Attaining Reasoning Without a Language Model

In the final put up we launched Minotaur, an AI that causes with no language mannequin and no GPU. This put up seems to be at why language fashions assume the way in which individuals do, at what was tried earlier than them and the place every try stopped, and at how Minotaur’s mind does the identical sorts of reasoning in another way: not by imitating how individuals write their considering down, however by engaged on the construction of the system it’s reasoning about.

Introduction

In the last post we launched Minotaur, an AI structure constructed on ELC that works out its solutions by measuring somewhat than predicting. We confirmed it checking actual rules in opposition to information of how they have been utilized: discovering which organisation broke a rule, by how a lot and the place, and telling a damaged rule from a badly recorded measurement.

Checking guidelines in opposition to information is one sort of considering. Today’s language fashions are admired for one thing broader: they will deduce, clarify, diagnose, weigh a speculation, think about what would have occurred if issues had been totally different, and alter their minds when advised one thing new. Can a machine with no language mannequin, and no GPU, do all of that?

The reply is sure:

• Minotaur’s mind does the principle sorts of reasoning normally credited to massive language fashions: deduction, induction, abduction, analysis, seeing in opposition to doing, selecting the experiment that settles a query, revising a belief, and reasoning about histories wherein asking a query modifications the reply.
• It does all of them with one operation on one community, on one odd processor core, at a couple of quarter of a second per query on a described system.
• On blind checks written by individuals who by no means noticed its code, it matched a present language mannequin reasoning in phrases, on a police interview the place the questions change the witness’s reminiscence, and on studying, diagnosing and predicting programs from their knowledge alone.

Why a Language Model Thinks Like a Person

A big language mannequin is educated to foretell the following phrase of a chunk of textual content. It learns this from an infinite quantity of writing, from books, papers, textbooks, boards and code, by doing it trillions of instances on 1000’s of GPUs.

To predict the following phrase of a proof, you need to observe the proof. To predict the following line of a physician’s notes, you need to observe the analysis. Much of the writing within the planet is individuals considering on paper, so a mannequin that will get excellent at persevering with it finally ends up excellent at imitating the considering. That’s why merely asking a mannequin to “think step by step” makes its solutions higher.

So a language mannequin mirrors the way in which individuals assume. It inherits our strengths: explaining, drawing analogies, utilizing frequent sense, taking a query in odd phrases. Researchers have additionally discovered that it inherits our weaknesses. Like individuals, language fashions discover a logical argument simpler to just accept when its conclusion sounds plausible, and tougher when it doesn’t. Their solutions can shift with how a query is framed or the order issues are put in. And like ours, its explanations are sometimes written after the actual fact, not a file of how the reply was reached.

The mannequin by no means sees the planet, solely our phrases about it.

It does imply that if you need reasoning that may be checked, replayed and licensed, imitating the way in which individuals write their considering down will not be the very best place to start out. Exact, checkable reasoning was what most approaches to AI earlier than language fashions have been making an attempt for.

Seventy Years of Trying Something Else

People tried to construct machines that motive for some sixty years earlier than language fashions. Each try labored up to a degree after which stopped, and the place they stopped explains why we ended up with language fashions.

Search and theorem proving

In 1956 Allen Newell, Cliff Shaw and Herbert Simon constructed the Logic Theorist, a program that proved theorems from Whitehead and Russell’s Principia Mathematica. Their subsequent program, the General Problem Solver, tried to do the identical for any downside that could possibly be written as a place to begin, a aim and a set of strikes. The thought was that reasoning is search: strive the attainable steps till you attain the aim.

It labored on puzzles. On actual issues the variety of attainable steps grows so quick that no laptop might search them. In 1973 the Lighthill report to the British authorities named this the combinatorial explosion, and funding in Britain was lower sharply.

Expert programs

The reply within the Nineteen Seventies was the expert system, which gave the machine data as a substitute of search: a whole lot or 1000’s of hand-written guidelines of the shape if this, then that, collected from human specialists. MYCIN, constructed at Stanford, beneficial antibiotics for severe infections, and in a 1979 analysis its recommendation in contrast effectively with that of specialists. XCON, at Digital Equipment Corporation, configured laptop orders and saved the corporate quite a lot of cash.

They stopped for 2 causes. Every rule needed to be obtained out of an professional’s head and written down by hand, which got here to be known as the knowledge acquisition bottleneck. And they have been brittle: simply exterior what the principles lined, the system had nothing wise to say. By the tip of the Eighties the marketplace for them had collapsed.

