Burhan Showkat

What's on My Mind

Notes, curiosities, and working margins.

A working notebook of research ideas, methods I am learning, and curiosities beyond pavement engineering. Some are developed; others are questions worth returning to.

Browse the Notebook

Follow a line of thought.

Choose a theme to gather related notes. Some ideas belong to more than one intellectual thread.

Notebook Threads Six paths

Showing all notes.

Works in Progress

Two Questions in Motion

Two active lines of inquiry, connected by one concern: turning material and system response into evidence that supports better infrastructure decisions.

Primary focus

Road Infrastructure Resilience

Research question How can the capacity of road infrastructure to endure, adapt, and recover be quantified in a meaningful and decision-useful way?

Conceptual infrastructure resilience trajectory A conceptual curve showing service disruption followed by adaptation and recovery. ENDURE ADAPT RECOVER disruption
  • Quantification
  • Recovery
  • Decision support
Active development

Reflective-Crack Propagation

Research question How do interlayers alter crack bifurcation, arrest, and lateral spreading in asphalt mixtures under monotonic loading?

Conceptual reflective-crack path through an interlayer system A crack bifurcates below an interlayer. One branch arrests at the lower interface, while the other crosses the interlayer, reaches the upper interface, and spreads slightly downward and to the right. INTERLAYER
  • Bifurcation
  • Crack arrest
  • Lateral spreading

A Technical Note

Ways of Reading Asphalt Binder

One material, nine analytical lenses. Start with the question, then choose the evidence.

Composition

What is in the binder?

  • SARA Fractionation

    Iatroscan

    Four chemical fractions

    Saturates, aromatics, resins and asphaltenes reveal the binder's compositional balance.

    How to read it

    Separation by solubility and polarity yields fractions used to calculate a colloidal index. This helps assess stability and compatibility, but is not a stand-alone prediction of phase separation or brittleness.

  • CHNS

    Elemental analysis

    The elemental profile

    Carbon, hydrogen, nitrogen and sulfur, measured as mass fractions.

    How to read it

    A small sample is combusted for analysis. Carbon-to-hydrogen ratios offer clues to aromaticity; sulfur can support comparisons between sources, but cannot identify a crude-oil origin on its own.

  • XRF

    X-ray fluorescence

    Trace-element fingerprints

    Characteristic X-ray signals reveal elements that bulk composition can miss.

    How to read it

    Secondary X-rays identify elements such as zinc, copper and molybdenum. Their patterns can help screen for recycled engine-oil bottoms or other additives; attribution needs calibration and corroborating evidence.

Molecular fingerprints

How are the molecules changing?

  • FTIR

    Fourier transform infrared spectroscopy

    Oxidation fingerprints

    Infrared absorption identifies functional groups and tracks chemical aging.

    How to read it

    Carbonyl and sulfoxide bands are commonly used to form oxidation indices. Compare spectra consistently to assess changes associated with heat and oxygen exposure.

  • GPC / SEC

    Gel permeation / size-exclusion chromatography

    Molecular-size distribution

    Size-based separation exposes shifts toward larger or smaller molecular species.

    How to read it

    Elution profiles can track aging-related large-molecular-size fractions and polymer contributions, including SBS. Apparent molecular weights depend on the solvent and calibration.

  • NMR

    Nuclear magnetic resonance

    Chemical environments

    Hydrogen or carbon signals distinguish aromatic and aliphatic environments.

    How to read it

    Nuclei respond to a magnetic field according to their local environment. Spectral regions reveal structural features beyond FTIR, rather than a complete map of every molecule in the binder.

Transitions & microstructure

What changes with heat and blending?

  • DSC

    Differential scanning calorimetry

    Thermal transitions

    Heat flow reveals the glass transition and wax-related melting or crystallization.

    How to read it

    Heating and cooling reveal changes in heat capacity and thermal events. Glass-transition temperature \(T_g\) informs low-temperature behavior; wax estimates require a suitable enthalpy reference.

  • AFM

    Atomic force microscopy

    A nanoscale surface map

    A scanning probe resolves topography and local surface heterogeneity.

    How to read it

    Features such as bee structures depend on composition, thermal history and preparation. They are not direct maps of wax or asphaltenes; adhesion interpretation requires complementary measurements.

  • Fluorescence Microscopy

    UV- or blue-light-excited imaging

    Polymer dispersion

    Optical contrast makes polymer-rich domains and uneven blending visible.

    How to read it

    Fluorescence contrast can reveal swollen polymer-rich phases, their distribution and agglomeration. Brightness depends on the binder, modifier and illumination, not simply polymer present versus absent.

Read the methods together. Chemistry and microscopy complement rheology; no single image or index tells the whole performance story.

All miniatures are conceptual, not measured data. Further reading: FHWA chemistry methods and polymer–bitumen interactions.

A Coupled-Materials Note

Chemistry Is Part of the Load Path

Chemical history becomes mechanical consequence.

State ↔ Response

Asphalt is not a mechanically active material with chemistry sitting quietly in the background. Oxidation, moisture, composition, and modification alter the molecular state that rheology and mixture mechanics are attempting to represent.

01 Chemical state oxidation · moisture · composition
02 Rheological response stiffness · phase · relaxation
03 Mixture performance adhesion · healing · cracking

Damage opens new pathways for water and oxygen, feeding exposure back into chemistry.

THE MINIMAL STATEMENT
\[ \text{Response}=\mathcal F(\text{load},T,\text{chemical state},\text{moisture}) \]

The same load need not produce the same response after the material state has changed.

WHAT MOVES THE STATE?
Oxidation Moisture Recycled binder Polymers Rejuvenators

Binder source, exposure history, and mixture structure keep any single chemical index from predicting mechanics alone.

EVIDENCE MUST CROSS SCALES Read both sides of the coupling.
Chemical lens FTIR · SARA · GPC · Microscopy

Molecular change, composition, size, and morphology.

Mechanical evidence DSR · BBR · Adhesion · Cracking

Rheology, low-temperature response, interface, and damage.

The objective is not merely to correlate two columns of data, but to explain how a changing molecular state becomes an engineering consequence. Studies of thermal and oxidative binder aging and chemical and rheological aging indices in mixtures illustrate why both sides of the coupling must be measured.

A Rheological Model

Five Elements, One Rheological Landscape

A compact architecture for a continuous spectrum.

2S 2P 1D

The name 2S2P1D is almost a mechanical inventory: two springs, two parabolic elements, and one dashpot. Together they describe the linear viscoelastic response of asphalt binders and mixtures from the glassy limit, through the broad transition region, to viscous flow at long loading times.

