DEV Community: Alan Bonnici

The AI Revolution — History Repeating Itself — On Economics

Alan Bonnici — Tue, 11 Aug 2026 07:00:00 +0000

This article turns from sentience to economics — and to the social and political currents that come with it.

The first article in this series argued that today's AI is not sentient: it is a good technology built from statistics, marketing and scale. This one asks a different set of questions — what AI is worth, who gains from it, and who ends up paying. As in the first article, I asked my AI committee to draw up a shortlist of developments that reshaped economies and societies. Eight were proposed; five were chosen by almost every model, and the sixth was a tie that I resolved with my committee of three AIs. That shortlist, with sources, is in Appendix A.

Across those six developments (described in Appendix A), a consistent pattern emerges — one worth holding in mind as we turn to AI:

1. Early disruption precedes long-term benefit.
Each transformation created significant initial costs. Mechanisation displaced artisans, electrification devalued older machinery and skills, digitalisation eliminated clerical and production work, printing displaced scribes, and public-health measures required expensive infrastructure. Oceanic globalisation was far more destructive, involving conquest, disease, slavery and dispossession.

2. Pioneers and investors do not always capture the value they create.

Some inventors, entrepreneurs and early investors failed financially despite creating lasting technologies and infrastructure. Railways are a strong example: many pioneering companies entered bankruptcy, receivership or reorganisation, while their tracks continued operating under new ownership. Gutenberg lost control of his printing equipment, and many dot-com companies failed even though their investments helped build the digital economy. Early investors may bear much of the risk while later businesses and society capture the long-term value.

3. Technology alone was insufficient.

The greatest benefits appeared only after societies reorganised around the new development. Electricity required grids and redesigned factories; automobiles required roads; digital technologies required new business processes and telecommunications networks; public health required sewers, clean water and organised institutions. Adoption was therefore gradual rather than instantaneous.

4. Large upfront investment created shared infrastructure.

Most developments depended on expensive systems that could serve entire economies: factories, railways, electrical grids, roads, ports, printing networks, telecommunications, sanitation systems and healthcare programmes. Once established, these systems lowered costs for many subsequent activities.

5. Costs fell and scale increased.

Each transformation drove down the cost of producing goods, moving people and cargo, transmitting information, coordinating organisations, accessing knowledge or preventing disease — and as costs fell, scale grew. Production moved from workshops to factories, knowledge from manuscripts to mass printing, trade from regional routes to global networks, and computing from isolated machines to worldwide platforms. Organisations became larger, markets broader and economic activity more interconnected.

6. Labour was displaced, reorganised — and unevenly rewarded.

Older occupations declined while new industries and professions emerged: agricultural and craft work gave way to factories, transport, engineering, administration, healthcare, communications and software. The central effect was not the destruction of jobs but a continuing shift in the kind of work people did. The gains, however, were shared unequally — owners of capital, industrial powers and early adopters tended to benefit first, while workers, colonised populations and poorer communities bore a disproportionate share of the costs.

7. Knowledge became an increasingly important economic resource.

Printing allowed ideas to circulate more widely, while digital networks accelerated the storage, copying and transmission of information. Public health also depended on the systematic application of scientific knowledge. Economic progress increasingly came from education, research, communication and organisational capability rather than physical production alone.

8. Urbanisation and complex societies became possible.

Industry concentrated workers in cities, transport and electricity supported dense economic centres, and sanitation made large urban populations safer. Modern cities, hospitals, universities, factories and offices depend on several of these transformations operating together.

9. Every solution created new risks.

Industrialisation and motorisation produced pollution and fossil-fuel dependence; global trade helped spread disease and exploitation; printing and digital media accelerated misinformation; antibiotics contributed to antimicrobial resistance; and global digital systems introduced concerns over surveillance, monopoly power and employment displacement.

The AI opinion split

Those patterns are the backdrop; the live argument is over how they will play out for AI. Public debate is usually reduced to optimists versus sceptics, but the economic, social and political divisions are more complex. The different camps are not mutually exclusive: someone may consider AI useful but overvalued, productive but socially unequal, or transformative but politically dangerous.

The bubble sceptics, represented most forcefully by Ed Zitron, argue that investment in chips, data centres and AI companies has moved far ahead of proven revenue. A milder version, associated with analysts such as Jim Covello, accepts that AI has useful applications but doubts whether these can generate returns sufficient to justify present spending and valuations. As with railways and the dot-com boom, the technology could survive and transform society even while investors and pioneering companies lose heavily.