Logic programming

Logic programming aimed to have the machine motive precisely, in a correct logic. That was the guess behind Prolog, created in Marseille in 1972, and behind Japan’s Fifth Generation Computer Systems challenge, which ran from 1982 to 1992 and was constructed round logic programming.

Logic offers actual solutions, and its descendants, the solvers that verify whether or not a set of circumstances can all be true without delay, are actually used on daily basis to confirm laptop chips. But logic solely causes about what somebody has written down as logic. The planet doesn’t arrive in that type, and the Fifth Generation challenge ended with out the considering machines it had promised.

Hand-built frequent sense

In 1984 Douglas Lenat began Cyc, an try to write down down, by hand, the frequent sense that professional programs lacked: that water runs downhill, that individuals have moms, {that a} factor can’t be in two locations without delay. Over the next many years its staff entered hundreds of thousands of such information.

It confirmed how massive the frequent sense individuals take as a right is, and that writing it down by hand by no means involves an finish.

Probability and causal fashions

From the late Eighties, led above all by Judea Pearl, reasoning below uncertainty obtained a correct mathematical footing. A Bayesian network says which issues have an effect on which and the way strongly, and lets a machine work out what an statement implies. Pearl went on to make exact the difference between seeing something and doing it: seeing a moist pavement suggests it rained, however hosing the pavement your self tells you nothing concerning the climate. In the identical years, Raymond Reiter, Johan de Kleer and Brian Williams turned diagnosis right into a exact technique for locating which element of a system has failed.

These are among the finest concepts within the area, and Minotaur reproduces them. But somebody has handy them the mannequin first: the causal graph, or an outline of each element. Given solely a column of numbers, they will’t start.

Deep studying

From 2012, neural networks educated on massive quantities of knowledge, often known as deep learning, started to beat the whole lot else at recognising photos and speech. They want nothing handed to them besides examples.

What they be taught is a sample, with no construction to motive on. A community that has realized how a manufacturing line behaves can predict its output. It can’t let you know that one in every of its inputs is a security cut-off, or reply the query “what if that cut-off have been caught open?”, as a result of there’s no cut-off anyplace inside it to be caught.

And then, language fashions

Language fashions, constructed on the transformer design revealed in 2017, reached the general public in 2022, and from 2024 started to be educated to motive step-by-step earlier than answering. Language fashions received as a result of they’re the one method that wants nothing handed to it. You don’t write a mannequin of your downside for it; you describe the issue in plain phrases, and it reads them.

The value is the one we mentioned within the final put up. The reasoning runs via a whole lot of billions of weights on data-centre GPUs, and the reply comes with no certificates of the way it was reached.

Each manner of getting a machine to motive has had a value:

• The ones that motive precisely want the mannequin handed to them.
• The ones that be taught from knowledge provide you with no construction to motive on.
• The one which reads the prose wants an information centre, and might’t certify its solutions.

Minotaur sits within the center. It might be given an outline or be taught one from knowledge, places it into typed roles it might motive on, and causes precisely sufficient to certify every reply, on one odd processor core.

Minotaur’s Brain

The concepts behind ELC are set out on the MINOS page, and the final put up gave the quick model. Every system, no matter it’s about, is learn in three roles:

• one thing drives it (Energy, or E);
• one thing constrains it (Control, or C);
• and the system has its personal manner of relating the 2 (Logic, or L).

E and C by no means join straight. A drive doesn’t say what’s allowed, and a restrict doesn’t say which drive it solutions. The relation between them all the time needs to be provided, and supplying it’s what the system’s logic is.

Each factor of a system is learn as a share. Of what arrives at it, some passes and a few is held again, and the 2 all the time add as much as the entire. At one finish the whole lot passes, on the different the whole lot is held, and in between there’s a steadiness level.

This is a minimal illustration of a system; MINOS stands for the minimal ordinal illustration of programs. An entire boiler might be learn as a single factor: one thing drives it, one thing constrains it, and its logic relates the 2. When extra element is required, the identical studying is utilized to its elements. The boiler turns into a pump, a fan, a burner and a fuel valve, every an ELC factor with its personal drive, constraint and logic, linked in order that what one factor passes on turns into the following one’s drive. The description will get finer, but it surely by no means leaves the ELC format: it’s the identical three roles at each scale, simply extra finely divided and extra interconnected.

