AI research in the 1990s
Hinton, neural networks and the long road to deep learning — what the field looked like before the data and the GPUs arrived.
Turning rigorous research into intelligent systems.
Resources · Blog
Notes & thoughts.
A monthly note on the technology we have built, seen, or simply been thinking about — drawn from three decades across AI research, capital markets and high-performance computing.
-
Foresight, Not Hindsight: Big Data, Algorithmic Decisions, and the Responsibility That Comes With Them
“Compliance is a floor, not a ceiling. The question is no longer whether we can mine and decide at this scale, but what we owe the people whose lives sit inside the data.”
The capacity to mine data and decide at scale is settled; the frameworks around it have matured, though they lag and diverge across borders. What matters now is whether the people who build these systems exercise foresight — rather than waiting for harm to arrive. The companion to last month’s piece on the infrastructure that made the scale easy. Read →
-
When Data Outgrew the Machine: The Infrastructure That Made Scale Easy
“A query over a petabyte now runs about as easily as a query over a spreadsheet did a generation ago. We watched that ease arrive, and it was not inevitable.”
Replicated storage, divided computation, and a steady climb in abstraction — from MapReduce to Spark to cloud data warehouses — turned big data from a specialist ordeal into something ordinary. A short history of the hardware and software that tamed scale, and what replaced the constraint once it dissolved. Read →
-
Reading the Signal in Time: How Forecasting Has Matured, and Why Finance Stays Hard
“The tools will keep improving, and they should be welcomed. They will not relieve the practitioner of the harder work, which is knowing the limits of what any forecast can honestly claim.”
Forecasting that once demanded a specialist now takes a competent developer an afternoon. We built a seasonal model in 2021 and watched that ease arrive — yet financial data, for reasons of codependency and regime change, stays among the hardest of all to forecast. The third of three notes on prediction in finance. Read →
-
What Should You See Next? The Long Evolution of Recommender Systems
“When a system decides not just which film to suggest but which posts, news, and viewpoints reach a person, it has stopped being a convenience and become an editor.”
From a narrow question about which film you might like, recommender systems became the quiet logic ordering much of what we read, buy and believe. We explored their early promise for institutional finance in 2014 — and have watched their reach, and their problems, grow ever since. Read →
-
The Long Way Round to a Simple Idea: Customer Value Modelling in Finance
“The model captures the statistical shadow that behaviour casts on data rather than the behaviour itself, and the shadow proves more regular than the behaviour.”
A handful of distributions taught to us as undergraduates — Poisson, gamma, negative binomial — turned out to predict customer value and defection in financial services with surprising robustness. Why the simple Buy Till You Die models still hold their own against deep learning. Read →
-
Owning the Whole Problem: How Software Practice Changed
“Modern practice makes that depth optional in a way the older practice never did, and the engineers who choose it anyway are the ones whose work we trust most.”
Requirements drawn up once, blocks of work owned for months, integration at the end. Agile and continuous integration corrected the real failures of that approach — but a particular kind of depth went with it, and it is worth recovering inside the modern structure. Read →
-
A Difference of Category: Hardware and Software Since the 1990s
“A server in 1998 and a cloud instance in 2026 are not two points on the same scale. They differ in kind, and the engineering each one demands differs in kind with them.”
Compute was a budget, memory was counted, and standard utilities had to be written from first principles. Abundance lifted every one of those constraints — and quietly removed an awareness the scarcity once made compulsory. Read →
-
Building Without the Scaffolding: What Finance Technology Achieved Before the Tools Caught Up
“The firms that grasped each new tool early took ground from those that waited, and the advantage compounded.”
Without hardened security libraries, reliable databases or open source, 1990s investment banks built straight-through processing anyway — treating emerging technology as a competitive edge, and reorganising how an entire industry moved information in a window of a few years. Read →
-
Why Financial NLP Needs Interpretability by Design, Not by Retrofit
“A regulator asking why a specific lending decision was made deserves an answer about the model, not an answer about a surrogate of the model. For high-stakes deployment, approximate transparency is not a stepping stone to real transparency. It is a category error.”
With the EU AI Act in force, interpretability, uncertainty quantification and lifecycle monitoring are no longer optional extras for financial NLP — they are operational requirements. Read →
-
A Tension Worth Watching: Optimisation, Cryptography, and the Problems Between Them
“Optimisation researchers want hard problems to fall. Cryptographers need them to stand.”
The same intractable problems optimisation wants to crack are the ones cryptography needs to hold. A thirty-year view of that tension — through Shor’s algorithm, quantum annealing, QUBO and the still-open P versus NP — and the unsettled mathematics beneath both. Read →
-
From Caesar to the Internet: Robert Churchhouse and the Long Arc of Cryptography
“The cryptography protecting a bank transaction today belongs to the same continuous intellectual tradition as the cipher that protected a message in Caesar’s time.”
He was taught by Turing and Newman, spent eleven years on ciphers inside Britain’s signals-intelligence service, and closed his career with a book that ran from Julius Caesar to the encryption of the internet. We were among his students — and his life is a lesson in the continuity of cryptographic security. Read →
-
Requests for Comments: How Open Standards Built the Network We Trust
“The network spent its first era assuming trust and its second engineering against its absence — and the tools it reached for had been waiting in the mathematics since well before the problems arrived.”
From plaintext Telnet on trusting research networks to SSH, TLS and the encrypted standards that secure everything today — how open standards are made by a request for comments rather than a decree, why their priorities shifted from connection to security, and the old mathematics waiting underneath. Read →
-
Closer to the Machine: What Decades of Programming Languages Have Taught Us
“The Python developer who knows what their data structures cost chooses differently from the one who does not, and the difference shows in the result.”
Thirty years across Ada, C, C++, Java and Python taught us that abstraction has been good for the field — and that it leaves a gap. Why every programmer should spend real time closer to the machine, not for nostalgia, but for the model of computation it builds. Read →
-
Framework, Not Metaphor: What a Metaheuristic Actually Is
“If any metaphor can license a new metaheuristic regardless of whether it optimises anything, the term describes nothing. Holding the framework definition is what keeps the word meaningful.”
A history of the field — co-written by Fred Glover, who coined the term — restates what a metaheuristic really is: a framework, not an algorithm, powered by memory and learning rather than randomness dressed in metaphor. Read →
-
Neural Networks in Optimisation: Patience, Not Hype
“A method can be both not-yet-competitive and worth sustained investment, and the metaheuristics community treated neural networks as exactly that for the better part of two decades.”
Long before the 2012 breakthrough, optimisation researchers studied, debated and hybridised neural networks — assessing them honestly rather than dismissing or overselling. Why that patience aged well. Read →
-
Metaheuristics at Fifty
“A discipline that began as the pragmatic fallback when exact methods proved intractable now schedules, routes, and assigns across much of the infrastructure we rely on.”
EURO marked its fiftieth anniversary with an invited review of the field. A brief marker of the milestone — fifty years on, the techniques are everywhere and the most interesting work is still ahead. Read →
-
Old Methods, New Combinations: Why 1980s Optimisation Still Runs the World
“Forty years on, the core techniques have aged remarkably well. Composition is where the interesting work now lies.”
Simulated annealing, genetic algorithms and tabu search still run much of the infrastructure around us. The frontier now lies in how we combine them — and in what machine learning adds to the mix. Read →
Coming soon
A rolling series, roughly one a month, on themes from across our work and the history we have lived through.
Programming languages across the decades
From BASIC on a ZX Spectrum and Ada in avionics to C, C++, Java and Python — how the tools shaped the systems.