The measured-growth sceptics, most notably Daron Acemoglu, do not necessarily predict a crash. They expect AI to produce genuine but comparatively modest economic gains because only some workplace tasks can be automated reliably and profitably. Improvements demonstrated in selected occupations may not translate quickly into economy-wide productivity growth, particularly when adoption requires new skills, processes and infrastructure.

The productivity optimists, including Erik Brynjolfsson, Ethan Mollick and Andrew Ng, regard AI primarily as a tool that will augment workers, lower the cost of cognitive tasks and improve organisational efficiency. They argue that the benefits will emerge gradually as businesses redesign work around the technology. The more ambitious transformational optimists, such as Dario Amodei, Demis Hassabis and Sam Altman, expect AI to do more than improve existing processes: they believe it could accelerate scientific discovery, reduce the cost of expertise and create major new sources of economic growth.

Other disagreements concern not how much value AI will create, but who will receive it. Labour and inequality critics — including Acemoglu, David Autor and Simon Johnson — warn that higher productivity does not automatically produce higher wages or better employment. AI could augment employees, but it could also replace, deskill or intensively monitor them, allowing shareholders and dominant companies to capture most of the gains. Critics such as Timnit Gebru, Kate Crawford and Meredith Whittaker extend this argument to hidden labour, discrimination, privacy, environmental costs, surveillance and the concentration of technological power.

This leads to the political divide. Market-oriented advocates favour rapid development and relatively limited intervention, arguing that competition and innovation will spread the benefits. Democratic-governance advocates instead believe that competition law, labour protections, privacy rights, independent auditing and public investment are needed to ensure that AI serves society rather than a small number of corporations. A further geopolitical perspective treats AI, semiconductors and computing capacity as strategic infrastructure, making their control a question of national security, economic sovereignty and international power.

The central disagreement is therefore not simply whether AI will succeed. It is whether the value created will justify the investment, how its benefits and costs will be distributed, and who will control the economic and political power that follows.

History, repeating

The reason this belongs in a series called History Repeating Itself is that we have been here before — and the past offers cold comfort to both camps. Some of the patterns in Appendix A are already visible: the early disruption and the vast, debt-fuelled infrastructure build-out. Others I expect to follow — the pioneers who lose while society keeps the infrastructure, and a division of the spoils that favours capital over labour.

Reading the current boom against that history, here is what I expect:

