Not true. Can we stop saying that building models of what you call "reasoning" and then showing that LLMs correlate with your ad-hoc definition, is not, in any way proof that these models "reason" as humans. Computational neuro-science by newbies is not science.

I'm working on visual demonstrations of the logical computations for the M=2 axioms of unsupervised evaluation for classifiers. This example is for 3 labels (a, b, c). Each illustration shows the decomposition that must be true for any pair of classifiers given assumed number of label questions.

For three or more labels, there are more error modes than correct responses. So there are many ways to be X% correct on any given label. In binary classification, there is only one. This is shown here for three label classification by three LLMs-as-Judges grading the output of two other LLMs.

ICE Raids: Myth Vs. Fact https://theonion.com/ice-raids-myth-vs-fact/

The central question in any logic of unsupervised evaluation is -- what are the group evaluations logically consistent with how we observe experts agreeing/disagreeing on a test for which we have no answer key? As this series of plots for a single binary classifier that labeled (Q=10) items shows.

Any dialogue framework would benefit from understanding the logic of unsupervised evaluation - what are the group evaluations logically consistent with how we observe experts agreeing/disagreeing in their decisions? Evaluate, then collectively decide.

Gonna pump my own paper here which is basically saying it’s the wrong problem to focus on for making good decisions in medicine 😅 www.sciencedirect.com/science/arti...

Someday it will seem strange that seminars on the theoretical foundations of machine learning did not discuss the logic of unsupervised evaluation and its axioms. Simple linear relations that anyone with knowledge of high-school math can comprehend. arxiv.org/abs/2510.00821

Evaluation is not the same as decision. This is demonstrated here in two ways. 1. "And you can never test whether your data was generated in the iid model, nor can you test if it will be generated in the iid model tomorrow." You can actually do the first part with evaluations but not the second.

Logical consistency between disagreeing experts can be used to keep us safer when we use their noisy decisions. My latest on this overlooked tool to tame AIs -- submitted to IEEE SaTML 2026 in Berlin. arxiv.org/abs/2510.00821

Formalization is considered the sine qua non of maturity in many fields. How far could you go with your work, @far.ai , if you understood and used the logic of unsupervised evaluation I have developed? How far could you inspire others? Anybody can try this logic. ntqr.readthedocs.io/en/latest

And we need @far.ai to understand that the verification of evaluations of classifiers exists. This is my work on the logic of unsupervised evaluation for classifiers. So simple anybody who understands high-school math will get it. ntqr.readthedocs.io/en/latest

Live from "war ravaged" Portland

Sumerians Look On In Confusion As God Creates World https://theonion.com/sumerians-look-on-in-confusion-as-god-creates-world-1819571221/

@adolfont.github.io, it may interest you to know that we can take concepts from theorem provers to formalize the computation of the group evaluations that are logically consistent with how we observe experts, human or robotic, agreeing/disagreeing on a test. ntqr.readthedocs.io/en/latest

Speaking of boxes of numbers. Here are the possible evaluations for a single binary classifier that labeled Q=10 items (left). If you understand that evaluations of classifiers obey linear axioms, you can restrict this space of 285 possible evaluations to just 35 after seeing a summary their labels.

I spent the week at the police surveillance convention and let me tell you my biggest observation: The name of the game now is consolidating as much information as humanely possible from surveillance devices, the internet, other governmental data, and literally a million other places. 🧵

Although I claim to have discovered that there is such a thing as a "logic of unsupervised evaluation for classifiers", I am not the first to use logic to derive equations that must be universally true for all classifiers in all domains. That was done by Platanaios, Blum and Mitchell.

I had a very small role to play in the initial spread of Python when I worked at Dragon Systems 1990s along with Tim Peters in Newton, MA. Tim was a software engineer, I was a scientist in the research group where most code was written in Perl. I loved Perl, I quickly learned to use it and still do,

🚀 We’re thrilled to announce the upcoming AI & Scientific Discovery online seminar! We have an amazing lineup of speakers. This series will dive into how AI is accelerating research, enabling breakthroughs, and shaping the future of research across disciplines. ai-scientific-discovery.github.io

My personal belief is that marketing in the aggregate, via statistical measures of groups, that respects the privacy of individuals via anonymous algorithms can allow us to have our cake and eat it too.

Logical consistency between disagreeing experts can be a powerful tool for understanding how correct they could possibly be.

It is astonishing to me that the AI research community is ignorant of even the axioms of evaluation for single classifiers, never mind for pairs, trios, etc.

I'm adding an appendix to my recent paper discussing the M=2 axioms for pairs of classifiers for R labels. The paper focuses on the M=1 axioms, the single R-labels classifier. But as the paper points out, M=2 logically possible evaluations must be a subset of the product of their M=1 evaluations!