On this illustration we constructed Minotaur’s mind.

Two sides

A mind that causes usually turned out to wish two sorts of college, finest saved aside, just a little just like the two hemispheres of a human mind.

The ELC facet holds the system being reasoned about, as a community of ELC parts. It has precisely one operation, settling, described under. It can even be taught a system it has by no means been advised about, from information of what that system did. By design, it comprises no counting, no actual arithmetic and no matching of textual content. Its outcomes are measured: the community got here to relaxation, or it didn’t, to inside a acknowledged tolerance.

The actual facet is a calculator. It reads an outline of a system, lists each case when each case needs to be checked, does actual arithmetic and checks outcomes. Its outcomes are actual: true in each case, or here’s a case the place it isn’t.

The crossings

Nothing passes between the 2 sides besides via a named, one-way crossing:

Crossing	Direction	What it does
DECLARE	actual to ELC	a regulation of the system is constructed into the community
LEARN	knowledge to ELC	the community learns part of the system from information of what it did
OBSERVE	actual to ELC	a truth is held mounted; the node nonetheless obeys its personal regulation
SUPPOSE	actual to ELC	a variant planet: a replica with some issues set in any other case
INTERVENE	actual to ELC	a node is compelled to a worth, and its personal regulation is lower
ASK	actual to ELC	a node is left free, for the community to settle
READ	ELC to actual	a settled worth is learn as passes, holds or undecided
REQUEST	ELC to actual	a sum the community can’t settle is shipped to the calculator

Every crossing is recorded, together with the certificates it carries. When a solution rests on a number of steps, it carries the weakest certificates amongst them.

That file can also be the mind’s chain of thought. When a language mannequin exhibits its reasoning, it writes a brand new piece of textual content. When Minotaur exhibits its reasoning, it exhibits the crossings it really made, so as.

Settling

Picture the community as a set of nodes, every holding a share between “all handed” and “all held”, joined alongside the traces ELC permits: drives into logic, constraints into logic, logic on to logic and to constraints. Never drives straight to constraints, or again.

To ask a query, you say what every node is doing:

• some are held: these are the information you’re given;
• some are lower: compelled to a worth, with their very own regulation switched off;
• some are left free: these are the holes, the belongings you wish to know.

The relaxation observe their legal guidelines. Then the community settles, just a little like water discovering its stage:

1. Every lawful node takes the worth its regulation offers it.
2. Minotaur measures the stress: how a lot of what was held fails to come back again because it was held.
3. That misfit is carried backwards, however solely via the logic nodes and solely to the holes, so the rule that drives and constraints by no means join straight applies backwards in addition to forwards.
4. Each gap is pushed to commit to 1 finish or the opposite: passes or holds. A gap that involves relaxation on the steadiness level hasn’t actually been answered.
5. The result’s learn, with its stress proven.

If the stress falls to virtually nothing and each truth reads again because it was held, the planet closes: there’s a manner for issues to be that matches the whole lot given. If no setting of the holes will shut it, the stress says the place the contradiction sits.

One community, many questions

Every sort of reasoning is identical settle, requested in another way. What modifications is just which nodes are held, that are lower and that are left free.

The ELC facet has no particular code for any of those sorts of reasoning. Where a solution wants each case checked, as deduction does, the precise facet lists the circumstances and the ELC facet settles each. So deduction will get an actual certificates, as a result of no case is ignored.

Following the Brain Through

All of the programs under have been written in a easy format that lists a system’s drives, constraints and legal guidelines. The first two have been written by us for improvement. The others have been written by individuals who by no means noticed Minotaur’s code, and we come to how they have been examined afterwards.

A analysis

Start with a home’s lighting. The provide is available in via a most important breaker and splits right into a corridor circuit and a kitchen circuit, every with its personal fuse, change and bulb, and the kitchen circuit additionally feeds a socket and a kettle. All the switches are on. The kitchen lamp is lit and the corridor lamp is darkish. What has failed?

The actual facet supposes, separately, that every half has failed, beginning with the smallest potentialities. For every supposition, the ELC facet settles the community and experiences whether or not the planet closes. Supposing nothing has failed, it doesn’t shut, and the stress is on the corridor lamp. Supposing the corridor bulb has failed, it closes. The kitchen bulb doesn’t shut. The corridor fuse closes. Larger mixtures that already comprise a type of solutions are skipped, as a result of a smaller rationalization is all the time higher. None of the remainder closes.