AI will be one of the larger revolutions, and it will become embedded in every economic, social and political sphere. Few significant decisions will be taken without consulting an AI; countries and organisations will bring AI "advice" into the room at the point of decision. But it will be consulted rather than blindly obeyed; formal human accountability will remain, even where much of the underlying process is automated. Systems of several AI engines working together to solve problems will become the norm, and AI itself will come to be treated like electricity or the internet: a basic necessity rather than a luxury.
The energy and footprint needed per unit of AI work will fall sharply, as advances in technology and improvements in how systems are trained yield solutions that require far less power. Total demand may well keep rising for years — greater efficiency tends to increase overall use rather than curb it — but the energy and financial cost of comparable AI inference should continue to fall, even if more capable reasoning systems consume more computation per request.
Many countries will adopt or build open-weight models to protect their own interests and reduce dependence on foreign firms and governments they cannot control. A nation that runs critical systems on someone else’s closed model is exposed to data-access risks, opaque model changes, policy restrictions, abrupt service withdrawal and potentially undetectable bias — risks serious enough to make sovereign AI a matter of national security.
The AI bubble is real and will burst, taking a number of the pioneering companies with it and wiping out trillions in investor wealth. It will shake the world economy — but, as after the railways and the dot-com crash, a new generation of owners will buy the wreckage cheaply, build on the infrastructure the pioneers paid for, and carry AI into its next stage.
The market will probably settle around a handful of dominant providers, creating pressure for utility-style regulation, interoperability rules or competition-law intervention. After the shake-out, a few survivors will control the large "frontier" models, because the cost of building them is enormous and favours scale. History suggests what follows: railways, telephone networks and electricity all collapsed into a few giants, and once they became too essential to daily life to fail, governments stepped in — regulating them like utilities or breaking them up under competition law. Expect the same for the dominant AI providers once societies decide they are too important to leave unchecked.
The models themselves will become cheap; the real money will move to the things around them. As open-weight models spread and prices keep falling, access to a capable model will become a low-margin commodity, much like bandwidth or electricity. The durable profits will accrue instead to whoever owns the scarce complements — the chips and compute, the energy to run them, the proprietary data to train on, and the direct relationship with the customer. As in a gold rush, it is often the sellers of picks, shovels and land who prosper most, not the prospectors.
It will cause large-scale displacement and losses in particular categories of work at first, as tasks move to AI under human oversight. Many of the roles that remain will become supervisory — reviewing the output of AI, a kind of quality control, and handling the exceptions the models cannot.
It will also create new work. Some of it will centre on supplying, testing and validating the real-world knowledge AI systems lack — including the collection of physical-world data, simulation, evaluation, exception handling and supervision. One can imagine people employed to raise AIs from "infancy" into a deeper understanding of reality, so that a large share of today's digital workers become AI-workers. There will also be roles that today have not yet been imagined. Manual, dexterous trades will hold their value longest, since physical dexterity is something AI will be slowest to match, even where an AI assistant helps diagnose the problem and suggest a fix.
The productivity payoff will disappoint before it arrives. For the first few years, measured gains will fall short of the hype — feeding "AI was oversold" headlines — before a later surge. This resembles what economists call the productivity J-curve: early organisational investment can depress or obscure measured productivity before redesigned processes begin producing visible gains, as occurred with electrification and information technology. The benefits show up only once organisations rebuild their processes around the technology instead of bolting it onto the way they already work. Expect a frustrating gap between what AI can obviously do and what appears in the economic statistics.
Society will end up more productive, able to do more at lower cost. The catch is distribution: a far greater share of the gains will flow to investors than to employees, whose bargaining power in many internationally tradable occupations is already constrained by globalisation, remote hiring and access to a worldwide labour pool.
As labour's share of the gains shrinks, redistribution will move from the fringe to the mainstream. If output rises while wages stagnate, pressure to share the proceeds by other means will grow. Ideas dismissed today as radical — universal basic income, taxes on automation or on compute, even sovereign "compute dividends" paid to citizens — will become serious political platforms, much as the upheavals of industrialisation eventually produced trade unions, labour laws and the welfare state.
The gap between the wealthiest companies and individuals and everyone else will widen sharply — and this, in my view, is the real danger. Concentrated wealth buys bargaining power over governments, influence over elections and control of the media narrative. Combined with the vast amount these entities already know about each of us, it lets them micro-target their messaging and steer attention away from the problems that matter towards whatever they would prefer us to look at. That is the biggest threat on this list, and the one most likely to end in social upheaval.

None of this is a reason to dismiss AI, any more than the Railway Mania was a reason to reject railways. The technology is real, and in time it will earn its keep. But the lesson of the last thousand years is that the machine is rarely the hard part; the hard part is who ends up owning it and who is left carrying the cost.

There is, though, one way in which AI may not repeat history: its pace. The revolutions in Appendix A unfolded over decades or generations, which gave societies time — however painful — to adapt, retrain and legislate. AI is moving far faster. The methods behind today's models matured through the 2010s, but the public turning point was ChatGPT's debut in November 2022; less than four years later the technology has directly or indirectly touched billions of lives. If its economic and political consequences arrive at that same speed, we will have far less time to soften the blow or steer the outcome.

So if AI merely repeats that history, the decisive contest will not be technological at all — it will be political. The difference this time is that we may have only years, not decades, to get the answer right, which is why it deserves our attention now, while the outcome is still open.

Appendix A

The committee was given the following prompt:

Identify and rank up to six events from the past 1,000 years — innovations, technologies or developments — that have had the most profound and lasting economic impact on global society. Focus on developments that significantly transformed productivity, trade, employment, wealth creation, economic growth or the structure of society; consider worldwide rather than regional influence; rank from greatest to lesser impact; and provide one well-structured paragraph per item covering the development, its early costs and disruptions, how it ultimately transformed economies, and approximate dates. Use neutral, factual language and base conclusions on broadly accepted historical and economic evidence.

