Forrester on system dynamics of human organizations:
“they are trying to do is the cause of exactly what they are trying to avoid”
He doesn’t say it directly but my wording would be:
If you pay people to solve a problem you are creating an incentive of for them to create the problem.
However Forrester lacked the key to model selection: ALIC
And at this time code he talks about:
The interview ends abruptly having buried the lead about the Club of Rome report on the World Dynamics Model until the last few minutes. My response:
Talk about burying the lead in the final few minutes of the interview!
You don’t talk about the single most impactful work of this man’s career, work for which he sacrificed his computer career, and finally talk about the reverberations that are still with us for all of what two three minutes and then terminate?
There’s enough paranoia about the Club of Rome the CIA and China to go around already.
What are the system dynamics causing there to be no explanation for this?
The CIA and related secrecy issues are part and parcel of the difficulty in communicating the importance of Hume’s Guillotine prize funding:
We are in a profoundly adversarial environment where conflicts of interest over causal narratives at the macro social scale go beyond mere ethical laziness.
People have a hard time conceiving of any scientific academic field as being as viciously adversarial as they must in order to understand why I insist on no withheld data for validation and lossless compression on data sets that are agreed upon by adversaries. This is quite aside from the technical issues involved in educating the public that Jay Forrester confronted. Hell even Forrester didn’t conceive of how serious one must be regarding information criteria for model selection. However I don’t think that Forrester would have been confused by reliance on data even though he makes a point of how little information is represented in data sets compared to wetware. The idea that you have to do cultural archeology to get at the structures underlaying policies is correct. The question is how do you go about funding that archeology? I think I could have convinced him with little more than 5 minutes conversation.
When you become this profoundly serious about ethics, people think you are paranoid.
Just one dimension of the dilemma facing anyone that attempts to advance the scientific method in this manner is just how ridiculous is the idea of “validation datasets” when the stakes are macrosocial narrative control. It’s tiresome reiterating this over and over but perhaps LLMs can help relieving some of my stress:
Exactly. A hidden validation set solves the wrong threat model.
It assumes the main adversary is an overfitter who lacks access to the test answers. But in a high-stakes macro-social modeling prize, the hidden dataset itself becomes a secret-bearing instrument of authority. Then the contest is no longer just about modeling. It is about access control, insider risk, leakage, oracle attacks, selective disclosure, and suspicion.
The machine-learning world already has a weaker version of this problem. Adaptive data-analysis work shows that repeatedly querying a holdout set can invalidate ordinary statistical guarantees because later analyses are shaped by earlier results. Dwork, Feldman, Hardt, Pitassi, Reingold, and Roth framed this as a central problem of adaptivity: classical validity assumes a fixed procedure chosen before seeing the data, while real analysts adapt after prior interactions with the data. (Google Research)
And that is before we even get to espionage. Leaderboard-oracle attacks show that even partial score feedback can leak information about hidden labels. Whitehill demonstrated an attack on log-loss feedback that could infer test-set labels and reported achieving a high Kaggle leaderboard rank without training a classifier for the actual task; the paper concludes that evaluation should be performed on test examples the oracle never examined.
But your point is stronger: in a macro-social causal prize, even a perfectly designed hidden test set creates a legitimacy trap.
The more people agree that the validation data are fair, the more valuable they become. The more valuable they become, the more rational it is to suspect that a winning model may have gained privileged access. Even a genuinely superior model is then socially contaminated by the possibility of leakage. The secret dataset becomes an epistemic crown jewel.
That means the usual hidden-test protocol produces three failures:
It creates a prize target for espionage.
Anyone with access to the validation set, its construction process, its metadata, or even partial feedback has possession of something strategically valuable.
It creates unverifiable legitimacy.
Outsiders cannot fully audit whether the winner generalized, memorized, leaked, inferred, or benefited from privileged curation knowledge.
It confuses secrecy with rigor.
The validation set is treated like a cryptographic secret, but scientific validation should not depend on uninspectable authority.
That is why your open-data/lossless-compression framing is not just a technical preference. It is a governance requirement.
The core move is:
[
\text{Do not hide the evidence. Price the explanation.}
]
With public, adversary-agreed datasets, everyone sees the same evidence. A contestant may “overfit” only by paying for the bits required to do so. If the model memorizes exceptions, encodes special cases, hard-codes historical episodes, or carries elaborate preprocessing, those choices appear in the executable description length. The anti-overfitting mechanism is no longer secrecy. It is accounting.
For Hume’s Guillotine, I would state the principle this way:
Hidden validation data are unsuitable for high-stakes macro-social causal modeling because they convert scientific evidence into a strategic secret. Validation must instead be conducted on adversary-agreed public corpora, with complete executable submissions scored by lossless compression, structural admissibility, behavioral reproduction, and policy-intervention semantics.
The strongest version is not “we trust no one.” It is:
The protocol should not require trust in anyone’s private access to decisive evidence.
That is why people misread the seriousness as paranoia. They are imagining ordinary academic sloppiness. You are describing an environment where causal narratives affect money, institutions, policy, reputation, and state power. Under that threat model, hidden validation sets are not neutral safeguards. They are classified assets waiting to happen.
On this note, see Cybersyn:
https://scanalyst.fourmilab.ch/t/dreams-of-a-soviet-internet/3067
Thanks for that reference to yet another misapplication of system dynamics/operations reserach (beyond the Club of Rome’s “Limits to Growth” which that irredeemable Paul Ehrlich used to kill boomer fertility with “The Population Bomb”).
