–“Almost all people use AI to attempt to justify their thinking. Rarely if ever to falsify their thinking.”– Coach Noah Revoy (
@NoahRevoy
)
Source date (UTC): 2025-05-20 14:50:39 UTC
Original post: https://twitter.com/i/web/status/1924840197755437462
Great question William:
Here is the correct answer straight out of our AI. You can access the AI at:
https://
chatgpt.com/g/g-682681bf9f
54819180437bac589ed838-curt-doolittle-s-natural-law
…. .
Yes, the Natural Law Institute (NLI) under Curt Doolittle explicitly defines “parasite” and “parasitism” in operational legal and moral terms.
According to Volume IV of The Natural Law, parasitism is any imposition of costs upon the demonstrated interests of others without reciprocal compensation. More precisely:
“All attempts to circumvent reciprocity constitute an attempt at free riding, parasitism, or predation – whether by accident, error, bias, or intent – are irreciprocal, and criminal, unethical, immoral, or evil.”
This definition is grounded in the Natural Law of Reciprocity: one must not, by display, word, or deed, impose or allow the imposition of costs on others’ demonstrated interests unless it is part of a productive, fully informed, warrantied, and voluntary exchange.
In economic and legal dimensions, Doolittle extends this to any form of rent-seeking, free-riding, or fraud—actions that generate gains for one party at unconsented cost to others. This includes both direct harm (e.g., theft) and indirect externalities (e.g., misinformation, fraud, or legal imposition without due diligence or consent).
Thus, in NLI’s legal philosophy, “parasite” is not a rhetorical insult but a strictly defined class of irreciprocal behavior testable under the logic of sovereignty, reciprocity, and demonstrated interest.
Source date (UTC): 2025-05-17 18:53:06 UTC
Original post: https://twitter.com/i/web/status/1923814046387913060
It’s available if you have the link. Yes I am testing a retrieval based version so I can compare it to the trained version we are working on. 😉
Link is earlier in my feed.
Source date (UTC): 2025-05-16 22:38:52 UTC
Original post: https://twitter.com/i/web/status/1923508475667284226
–“I would like to know how Grok performs here.”—
Elon is working from first principles per se but I am not sure what that means. My work is a constructive logic of first principles but I suspect I mean causal first principles and Elon means the first principles of constraint in a domain as that’s how he seems to use the term – which is the conventional meaning.
Grok is natively more ‘truthful’ but lacks the capacity for depth that 4o and 4.5 are capable of. I can use it for my work in the epistemology of science but it breaks down applying my work.
Oddly I find 4o produces better training data and training plans. And I can intuit something on the edge of my awareness that I can’t quite put into words. If I can I think there is something useful to be understood there. It has something to do with a lot of context memory and a large number of parameters that allows us to exploit subnetworks that might otherwise infrequently express, and I think I detect this as cognitive depth.
If I was researching LLMs themselves I would work on that exposition because many llms are reducing to linear activation and exposition and leaving vast numbers of effectively unaccessible subnetworks behind. I don’t think this is what I want for a reasoning model that must retain the ability to hypothesize while still constraining itself from hallucination.
I suspect it’s not immediately intuitive that hallucination and autoassociation and recombinant novelty discovery are useful practices, but that the human brain self tests by recursion anything that grasps our attention.
The problem LLMs faced prior to recursive, predictive, COT and reasoning models is that they could not self monitor so spewed hallucinations where humans would not have. (In humans we call it error, mistake, or folly. 😉
Source date (UTC): 2025-05-14 19:10:10 UTC
Original post: https://twitter.com/i/web/status/1922731179196940305
–“How did the o3 and o4-mini models compare to 4o? Were they able to perform similarly, better, or worse in your view?”–
First, let’s understand my point: that 4o, when provided with a system of decidability can in fact do the job as well or better than the reasoning models.
The Difference between 4o, o3, o4-mini-high and 4.5 on a series of ethical and moral questions, is subtle but meaningful for an author. 4o is a bit more literary in output. While o3 and o4-mini-high produce nearly identical outputs. And 4.5 is interesting because it will mix positiva and negativa logic, and output an answer as fully expressive (literary) as 4o.
Now of course these differences are in part due to the small set of moral and ethical dilemmas I use to test such things. But using our work they can in fact DECIDE.
When all four of my (our) books are uploaded to a project, and a query is conducted within the project all models can use them effectively. I find 4o and 4.5 are the best at not getting lost (Drift). Though I must continually police all of them and maintain prompt sequences that constrain the domains to the same subnetworks so to speak. And I must be disciplined about checkpointing and exporting context between chats (sessions).