Mechanical analogue of the 2S2P1D model, with spring G g in parallel with a series branch containing spring G zero minus G g, parabolic elements alpha k and h, and dashpot eta.
Elastic limits, distributed transition, and terminal viscous flow represented in one analogue.
01 Glassy bound

\(G_0\) anchors the high-frequency or low-temperature limit.

02 Distributed transition

\(k\), \(h\), and \(\alpha\) shape the broad viscoelastic passage.

03 Long-time response

\(G_g\), \(\beta\), and \(\eta\) govern the equilibrium and viscous scales.

COMPLEX SHEAR FORM Frequency and temperature meet in \(z=i\omega\tau(T)\).
\[ G^*=G_g+\frac{G_0-G_g}{D(\omega,T)}, \] \[ D(\omega,T)=1+\alpha z^{-k}+z^{-h}+(\beta z)^{-1}, \qquad 0<k<h<1. \]
\(G_0,\ G_g\) Elastic limits
\(k,\ h,\ \alpha\) Transition shape
\(\eta=(G_0-G_g)\beta\tau\) Viscous scale
\(\tau(T)=a_T\tau_0\) Temperature shift

What I find appealing is its economy: five idealised elements produce one continuous map of stiffness and phase across loading rate and temperature. Yet an excellent fit is not, by itself, a unique molecular explanation; parameter stability, experimental range, and thermorheological simplicity still have to be tested. The model was introduced for bituminous binders and mixtures by Olard and Di Benedetto.

A Rheological Fingerprint

The Cole–Cole Plane: Two Moduli, One Binder

Storage and loss traced together.

G'' versus G'

A Cole–Cole diagram plots the loss modulus (G'') against the storage modulus (G'). Each DSR observation becomes a point in the complex-modulus plane; changing temperature or frequency traces a path that serves as a compact viscoelastic fingerprint of the asphalt binder.

Original Cole-Cole diagram for an asphalt binder, showing normalized loss modulus against storage modulus, the equal-modulus boundary, and a trajectory from warm slow loading toward cold fast loading.
The trajectory is illustrative: its position and shape should be compared across binders or conditions, not treated as a standalone performance threshold.
COMPLEX MODULUS
\[G^*=G'+iG''.\]
VISCOELASTIC BALANCE
\[G'=G''\quad\Longleftrightarrow\quad\delta=45^\circ.\]
01Position

(G') records stored elastic energy; (G'') records viscous dissipation.

02Path

Temperature and loading rate order the observations along the trajectory.

03Change

Shifts or altered curvature can reveal a changed relaxation architecture.

I like the diagram because it refuses to split elasticity and viscosity into separate stories. Aging, foaming, or modification may move or reshape the path, but the interpretation is strongest when read alongside master curves, Black diagrams, and model parameters rather than in isolation.

Technical examples: water-foamed bitumen and polymer-modified asphalt binders.

A Rheological Fingerprint

The Black Diagram: A Binder Without Temperature

Complex modulus meets phase angle.

|G*| ↔ δ

A Black diagram removes temperature and loading frequency from the visible axes and places every DSR observation in a plane defined by phase angle \(\delta\) and complex shear modulus \(|G^*|\). The resulting path becomes a compact rheological fingerprint: stiffness, viscoelastic balance, and departures from simple time-temperature equivalence can be read together.

Illustrative Black diagram showing one asphalt-binder trajectory from a stiff elastic response toward a soft viscous response as phase angle increases.
Every point represents a DSR measurement; temperature and frequency are encoded along the path rather than shown as axes.
\[ \mathcal{B}=\left\{\left(\delta_i,\log_{10}|G_i^*|\right)\right\}_{i=1}^{n} \]
01 Read position

Up and left is stiffer and more elastic; down and right is softer and more viscous.

02 Read shape

A smooth collapse supports simple shifting; shoulders, loops, or scatter invite investigation.

03 Read movement

Ageing and modification can relocate or reshape the path, revealing a changed response architecture.

What appeals to me is the economy of the representation. Two familiar rheological quantities turn many temperatures and frequencies into one visual argument. The diagram does not replace constitutive modelling or performance testing, but it can reveal where the binder behaves simply, where its structure intervenes, and where the data themselves deserve another look.

Technical basis: Airey, “Use of Black Diagrams to Identify Inconsistencies in Rheological Data”.

A Rheological Test

A Binder, One Cycle at a Time

A time sweep asks a deliberately simple question: what happens when an asphalt binder is sheared again and again under the same nominal conditions? In a dynamic shear rheometer, temperature, frequency, and loading amplitude are held fixed while the material response is recorded cycle by cycle.

A dynamic shear rheometer applies repeated sinusoidal shear at fixed temperature, frequency, and strain amplitude while normalized complex modulus and phase angle are tracked against loading cycles.
The familiar 50% modulus point is a useful convention, not an automatic proof that every preceding change was fatigue damage.
01 Hold fixed

Temperature \(T\), frequency \(f\), and strain \(\gamma_0\) or stress amplitude.

02 Track

Complex modulus \(G^*\), phase angle \(\delta\), and dissipated energy through cycle \(N\).

03 Interpret

Rate of change, failure criterion, strain sensitivity, and any recovery after rest.

\[ \begin{aligned} \gamma(t)&=\gamma_0\sin(2\pi f t),\\ W_d(N)&=\pi\gamma_0^2G^*(N)\sin\delta(N). \end{aligned} \]

Under strain-controlled loading, a larger imposed amplitude generally drives a faster reduction in \(G^*\) and a shorter apparent fatigue life. A common definition, \(N_{f,50}\), is the cycle at which \(G^*\) reaches 50% of its initial value. Other interpretations use peaks in phase angle or stiffness multiplied by cycle count, dissipated-energy ratios, or continuum-damage concepts.

The caution is as important as the curve. Self-heating, thixotropy, nonlinear response, radial flow, and edge fracture can reduce the measured modulus alongside genuine damage. A binder time sweep is therefore a rich comparative rheological probe, but not a direct translation of pavement fatigue life.

Useful discussions include the comparison of time-sweep failure definitions and the assessment of strain-controlled fatigue analysis methods.

A Sustainability Note

A Pavement Does Not Begin at the Paver

LCA whole life

Sustainability in flexible pavement engineering cannot be judged from the asphalt mixture alone. A pavement begins with aggregate extraction and binder production, remains environmentally active through construction and use, and continues beyond milling into recycling, recovery, or disposal.