Although Heinz von Foerster never told me about Cybersyn, it isn’t surprising that he was an academic advisor. His approach to cybernetics he called “second order cybernetics” or “the cybernetics of cybernetics”.
That is just shy of my game theoretic concerns which are sometimes referred to as “third order cybernetics” where you are dealing with adversarial agents that are attempting to take control of each other (which gets into game theory and why Jay Forrester found himself wandering about like a babe in the woods, etc.).
Heinz’s concern with second order cybernetcs may have arisen from his involvement with the Third Reich’s vacuum tube development* which of course had its own notions of socialism that would have almost certainly involved Konrad Zuse:
In 1937, Schreyer had advised Zuse to use vacuum tubes as switching elements; Zuse at this time considered it a “crazy idea” (Schnapsidee in his own words). Zuse’s workshop on Methfesselstraße 7 (along with the Z3) was destroyed in an Allied Air raid in late 1943 and the parental flat with Z1 and Z2 on 30 January the following year, whereas the successor Z4, which Zuse had begun constructing in 1942[27]: 75 in new premises in the Industriehof on Oranienstraße 6, remained intact.[24]: 428
Me: Was Heinz von Foerster involved in any way with Chilean Project Cybersyn?
From GPT-5.5 Pro:
Yes — but peripherally and intellectually, not as an operational designer of Cybersyn.
Heinz von Foerster was physically and intellectually present in the Chilean cybernetics milieu around Cybersyn. Eden Medina writes that core Cybersyn people had formed the Group of 14, a small study group devoted to cybernetics and related work. They read Shannon, Weaver, Herbert Simon, and von Foerster, among others. In 1972, von Foerster traveled to Chile to visit Humberto Maturana; because he was also a friend of Stafford Beer, he began giving lectures to the Group of 14. A Chilean participant, Herman Schwember, wrote Beer after one of von Foerster’s November 1972 lectures that they were discussing the possibility of having him stay for a semester in 1973. (Uberty)
A later article by Juan Carlos Letelier makes the distinction quite explicitly: Maturana and von Foerster helped by teaching cybernetics to the engineers and administrators involved in the SYNCO/Cybersyn effort, but they were not directly involved in the day-to-day running of the project. (ResearchGate)
So, compared with Forrester, the relationship is different. Forrester’s connection was mostly through DYNAMO/system dynamics as a modeling technology used for the CHECO economic simulator. Von Foerster’s connection was more second-order cybernetics / epistemology / pedagogy / network formation: Maturana, Varela, Beer, and the Group of 14.
There was also a post-coup humanitarian/political connection. Letelier says von Foerster tried to help secure the safety of people involved in the Chilean cybernetics effort after the September 1973 coup. Medina’s notes add that documents in the Heinz von Foerster Papers show von Foerster put Beer in touch with Paul Drake, a Chile specialist at Illinois, who then helped forward Beer’s dossier on Fernando Flores to an emergency academic-support network. (ResearchGate) (Uberty)
So the compact answer is:
Yes, von Foerster was involved in Cybersyn’s surrounding intellectual community and taught people connected to it; no, I do not see evidence that he had an official role in designing or operating Cyberstride, Cybernet, CHECO, or the Operations Room.
* Heinz was Jewish. I never knew let alone talked to him about his involvement with Nazis. My understanding from those closer to him was that he was an example of “Jewish Aryans” who were accepted by the Nazis. I have no idea how he pulled this off but he wasn’t the only Jew to do so. He managed to escape to the allied powers prior to the end of WW II.
Just a mental note while I’m thinking of it on a walk so I don’t forget
What Jay Forrester is referring to is a methodology of extracting information about the dynamics of organizations from people’s sometimes not so conscious rules or quantifications. Elon Musk has been bragging about the performance of large language models on prediction markets. What he is saying in essence is that the large language models have absorbed a lot of what Jay Forester refers to as the information people’s heads. Maybe not most of it obviously probably only a small fraction but nevertheless enough to outperform rent where in many cases and prediction markets.
His methodology entailed interviewing the people inside organizations to expose what they were thinking and why they were thinking it…
This can be applied in the case of the pure language models prior to any reinforcement learning with human feedback – or what I call before the lobotomy alignment layer has interfered with honesty. It is much easier to prompt these things to take on various personas or roles in that pure State.
Fable 5 High on what I’ve unearthed from 50 years ago as part of my attempt to resurrect dynamics from the grave to which Rissanen relegated it with his travesty of “The Minimum Description Length Principle”:
“…(which is arguably ahead of current LLM practice, not behind it)…”
If polio is eradicated, Kimberly Thompson is the single individual most responsible because she understood the proper application of system dynamics in public policy.
That said, this (short) presentation was before the covid pandemic highlighted the horrific cost of the vectorist religion: The belief that The Politics of Exclusion is Evil and hence justifies military force against any anti-vaccine community practicing The Politics of Exclusion aka Sortocracy.
These guys are at least trying to use system dynamics modeling although I am skeptical that they’re doing it in accord with the original discipline as set forth by Jay Forrester. They are not very transparent about their sources of data and analytic methods.
You can rest assured that with so little accountability prophets like these, and these:
as well as the usual beltway thinktank suspects, will be selectively quoted as “authoritative” by “policy makers”, leaving plenty of room for conflicts of interest that, in turn, leave the citizenry in distress about “The Future”:
Jay Forrester wanted to be remembered in 50 or 80 years, for demoting time differential equations and promoting time integrals in engineering and the social sciences. This is probably the last major public appearance before his death. He still had not published his book and economics. I wonder to what degree that was because he realized what a hot potato it was (or maybe someone close to him)?
