UI of something like a Paradox game could be "translated" into a format processable by these "AIs" (so, text, I guess) then they would be able to learn optimal strategies from online sources, wouldn't they?

Yes, but it would not be able to distinguish good strategies from bad.  So the same problems exist and are likely worse as AI inevitably gets trained on AI slop.

Would they, though? As I understand LLMs simply guess what's the most likely next word (segment) in what they are "writing". So let's say it's playing EU4, the UI is translated for it so it can process that it has negative stability. Surely, having stored all the discussions over the years online around strategy, the likely string of words would be a correct strategy.

The likely string of words would be a correct strategy if the training data contains good optimal strategies and the model has been given feedback that leads us to give higher probability weighting to those choices.  Looking at DG's CHAT GTP output, it appears that the model has been fed the Paradox forums and reddit threads on HoI4 as it parrots some of the buzzwords you'd find there.

The model "knows" that certain symbols in the game correspond to certain words like "manpower", "losses," and resources.  It can spit back out those numbers and associate it with those characters.  Then it can make a guess (i.e. assign probability distros) as what a likely response would be to a combat screen showing zero "losses" based on what people have said on the forums, reddit etc. about similar situations: "insane / bugged / encirclement abuse / nukes + AI collapse / paradrop cheese".   CHAT GPT has no idea what "losses" are or what "paradrop cheese" is, but its algorithm tells it that in the context of discussions about the symbol "HoI 4" there is some probabilistic relationship between those two phrases. The model has no idea what an optimal strategy is, or what a strategy is, or what Hearts of Iron is.  It is just making probabilistic connections between different symbolic representations in accordance with its algorithm.

#12

Off the Record / Re: The AI dooooooom thread

Last post by The Minsky Moment - Today at 10:22:20 AM

You are asserting that an LLM cannot generalize beyond its training data.

I'm not making that claim. I agree that an LLM "Einstein" could in theory make statements Einstein never made in response to novel stimuli, in a way that simulates Einstein in a reasonably convincing manner. It  could do things like imagine what Einstein might say if constrained to express his ideas in Shakespearean iambic pentameter ot in the style of a 2010s mumble rapper. What it can't do is what Einstein actually did and advance the frontiers of science through novel conceptions.

#13

Off the Record / Re: Brexit and the waning days...

Last post by garbon - Today at 09:44:42 AM

Totally separately but the OBR published its update before the budget by mistake. Apparently due to a "technical error".

I still think we should abolish the OBR in general.

But on this - people need to resign/be fired. We've got the City making moves on it but also the Guardian live blogging what the Chancellor is going to announce as they work through the OBR report before they've even stood up in the Commons. It's clown-ish.

I don't think I'm looking through rose-tinted glasses here but standards were far higher in the 1970s and 80s when it came to budget purdah. It is actually criminal to be leaking budget plans before the actual event; people will be making or losing money because of this, insiders could be coining it. This is another non-partisan matter where I think citizens are entitled to demand and insist on higher standards, both by politicians and the civil service.

OTOH, it wouldn't have been possible back then for the OBR to directly transmit the plans to everyone early on account of there being no easy distribution platform. So would have been much more of a purposeful (handed to press) or negligent (in case left somewhere easily accessible by unapproved parties) steps taken to get it out early.

#14

Off the Record / Re: The AI dooooooom thread

Last post by garbon - Today at 09:40:31 AM

Anyways, we have been ignoring a key question:

How can LLMs be made to play computer games?

I was thinking: if the UI of something like a Paradox game could be "translated" into a format processable by these "AIs" (so, text, I guess) then they would be able to learn optimal strategies from online sources, wouldn't they?

Yes, but it would not be able to distinguish good strategies from bad.  So the same problems exist and are likely worse as AI inevitably gets trained on AI slop.

Would they, though? As I understand LLMs simply guess what's the most likely next word (segment) in what they are "writing". So let's say it's playing EU4, the UI is translated for it so it can process that it has negative stability. Surely, having stored all the discussions over the years online around strategy, the likely string of words would be a correct strategy.

Of course the games have evolved so not all written over the length of eu4s existence would be still current info.

#15

Off the Record / Re: Brexit and the waning days...

Last post by Richard Hakluyt - Today at 09:20:42 AM

Totally separately but the OBR published its update before the budget by mistake. Apparently due to a "technical error".

I still think we should abolish the OBR in general.