Why is autism really on the rise? What the science says - As Trump blames Tylenol, Nature looks into the decades of research on the causes of autism. www.nature.com/articles/d41...

👀 www.arxiv.org/abs/2509.18234

Next week I will be posting to ArXiv this paper that I submitted today to IEEE SaTML 2026. It will include two appendices I was not able to write by the deadline: 1. If you solve the logic for single question summaries, you solve the logic for summaries of any sequence of questions. 2. ...

"It's less like pressing a button for an answer and more like mentoring a junior colleague—asking the right questions, pushing back on shallow reasoning, and nudging the system toward sharper insights."

My latest paper explaining how logical consistency can be used to evaluate experts smarter than us on tests for which we have no answer keys. Submitted to IEEE SaTML (Secure and Trustworthy ML) 2026 in Berlin.

🐋🦑

No-Knowledge alarms built using the logic of unsupervised evaluation for classifiers are based on detecting that no classifier, at any assumed answer key, can be greater than x% correct on any one label. Here that is shown in my current working paper. Two judges (humans and gpt4) have graded

The military, via the Naval Research Laboratory, is the only institution that has supported my work on the logic of unsupervised evaluation. Last year I was awarded the "Scientific Tip of the Spear" award ("Scientific Extremum Hastae"!) for my perseverance in developing the logic and NTQR.

I disagree. That was not my experimental evidence building rare-disease patients models that had about 50% accuracy but diagnosed 98% faster than the current US health system. Nor is it the case with the speech recognition system on your phone that you used today.

The logic of unsupervised evaluation (grading tests without answer keys) proceeds as a chain of assumptions. Its simplest form form is, for example: 1. If (Q_a=n_a,Q_b=n_b,Q_c=n_c), and given the label counts in each of the expert's test responses, these evaluations are possible. As shown here.

In unsupervised evaluation of classifiers we do not have the answer key for the correct labels of the items. So the computation of the possible evaluations must start by assuming something about that key. The simplest summary is just the count of labels in the key. That is where the logic starts!

"This line of thinking takes as a given that intelligence is a discrete, measurable concept, and that increasing it is a matter of resources and processing power. But intelligence doesn’t work like that." www.theatlantic.com/books/archiv...

That feeling when your entire division + funding gets cut and someone like RFK Jr. says they will find the cause for autism by "September". I agree with Erin McCanlies that when someone arrogantly & ignorantly says that the work has not been done or it can easily be done. via🧪 @propublica.org

The picture on the left is your brain before you understand the logic of unsupervised evaluation for binary classifiers. The one on the right is after you understand it. The only possible evaluations for a classifier that answered (R_a=4,R_b=6) when labeling 10 items are on the right.

If you want to understand how to use logical consistency as a shield against the oppression of those smarter than you, consider using my Open Source Python package, NTQR on the logic of grading disagreeing experts by using the counts of how they agree/disagree. ntqr.readthedocs.io/en/latest

Checking axioms in the logic of unsupervised evaluation for classifiers is helped by the fact that they have to be symmetric in permutations of the labels. This symmetry, imposed by consistency not entailing correctness, has different algebraic implications depending on the space you use for evals!

I'm no fan of Turing when it comes to his views on intelligence and the problem of having beings smarter than you around. In comparison to his other work, it leaves a lot to be desired. It has no formalism, for example. BTW, if we want to get ideas on how to deal with beings smarter than us ...

I love this whole wonderfully evocative thread about Ötzi the Iceman and his final journey so much 😍😭🧪

Formulating the logic of unsupervised evaluation as a Linear Programming problem can only be visualized for a single binary classifier. In general, possible evaluations exist in spaces greater than 3-D. For a single classifier, that space is (R-1)(R+1) when doing R labels and this is = 3 for R=2.

If we survive authoritarianism, trivia-night at the pub is going to be lit

The logic of unsupervised evaluation for classifiers is the dumbest logic you could possibly have for them. It is the logic of evaluating bubble-sheet exams. But that means the same algebra applies to classifiers as to evaluating anybody that takes a multiple-choice exam in ANY subject. Why?

Logics of unsupervised evaluation are data streaming algorithms. And like them, they do not encode anything about what is producing the stream they observe, they just use some observed statistics of the stream to compute some other statistic, observed or unobserved. A 🧵on the forest, not the trees

Did you know that over 50% of ransomware victims who pay the ransom still don’t get their data back? #CyberSecurity #Ransomware

When we give tests to experts in unsupervised settings (no answer key), we should suspect everyone in the room. The test could be wrong, the answer key wrong or non-existent, the experts may be faulty. Logical consistency can warn us something is wrong, but not who the suspect is.