A circular pavement life-cycle diagram connecting materials, asphalt production, construction, use, maintenance, and recovery around a lane-kilometre functional unit.
The system boundary follows the pavement from raw material to service, intervention, and its next material life.

\[ I_k=\sum_{s\in\mathcal S}\sum_i q_{i,s}\,CF_{i,k}, \]

\(q_{i,s}\) flow quantity \(CF_{i,k}\) impact factor \(\mathcal S\) life-cycle stages
01 Count the whole system

Materials, transport, construction, use, maintenance, rehabilitation, and end of life all belong inside the ledger.

02 Compare equal service

A lower plant impact is not an improvement if it produces earlier intervention or poorer performance elsewhere.

03 Test circular claims

RAP, warm mix, durable binders, and preservation matter through their total effects on demand, energy, service life, users, and recovery.

LCA is not a label of greenness. It is a transparent accounting system for asking whether a pavement delivers more service with less environmental burden.

Start with the FHWA Pavement Life Cycle Assessment Framework for the broader methodology, then explore alternatives using LCA Pave in practice.

A Foundational Note

When Does a Machine Actually Learn?

Machine learning becomes much clearer when it is framed as a measurable change rather than as computational mystique. A system is given experience (E), asked to perform a task (T), and judged by a performance measure (P). Learning has occurred only if further experience improves that measured performance.

Experience, task, and performance feed a learning system; beside it, prior belief and data combine into a posterior distribution that guides a decision.
Two complementary views: learning must improve something we can measure, while probabilistic modelling keeps uncertainty visible.
E Experience

The observations, examples, feedback, or interactions from which the system can change.

T Task

The operation to be learned: classification, prediction, ranking, control, or another defined objective.

P Performance

The criterion that makes improvement testable rather than impressionistic.

\[ \operatorname{Perf}(T;E_{\mathrm{new}}) > \operatorname{Perf}(T;E_{\mathrm{old}}). \]

The probabilistic lens

Replace certainty with calibrated belief.

Predictions, parameters, and unobserved states are rarely known exactly. Representing them with probability distributions records both what the model expects and how strongly the available evidence supports that expectation.

prior belief observed data updated belief

\[ p(\theta\mid D)\propto p(D\mid\theta)\,p(\theta). \]

Data (D) updates uncertainty about ( heta); the resulting posterior can then support a decision under uncertainty.

I like this formulation because it imposes discipline before architecture: define the task, choose an honest measure, state what constitutes experience, and preserve uncertainty in the conclusion. Probability then becomes a common language linking machine learning with statistics, engineering, control, and decision science.

Reading foundations: Tom Mitchell's definition of learning and Kevin Murphy's probabilistic perspective.

A Learning Paradigm

Supervised Learning: Learning With an Answer Key

Examples arrive as feature-output pairs.

x → y

In supervised learning, the experience is a collection of examples for which the desired output is already known. The task is to learn a mapping from an input x ∈ 𝒳 to an output y ∈ 𝒴, then use that mapping when a new input arrives without its answer attached.

An original supervised-learning diagram in which labelled Iris feature vectors train a mapping that predicts the class of an unseen flower.
The Iris example turns four measurements into one of three species labels; the principle extends to both classification and regression.
THE DATA
\[ \begin{aligned} \mathcal D&=\{(x_n,y_n)\}_{n=1}^{N},\\ x_n&\in\mathbb R^D,\qquad y_n\in\mathcal Y. \end{aligned} \]
THE FIT
\[ \begin{aligned} \widehat R_N(\theta)&=\frac{1}{N} \sum_{n=1}^{N}\mathcal L(f_\theta(x_n),y_n),\\ \widehat\theta&\in\arg\min_{\theta}\widehat R_N(\theta). \end{aligned} \]
01 Represent

Turn each observation into informative and consistently measured features.

02 Label

Attach the class, value, or response that the model is expected to predict.

03 Fit

Choose parameters that reduce a task-appropriate loss on the training examples.

04 Generalize

Judge the learned rule on untouched examples, not by its memory of the training set.

I like supervised learning because its contract is unusually explicit: specify what is observed, state what must be predicted, define the penalty for being wrong, and reserve evidence that can challenge the fitted model. The sophistication of the algorithm never removes the need for trustworthy labels and an honest test.

Reading source: Kevin Murphy, Probabilistic Machine Learning: An Introduction.

A Model-Fitting Principle

Empirical Risk Minimization: Turning Error into an Objective

Define the cost. Average it. Minimize it.

min R̂

A model cannot improve until “wrong” has a numerical meaning. A loss function assigns a cost to each prediction; empirical risk is the average of those costs over the observed training set. Empirical risk minimization chooses the model parameters that make this average as small as possible.

An original empirical-risk diagram showing individual training losses, their average, and the parameter selected at the minimum of a risk curve.
The loss function creates the landscape; optimization searches that landscape for a parameter setting with lower average cost.
EMPIRICAL RISK Average observed loss
\[ \begin{aligned} \widehat R_N(\theta)&=\frac{1}{N} \sum_{n=1}^{N}\ell\!\left(y_n,f_\theta(x_n)\right),\\ \widehat\theta&\in\arg\min_\theta\widehat R_N(\theta). \end{aligned} \]
EVERY MISS COSTS ONE Zero-one loss
\[\ell_{01}(y,\widehat y)=\mathbb I(y\ne\widehat y).\]

Useful for misclassification rate when all wrong labels are treated equally.

THE CONSEQUENCE MATTERS Cost-sensitive loss
\[\ell_C(y,\widehat y)=C_{y,\widehat y}.\]

Appropriate when one kind of mistake is more damaging than another.

01Predict

Apply fθ to every training input.

02Penalize

Translate each discrepancy into a loss.

03Average

Summarize the sample as R̂N(θ).

04Optimize

Search for parameters with lower empirical risk.

I find ERM appealing because it exposes every modelling choice: the examples define the evidence, the loss defines what matters, and the optimizer follows the geometry produced by those two decisions. Before asking which algorithm is best, one should ask whether the chosen loss represents the real consequence of error.

Reading source: Kevin Murphy, Probabilistic Machine Learning: An Introduction.

A Statistical Bridge

From Misclassification Rate to Empirical Risk

First count the mistakes. Then ask what each mistake costs.

loss / N

Misclassification rate treats every incorrect label identically: a correct prediction contributes zero and an incorrect prediction contributes one. Empirical risk keeps the same averaging structure but replaces this fixed penalty with a loss function capable of representing severity, asymmetry, or scientific consequence.