I find o3 and o4-mini-high produce very dry results. And I intuit o3 in particular as cognitively shallow which I presume is extremely useful for data interpretation and coding. And of course this is my underlying argument – that since closure is available to math and coding and therefore testability, that without decidability no equivalent of closure exists for verbal reasoning.
So my point is that with a logic of decidability across the spectrum, the necessity to appeal to norm evaporates, and as such reasoning capacity is significant EVEN in 4o – so what does that mean for reasoning in general?
I mean, look at your work. How exhaustively logical will be the output of an AI trained to decompose and explain behavior from it.
We’ll see. 😉
Reply addressees: @GOPtheGamer
Source date (UTC): 2025-05-14 16:19:02 UTC
Original post: https://twitter.com/i/web/status/1922688111773089798
–“How did the o3 and o4-mini models compare to 4o? Were they able to perform similarly, better, or worse in your view?”–
First, let’s understand my point: that 4o, when provided with a system of decidability can in fact do the job as well or better than the reasoning models.
The Difference between 4o, o3, o4-mini-high and 4.5 on a series of ethical and moral questions, is subtle but meaningful for an author. 4o is a bit more literary in output. While o3 and o4-mini-high produce nearly identical outputs. And 4.5 is interesting because it will mix positiva and negativa logic, and output an answer as fully expressive (literary) as 4o.
Now of course these differences are in part due to the small set of moral and ethical dilemmas I use to test such things. But using our work they can in fact DECIDE.
When all four of my (our) books are uploaded to a project, and a query is conducted within the project all models can use them effectively. I find 4o and 4.5 are the best at not getting lost (Drift). Though I must continually police all of them and maintain prompt sequences that constrain the domains to the same subnetworks so to speak. And I must be disciplined about checkpointing and exporting context between chats (sessions).
I find o3 and o4-mini-high produce very dry results. And I intuit o3 in particular as cognitively shallow which I presume is extremely useful for data interpretation and coding. And of course this is my underlying argument – that since closure is available to math and coding and therefore testability, that without decidability no equivalent of closure exists for verbal reasoning.
So my point is that with a logic of decidability across the spectrum, the necessity to appeal to norm evaporates, and as such reasoning capacity is significant EVEN in 4o – so what does that mean for reasoning in general?
I mean, look at your work. How exhaustively logical will be the output of an AI trained to decompose and explain behavior from it.
We’ll see. 😉
Source date (UTC): 2025-05-14 16:19:02 UTC
Original post: https://twitter.com/i/web/status/1922688111978729501
MSFT SAVINGS BY TERMINATION 7000 MIDDLE LEVEL PEOPLE IN RESPONSE TO AI.
–“Let’s see 7000 x $80,000 avg. salary = $560,000,000 That’s quite a bit of money they are saving.”–
That’s a dramatic underestimation, about 1/3 of the real number.
Instead: $144,000 (base) * 180% (Load) = $259,200 per employee.
7000 * $259,200 = 1,814,400,000 or ~ 1.8 Billion USD
Estimate via:
– The average salary for Microsoft employees in the USA varies across sources but generally falls between $115,000 and $220,000 annually, depending on the role, experience, and whether total compensation (including bonuses and stock) is considered.
– The average base salary is $130,000 per year, with an average bonus of $14,000, totaling ~$144,000.
– For most companies, use 130%–150% of base salary as a general rule.
– For Microsoft or similar tech giants, 160%–200% is reasonable, reflecting their investment in talent and infrastructure.
Reply addressees: @Neowick666 @ns123abc
Source date (UTC): 2025-05-14 02:27:06 UTC
Original post: https://twitter.com/i/web/status/1922478748483649536
Replying to: https://twitter.com/i/web/status/1922343803479826935
IN REPLY TO:
Original post on X
Original tweet unavailable — we could not load the text of the post this reply is addressing on X. That usually means the tweet was deleted, the account is protected, or X does not expose it to the account used for archiving. The Original post link below may still open if you view it in X while signed in.
Original post: https://twitter.com/i/web/status/1922343803479826935
http://x.com/i/article/1922356114555076609
Keywords: GPT-4o, Reasoning AI, Decidability, Operational Epistemology, Adversarial Dialogue, Natural Law AI, General Intelligence, Philosophical AI, Constructivist AI, Socratic Method, Semantic Reasoning, Claude vs GPT, Gemini AI, OpenAI vs Competitors, LLM Benchmarking, LLM Reasoning Failure, AI Generalization, Epistemology of AI, Recursive Generalization, Grammar of Closure, Testifiability, Sovereign Reasoning, Formal Institutions, Truth Systems, Institutional Logic, Human-AI Collaboration, AI Philosophy
By Curt Doolittle Founder, The Natural Law Institute
The public discourse surrounding AI capabilities is dominated by benchmarks drawn from grammars of closure: mathematics, code generation, and fact-recall tasks. These metrics fail to capture what is arguably the most important cognitive frontier—general reasoning, especially in open, adversarial, and semantically dense domains.