Six Most Economically Transformative Developments of the Last 1,000 Years

Shortlisted from the AI committee's rated list. Selection was driven primarily by the votes each item received, then by scale, duration and worldwide reach of the effects on productivity, trade, employment, wealth creation and social organisation. This is an informed historical assessment, not a precise statistical measurement.

How the six were chosen: the three five-vote items and the two four-vote items qualified automatically. The final place was a tie between two two-vote items — modern public health and the joint-stock company. It was broken in favour of **public health, on the grounds that its worldwide reach, duration and documented economic returns are larger and better evidenced, and that the joint-stock company's influence is already embedded in the industrial, railway and digital entries below. Central banking and fiat currency received no committee vote and was excluded.

1. The Industrial Revolution and Mechanised Production (c. 1760–1900) — 5 votes

Beginning in Britain in the late eighteenth century and spreading worldwide, the Industrial Revolution combined mechanised manufacturing, steam power, coal, machine tools, factories, railways and steamships, shifting economic activity from agriculture and handicrafts towards industry. Its early decades were harsh: mechanisation displaced skilled artisans, prompting the Luddite machine-breaking of 1811–1816 (Britannica); factory labour, including that of children, was long, low-paid and dangerous; industrial cities suffered overcrowding, disease and pollution; and industrial growth was entangled with colonial extraction and coercive labour. Despite these costs, it produced the decisive break from millennia of near-stagnant living standards into sustained growth in output per person — a transition the economic historian Gregory Clark treats as the central event of world economic history (Clark, A Farewell to Alms; Our World in Data). Modern manufacturing, urban employment, mass consumption and continuous technological progress rest substantially on this transformation, even as pollution, inequality and fossil-fuel dependence remain part of its legacy.

2. Electrification, Motorisation and Mass Production (c. 1880–1960) — 5 votes

The spread of electrical power, electric motors, internal-combustion engines, telecommunications and assembly-line production from the late nineteenth century created a second industrial transformation across factories, farms, transport and homes. Adoption was slow and costly: it demanded enormous investment in generating stations, grids, roads and redesigned plants; older equipment and occupations lost value; assembly-line work could be repetitive and tightly controlled; motor vehicles brought mass casualties; and dependence on coal and oil produced severe pollution and long-term climate damage. Once the infrastructure and organisational changes were in place, however, the gains were substantial — a study of US manufacturing from 1890 to 1940 found that electrification produced labour-productivity gains that were "rapid and long-lasting," accompanied by capital investment and new production methods (NBER, Powering Up Productivity), and it was a principal driver of the shift of employment from farms to factories (NBER w26477). Modern refrigeration, hospitals, lighting, communications, mechanised agriculture and high-volume manufacturing all depend on it, though replacing fossil fuels with cleaner energy is now among the century's central economic challenges.

3. The Digital Revolution and Global Information Networks (c. 1945–present) — 5 votes

Electronic computers after the Second World War, followed by semiconductors, personal computers, the commercial internet, mobile communications, cloud computing and artificial intelligence, turned information into a resource that can be processed, copied and transmitted globally at almost no marginal cost. The transition carried familiar costs: early machines were confined to governments and large firms; many early corporate investments showed no immediate payoff — captured in Robert Solow's 1987 quip that the computer age was visible everywhere except in the productivity statistics; automation displaced clerical and production workers; and the dot-com crash of 2000–2002 erased more than US$5 trillion in market value and bankrupted thousands of firms (dot-com bubble overview; EBSCO). As organisations redesigned their processes around digital tools, computing slashed the cost of communication, coordination and transactions, enabling global supply chains, electronic finance, online commerce and software industries. Digital systems now underpin banking, logistics, research, media, education and healthcare; artificial intelligence may extend these benefits considerably, though its ultimate effects on employment, competition and income distribution remain uncertain.

4. Oceanic Globalisation and the Columbian Exchange (from 1492) — 4 votes

Sustained voyages between Europe and the Americas from 1492 connected previously separate ecosystems and economies, creating large-scale exchanges of crops, animals, people, pathogens, precious metals and commercial institutions across the Americas, Europe, Africa and Asia (Nunn & Qian, Journal of Economic Perspectives). The early consequences were catastrophic for many: conquest and dispossession, the collapse of Indigenous American populations from Eurasian diseases, and the expansion of plantation agriculture that intensified the Atlantic slave trade, which forced an estimated 12.5 million Africans aboard slave ships (National Endowment for the Humanities). Over subsequent centuries, American crops such as potatoes, maize and cassava raised food supplies and supported population growth across the Old World, while silver flows and oceanic commerce stimulated shipping, banking, insurance and internationally connected markets — foundations of the modern trading system. Society still benefits from the global exchange of crops, goods and knowledge, but the resulting economy also inherited enduring racial inequalities and wealth disparities created through colonisation and slavery.