The reply is both the corridor bulb or the corridor fuse, which is what a great electrician would say. Asked for an experiment to inform the 2 aside, Minotaur says that no setting of the switches can: each faults go away the corridor darkish no matter you do. That can also be right.

The identical lighting circuit answered 29 questions of each form, and all 29 matched the solutions of an impartial checker. On 40 planted faults it discovered the precise set of attainable explanations 40 instances out of 40, with no false ones. The entire query file took a couple of second and a half.

Seeing is just not doing

The second system is a greenhouse watering controller, with a hidden trigger in it. When it rains, the beds get moist, and the rain gauge additionally reads moist. When the gauge reads moist, the controller stops the pump.

Ask: if we see the gauge studying moist, are the beds moist? Minotaur holds the gauge as an noticed truth, with its regulation intact. The misfit is carried again via the gauge to its trigger, rain, and the reply is sure: the gauge says it rained, so the beds are moist.

Now ask: if we power the gauge to learn moist, are the beds moist? This time the gauge is held and its personal regulation is lower, as a result of forcing it says nothing about what induced it. Nothing is carried again via it. The pump stops, rain stays unknown, and the reply is undetermined.

This is Pearl’s distinction between seeing and doing. In Minotaur it comes from the distinction between holding a node and chopping it, not from a particular rule.

A boiler, from deduction to induction

The third system was written by somebody exterior the challenge, from the format alone: a house mixture boiler with 30 elements, together with a timer, a pump, a fan, an igniter, a flame sensor, a fuel valve, and a mains provide that feeds each the fan and the pump. Here are among the 36 questions it was requested, with Minotaur’s solutions. Each one matched the impartial reply key.

Question	Minotaur’s reply
Mains on, timer on, scorching faucet closed, room studying 0.40, strain studying 0.50. Must the radiators be heat?	No. Counterexample: the fuel provide is off. (Exact: each case checked.)
The radiators are chilly. Which of mains, fuel or pump would clarify it?	Any one of many three: mains, fuel or pump.
The radiators are chilly and the lockout gentle is on. Now what explains it?	The flame sensor, or the fuel provide. The lockout gentle guidelines the others out.
The radiators are chilly, the lockout gentle is off and the pump is working. Which elements might have failed?	The fan, the fuse, the igniter or the overheat stat.
Has the diverter valve failed, or the pump? What take a look at would inform?	Close the new faucet and change the timer on.
With the mains off: has the igniter failed, or the fuel valve? What take a look at would inform?	None. With the mains off, no take a look at can.
Pressure gauge readings of 0.20 (low), 0.35 (tremendous) and 0.28 (low). Where’s the road?	Above 0.28 and at most 0.35. (Exact: the entire vary the readings permit.)

Saying no take a look at exists, as within the second-to-last row, is a sound reply, and a reasoner that all the time proposes a take a look at will ultimately suggest a ineffective one. In the final row, Minotaur doesn’t invent a single threshold; it experiences the precise vary the circumstances seen permit, and the place to look subsequent to slender it.

When asking modifications the reply

Ordinary logic assumes that information don’t change while you have a look at them, however a witness’s reminiscence can. Ask a witness a leading question, “did he have a beard?”, and you may plant the beard you requested about.

The fourth system, once more written by somebody exterior the challenge, was a police interview. It has a theft, a witness, a line-up, a detective who could or could not ask main questions, a lawyer who could or will not be current, and a video recorder. And it has a listing of what every query does to the witness’s reminiscence: a number one query a couple of beard crops one, a number one query a couple of jacket can wipe the jacket out, and choosing somebody in a line-up makes the witness extra assured.

Minotaur handles this by reasoning over a chain of worlds. The first planet is issues as they occurred. Each query produces the following planet: the reply is learn within the planet earlier than, after which regardless of the query modifications is reset. The entire chain is became one community and answered by the identical settle.

In the diagram, the witness was requested twice, with main questions, whether or not the robber had a beard, and answered no after which sure. Did the robber have a beard? Minotaur solutions no. The first reply was the true reminiscence, and the primary query planted the beard behind the second.

Reasoning over mounted information can’t do that. To it, “no” after which “sure” about the identical truth is a contradiction, so it calls the story unimaginable. On the 15 questions whose solutions rely upon the order of asking, fixed-fact reasoning obtained 0 of 15. Minotaur obtained 15 of 15, and all 30 questions in all, together with one wherein main questions led to a wrongful cost. The entire set took seven and a half seconds.

