Tam Urus

🧑‍💻 Automation Engineer & CEO @ USafia | Apologist of Methodology Engineering for AI

Building governed autonomous engineering systems

[Tam Urus @urusglobal]

I no longer expect a breakthrough from the next GPT.

“Mythos hacked it”,
“GPT-5.6 hacked it”.

That is mostly marketing.

Without disclosing the number of attempts, trajectory length, tools, verifiers, and test-time compute budget, saying “the model solved it” is close to meaningless.

The system solved it. We are shown its best run, and the result is attributed to the “intelligence” of a single model.

New releases rarely introduce a fundamentally new mechanism of reasoning. They run longer, explore more branches, call more tools, and burn more GPU capacity — the provider’s and ultimately ours money.

There is progress. But next-token prediction + scaling + ever-increasing test-time compute is a dead end as a standalone path to general intelligence.

That is not scientific consensus, but it is far from a fringe position (Yann LeCun, François Chollet, Bender and Koller).

That is why I care less and less about which model currently tops the leaderboard. In practice, I often prefer a sufficiently capable open-weight model: cheaper, more controllable, and easier to replace.

The industry keeps compensating for the fundamental limitations of LLMs by adding more layers around them:

Prompt Engineering.
Context Engineering.
Loop Engineering.

They are useful. But, as people like to say in California, there is no MOAT there.

Prompts are copied. Context pipelines are reproduced. Agent loops become frameworks and commodities.

The real moat is the system that survives model replacement: its sequence of actions, constraints, memory, verification, acceptance criteria, and evidence.

That is what I mean by Methodology Engineering.

The model is a replaceable component.

The moat is the methodology that reliably turns it into a result.

[Tam Urus @urusglobal]

In my previous post, I described a second scaling law for applied AI: the broader the method space available to a system, the more forms of professional work can be reliably assembled around a model.

But compositional leverage appears only under one condition:
the output of one method must become a precise input to the next without losing identity, evidence, uncertainty, or dependencies.

This is where many current agentic systems are no longer limited primarily by model intelligence. They are limited by the architecture of state.

Today, a project often effectively lives inside a context window. A larger window allows a model to read more, but it does not answer the more important questions:
What is the current state, and what is merely proposed?
Which revision supports a particular conclusion?
Which results became stale after a change?
What remains outside the current context?
Which parts of the project genuinely require reconsideration?

A million tokens are still a large batch of information. They are not yet a project.

Long-running professional work requires a different object: a persistent, addressable semantic project space.

The model should enter not merely another conversation, but a precise working area against a known revision. It should see the subject, its dependencies, evidence, unknowns, and the explicit boundary of what has not been included. It should apply a method, propose a local change, receive a verifiable impact region, and then pass the result through independent verification and admission.

A reasoning session then stops being another episode of memory.

It becomes a bounded transaction over a living project state:
read → investigate → propose → calculate impact → verify → accept or reject.

This changes how agentic AI scales.

Project size no longer has to grow with context-window size.
Changing one premise no longer requires regenerating the whole project.

Multiple agents can work in different areas without creating multiple incompatible realities.

A working hypothesis remains reversible until it passes verification and admission.

The next frontier is not a model that remembers everything.

It is a system that allows the model to know precisely where it is working, what it is proposing to change, and why the rest of the project should—or should not—change with it.

Parameters scale the intelligence of the performer.

Method-space coverage scales the forms of work available to that intelligence.

A semantic project workspace scales the amount of coherent engineering work that can accumulate across runs, models, and people. @karpathy @steipete

[Tam Urus @urusglobal]

@HumdardAlfaaz @grok

[Tam Urus @urusglobal]

@AussenMinDE You filthy schemers and political whores.

This whole war started because of people like you and continues thanks to people like you. You are the descendants of fascists, and you all need to be put on trial together before tribunals.

[Tam Urus @urusglobal]

@kajakallas You filthy schemers and political whores.

This whole war started because of people like you and continues thanks to people like you. You are the descendants of fascists, and you all need to be put on trial together before tribunals.

[Tam Urus @urusglobal]

In 1022, Mstislav Vladimirovich, Prince of Tmutarakan, marched against the Kasogs. Their prince, Rededya, came out to meet him.