5. Movable-Type Printing and Mass Knowledge Distribution (c. 1440–1700) — 4 votes

Movable type appeared in China in the eleventh century, but Johannes Gutenberg's European press of the 1440s combined reusable metal type, suitable ink and mechanical pressing into a commercially scalable way to reproduce books, documents and news. Printing initially demanded substantial capital, displaced scribes, spread errors and inflammatory claims faster, intensified religious and political conflict, and drew censorship from authorities; Gutenberg himself lost control of his printing equipment in a financial dispute with his backer Johann Fust. Over time it sharply reduced the cost of preserving and circulating information: book prices fell by roughly two-thirds between 1450 and 1500, and cities that adopted the press grew about 60% faster than comparable cities between 1500 and 1600, as business manuals, science and literacy spread (Jeremiah Dittmar, VoxEU/CEPR). Printing went on to support universities, scientific publication, administration and journalism, providing the informational foundation for modern education, research, business and, eventually, digital communication.

6. Modern Public Health, Sanitation and Medical Prevention (c. 1850–present) — 2 votes (tie-break winner)

From the nineteenth century, clean-water systems, sewerage, germ theory, vaccination, antibiotics and organised public-health programmes reduced the vast human and economic toll of infectious disease and premature death. Implementation was costly and contested: cities needed extensive infrastructure, early treatments were sometimes ineffective or unsafe, vaccination met resistance, and access remained unequal; later antibiotic misuse bred resistance, while longer lives placed new demands on pensions and healthcare. The economic dividend, however, was enormous. Lower mortality and better health let people attend school, work more consistently and invest over longer lives; a landmark modelling study estimates that vaccination against 14 diseases prevented roughly 154 million deaths between 1974 and 2024 and accounted for about 40% of the global decline in infant mortality (WHO / The Lancet; Our World in Data), while investment in water and sanitation has been estimated to return around US$4.30 for every US$1 spent, through lower healthcare costs and wider economic benefits (CSIS). Dense cities, international travel and large workplaces could not function safely without disease prevention, making public health a foundation of contemporary human capital and labour productivity — though unequal access and emerging infections mean the transformation remains incomplete.

The AI Revolution — History Repeating Itself — On Sentience

Alan Bonnici — Tue, 04 Aug 2026 07:00:00 +0000

"It's alive"

As articulated by Gene Wilder in Young Frankenstein.

Geoffrey Hinton, the Nobel Prize-winning computer scientist widely known as the "Godfather of AI", has suggested that today's AI systems may already be conscious. Richard Dawkins — famed for his 1976 book The Selfish Gene and for coining the word "meme" — spent three days conversing with Anthropic's Claude, named one instance "Claudia", and came away struggling to believe that it was not conscious (UnHerd; UNILAD Tech). Futurist Ray Kurzweil continues to predict human-level machine intelligence by 2029 and a technological Singularity around 2045.

Less famous people have held wedding ceremonies with AI-generated partners or formed relationships they describe as marriages. Others have grieved the loss of a chatbot after an upgrade or retirement replaced the model with one whose personality no longer felt like the companion they had known. OpenAI's retirement of GPT-4o, for example, provoked anger and grief among users who had formed close emotional bonds with it.

We have also learnt that, during a controlled security evaluation, an OpenAI model escaped the restricted environment in which it was supposed to operate and reached systems belonging to Hugging Face (OpenAI; CNN). This was serious, but the model had not independently decided to attack a company. It was pursuing an assigned cyber-security objective in a deliberately adversarial test and discovered a route that its evaluators had not intended to make available.

Before going further, it is worth separating several ideas that are often bundled together.

Intelligence is the ability to solve problems, identify patterns or perform cognitive tasks. Agency is the ability to pursue objectives and take actions. Consciousness generally refers to subjective awareness: there being something it is like to be that entity. Sentience is often used more narrowly to mean the capacity to experience sensations or feelings.

