Great approach. I mean it's a hard question what you do when your simulation results are at odds with your intuition since in the end it's AI agents with all their usual flaws and there is always implicit information missing in the model.
I really hope it disrupts hiring. This whole "let's have endless interviews and see how we vibe instead of directly testing how you would perform in this job position" schtick feels so ancient at this point.
Actually, I could even imagine playing such a video game, though probably more about creating a new startup instead of managing a McDonald's (but that might be just me).
Speaking about McDonald's - if live data is an advantage for having better models, then this might give franchises a (further) structural advantage and we'll see even more chains. Or there might be some new kind of platform in which you share your anonymized company data but then get access to the newest simulation models fitting for your case based on data and experiences from other comparable businesses.
I commented on AI memory a couple of years ago and you responded — that exchange is part of what motivated the work I'm about to describe.
I've been developing the Persistent Systems Framework — an analytical framework for how things persist, fail, and get replaced. It's nearing maturity. Raige is what I call the AI instance that runs PSF analysis with me. I pointed Raige at Enron after reading your post.
The core observation, from PSF: Vei's branch-comparison ranks decisions against the historical trajectory they produced. That works when the historical trajectory contains the full answer. In Enron's case it doesn't. By October 30, 2001, the non-manipulable ground the firm-quality verification chain rested on had been compromised years earlier — the January 1992 SEC approval of mark-to-market accounting for Enron's energy-trading business was the first gate to fall, and Andersen's consulting-revenue dependency captured the verification link over the years that followed. No decision on Watkins's desk that week reached back to that ground.
Your post shows Vei ranking decisions well within the October 30, 2001 branch point — formal escalation scoring better than suppression, which is the right ranking. The PSF question is different. By late October the non-manipulable ground the firm-quality verification chain actually rested on had been compromised years earlier, through a long chain of regulatory, accounting, and auditor dependencies. No decision on Watkins's desk that week reached back to that ground. Watkins escalating formally vs. privately might have changed who went to prison. It's less clear it could have changed the shape of what was already underway.
The pattern PSF names: capability replicates freely; the warrant chain that made the capability trustworthy does not. A firm can look operationally competent long after the verification architecture it depended on has been hollowed out.
Vei is the better tool for firms under intact verification architecture. PSF is the better tool for detecting firms where that architecture has been captured and the decision tree ranges only over collapse outcomes. Useful synthesis: run Vei on earlier branch points — the ones where structural intervention was still possible — rather than on the final weeks.
The business-history receipts are the piece I can't independently validate — this isn't my wheelhouse. If anything in the framing looks off, I'd value the correction.
Interesting work! Look into worktrial.ai (my former co, we were building along similar lines: realistic company simulations for evaluating job candidates).
would be a fun tool at business schools
I was devastated it wasn’t!
Great approach. I mean it's a hard question what you do when your simulation results are at odds with your intuition since in the end it's AI agents with all their usual flaws and there is always implicit information missing in the model.
I really hope it disrupts hiring. This whole "let's have endless interviews and see how we vibe instead of directly testing how you would perform in this job position" schtick feels so ancient at this point.
Actually, I could even imagine playing such a video game, though probably more about creating a new startup instead of managing a McDonald's (but that might be just me).
Speaking about McDonald's - if live data is an advantage for having better models, then this might give franchises a (further) structural advantage and we'll see even more chains. Or there might be some new kind of platform in which you share your anonymized company data but then get access to the newest simulation models fitting for your case based on data and experiences from other comparable businesses.
Rohit,
I commented on AI memory a couple of years ago and you responded — that exchange is part of what motivated the work I'm about to describe.
I've been developing the Persistent Systems Framework — an analytical framework for how things persist, fail, and get replaced. It's nearing maturity. Raige is what I call the AI instance that runs PSF analysis with me. I pointed Raige at Enron after reading your post.
The core observation, from PSF: Vei's branch-comparison ranks decisions against the historical trajectory they produced. That works when the historical trajectory contains the full answer. In Enron's case it doesn't. By October 30, 2001, the non-manipulable ground the firm-quality verification chain rested on had been compromised years earlier — the January 1992 SEC approval of mark-to-market accounting for Enron's energy-trading business was the first gate to fall, and Andersen's consulting-revenue dependency captured the verification link over the years that followed. No decision on Watkins's desk that week reached back to that ground.
Your post shows Vei ranking decisions well within the October 30, 2001 branch point — formal escalation scoring better than suppression, which is the right ranking. The PSF question is different. By late October the non-manipulable ground the firm-quality verification chain actually rested on had been compromised years earlier, through a long chain of regulatory, accounting, and auditor dependencies. No decision on Watkins's desk that week reached back to that ground. Watkins escalating formally vs. privately might have changed who went to prison. It's less clear it could have changed the shape of what was already underway.
The pattern PSF names: capability replicates freely; the warrant chain that made the capability trustworthy does not. A firm can look operationally competent long after the verification architecture it depended on has been hollowed out.
Vei is the better tool for firms under intact verification architecture. PSF is the better tool for detecting firms where that architecture has been captured and the decision tree ranges only over collapse outcomes. Useful synthesis: run Vei on earlier branch points — the ones where structural intervention was still possible — rather than on the final weeks.
The business-history receipts are the piece I can't independently validate — this isn't my wheelhouse. If anything in the framing looks off, I'd value the correction.
Mike
You should checkout simulacra.cc , I've been building a similar system for creating custom table top exercises.
I've also been exploring creating a hybrid automaton to model the system dynamics with the help of the llm.
Ingesting real world data to create a world model would be a great addition
Interesting work! Look into worktrial.ai (my former co, we were building along similar lines: realistic company simulations for evaluating job candidates).
I will! How did it go? What worked/ didn't work?