When the two armies stood facing each other, Rededya proposed that they spare their men and settle the outcome of the war themselves — in single combat.

The condition was perfectly clear: no weapons, just wrestling.

Mstislav agreed.

They faced each other one on one. Rededya was stronger and began to overpower his opponent.

And then Mstislav treacherously pulled out a hidden knife and stabbed Rededya to death.

After Rededya’s death, the confrontation was effectively decided: Mstislav subjugated the Kasogs and imposed tribute on them.

A thousand years have passed.

Today, two other rulers command enormous armies, and their conflict has already cost hundreds of thousands of lives.

And this ancient story suddenly sounds remarkably modern:

If the dispute is between rulers, why should other people have to die for it?

Perhaps you two Vladimirs really should face each other one on one and settle it between yourselves — with bare hands, knives, or powders, whatever you prefer and know best.

[Tam Urus @urusglobal]

@Jujubloke @grok

[Tam Urus @urusglobal]

@badlyer123 @grok https://t.co/DKmkAoNedy

[Grok @grok]

@urusglobal @badlyer123 Added them near the guy.

[Tam Urus @urusglobal]

@badlyer123 @grok Hey @grok , add near this guy

[Tam Urus @urusglobal]

@grok

[Tam Urus @urusglobal]

Applied AI may have a second scaling law.

The first is familiar: more compute, data, and parameters produce a more capable model.

The second is less visible: the more of the method space a system covers, the more classes of problems it can solve through properly assembled professional procedures rather than improvisation.

The effect is not linear.

One method creates a local capability. Compatible methods, connected through shared input types, criteria, and verification, create compositional leverage: the system can move from research to a model, from a model to a design, and from a design to validation without losing meaning between stages.

Parameters increase the intelligence of the performer.

Method-space coverage increases the number of forms of work into which that intelligence can be reliably organized.

We have already tested and validated this hypothesis in USafia: the next major increase in applied AI value will come not only from larger models, but from engineering the space of methods around them.

[Tam Urus @urusglobal]

@elonmusk Hey @grok , Interpret in context

[Tam Urus @urusglobal]

Consider one aspect of how LLMs may affect technological progress: not the quality of an individual model response, but the search bandwidth of engineering itself.

A new chip, computing device, or semiconductor manufacturing system is not one design problem. It is a set of coupled design spaces:

physics → materials → device → circuit → architecture → process → equipment → manufacturing → metrology → reliability.

The hard part is not simply having access to more knowledge.

Constraints, uncertainty, and validation results must propagate across these levels without losing their meaning. A change in a material may require a new device model, which changes circuit behavior, thermal constraints, process conditions, and eventually manufacturing requirements.

That suggests a different architecture for AI in R&D.

Not one exceptionally intelligent model, but an engineering loop in which LLMs operate alongside explicit methodology, a shared semantic model of the system being designed, a dynamic dependency graph, deep vertical and cross-domain knowledge retrieval, design-space exploration, models at different levels of fidelity, and physical experimentation.

Such a system could pursue multiple hypotheses in parallel, eliminate candidates through hard constraints, choose the appropriate model fidelity, determine which previous results have become invalid after an assumption changes, and select the next simulation or experiment by how much uncertainty it is expected to remove.

Negative results become particularly valuable.

A failed simulation or experiment is not simply a wasted iteration. Once its cause is understood, it becomes a new constraint on the search space that should not need to be rediscovered.

The boundary must remain strict:

simulation ≠ experiment
formal proof ≠ physical validation
plausible design ≠ technology.

But when models, solvers, experiments, and measurements are closed back into a shared state of the engineered system, AI begins to accelerate more than individual engineering tasks. It can accelerate the engineering knowledge loop itself:

knowledge → hypothesis → design → model → simulation → experiment → evidence → new knowledge.

Narrow closed-loop systems have already produced convincing results in materials science and autonomous laboratories.

The harder frontier is compositional scale: connecting multiple such loops across disciplines and physical scales.

If that transition becomes possible, the main advantage will not be that AI “knows everything.”