An original three-stage diagram showing mistakes counted with zero-one loss, assigned unequal costs, and averaged into empirical risk.
The denominator does not change. What changes is the information carried by each term in the numerator.
SPECIAL CASE Count every error equally
\[ \begin{aligned} \widehat y_n&=f_\theta(x_n),\\ \operatorname{MCR}(\theta)&=\frac{1}{N}\sum_{n=1}^{N} \mathbb I(y_n\ne\widehat y_n). \end{aligned} \]
GENERAL CASE Measure the cost of each error
\[ \widehat R_N(\theta)=\frac{1}{N}\sum_{n=1}^{N} \ell(y_n,\widehat y_n). \]
THE KEY IDENTITY Zero-one loss makes the two quantities identical.
\[ \ell=\ell_{01}\quad\Longrightarrow\quad \widehat R_N(\theta)=\operatorname{MCR}(\theta). \]
01Observe

Compare each predicted label with the corresponding truth.

02Value

Use a loss function to encode the consequence of the discrepancy.

03Average

Combine the individual losses into one sample-level measure.

This small mathematical substitution changes the question from “How often is the model wrong?” to “How costly are its errors on the evidence we observed?” The second question is often closer to the decision we actually need to make.

Reading source: Kevin Murphy, Probabilistic Machine Learning: An Introduction.

A Useful Equivalence

From Loss Function to Maximum Likelihood

Reward probability assigned to what actually occurred.

MLE

For a probabilistic model, a natural loss is the negative log of the probability assigned to the observed outcome. A confident, correct prediction receives a small penalty; assigning little probability to the truth receives a large one.

PROBABILITY OF THE OBSERVED TRUTH The negative logarithm turns confidence into cost.
HIGH SUPPORTp = 0.90
−log(0.90) = 0.105
LOW SUPPORTp = 0.10
−log(0.10) = 2.303
01 · ONE OBSERVATION Negative log loss
\[\ell_n(\theta)=-\log p_\theta(y_n\mid x_n).\]
02 · THE DATASET Average negative log-likelihood
\[ \begin{aligned} \mathcal L_N(\theta)&=\prod_{n=1}^{N}p_\theta(y_n\mid x_n),\\ \operatorname{NLL}_N(\theta)&=-\frac{1}{N}\log\mathcal L_N(\theta). \end{aligned} \]
THE SAME OPTIMUM, VIEWED FROM TWO DIRECTIONS
\[ \begin{aligned} \widehat\theta_{\mathrm{MLE}} &=\arg\max_\theta \mathcal L_N(\theta)\\ &=\arg\min_\theta \operatorname{NLL}_N(\theta). \end{aligned} \]

Negation reverses the direction; the logarithm turns a product into a sum without changing its maximizer.

Thus, empirical-risk minimization with negative log loss is not a separate fitting principle: under the probabilistic model, it is maximum-likelihood estimation written in the language of loss.

Reading source: Kevin Murphy, Probabilistic Machine Learning: An Introduction.

A Generalization Note

When a Perfect Fit Learns Too Much

Training performance is evidence of fit, not proof of generalization.

Δgen

A sufficiently flexible model can drive training loss toward zero by capturing both structure and accidental detail. Generalization asks a harder question: does the fitted rule remain accurate on observations that did not participate in fitting or model selection?

Original diagram showing training loss decreasing with model complexity, validation loss following a U-shaped curve, and separate train, validation, and test roles.
The lowest training loss lies at the far right; the preferred model lies near the minimum validation loss, before complexity begins to fit sample-specific noise.
WHAT WE CAN MINIMIZE Training risk
\[ \widehat R_{\mathrm{train}}(\theta) =\frac{1}{n}\sum_{i=1}^{n}\ell(y_i,f_\theta(x_i)). \]
WHAT WE CARE ABOUT Population risk
\[ R_{\mathrm{pop}}(\theta) =\mathbb E_{(x,y)\sim p^*}\!\left[\ell(y,f_\theta(x))\right]. \]
GENERALIZATION GAP The distance between fitting the sample and predicting the population.
\[ \begin{aligned} \Delta_{\mathrm{gen}}(\theta) &=R_{\mathrm{pop}}(\theta)-\widehat R_{\mathrm{train}}(\theta),\\ \widehat R_{\mathrm{test}}&\approx R_{\mathrm{pop}}. \end{aligned} \]
01Fit on training data

Estimate parameters and gradients using the training observations.

02Select on validation data

Choose capacity, hyperparameters, and stopping rules without consulting the test set.

03Report on test data

Estimate final out-of-sample performance only after model choices are fixed.

The point of learning is not to explain the examples already in hand as intricately as possible. It is to discover enough stable structure that the rule survives contact with the next example.

Reading source: Kevin Murphy, Probabilistic Machine Learning: An Introduction.

A Modelling Maxim

Wrong, Useful, Revisable

A model earns trust by surviving a purpose, not by becoming reality.

model ≠ world

All models are wrong, but some models are useful.

George E. P. Box

A model is a deliberate compression of the world. It ignores detail so that a pattern can be calculated, communicated, or used to decide. Its omissions are not automatically defects; they become defects when they erase what matters for the question at hand.

A SMALL FORMALISM Adequacy is conditional on task, domain, and tolerance.
\[ \mathbb E_{(x,y)\sim D}\!\left[\ell_T(y,m(x))\right] \le \varepsilon_T. \]
01 Useful for what?

Name the prediction, explanation, comparison, or decision the model must support.

02 Useful where?

State the population, loading regime, scale, and conditions that define its domain.

03 Wrong in which way?

Identify the errors that are tolerable and those that would invalidate the intended use.

01 Simplify 02 Confront 03 Use 04 Revise

The sentence is not an argument against modelling. It is an argument for intellectual humility: make assumptions visible, test consequences against evidence, respect the boundary of the model’s domain, and revise it when reality refuses to cooperate.

A Probabilistic Note

Uncertainty: Knowing What the Model Does Not Know

A probability is a degree of belief, not a guarantee.

p(y | x)

Exact prediction is often impossible. Sometimes the model lacks knowledge; sometimes the observation itself is noisy or compatible with several outcomes. A responsible model represents this ambiguity as a distribution over possible outputs rather than concealing it behind a single confident label.

An original uncertainty diagram showing a flower observation transformed into class probabilities, with epistemic and aleatoric uncertainty distinguished below.
One prediction can distribute belief across several outcomes; the reason for that spread determines what should happen next.
FROM SCORES TO PROBABILITIES Softmax normalizes the class evidence.
\[ \begin{aligned} p(y=c\mid x;\theta) &=\operatorname{softmax}_c(f_\theta(x)),\\ &=\frac{e^{f_c(x;\theta)}}{\sum_{c'=1}^{C}e^{f_{c'}(x;\theta)}}. \end{aligned} \]
EPISTEMIC · MODEL UNCERTAINTY Uncertainty from limited knowledge.

Sparse coverage, unfamiliar inputs, uncertain parameters, or an inadequate model can make several explanations plausible. Better-targeted data or a better model may reduce it.