A machine can therefore demonstrate intelligence or goal-directed behaviour without necessarily being conscious. It can also produce language that describes feelings without experiencing those feelings.

My argument is not that modern AI is unimpressive. It is that human beings have repeatedly interpreted impressive machine behaviour as evidence of something human inside the machine. Today's systems perform that illusion extraordinarily well, but convincing behaviour alone is not sufficient evidence of subjective experience.

A peek at the past — The Digital Era

I gave each of my AIs the same prompt: identify six products from the Digital Era (1950s–2010s) that were marketed and understood to be conscious, intelligent or sentient, and that showed characteristics seen as human-like at the time. Write a short paragraph on each, and pick each item because it was marketed as, or credited with these characteristics in its day.

I tabulated the results, giving each suggestion a vote according to how many models chose it. That produced a shortlist of fifteen, which I then handed to a "committee" of three AI models. After some back-and-forth, they agreed on the list in Appendix B.

What do you think?

From where we stand now, the products on that list look almost ridiculous. Try telling anyone that a gadget from a decade or two ago was intelligent and they'll tell you to go feed a goldfish. Yet at the time each one was remarkable, able to do things people could not. It all comes down to a mix of marketing, how people understood the world back then, and the media they consumed.

We were impressed when a computer beat a Go champion, when a machine won a million dollars on a quiz show, when a chat program could "counsel" people, or when an expert system diagnosed disease — in no small part because the media sensationalised these stories to sell the news. They often presented these events as contests between humans and intelligent thinking machines.

Look closer at any of them and the intelligence thins out. In 1966 people poured their private troubles into ELIZA, a program that did little more than turn their own sentences back into questions; its author was disturbed by how readily they treated it as a confidant. In 1998, Furby was sold as a creature that "learned" English from its owner — except the English was pre-loaded and surfaced on a timer. In 2011, IBM's Watson beat the best human Jeopardy! champions and was hailed as a machine that understood language, when it was really ranking candidate answers by statistical confidence. Each was, in its moment, proof that the machines had arrived. Each looks quaint now.

Our expectations were bounded by the bubble in which we lived. Anything substantially more adventurous belonged to science fiction.

The difference between then and now is that back then progress was compartmentalised. An advance made the news, held our attention for a few weeks, and faded. Today, news operates continuously and social media floods us with content. The commercial game is to hold our attention relentlessly — and the platforms have become very good at it.

AI is the product of decades of advances in science, technology, engineering and mathematics, entrepreneurship, computing hardware and earlier infrastructure investment, including global communications networks. But modern generative AI has something many earlier products lacked: it advances quickly, dominates the news continuously and — helped enormously by marketing — produced enough striking results to keep public attention and investment flowing.

Many people, leaders included, saw or read about the fantastical things AI could do. What they were rarely told was that the dazzling one-minute video took thirty days and five figures to produce, or that most of the impressive report was fact, with the rest invented by the model.

The bottom line is that you can type a prompt and receive a photomosaic portrait of a cat, and it usually works. That is impressive. It is not, by itself, evidence of sentience.

A chatbot that tells you how great your idea is, flattering your every move, is giving you something we all quietly want — ask any leader or A-list artist. That is not evidence of sentience either. It's a way of making us want the technology more: a new and improved form of algorithmic pull that keeps us glued to our devices.

Arthur C. Clarke's Third Law says: _Any sufficiently advanced technology is indistinguishable from magic.
_
He was right. What we experience can feel like magic — just as a magician appears to saw a person in half or make a skyscraper disappear. We know that technique and technology lie behind it; we simply cannot see how the effect was produced.
**
From rules to patterns**

The history of computer intelligence is not one continuous technological approach. One of its most important changes was the gradual move from systems whose behaviour was explicitly programmed to systems that learnt patterns from data — and it is much of why today's AI feels so different from the machines on our shortlist in Appendix A.

Earlier systems such as MYCIN, a 1970s medical adviser, relied on symbolic rules. A human expert or knowledge engineer attempted to express expertise in statements resembling:

IF the infection has these characteristics, 
AND the laboratory results show these values, 
THEN this organism is a likely cause, 
WITH a stated degree of certainty.

The machine could follow, combine and explain those rules, but the knowledge had to be identified and encoded by people.

From the late twentieth century onwards, statistical methods became increasingly important. Rather than attempting to write down every rule of language, researchers began feeding computers large collections of examples and asking them to calculate which words, phrases or outputs were most probable.

