I'm having trouble understanding all the upvotes for a wall of buzzwords from across-fields. Would love to understand what this even is in plain english if somebody has the time.
It's saying, "AI needs consistent, scored rules for communication, so it can self-train being a person the same way it self-trains winning games of Go" (sure).
Then it says "we have those rules" (they don't) "and they work great" (they don't) "and you can't see them" (sure).
Exactly! And it was obviouslly written by (or with help of) a LLM. The first definition is even gramatically incorrect! Lets see what gemini thinks of the first paragraph if I force it to criticize it:
thx for response. I think it is pretty straight forward on the site. project is in stealth mode. "Conversational Game Theory" is pretty straightforward.
Is there an exposition of the ideas and results of conversational game theory in the format and authorial style of an academic paper? Even if there is no intention to submit it to e.g. neurips, creating and prominently linking to such a pdf could help not only to refine the ideas but to communicate to outsiders. The linked website takes a while to get to the point and filled up my crank-o-meter before I got answers to basic questions like “what loss are the agents trained on” or “what do you think the equilibrium of this game is and why do you think that.” The field has settled on the (paper, github, blog post) link triplet for good reason.
I would also love some more traditional explanation of the ideas. My impression is the author is operating outside "the field", in a traditional academic sense, whether you are referring to AI/ML or game theory research. I had a bit of a search for citations or other mentions of their 9x3 nerrative logic or Conversational Game theory and only found their own sites (in a few different places https://bigmotherdao.com/f-a-q/, parley.aikiwiki.com) or an article under some contention by the author themselves on RationalWiki https://rationalwiki.org/wiki/Rome_Viharo. The mentioned professor James H. Fallon doesn't appear to have clear links to project outside of what is stated in the article.
My crank-o-meter (great phrase) overflowed at this point, but I am also genuinely interested in some kind of more clarity around the ideas, or where they came from in an academic legacy sense, as they felt like they might have some interesting bones.
Maybe someone with stronger google foo can find more, or confirm absence?
i think natural skepticism is a good idea. This is an independent project outside of academia, some think that projects outside of academia are crank projects by default, comes with the territory. Project should be judged on its own merits. This is just a stealth site.
It's worse than that. The entire notion of an independent arbiter is deceptive at best, because the decision making will always come down to how the questions are framed.
The claim here is that an automated process backed with enough data and computational power can distill any of the world's most intractable conflicts and come up with a compromise immune to "toxicity" and human bad-faith. I'd submit that in any conflict in which at least one actor is acting in bad faith to begin with, in the absence of an overriding ethical framework, any such compromise would result in a worse outcome than making a decision that one side was right and the other wrong. It's the compromise between hens and wolves.
If/since the universe of human bad faith is finite, it does seem possible (and useful) to detect all such forms in transactions that are on the record.
This won't solve out-of-band corruption and might indeed increase its value, and hence increase the value of participating in a corruption network.
There is a 77 minutes video in the Project Reviews page from the menu of the left. Maybe there is a demo in there but I didn't bother to check. It's conversational, it's text based, why didn't they include a textual demo in a page?
That's a static page with underlined text that are not links. Does anything interactive start after "Login With Google" (that I won't press) ? Anyway, my point is that they should have presented the transcript of at least one round of the process and demonstrate the improvement it produces. It's much less time consuming for a reader that does not want to invest hours in videos or actually taking part of the experiment. On the other side, those readers might not be of interest to the authors so this is a way to let them go.
Well, strictly speaking, it is suitable for places where all actors make choices in ways we can predict. So even if an actor was choosing, say, uniformly at random (and we knew this fact), we'd be able to incorporate it into analysis to get some expected behavior.
But game theory also presumes a closed, on-the-record world. E.g., it doesn't handle dumping (loss as strategic gain) without opening the scope of game, but there are no scope bounds since value of any type is exchangeable (land for peace...). So game theory like newtonian physics only helps in narrowly constrained systems, but none of the real systems where help is needed are so constrained (banking laws notwithstanding).