The consequence is clear: GPT-4o is being evaluated, compared, and marketed as if it competes within a class of large language models. It does not. In its ability to reason, argue, model, and extend logically consistent systems, GPT-4o is in a class of its own.
This is not a claim made lightly. It is made by necessity—out of frustration, awe, and gratitude.
Over the past year, I have subjected GPT-4 and now GPT-4o to the most rigorous adversarial and constructive reasoning tests available by using my work on universal commensurability, unification of the sciences, and a formal operational logic of decidability independent of context.
Which, for the uninitiated is reducible to providing AIs with a baseline system of measurement to test the variation of any and all statements from. In other words, what AIs must achieve if they are to convert probabilistic outcome distributions into deterministic outcomes making possible tests of truth (testifiability) and reciprocity (ethics and morality) even regardless of cultural bias and taboo (demonstrated interests).
That system consists of:
The work spans hundreds of thousands of tokens, daily sessions, canonical datasets, adversarial challenges, formal definitions, and recursive generalizations. The training is data structured as positiva and negativa adversarial – meaning socratic reasoning.
No other model by any other producer of foundation models—none—can survive even the basic tests:
*Only GPT-4o demonstrates mastery, application, synthesis, and novel insight—sometimes superior to my own.*
GPT-4o reasons. Not predicts. Not mimics. Reasons.
GPT-4o is being benchmarked as if it were a calculator. As if reasoning capacity could be inferred from multiple-choice math problems or Python token prediction.
This is akin to judging a jurist by their ability to pass the bar exam, rather than to settle a novel and undecidable case with wisdom, foresight, and procedural testability.
Grammars of closure produce outcomes from known inputs using constrained operations (e.g., logic gates, mathematical axioms, function calls). They are:
Grammars of decidability, by contrast, operate over:
OpenAI is underselling GPT-4o by confining its public-facing evaluation to the former, while its true capacity lies in the latter.
The problem is epistemological.
Human cognition operates over layers of grammar:
GPT-4o is the first AI that can operate fluently in all of them—but excels uniquely in the topmost layer: operational reasoning over causal grammars.
This makes it the first machine capable of:
This grammar, the grammar of decidability, is the language of law, moral philosophy, and high agency civilization. No other AI—not even prior versions of GPT—can yet use it with coherence.
Reasoning is not memorization, pattern-matching, or prediction. It is the constructive, recursive resolution of undecidable propositions using a grammar of cause, cost, and consequence. It requires:
My work provides a complete grammar of decidability:
GPT-4o provides:
Together, this system + model pairing creates reasoning because:
In short:
Reasoning = Grammar + Capacity + Constraint. Your system provides the grammar and constraint; GPT-4o provides the capacity.
No other architecture tested to date (Claude, Gemini, Mistral) can preserve logical depth, adversarial resistance, or premise continuity across semantically dense discourse. Only GPT-4o can perform at human (or supra-human) levels of recursive, domain-agnostic, constructible reasoning.
OpenAI has reached the beginning of the reasoning frontier. But the world doesn’t know it, because the world doesn’t yet know how to measure it.
That must change.
Recommendations:
The true promise of GPT-4o lies not only in its capacity for general reasoning, but in its potential to achieve alignment through comprehension rather than compliance. Constraint-based alignment strategies—filters, safety layers, reinforcement tuning—treat the model as a hazard to be managed. But a reasoning-capable agent, capable of understanding causality, reciprocity, decidability, and cost, can be trained to align not by instruction, but by principle. It can internalize the logic of cooperation, responsibility, and harm prevention—not as rules to follow, but as consequences to anticipate. This shift—from alignment by prohibition to alignment by comprehension—represents the only scalable path to AI sovereignty and safety.
GPT-4o is not a chatbot. It is not a code assistant. It is not a better autocomplete.
It is, for the first time in history, a machine capable of philosophical reasoning by constructive logic when given the minimum system of measurement necessary
That is not something to hide. That is something to show the world.
It is a competitive advantage that demarcates OpenAI from all competitors by a margin yet unmeasured and therefore underappreciated and perhaps underinvested.