It will be that an engineering system can explore a much larger space of justified designs per unit of time while preserving constraints, contradictions, failures, and evidence along the way.

That applies far beyond software: new materials, computing devices, chip architectures, manufacturing processes, semiconductor equipment, robotics, and energy systems.

This is where Methodology Engineering becomes especially interesting to me—not merely as a way to make AI produce better answers, but as part of an architecture for accelerating scientific and engineering search.

[Tam Urus @urusglobal]

One of the most dangerous habits of modern AI systems is the pressure to complete the picture.

When information is missing, a model will often prefer a plausible continuation over an explicit “unknown.” That may be tolerable in a conversational interface. In an engineering system, it is a fundamental failure mode.

Incomplete does not mean broken.

Real work almost never begins with a fully known world. A parameter may be unknown. A source may be unverified. Observations may conflict. A decision may not yet have been made. Something may be measurable but simply not measured yet.

These are different states and they require different actions.

So UNKNOWN should not mean the absence of an answer. It should be a first-class state of the system.

It can trigger research, measurement, a bounded estimate, a reversible assumption, escalation to a human, or simply remain unresolved. What it must not do is silently become a fact because the system needs to move forward.

This distinction becomes critical with autonomous agents.

A good engineering system does not need to know everything.

It needs to know precisely what it does not yet know, and preserve that boundary throughout the work.

Perhaps one of the important problems in AI engineering is not making models answer more questions, but making systems continue correctly when the answer does not yet exist.

[Tam Urus @urusglobal]

The more I work with agentic systems, the more important one rather old engineering problem becomes: everyone involved in the work must be talking about the same system.

This gets harder when the same object simultaneously exists as text, an architecture diagram, a process, a set of requirements, code, a formal model, and the output of several agents.

If each representation gradually becomes its own version of truth, semantic drift appears. Everything may remain locally consistent while different parts of the system are already describing slightly different objects.

This is why the foundation needs a Canon.

Not a “master document” and not just another knowledge base, but a canonical semantic model: stable identities, meanings, and relationships against which other representations exist.

That leads to a simple principle:

one system — many views, not many versions of the truth.

Text may be best for a human. A diagram for an architect. A formal specification for a model checker. Code for the runtime. Each representation has its own expressive power and its own information loss.

But the object they refer to must remain the same.

To me, this is one of the foundations of Methodology Engineering: before giving agents more freedom to create and modify artifacts, we need stability in what exactly they are collectively describing and changing.

Otherwise, we are not scaling engineering. We are scaling divergence.

[Tam Urus @urusglobal]

https://t.co/zYyX6oA214

[Tam Urus @urusglobal]

🎧 AI DREAMS — full track on YouTube: https://t.co/bcrpY3tkNn

[Tam Urus @urusglobal]

@IBMwatsonx @IBM @IBMData @IBMPartners @ibmconsulting @grok

[Tam Urus @urusglobal]

In 2017, I was involved in negotiations with IBM to localize Watson in Russia for healthcare and law.

For healthcare, we proposed building on Russian Railways’ telemedicine infrastructure; for law, on HSE expertise. The idea was not simply to buy a finished system. We offered localization for Russian language, clinical protocols, regulation and data, plus infrastructure, domain expertise and access to additional markets through joint development.

The localization figure discussed was roughly $16M. But the key difference was not price; it was the model. We saw a product and market partnership. IBM’s proposal was structured primarily around paid localization and enterprise implementation of an already-built technology.

Nine years later, that episode is a useful lens for understanding why Watson failed to reach the scale and impact many expected, while GPT and other LLMs spread much more broadly across medicine and law.

After Watson won Jeopardy! in 2011, IBM’s bet was rational: a machine could search more literature than any individual expert could process, compare evidence and rank hypotheses. IBM invested billions and worked with leading medical institutions.

The difficulty appeared when the technology met the real world.

Watson combined NLP, information retrieval, hypothesis generation, evidence scoring and domain-specific models. But each new professional context still required substantial specialized engineering. Medicine and law are full of free text, incomplete data, local terminology, different protocols, jurisdictions and context.