Ask: what evidence would change the model?
ALEATORIC · DATA UNCERTAINTY Uncertainty within the observation process.

Measurement noise, class overlap, and inherent stochasticity can leave different outcomes plausible even with a well-learned model. Repeating the same information does not remove it.

Ask: what ambiguity remains at fixed information?
01Act

When uncertainty and consequences are both acceptably low.

02Investigate

Gather a test, measurement, or expert review when information may help.

03Abstain

Defer the decision when uncertainty and the cost of error are both high.

What matters is not merely whether the model is uncertain, but why. Epistemic uncertainty invites learning; aleatoric uncertainty invites caution; decision theory asks what each possible mistake would cost. A model that can express doubt is often more useful than one that is confidently wrong.

Reading source: Kevin Murphy, Probabilistic Machine Learning: An Introduction.

Current Notes

Machine Learning

Reading for current work

Convolutional neural networks: learning structure from images.

I am studying CNNs because I expect to use image-based pattern recognition in my current work. Their appeal lies in a disciplined progression: learn locally, reuse what is learned, and assemble simple features into increasingly meaningful representations.

A convolutional neural network pipeline in which an image passes through a shared filter, produces feature maps, is pooled, and yields pattern evidence.
A conceptual CNN pipeline: local filters transform pixels into a hierarchy of learned evidence.
01 Local perception

Small kernels inspect neighborhoods rather than treating every pixel independently.

02 Shared weights

The same filter searches across the image, reducing parameters and preserving spatial logic.

03 Feature hierarchy

Edges and textures can combine into shapes, objects, and task-specific patterns.

CNNs are among the most influential architectures in modern machine learning because they make image structure part of the model itself. Convolution, nonlinear activation, and pooling provide a relatively clear entry point into how neural networks turn visual information into a prediction.

My aim is to understand this architecture from the foundations upward rather than treat it as a black box. That means asking not only whether a network is accurate, but what evidence it learned, whether the dataset permits that claim, and whether the result remains stable beyond the images used for training.

Starting point

A gentle introduction for readers who want an entry into CNNs is arXiv:1511.08458.

An Intellectual North Star

Grigori Perelman

Depth without spectacle.

RIGOR

Grigori Perelman remains very much on my mind. His intellectual courage, depth, and refusal of spectacle form a rare model of what it means to pursue an idea with absolute seriousness and without compromise.

Portrait of Grigori Perelman
A mathematician whose work and choices resist spectacle.
01 Depth

Stay with the problem beyond its fashionable surface.

02 Independence

Let the work, rather than spectacle, establish its seriousness.

03 Integrity

Hold intellectual standards without bargaining them away.

GEOMETRY IN MOTION

\[ \frac{\partial g}{\partial t}=-2\,\operatorname{Ric}(g). \]

Ricci flow lets a metric evolve in response to curvature. Perelman's work made this evolution part of a profound route through topology, singularity, and geometric structure.

curvature entropy surgery topology

What stays with me is the union of imagination and discipline: an immense problem approached without theatricality, and a life that makes intellectual seriousness feel almost physical.

Current Curiosity

Rayleigh Distribution and the Stress-Strain Curve

A statistical lens for nonnegative response features.

\(R\geq0\)

I want to understand whether the Rayleigh distribution can offer a useful language for selected features extracted from families of stress–strain responses. The attraction lies in connecting a disciplined probability model to visible mechanical variability, without confusing a distribution with a constitutive law.

A Rayleigh probability density connected to an ensemble of stress-strain curves, showing how a nonnegative response feature could be extracted and studied statistically.
A possible workflow: define a physically meaningful nonnegative feature, observe it across specimens, then test whether the Rayleigh assumptions survive the data.
THE GEOMETRIC ORIGIN The magnitude of two independent Gaussian components
\[ R=\sqrt{X^2+Y^2}, \qquad f_R(r)=\frac{r}{\sigma^2} e^{-r^2/(2\sigma^2)},\quad r\geq0. \]
01 Shape

The density begins at zero, rises to a mode at \(r=\sigma\), and then decays.

02 Scale

\(\sigma\) controls both the characteristic magnitude and the spread.

03 Mechanical question

Could peak stress, strain at a feature, or a response amplitude behave like \(R\)?

What interests me is the bridge itself: moving from the geometry of a probability distribution to a falsifiable question about material response, while keeping statistical description and mechanical interpretation rigorously distinct.

A Thought

Rheology, Theology, and the Deborah Number

De

The Deborah number asks a deceptively simple question: how long does a material need to relax compared with how long we choose to observe it? What appears permanent may simply be changing more slowly than our experiment allows us to see.

The same material retaining its shape during a short observation and gradually spreading during a long observation, illustrating high and low Deborah numbers.
The material has not changed identity; only the competition between its relaxation time and our observation time has changed.

\[ De=\frac{t_{\mathrm{relax}}}{t_{\mathrm{obs}}}. \]

HIGH \(De\) Shape persists

Observation is fast relative to relaxation, so the response appears solid-like.

LOW \(De\) Shape yields to time

Observation is long relative to relaxation, so flow becomes visible.

This is where rheology meets a theological image. The biblical line often paraphrased as “the mountains flowed before the Lord” gains a material meaning: wait long enough and even mountains move. The Deborah number is therefore more than a ratio to me; it is a reminder that permanence is often a matter of timescale.

Howard A. Barnes and his colleagues discuss this beautifully in An Introduction to Rheology, which is a book I love and keep returning to.

A Mathematical Aside

Erdős Number

Collaboration measured as distance.

distance 3

The Erdős number turns collaboration into something almost topological. It is playful, but it also says something serious about how mathematical work travels through people, papers, and shared intellectual lineages.

A collaboration network with a highlighted three-edge shortest path from an author v to Paul Erdos.
A path of length three means three coauthorship links connect the author \(v\) to Paul Erdős.
V Authors

Each scholar is a vertex.

E Papers

A coauthored paper creates an edge.

d Distance

The shortest collaboration chain.

\[ \begin{aligned} G&=(V,E),\\ \operatorname{Erdos}(v)&=d_G\!\left(v,\mathrm{Erdos}\right). \end{aligned} \]

In that sense, a whimsical academic fact becomes a clean expression of connectedness inside a scholarly network. It also reminds me why graph theory fascinates me: intellectual life can be read as relation, path, reachability, and structure.

In that spirit, Network Science by Albert-László Barabási is a wonderful read.

A Mathematical Aside

Poisson Distribution

Order hidden inside scattered events.

The Poisson distribution has an austere elegance: it counts how often an event occurs when arrivals are independent and governed by a steady average rate. One parameter is enough to give apparently scattered events a disciplined shape.