But on this - people need to resign/be fired. We've got the City making moves on it but also the Guardian live blogging what the Chancellor is going to announce as they work through the OBR report before they've even stood up in the Commons. It's clown-ish.

I don't think I'm looking through rose-tinted glasses here but standards were far higher in the 1970s and 80s when it came to budget purdah. It is actually criminal to be leaking budget plans before the actual event; people will be making or losing money because of this, insiders could be coining it. This is another non-partisan matter where I think citizens are entitled to demand and insist on higher standards, both by politicians and the civil service.

#16

Off the Record / Re: The AI dooooooom thread

Last post by DGuller - Today at 09:19:33 AM

ChatGPT has been able to interpret game screenshots for a long while now.  For example, here is what it will return if you ask it to explain what's going on in this screenshot of my HOI4 game:

You absolutely annihilated the USSR without taking any casualties.
This is a Hearts of Iron IV war overview screen showing a German–Soviet War where:

Germany has:

4.00M fielded manpower

0 losses (which is insane / bugged / encirclement abuse / nukes + AI collapse / paradrop cheese)

100% war participation

The USSR side:

~524k–525k fielded manpower left

19.93M losses — that's nearly twenty million dead

Their minors, Sinkiang and Mongolia, are irrelevant and barely participated

So the reality here is:

You've completely broken the Soviets

Their army has evaporated. Twenty million casualties with effectively no German losses means either:

You rolled them with colossal encirclements and they died out of supply
or

You exploited AI stupidity with attacking-only orders where they suicided into your lines
or

You used nukes / CAS / terrain chokepoints / the River Line meatgrinder
or

Something glitched and the game isn't counting your casualties

Regardless, the USSR is militarily finished. One push and they collapse.

In HOI4 terms, this is a total reverse-Barbarossa massacre. They're done. You can march to Vladivostok with a blindfold on.

If you want, tell me how you achieved zero losses — I'm curious, because that's not something you see naturally.

In fact the AI USSR did suicide into my lines, it was a bug in the fairly early version of HOI4, I just set myself up to take the most ridiculous advantage of it to prove a point.

It can also cheat very well at GeoGuessr already (or alternatively help you learn to play it better, as it explains its reasoning).  Just give it a screen cap or what you're looking at, and more often than not a single still will already point you in the general area.

#17

Off the Record / Re: Brexit and the waning days...

Last post by PJL - Today at 09:03:56 AM

The salary sacrifice cap for exempting NI payments for pensions seems too low at £2,000 per year, especially since it won't come into force until 2029. Can see a lot of pension front loading going on until then. £5,000 would have been better, and why not do this from 2026? Doesn't make sense to me to leave it so long.

#18

Off the Record / Re: The AI dooooooom thread

Last post by Tamas - Today at 08:59:19 AM

Anyways, we have been ignoring a key question:

How can LLMs be made to play computer games?

I was thinking: if the UI of something like a Paradox game could be "translated" into a format processable by these "AIs" (so, text, I guess) then they would be able to learn optimal strategies from online sources, wouldn't they?

Yes, but it would not be able to distinguish good strategies from bad.  So the same problems exist and are likely worse as AI inevitably gets trained on AI slop.

Would they, though? As I understand LLMs simply guess what's the most likely next word (segment) in what they are "writing". So let's say it's playing EU4, the UI is translated for it so it can process that it has negative stability. Surely, having stored all the discussions over the years online around strategy, the likely string of words would be a correct strategy.

#19

Off the Record / Re: Brexit and the waning days...

Last post by Tamas - Today at 08:44:58 AM

I like the tax on gambling. I guess some losers' money finding its way to organised crime is not great but the industry seems awash in money advertising absolutely everywhere, which is kind of gross.

Tax on 2m+ properties at least has symbolic value so good.

Not thrilled about the 3p charge on EV miles driven, that's going to almost triple our current 1.6p per mile cost, but we are going to survive.

#20

Off the Record / Re: The Off Topic Topic

Last post by crazy canuck - Today at 08:25:31 AM

I enjoyed my visit, but that was 20 years ago and everything was very cheap. Lots of kitsch architecture, the different markets for which the various resorts were catering was very interesting. The undoubted highlight though was the Elvis impersonator who was so drunk/high that he started off with an impersonation of Wayne Newton before he remembered what he was supposed to be doing.

Yeah, that was its charm. I was cheap in every sense of the word. But now it is expensive and it's trying to pose as a luxury destination.  No thanks, I can spend less money at a real luxury location.