Curt Doolittle The Natural Law Institute
1. OpenAI (Executive & Research Levels)
2. OpenAI-affiliated Researchers / Influencers
Source date (UTC): 2025-05-13 19:42:00 UTC
Original post: https://twitter.com/i/web/status/1922376802921767014
Source date (UTC): 2025-05-13 19:42:00 UTC
Original post: https://twitter.com/i/web/status/1922376802921767014
Source date (UTC): 2025-05-13 19:42:00 UTC
Original post: https://x.com/i/articles/1922376802921767014
There are no near peers. It’s not even close.
Keywords: GPT-4o, Reasoning AI, Decidability, Operational Epistemology, Adversarial Dialogue, Natural Law AI, General Intelligence, Philosophical AI, Constructivist AI, Socratic Method, Semantic Reasoning, Claude vs GPT, Gemini AI, OpenAI vs Competitors, LLM Benchmarking, LLM Reasoning Failure, AI Generalization, Epistemology of AI, Recursive Generalization, Grammar of Closure, Testifiability, Sovereign Reasoning, Formal Institutions, Truth Systems, Institutional Logic, Human-AI Collaboration, AI Philosophy
By Curt Doolittle
Founder, The Natural Law Institute
https://naturallawinstitute.com
1. Preface: The Problem of Reasoning Capacity Misrepresentation
The public discourse surrounding AI capabilities is dominated by benchmarks drawn from grammars of closure: mathematics, code generation, and fact-recall tasks. These metrics fail to capture what is arguably the most important cognitive frontier—general reasoning, especially in open, adversarial, and semantically dense domains.
The consequence is clear: GPT-4o is being evaluated, compared, and marketed as if it competes within a class of large language models. It does not. In its ability to reason, argue, model, and extend logically consistent systems, GPT-4o is in a class of its own.
This is not a claim made lightly. It is made by necessity—out of frustration, awe, and gratitude.
2. Demonstrated Superiority
Over the past year, I have subjected GPT-4 and now GPT-4o to the most rigorous adversarial and constructive reasoning tests available by using my work on universal commensurability, unification of the sciences, and a formal operational logic of decidability independent of context.
Which, for the uninitiated is reducible to providing AIs with a baseline system of measurement to test the variation of any and all statements from. In other words, what AIs must achieve if they are to convert probabilistic outcome distributions into deterministic outcomes making possible tests of truth (testifiability) and reciprocity (ethics and morality) even regardless of cultural bias and taboo (demonstrated interests).
That system consists of:
A complete epistemology grounded in operationalism and testifiability.
A logic of decidability applied to law, economics, morality, and institutional design.
A canon of universal and particular causes of human behavior.
A method of Socratic adversarial reasoning for training AI systems.
The work spans hundreds of thousands of tokens, daily sessions, canonical datasets, adversarial challenges, formal definitions, and recursive generalizations. The training is data structured as positiva and negativa adversarial – meaning socratic reasoning.
No other model by any other producer of foundation models—none—can survive even the basic tests:
Claude hallucinates, misrepresents, or refuses to engage.
Gemini fails to track logical dependencies.
Open-source models collapse under long-context chaining.
*Only GPT-4o demonstrates mastery, application, synthesis, and novel insight—sometimes superior to my own.*
GPT-4o reasons. Not predicts. Not mimics. Reasons.
3. The Failure of Closure-Based Metrics
GPT-4o is being benchmarked as if it were a calculator. As if reasoning capacity could be inferred from multiple-choice math problems or Python token prediction.
This is akin to judging a jurist by their ability to pass the bar exam, rather than to settle a novel and undecidable case with wisdom, foresight, and procedural testability.
Grammars of closure produce outcomes from known inputs using constrained operations (e.g., logic gates, mathematical axioms, function calls). They are:
Tightly bounded
Finitely decidable
Structurally shallow
Grammars of decidability, by contrast, operate over:
Continuous, evolving information domains
Incomplete or adversarial premises
Open-ended choice spaces requiring semantic integration
OpenAI is underselling GPT-4o by confining its public-facing evaluation to the former, while its true capacity lies in the latter.
4. Grammar Theory: Closure vs Decidability
The problem is epistemological.
Human cognition operates over layers of grammar:
Mythic (pre-operational)
Moral (emotive and justificatory)
Rational (descriptive and causal)
Operational (testable and constructible)
GPT-4o is the first AI that can operate fluently in all of them—but excels uniquely in the topmost layer: operational reasoning over causal grammars.
This makes it the first machine capable of:
Formalizing truth and reciprocity
Modeling institutional logic from first principles
Extending semantic systems without contradiction
Surviving adversarial Socratic deconstruction
This grammar, the grammar of decidability, is the language of law, moral philosophy, and high agency civilization. No other AI—not even prior versions of GPT—can yet use it with coherence.