Yea, and so IMO this idea makes sense if you assume that bad conversation is predictable conversation (this is a simplified way of saying this but this idea was formalized within the field of literary theory by a man named Wolfgang Iser as a response to the issue of the hermeneutic circle and the merger of horizons). A conversation is evolved to its best possible outcome when it encourages the human in the loop to face their preconceived assumptions about the world and effectively do the work to come to their own conclusions rather than act as a follower
Once again an entirely backwards, upside down perspective. On Twitter he called politics the easiest job in the world.
Xenophon said that man is the hardest animal to tame. And the author’s “make AI great again” mentality is a perfect example of this untameability.
If it’s not actual AGI don’t tell me how a tool can actually fix the problems it creates. It’s like the author doesn’t have the most basic concept of what a tool is.
Desktop, as opposed to mobile. (on mobile the site looks pretty similar to most others). On desktop it feels like a pdf, but extremely light weight (in a very nice way; loads quickly, scrolls perfectly [unlike pdfs], and isn't cluttered). Just a crisp experience.
I'm having trouble understanding all the upvotes for a wall of buzzwords from across-fields. Would love to understand what this even is in plain english if somebody has the time.
It's saying, "AI needs consistent, scored rules for communication, so it can self-train being a person the same way it self-trains winning games of Go" (sure).
Then it says "we have those rules" (they don't) "and they work great" (they don't) "and you can't see them" (sure).
Exactly! And it was obviouslly written by (or with help of) a LLM. The first definition is even gramatically incorrect! Lets see what gemini thinks of the first paragraph if I force it to criticize it:
https://g.co/gemini/share/97c1d475aae1
Besides the obvious irony of criticizing a suspect LLM work and then asking another LLM to criticize it to prove it, your question was loaded:
"Explain why this paragraph is contrived and grammatically incorrect: [..]"
Current LLMs are trained to follow your leash.
thx for response. I think it is pretty straight forward on the site. project is in stealth mode. "Conversational Game Theory" is pretty straightforward.
Stealth mode project that is publishing whitepapers to HN and replying in comments about its stealthiness?
Double-secret misdirection. The new stealth!
Is there an exposition of the ideas and results of conversational game theory in the format and authorial style of an academic paper? Even if there is no intention to submit it to e.g. neurips, creating and prominently linking to such a pdf could help not only to refine the ideas but to communicate to outsiders. The linked website takes a while to get to the point and filled up my crank-o-meter before I got answers to basic questions like “what loss are the agents trained on” or “what do you think the equilibrium of this game is and why do you think that.” The field has settled on the (paper, github, blog post) link triplet for good reason.
I would also love some more traditional explanation of the ideas. My impression is the author is operating outside "the field", in a traditional academic sense, whether you are referring to AI/ML or game theory research. I had a bit of a search for citations or other mentions of their 9x3 nerrative logic or Conversational Game theory and only found their own sites (in a few different places https://bigmotherdao.com/f-a-q/, parley.aikiwiki.com) or an article under some contention by the author themselves on RationalWiki https://rationalwiki.org/wiki/Rome_Viharo. The mentioned professor James H. Fallon doesn't appear to have clear links to project outside of what is stated in the article.
My crank-o-meter (great phrase) overflowed at this point, but I am also genuinely interested in some kind of more clarity around the ideas, or where they came from in an academic legacy sense, as they felt like they might have some interesting bones.
Maybe someone with stronger google foo can find more, or confirm absence?
thx for feedback. project is in stealth mode. always a good idea to be skeptical.
I'm surprised something where the crankmeter goes off so strongly made it to the front page.
i think natural skepticism is a good idea. This is an independent project outside of academia, some think that projects outside of academia are crank projects by default, comes with the territory. Project should be judged on its own merits. This is just a stealth site.
I agree, but this is obvious crankery.