Four events occurring along an interval above a Poisson probability profile with rate lambda equal to four.
A continuous interval, a discrete count: here \(X=4\) is one possible realization when \(\lambda=4\).
rare events independent arrivals steady rate

\[ \Pr(X=k)=\frac{e^{-\lambda}\lambda^k}{k!}, \qquad k=0,1,2,\dots \]

rate \(\lambda\)

Expected events in the chosen interval.

centre \(\mathbb E[X]=\lambda\)

The count gathers around the rate.

spread \(\operatorname{Var}(X)=\lambda\)

The same parameter governs uncertainty.

This equality of mean and variance is the signature I find most memorable. The model says something surprisingly economical: randomness can have structure without losing its randomness.

A Mathematical Infatuation

Graph Theory

The geometry of connection.

G = (V, E)

I am quite genuinely infatuated with graph theory. There is something deeply satisfying about the way it reduces structure, relation, and complexity to a language that is both austere and expansive. A few vertices and edges can contain distance, flow, centrality, cycles, spanning trees, and entire geometries of interaction.

An undirected graph with a highlighted shortest path from u to v and its corresponding symmetric adjacency matrix.
One object, several readings: a drawing reveals relation, a path reveals distance, and a matrix makes the structure computable.
01

The object

A finite vertex set \(V\) and a set \(E\) of unordered pairs define a simple undirected graph.

\[ \begin{aligned} G&=(V,E),\\ E&\subseteq\bigl\{\{u,v\}:u,v\in V,\ u\ne v\bigr\}. \end{aligned} \]

02

Local structure

The degree counts how many immediate relations meet at a vertex.

\[ \deg(v)=\bigl|\{u\in V:\{u,v\}\in E\}\bigr|. \]

03

Distance and reachability

Distance is the length of a shortest path; connectedness means every pair is reachable.

\[ \begin{aligned} d_G(u,v)&=\min\{\ell:\exists\text{ a path}\\ &\qquad u=v_0\sim\cdots\sim v_\ell=v\},\\ G\text{ connected}&\iff d_G(u,v)<\infty \quad\forall\,u,v\in V. \end{aligned} \]

04

Algebra and economy

The adjacency matrix records every edge, while a tree connects all vertices without a cycle.

\[ \begin{aligned} a_{ij}&=\mathbf 1_{\{\{v_i,v_j\}\in E\}},\\ T\text{ a tree}&\Longrightarrow |E(T)|=|V(T)|-1. \end{aligned} \]

What I love here is not only the formalism, though the formalism is beautiful, but its persistence: collaboration networks, infrastructure systems, transport, diffusion, reliability, and the movement of ideas all become legible through relations. Graph theory feels less like a chapter of mathematics and more like a durable way of seeing.

Topology by Gluing

A Square That Remembers Two Circles

The torus is periodicity made visible.

T²

A torus is the mathematical surface of a doughnut, but its most revealing construction begins with a square. Identify the left edge with the right edge and the square becomes a cylinder; identify the cylinder's two circular ends and it closes into a torus.

A square becoming a cylinder and then a torus as opposite edges are identified, with arrows showing the torus's two circular directions.
Gluing does not erase direction: it turns each independent translation into a closed loop.
One Direction

A line closes into a circle

Positions separated by an integer are treated as the same point.

\[ \begin{gathered} x\sim x+n,\quad n\in\mathbb Z,\\ \mathbb R/\mathbb Z\cong S^1. \end{gathered} \]

Two Directions

A plane closes into a torus

Both coordinates repeat independently across the integer lattice.

\[ \begin{gathered} (x,y)\sim(x+m,y+n),\\ (m,n)\in\mathbb Z^2. \end{gathered} \]

Two periodic directions

\[ \boxed{T^2\cong S^1\times S^1\cong\mathbb R^2/\mathbb Z^2} \]

One square contains the plane

\[ \begin{aligned} 0\le x<1,&\quad 0\le y<1,\\ (0,y)\sim(1,y),&\quad(x,0)\sim(x,1). \end{aligned} \]

The square is a fundamental domain: copies tile the infinite plane, while edge identification records how one copy wraps back onto itself.

The Same Idea, Two Complex Dimensions If \(\Lambda\subset\mathbb C^2\) is a rank-four lattice, then \(\mathbb C^2/\Lambda\) is a complex two-torus—four real dimensions built by making four independent directions periodic. The doughnut is only the first picture of a much larger idea.

When Abstraction Finds the World

The Unreasonable Effectiveness of Mathematics

Why does an invented language fit nature so precisely?

Wigner · 1960

Wigner's puzzle is not simply that equations can describe measurements. It is that structures cultivated in relative independence from experience—complex numbers, operators, and abstract spaces—later become the precise language of phenomena they were not designed to explain.

\[ \mathcal M \xrightarrow{\ \Phi\ } \widehat{\mathcal P}, \qquad \left\|\widehat{\mathcal P}-\mathcal P_{\mathrm{obs}}\right\|\ll 1. \]
01 Unexpected transfer

Ideas formed in one mathematical setting become useful in another physical domain.

02 Unreasonable precision

The agreement can reach far beyond the observations that first suggested the theory.

03 Predictive reach

The formalism reveals consequences that were not explicitly placed in its starting data.

What stays with me is Wigner's restraint. Mathematical success is not proof that a model is reality, and every law remains bounded by its domain. It is a recurring scientific gift: powerful enough to use, mysterious enough to keep questioning.

Read Eugene Wigner's original essay in Communications on Pure and Applied Mathematics.

A Physical Reverie

Dirac and the Other Half

When mathematical symmetry enlarged the material world.

e− ↔ e+

Dirac trusted an equation far enough to let it describe matter that no one had yet seen. In bringing quantum mechanics into accord with special relativity, he obtained not only a theory of the electron, but a mathematical mirror that the known inventory of nature could not yet fill.

Mirrored positive and negative relativistic energy branches beside an electron and positron with equal mass and spin but opposite electric charge.
The relativistic spectrum carries two signs; physical interpretation turns the apparent surplus into matter–antimatter symmetry.

\[ \bigl(i\hbar c\,\gamma^\mu\partial_\mu-mc^2\bigr)\psi=0, \qquad E_\pm(\mathbf p)=\pm\sqrt{p^2c^2+m^2c^4}. \]

01

A required union

A first-order wave equation had to respect both quantum mechanics and special relativity while accounting naturally for electron spin.

02

The mathematical surplus

The relativistic energy relation admitted positive and negative branches. The unwanted solutions could not simply be wished away.