Learning a system it was by no means advised about

Everything up to now began from an outline of the system. The tougher take a look at is a system no one has described.

So we gave Minotaur 400 rows of what a manufacturing line did. Each row had 4 readings, demand, feed, strain and the ambient temperature, plus the road’s output. It obtained no legal guidelines, no diagram and no labels saying which column did what. Behind the scenes the road was pushed by demand and feed, shut off by a security cut-off when the strain rose too excessive, and unaffected by the temperature. Then we broke the road, with out saying how, and gave Minotaur 80 rows of what it did afterwards.

First Minotaur realized the road, into ELC’s typed roles. It discovered that demand and feed have been drives, that strain was a constraint that shut the road off above a sure stage, and that the temperature was inert. When the community learns, the drive half and the constraint half can come out the fallacious manner spherical, every becoming the info equally effectively. ELC’s personal algebra says how one can put them again the best manner spherical with out altering something the community does.

Then it supposed, separately, every sort of fault a system with these roles can have: the cut-off has moved, the cut-off is caught open, it’s caught shut, a drive has stopped counting. It fitted every supposition to the 80 new rows. One fitted much better than the remainder: the strain cut-off has moved to 0.557. We had planted it at 0.55.

Finally it went again out. The altered community predicted what the damaged line would now do below new circumstances, with a mean error of 0.023, in opposition to 0.080 for the community earlier than the fault was discovered.

On ten recent traces, every with its personal planted fault, Minotaur named the fault ten instances out of ten and predicted the damaged line’s behaviour each time, with errors between 0.002 and 0.023. Each line took about 10 to fifteen seconds on one processor core, studying included.

The faults it might suppose, the cut-off has moved or a drive has gone lifeless, are faults of typed roles. A generic learner can match the identical 400 rows simply as effectively, but it surely has no cut-off inside it to maneuver and no drive to lose, so “what if the cut-off have been caught open?” isn’t a query it may be requested. The ELC neural community that does this studying is patent pending.

Carrying a conclusion right into a second system

Our manufacturing line feeds a packing station, and the packing station retains its personal books, in kilograms a minute somewhat than shares. It was realized individually, from its personal information alone.

MINOS holds that two programs framed independently can solely be joined via a provided relation between their frames, which we name a transport. So the precise facet fitted one from 30 rows logged on each programs: how the road’s output exhibits up on the station, within the station’s models. Then Minotaur ran its conclusion concerning the damaged line ahead, via the transport, into the station: what’s going to the station pack now?

On ten recent pairs it predicted the station’s output ten instances out of ten, with errors between 0.010 and 0.034. Leaving the transport out, and pretending the 2 programs converse the identical models, made the predictions 2.6 to 12 instances worse, each time. Asked to discover a single fault that could possibly be anyplace in both system, it discovered the best system and the best fault eight instances out of ten.

Aristotle, recovered

Any new manner of reasoning ought to first give again the outdated logic that we all know is true. Physicists name this the correspondence principle: a brand new principle has to agree with the outdated one the place the outdated one is safe.

So we learn Aristotle’s four classical statements in ELC’s personal phrases. “S is P” turns into the share of S that passes into P, which MINOS writes as an angle. Every S is P is the tip the place the whole lot passes. No S is P is the tip the place the whole lot is held. Some S is P and some S is just not P are the whole lot wanting one finish or the opposite. Checked on each small case, 1,302 pairs of phrases, each one got here out proper.

Then the syllogism, Aristotle’s two-premise argument comparable to each M is P, each S is M, so each S is P. We gave Minotaur not one of the classical guidelines. It had solely the angle, the truth that what passes and what’s held all the time add as much as the entire, and the MINOS rule for chaining two parts one after one other. Out of the 256 attainable kinds it proved precisely the 24 which are legitimate, proved nothing false, and refused the opposite 232, every with a motive. It even mentioned, for every argument that wants it, which time period needs to be non-empty for the argument to carry, which matches the classical account time period by time period.

One of the oldest errors in logic, the undistributed middle (each S is M, each P is M, so each S is P), is, on this studying, an try and chain two parts with no transport between their frames. It’s the identical mistake, on the scale of a sentence, as leaving the transport out between the manufacturing line and the packing station.