> Perfect conversations are made possible through Conversational Game Theory.
this is not the sort of thing someone who has thought critically about their own work says.
Hi, thanks for feedback. Project is in stealth mode at the moment, I was not expecting the response.
Game Theory and Agent Reasoning?
It seems like the AI decision making is 'immutable' but in the end it will always come down to training data
It's worse than that. The entire notion of an independent arbiter is deceptive at best, because the decision making will always come down to how the questions are framed.
The claim here is that an automated process backed with enough data and computational power can distill any of the world's most intractable conflicts and come up with a compromise immune to "toxicity" and human bad-faith. I'd submit that in any conflict in which at least one actor is acting in bad faith to begin with, in the absence of an overriding ethical framework, any such compromise would result in a worse outcome than making a decision that one side was right and the other wrong. It's the compromise between hens and wolves.
If/since the universe of human bad faith is finite, it does seem possible (and useful) to detect all such forms in transactions that are on the record.
This won't solve out-of-band corruption and might indeed increase its value, and hence increase the value of participating in a corruption network.
gosh darn unintended consequences...
In CGT, there is no third party arbitrator, no admin, the resolution is reached player to player. thx for response
in CGT, humans make the final decisions, it is a game designed for humans first, AI just came later
this is a so-often overlooked point.
There is a 77 minutes video in the Project Reviews page from the menu of the left. Maybe there is a demo in there but I didn't bother to check. It's conversational, it's text based, why didn't they include a textual demo in a page?
parley.aikiwiki.com. ?
That's a static page with underlined text that are not links. Does anything interactive start after "Login With Google" (that I won't press) ? Anyway, my point is that they should have presented the transcript of at least one round of the process and demonstrate the improvement it produces. It's much less time consuming for a reader that does not want to invest hours in videos or actually taking part of the experiment. On the other side, those readers might not be of interest to the authors so this is a way to let them go.
thx for feedback. wasn't expecting such a big response. project is in stealth mode
Isn't game theory suitable only for places where all the actors make rational decisions?
Well, strictly speaking, it is suitable for places where all actors make choices in ways we can predict. So even if an actor was choosing, say, uniformly at random (and we knew this fact), we'd be able to incorporate it into analysis to get some expected behavior.
But game theory also presumes a closed, on-the-record world. E.g., it doesn't handle dumping (loss as strategic gain) without opening the scope of game, but there are no scope bounds since value of any type is exchangeable (land for peace...). So game theory like newtonian physics only helps in narrowly constrained systems, but none of the real systems where help is needed are so constrained (banking laws notwithstanding).
Yea, and so IMO this idea makes sense if you assume that bad conversation is predictable conversation (this is a simplified way of saying this but this idea was formalized within the field of literary theory by a man named Wolfgang Iser as a response to the issue of the hermeneutic circle and the merger of horizons). A conversation is evolved to its best possible outcome when it encourages the human in the loop to face their preconceived assumptions about the world and effectively do the work to come to their own conclusions rather than act as a follower
That is in classical game theory, in CGT, it is assumed the actors are both rational and irrational
Once again an entirely backwards, upside down perspective. On Twitter he called politics the easiest job in the world.
Xenophon said that man is the hardest animal to tame. And the author’s “make AI great again” mentality is a perfect example of this untameability.
If it’s not actual AGI don’t tell me how a tool can actually fix the problems it creates. It’s like the author doesn’t have the most basic concept of what a tool is.
Anyway happy Thanksgiving to the Americans
Cheers, friend! Thankful for HN and all the great perspectives from around the world. Also thanks for the phrase "Make AI Great Again"! :D
Quite like the desktop website. Feels like a pdf without the cruft.
What is a "desktop website"?
Desktop, as opposed to mobile. (on mobile the site looks pretty similar to most others). On desktop it feels like a pdf, but extremely light weight (in a very nice way; loads quickly, scrolls perfectly [unlike pdfs], and isn't cluttered). Just a crisp experience.
Ah I understand now. Thanks.