03

Nature answers

The anti-electron would have the electron’s mass and spin but the opposite charge. Anderson’s positron supplied the experimental answer.

In that sense, Dirac conceived half of the material universe in his mind before experiment revealed it. It remains one of the loveliest examples of disciplined imagination: mathematics did not decorate an observation; it insisted that nature was larger, stranger, and more symmetrical than observation had yet shown.

Historical reference: Nobel Prize account of Anderson’s positron discovery.

A Reading Note

On the Importance of Reading

READ × 3

I keep returning to reading not merely as an academic obligation, but as a practice that enlarges inner life, sharpens judgment, and develops seriousness of thought. Information may be collected quickly; intellectual depth usually has to be read into being.

The word Read repeated beside an open book whose pages lead toward film and research.
Film and research may ask different questions, but both depend on sustained attention before expression.
01 Enlarge

Build an inner archive of voices, histories, methods, and unfamiliar ways of seeing.

02 Sharpen

Develop the judgment to distinguish an interesting claim from a merely fashionable one.

03 Deepen

Stay with a problem long enough for better questions to replace the first easy answers.

Werner Herzog’s repeated insistence to read, read, read feels demanding and liberating. He says it in the context of filmmaking, but it applies just as strongly to research, where depth of thought is often inseparable from depth of reading.

Watch the compilation of Herzog on reading.

A Playful Formalism

The Odyssey: A Very Long Way Home

The destination is simple. Preserving the self is not.

NOSTOS

Homer gives us a journey whose endpoint is known almost from the beginning: Ithaca. The difficulty is not identifying home, but passing through appetite, pride, grief, delay, and enchantment without forgetting why one wished to return.

A conceptual, non-geographical route from Troy to Ithaca through selected episodes of the Odyssey, with a memory-of-home signal that bends but never reaches zero.
Selected episodes, arranged conceptually rather than geographically: the route wanders, while the idea of home remains the hidden coordinate.
01 · WORLD \(G=(V,E)\)

Islands, thresholds, and seas become vertices and passages in a directed graph.

02 · TRAVELLER \(s_t=(x_t,m_t)\)

\(x_t\) is location; \(m_t\in[0,1]\) is memory of home and purpose.

03 · HOMECOMING \(x_T=\mathrm{Ithaca}\ \land\ m_T>0\)

Arrival alone is insufficient: the traveller must still know what arrival means.

THE NON-SHORTEST PATH A successful route need not be an efficient one.
\[ P=(v_0,\ldots,v_T),\qquad v_0=\mathrm{Troy},\quad v_T=\mathrm{Ithaca}, \] \[ \text{success}(P)= \mathbf 1\!\left[x_T=\mathrm{Ithaca}\ \land\ m_T>0\right]. \]

Perhaps that is why the poem remains so alive. Every life has an Ithaca, but the more interesting variable is what survives in us while we are trying to reach it.

A Note on the Inner Landscape

The Old Patterns Beneath New Stories

Carl Jung, archetypes, and the work of becoming whole.

Ψ

For Jung, an archetype is not a stock character, a personality label, or an inherited picture. It is a recurring organizing pattern within the collective unconscious—a form we encounter indirectly through dreams, myths, art, relationships, and the symbols a culture gives to difficult human experience.

A mandala-like map of Jungian motifs beside a sequence in which an archetypal form becomes a symbolic image, a lived encounter, and an opportunity for integration.
The archetype is not the image itself. It is the recurring pattern around which many different images can gather.
Archetype the organizing pattern

Never encountered in a final, culturally neutral form.

Symbol the pattern made imaginable

Historically situated, personal, revisable, and alive.

01

Persona

The social face that helps a person meet the world. Necessary, but dangerous when mistaken for the whole self.

02

Shadow

Qualities and possibilities the conscious ego does not readily own. The shadow is not simply evil; it may also contain neglected vitality.

03

Anima / Animus

Jung's historically gendered language for an inner other. It is most useful today as symbolic polarity, not a rigid identity rule.

04

Self

The organizing symbol of psychic wholeness: larger than the conscious ego, and capable of holding tension without pretending it has vanished.

Archetype ≠ fixed character

one pattern → many symbolic forms

Individuation, then, is not the production of a flawless self. It is the difficult practice of becoming less divided and less governed by what remains unexamined—bringing persona, shadow, inner other, and conscious purpose into a more honest relation.

An Eclipse, A Verdict

Eddington and the Bent Light

1.75 arcsec

Einstein's general theory of relativity made a precise and audacious prediction: the Sun's gravity should bend passing starlight, shifting the apparent position of a star near the solar limb. The eclipse supplied a way to see the otherwise invisible geometry.

A total solar eclipse revealing background stars, a curved light path near the Sun, and a photographic plate comparison showing a small angular displacement.
Hide the Sun, photograph the surrounding stars, then compare their eclipse positions with a reference plate.

\[ \Delta\theta=\frac{4GM_{\odot}}{c^2b}\approx1.75^{\prime\prime}. \]

01 Predict

Relativity specifies how much starlight grazing the solar limb should bend.

02 Reveal

On 29 May 1919, totality made nearby stars photographable from Príncipe and Sobral.

03 Compare

Eclipse plates and reference plates turn apparent stellar displacement into a test.

The results announced by Frank Dyson, Eddington, and Andrew Crommelin were consistent with Einstein's prediction and became the celebrated first observational confirmation of general relativity. I find the scene irresistible: an eclipse made the geometry of spacetime visible, and an idea formed in equations left a measurable mark upon the sky.

Read the Royal Astronomical Society's account of Eddington's total-eclipse expedition.

A Relativistic Note

When One Hour Becomes Seven Years

1 h 7 y

Interstellar makes time dilation emotional before it makes it mathematical. On Miller's planet, deep within the gravitational environment of Gargantua, one hour for the landing party corresponds to seven years far from the black hole.

A schematic of Miller's planet near the rotating black hole Gargantua, comparing one elapsed hour near the black hole with seven years far away.
The observers separate in spacetime, not merely in space; their clocks reunite carrying different histories.

\[ \begin{gathered} \mathcal R =\frac{\Delta t_{\mathrm{far}}}{\Delta \tau_{\mathrm{Miller}}} \approx 6.14\times10^4,\\ 1\ \mathrm{hour}\longleftrightarrow 7\ \mathrm{years}. \end{gathered} \]

01

Gravity changes the clock

A simple Schwarzschild clock at radius \(r\) accumulates proper time as

\[ d\tau=dt\sqrt{1-\frac{2GM}{rc^2}}. \]

02

Motion changes it too

Relative motion contributes the familiar special-relativistic factor

\[ d\tau=dt\sqrt{1-\frac{v^2}{c^2}}. \]

That is what gives the sequence its force. Cooper does not merely travel far from home; he returns displaced in time from the people he loves. Relativity becomes loss measured not in distance, but in years. There is no universal clock shared by every observer.