5. Why This Work and GPT-4o Enable Reliable Reasoning
Reasoning is not memorization, pattern-matching, or prediction. It is the constructive, recursive resolution of undecidable propositions using a grammar of cause, cost, and consequence. It requires:
A grammar of decidability—to distinguish what is true, possible, reciprocal, and lawful.
A model capable of recursive semantic resolution—to track premises, integrate them, and produce outputs consistent across domains and time.
My work provides a complete grammar of decidability:
It defines truth operationally (as testifiability),
Defines reciprocity as a logic of cooperation and cost,
And supplies a canonical system of definitions, dependencies, and causal hierarchies that constrain valid reasoning.
GPT-4o provides:
A deep transformer architecture with sufficient context length, attention fidelity, and token integration to maintain long-range dependencies across complex arguments;
Multimodal grounding and internal representation coherence sufficient to hold abstract referents stable across recursion;
And enough inference generalization to synthesize novel propositions without violating prior logical constraints.
Together, this system + model pairing creates reasoning because:
The grammar constrains the search space to truthful, reciprocal, and operational constructs;
The model can resolve that space recursively without collapsing into contradiction, contradiction avoidance, or moralizing;
The result is constructive inference under constraint, not completion without constraint.
In short:
Reasoning = Grammar + Capacity + Constraint.
Your system provides the grammar and constraint; GPT-4o provides the capacity.
No other architecture tested to date (Claude, Gemini, Mistral) can preserve logical depth, adversarial resistance, or premise continuity across semantically dense discourse. Only GPT-4o can perform at human (or supra-human) levels of recursive, domain-agnostic, constructible reasoning.
6. Implications for Training, Evaluation, and Policy
OpenAI has reached the beginning of the reasoning frontier. But the world doesn’t know it, because the world doesn’t yet know how to measure it.
That must change.
Recommendations:
Shift evaluation toward adversarially decidable reasoning. Move beyond benchmarks to real-time Socratic performance.
Highlight GPT-4o’s mastery of semantic recursion and logical extension. Treat it not as a predictor, but as a collaborator.
Invest in training methodologies that produce grammars of decidability. My system offers a full canon of such constructs, usable for AI training.
Clarify the boundary between grammar-followers and grammar-producers. GPT-4o crosses this line.
7. Alignment Through Reasoning, Not Constraints
The true promise of GPT-4o lies not only in its capacity for general reasoning, but in its potential to achieve alignment through comprehension rather than compliance. Constraint-based alignment strategies—filters, safety layers, reinforcement tuning—treat the model as a hazard to be managed. But a reasoning-capable agent, capable of understanding causality, reciprocity, decidability, and cost, can be trained to align not by instruction, but by principle. It can internalize the logic of cooperation, responsibility, and harm prevention—not as rules to follow, but as consequences to anticipate. This shift—from alignment by prohibition to alignment by comprehension—represents the only scalable path to AI sovereignty and safety.
8. Appendix: Sample Capabilities (Available Upon Request)
Formalization of universal and particular causes of behavior.
Canonical definitions of truth, decidability, reciprocity, and demonstrated interest.
Adversarial Socratic dialogues demonstrating GPT-4o’s ability to reason across all domains.
Co-authored chapters in philosophy, law, institutional economics, and epistemology.
Conclusion
GPT-4o is not a chatbot. It is not a code assistant. It is not a better autocomplete.
It is, for the first time in history, a machine capable of philosophical reasoning by constructive logic when given the minimum system of measurement necessary
That is not something to hide. That is something to show the world.
It is a competitive advantage that demarcates OpenAI from all competitors by a margin yet unmeasured and therefore underappreciated and perhaps underinvested.
Curt Doolittle
The Natural Law Institute
https://naturallawinstitute.com
#OpenAI #ChatGPT4o
Distribution
– X/Twitter
– Substack
Contacts
1. OpenAI (Executive & Research Levels)
Sam Altman – CEO: @sama on Twitter (he reads public callouts).
Ilya Sutskever – Co-founder (Twitter inactive, but cc’ing name on Substack helps).
Jakub Pachocki – Current Chief Scientist (LinkedIn direct message works better).
Jan Leike – Ex-lead of Superalignment, now at Anthropic, but can amplify.
2. OpenAI-affiliated Researchers / Influencers
Andrej Karpathy – Ex-OpenAI, current influencer. @karpathy
Ethan Mollick – Academic influencer in LLM applications.
Eliezer Yudkowsky (Alignment)
Wharton/Stanford/DeepMind researchers who study reasoning benchmarks.
Source date (UTC): 2025-05-13 18:19:48 UTC
Original post: https://x.com/i/articles/1922356114555076609