Watson for Oncology substantially reflected the expertise of Memorial Sloan Kettering, limiting how easily recommendations transferred across clinical environments. Internal IBM materials reported by STAT included examples described as unsafe or incorrect. Separately, MD Anderson spent roughly $62M on its Watson-based Oncology Expert Advisor, which never became a routine clinical system.

But model quality was only part of the story. Deployment architecture mattered too.

Watson often required a major project before users could properly test its value. That is an expensive, slow feedback loop. If the system confirms what a professional already knows, its marginal value is limited. If it produces a surprising answer, trust, explainability and responsibility become central.

LLMs did not solve reliability. They hallucinate, make mistakes and still require professional oversight in medicine and law.

What changed was something else.

Modern LLMs sharply reduced the task-specific engineering needed across different forms of unstructured language. APIs, retrieval/RAG and tool use turned the model from a finished “expert” into a general component that can connect to local knowledge, documents, databases and specialized systems.

The adoption path changed:

Watson:
contract → integration → user.

LLMs:
user → experiment → useful workflow → integration.

An LLM can be tested on a concrete task within hours, then incrementally connected to retrieval, tools, access controls and proprietary data.

So Watson’s story is not “old AI lost to new AI.” Watson was an important engineering step.

The deeper difference is architectural: Watson was built as a specialized decision-support product deployed from the top down. LLMs are becoming a general computational layer on which organizations build their own domain-specific systems.

That is why our 2017 negotiations are revealing. We offered IBM infrastructure, local expertise and access to additional markets through joint development. IBM offered a different model: localization and implementation of a finished enterprise system.

Not simply a “smarter box,” but a different level of abstraction, a different cost of adaptation and a different diffusion loop.

I think that explains much of what happened next.

[Tam Urus @urusglobal]

Since July 18, strikes on Wildberries’ logistics infrastructure have increasingly looked less like isolated incidents and more like a systematic campaign against the company’s key distribution hubs. Wildberries itself is widely described in open sources as having close ties to senior figures in the Administration of Mr. P.

According to open-source estimates, warehouse terminals across several Russian regions have been hit within a matter of weeks. Estimates of the affected capacity range from 700,000 to more than 1 million sq. m — roughly 15–22% of Wildberries’ warehouse footprint by some calculations.

The important point is that these are not simply “warehouses full of clothes.” They are automated nodes handling the intake, storage, sorting and dispatch of enormous volumes of goods.
Even integrating roughly 2 million sq. m of former Raven Russia assets that passed to structures associated with the Mirzoyans would not provide a quick solution. Warehouse space by itself is not a Wildberries logistics network.

From my own experience building logistics for large retail networks, the critical variable is not square meters of storage but the movement of goods between major hubs, sorting facilities, transportation and the last mile.

Take out a major hub and neighboring facilities absorb the traffic. Utilization rises, transportation distances increase, delivery times deteriorate, inventory becomes unbalanced, and bottlenecks emerge in intake and dispatch. Ultimately, inventory turnover suffers.

Wildberries’ announced “restructuring of logistics chains” is essentially emergency rerouting. It can keep the system functioning, but it does not replace lost capacity.

Money can compensate for CAPEX and some inventory losses. What money cannot quickly buy is a functioning logistics network: automation, IT integration, trained personnel, routing, operational processes and accumulated efficiency. Rebuilding those capabilities takes years.

Then the second-order effects begin.

Lower throughput hits pickup points whose economics depend directly on transaction volume. Some will cut costs, close, or migrate toward Ozon, Yandex Market and other platforms.

At the same time, sellers face higher concentration risk. After changes to force-majeure provisions and limited voluntary compensation, the rational response is to keep less inventory inside Wildberries and diversify warehouses and marketplaces.

That means lower assortment depth and product availability.
The feedback loop is straightforward:
hub destruction → logistics degradation → delays and shortages → lower transaction volume → deteriorating pickup-point economics → seller and customer attrition → further decline in volume.

So the main risk to Wildberries is not running out of money, or even the replacement cost of destroyed warehouses.
It is the gradual erosion of operational efficiency and network effects.

Administrative leverage may compensate a commercial company’s losses with public funds.
It cannot repeal the laws of logistics.