Explore the scientific construction of Gargantua in the paper by Oliver James, Eugénie von Tunzelmann, Paul Franklin, and Kip Thorne.

A Companion Relativistic Note

When Motion Shortens the Ruler

L < L0

Time dilation has a spatial companion. An object has its greatest length L0 in the frame where it is at rest. An observer who sees it move at speed v measures a shorter longitudinal length L, provided the positions of both endpoints are recorded at the same time in that observer’s frame.

Two-frame length-contraction diagram showing a rod at its proper length in its rest frame and a shorter measured longitudinal length in a frame where it moves.
The transverse size is unchanged in the illustration; only the dimension parallel to the relative velocity contracts.
\[ \gamma=\frac{1}{\sqrt{1-v^2/c^2}}, \qquad L=\frac{L_0}{\gamma}=L_0\sqrt{1-\frac{v^2}{c^2}}. \]
01

Proper length L0

Measure the endpoints in the object’s rest frame. This is the maximum spatial length assigned to the object.

02

Contracted length L

Measure both moving endpoints simultaneously in the observer’s frame. The result is L0/γ.

0.60cL = 0.800L0
0.80cL = 0.600L0
0.99cL ≈ 0.141L0

The lesson is quietly radical: there is no observer-independent answer to “how long is it?” until the frame and the simultaneity convention are specified. Space, like time, participates in the geometry of relative motion.

A Relativistic Thought Experiment

The Twin Who Returned Younger

20 y12 y

Two twins synchronize their clocks. One remains on Earth; the other travels to a distant star at high speed, turns around, and returns. At reunion, less time has elapsed for the traveller. The clocks are not confused: they have measured two different paths between the same meetings.

Cartoon of twins saying goodbye, a travelling twin turning around in a rocket, and the twins reuniting with different elapsed times, followed by their two worldlines through spacetime.
The Earth twin follows one inertial worldline; the travelling twin changes inertial frames at the turnaround.
\[ \Delta\tau=\int\sqrt{1-\frac{v^2}{c^2}}\,dt, \qquad v=0.8c\Rightarrow\gamma=\frac{5}{3}. \] \[ \Delta\tau_{\rm Earth}=20\,\mathrm{y}, \qquad \Delta\tau_{\rm traveller}=\frac{20}{\gamma}=12\,\mathrm{y}. \]
THE APPARENT SYMMETRY

Each sees the other move

During either constant-speed leg, each twin may describe the other clock as running slowly.

THE MISSING DISTINCTION

Only one twin changes frame

The traveller switches from the outbound frame to the inbound frame; the Earth twin does not.

Neither twin owns the universal clock because there is no universal clock. In relativity, elapsed time belongs to a path through spacetime, not merely to its starting and ending events.

The Grand Unfinished Question

Hawking and the Unfinished Equation

Can one framework speak for the very large and the very small?

GR + QM → ?

Modern physics rests on two astonishingly successful languages. General relativity describes gravity as the geometry of spacetime; quantum theory describes matter and fields through quantized states and probabilities. Each works beautifully in its own domain. The difficulty begins where both must speak at once.

General relativity and quantum theory shown as two successful frameworks separated by the unresolved problem of quantum gravity, with Hawking radiation from a black hole below as a clue connecting gravity, quantum fields, and thermodynamics.
Black holes force the two theories into the same room. Hawking radiation is a profound clue, but not yet the completed union.
\[ \begin{gathered} G_{\mu\nu}+\Lambda g_{\mu\nu} =\dfrac{8\pi G}{c^4}T_{\mu\nu} \qquad \text{gravity},\\[4pt] \Updownarrow\;?\\[-1pt] i\hbar\,\partial_t|\psi\rangle =\widehat H|\psi\rangle \qquad \text{quantum theory}. \end{gathered} \]
01 · THE AIM

One consistent foundation

A deeper framework should contain gravity and quantum physics without contradiction and recover their successful predictions where each already works.

02 · THE OBSTACLE

Spacetime must become quantum

Quantum fields usually evolve on a spacetime background; relativity says that background is itself dynamical. At extreme scales, that separation fails.

03 · HAWKING'S CLUE

Black holes are laboratories

Hawking showed that quantum fields near a horizon produce thermal radiation, binding gravity, quantum theory, and thermodynamics into one unavoidable puzzle.

What I find moving in Hawking's search is its scale of ambition joined to intellectual humility. The dream was not merely to find a shorter equation, but to understand why the universe can be described coherently from quantum fluctuation to cosmic history. The unfinished center is not a failure; it is the frontier.

Read Hawking's reflections in Gödel and the End of Physics.

A Scenario Worth Stress-Testing

AI 2040: A Future Delayed on Purpose

Scenario, not forecast

The AI Futures Project's AI 2040: Plan A is not presented as its authors' most likely forecast. It is a positive policy scenario: slow the race to superintelligence, make frontier research inspectable, and preserve human control long enough to build a credible safety case.

Read AI 2040
  1. 2029

    Make the deal

    The United States and China agree to a verified slowdown and declare the frontier compute they control.

  2. 2030

    Brake the loop

    The agreement blocks automated AI research from turning capability gains into a rapid, self-reinforcing takeoff.

  3. 2035

    Hold the line

    Development pauses near top-human-expert capability, buying time for alignment, oversight, and institutional learning.

  4. 2040

    Conditional handoff

    Scaling resumes only when the safety case is judged stronger than the risks of continued delay.

Verify the compute

Chip accounting, mutual declarations, audits, inference-only controls, and secured research clusters make the slowdown enforceable without relying on trust alone.

Open the research

Algorithms, experiments, code, and safety cases become open to scrutiny, while dangerous model weights and sensitive data remain protected.

Titrate capability

Scale toward the strongest AI that remains confidently controllable, then slow almost to a halt rather than racing through an intelligence explosion.

The five pressures Plan A confronts

  • Loss of control
  • Concentrated power
  • Great-power conflict
  • Labor displacement
  • Catastrophic misuse

What interests me most is the discipline of scenario scrutiny. Instead of saying that governance will somehow catch up, the proposal has to survive dates, incentives, verification limits, institutional weakness, and the possibility of defection. Whether or not one accepts Plan A, that is a serious way to think about a future arriving faster than our institutions can comfortably reason about it.

Are You Up?

A small numbers game.

There is a hidden number between 1 and 31. You have six attempts. I'll tell you if the guess should go higher or lower.

Six attempts remain.

Back to the top