How It Did Against a Language Model

How it was examined

Every outcome on this put up was produced below the identical guidelines:

1. The go marks have been written down earlier than every run. Nothing could possibly be tuned afterwards to suit.
2. The proper solutions got here from another person. Each reply was scored in opposition to a checker written individually, from the format’s guidelines alone, sharing no code with Minotaur.
3. The checks have been blind. The boiler and the police interview have been written by AI brokers that noticed solely the format. Minotaur’s code was fingerprinted and frozen earlier than each was opened.
4. There was a competitor. A recent language mannequin, given the identical system described in plain phrases, answered each query by reasoning in phrases, with no code allowed, and was scored by the identical checker.
5. The claims are made solely on recent checks. Development was carried out on separate knowledge.

The outcomes

Test	What it asks	Minotaur	Language mannequin	Reasoning over mounted information
A police interview, 30 questions	reasoning when asking modifications the reply	30 of 30	30 of 30	15 of 30; 0 of the 15 that rely upon order
A manufacturing line, ten recent traces	be taught it from knowledge, discover the fault, predict it again	10 of 10 faults named; errors 0.002 to 0.023	3 of three (three traces); errors about the identical	can’t begin: nothing described
A line feeding a packing station, ten recent pairs	carry a conclusion right into a second system in different models	10 of 10 predictions; 8 of 10 faults discovered	3 of three, on an earlier model of the take a look at; errors about the identical	can’t begin
Aristotle’s logic	does the classical logic come again out?	1,302 of 1,302; 24 of 24 legitimate kinds, 0 false	—	—

On the police interview and the 2 realized programs, Minotaur matched the language mannequin, on one odd processor core and with each reply licensed. Where there was nothing described to start out from, reasoning over mounted information couldn’t start.

Where ELC Earns Its Keep

Each of the older strategies does one sort of reasoning effectively, in its personal specialty, as soon as it’s handed an outline. Minotaur does all types with one operation, typed and licensed, on one processor.

And the place the system isn’t described, ELC can’t get replaced. For each different method, the query is what it must be handed to do what Minotaur did.

Approach	Good at	Would must be handed
Language fashions	reasoning in phrases	nothing, but it surely wants an information centre, and its solutions can’t be licensed or replayed precisely
Logic solvers	actual penalties and counterexamples	a written mannequin; can’t begin from knowledge
Diagnosis engines	discovering failed parts	an outline of each element
Causal fashions	seeing in opposition to doing	the causal graph and its features; can’t be taught the roles from knowledge alone
Bayesian networks and causal discovery	studying which issues have an effect on which	the sort of relation (a cut-off, not simply an affect) and the sorts of fault to suppose
Logics of actions	histories of occasions	the actions and their results, written down
Neural networks	becoming knowledge	the whole lot wanted to motive: there’s no cut-off inside to suppose caught

ELC provides, from its type alone, what the others must be handed: the roles of drive, constraint and logic, the locations the place a fault might be, and the frames and transports between programs. In the realized checks, ELC’s personal algebra determined which realized half was the drive and which the constraint. Every analysis was a fault of a typed function. And leaving the transport out made the predictions as much as twelve instances worse.

Where This Leaves Us

A language mannequin causes the way in which individuals do as a result of it realized from individuals writing. That has made it the very best basic reasoner we have now, and in addition a reasoner that inherits our habits: assured when it shouldn’t be, swayed by framing, and unable to indicate what its solutions relaxation on. The alternate options of the earlier seventy years both needed to be handed the mannequin or had nothing to motive on.

Minotaur’s mind causes otherwise. It doesn’t imitate how individuals write a couple of system; it holds the system itself in ELC’s typed roles, whether or not described or realized from knowledge, and solutions each sort of query with one operation: maintain some issues, lower some, go away some free, and settle. Every reply comes with the chain of steps it really took and a certificates of what sort of reply it’s, and offers the identical outcome each time it’s run. Head to move on blind checks, it reached the extent of a present language mannequin, on one odd processor.

One community, one operation, each sort of reasoning: and the working is all the time proven.

The concepts behind Minotaur, from the three roles of Energy, Logic and Control to the arithmetic of MINOS, are set out on the MINOS page. The arithmetic itself is within the paper An Operational Representation of Systems: A Worked Construction in Lattice Gauge Theory (doi:10.5281/zenodo.23491024). The first put up, Introducing Minotaur: A GPU’less AI Architecture, describes the ELC adapter and its checks on actual rules. The ELC adapter and the ELC neural community are patent pending.