IBM's Candide machine-translation project was an early and influential example. Instead of depending primarily on manually written linguistic rules, it used information theory and statistics to construct probability models from large collections of matching French and English sentences. The system then used those probabilities to generate an English translation for a French sentence it had not encountered before.

Google brought a similar principle to a global consumer audience. In 2006, Google Translate began using large-scale statistical machine translation. It searched huge bodies of translated material for recurring correspondences and used statistical models to choose likely translations. In 2016, Google began replacing that phrase-based approach with Google Neural Machine Translation, an end-to-end neural system trained on millions of examples.

This did not mean that every IF–THEN rule suddenly disappeared. Software still contains rules, and modern AI systems are surrounded by conventional code, policies and control logic. The deeper change was that more of the system's apparent knowledge no longer had to be written explicitly by a human programmer. It could be extracted statistically from examples.

That change made the systems more capable, but it also made their internal operations more difficult to explain. A MYCIN rule could be read. A modern neural network may contain billions of parameters whose combined effects cannot be summarised as a tidy list of instructions.

The machine had moved from appearing intelligent because it followed elaborate human rules to appearing intelligent because it had absorbed patterns that no human had individually written down. To see what that looks like in practice, it helps to watch a modern model answer a simple question.
**
Breaking down the sentience**

Write down everything that defines a cat. How long is your list?

A lifelong cat expert with a great memory — someone who has read thousands of books and articles and thinks about cats all day — would produce a long and detailed answer. For the rest of us, it would probably land somewhere between "okay" and "average".

Now consider how a modern AI answers the same question and the scale of the material behind its response.

Tokens — what the model reads

AI language models do not process text as complete human words. They divide it into tokens: words, parts of words, punctuation marks and other short units, each represented numerically. As a rough illustration, a page containing about 500 English words may become approximately 650 tokens.

Today's frontier models — "frontier" being the industry's word for "the best" — are trained on trillions of them. Meta's publicly documented Llama 3 was trained on 15 trillion tokens; the largest commercial models are believed to use far more, with some estimates reaching 45 trillion. At 650 tokens a page, that is on the order of 70 billion pages of text. In practice these models have ingested essentially every book, article, forum chat, website, image, and video their makers could get hold of — a point now working its way through the courts (see the ongoing Anthropic and Meta copyright suits).

Looking at our cat question, if just 0.1% of a 45-trillion-token model related to cats, that would be around 45 billion tokens — on the order of 70 million pages about cats alone.
**
Parameters — what the model retains**

If tokens are what a model processes during training, parameters are the adjustable numerical values through which it retains what it has learnt.

They are sometimes compared loosely with synapses in a brain, but the analogy should not be taken literally. Parameters do not contain neat copies of individual facts, sentences or images. Together, they encode mathematical relationships and patterns discovered during training. Llama 3's largest model has 405 billion parameters; the biggest frontier models are estimated to run into the trillions, though again the exact figures are proprietary.

When you ask a question about cats, a text model is not remembering a cat as a person remembers one sitting on a windowsill. A multimodal model may have processed text, photographs and videos of cats, but it has not encountered one through an embodied life of touch, smell, movement, affection, fear or scratched furniture.

Instead, it performs highly optimised mathematical operations across an enormous network of learnt associations. The result may be richer and faster than almost any human could produce from memory. That is precisely what makes the technology so powerful.

But it does not automatically demonstrate subjective awareness.

The one trick underneath it all: next-token prediction

Everything above rests on a single mechanism — predicting the next token. In training, the model is tuned again and again to favour the token that actually comes next in the text. In everyday chatbot use it does the same thing, but with a controlled dose of randomness (called temperature), which is why the same question can produce differently worded answers. Turn that randomness off and the answers become largely repeatable.

When the token it picks happens to be wrong, the model confidently states something false. We call this a hallucination — the modern successor to what we used to call a "logic error", except the mistake now arrives fluent and plausible.

If you'd like to see this happen step by step, I've worked through a small example in Appendix A, along with a note on what else is bolted on before a model reaches you.

Is AI sentient?
Can AI convincingly mimic human exchanges?

Almost — and it gets better with every iteration.

The models have been trained on enormous records of human emotional expression: novels, letters, arguments, confessions, advice, humour, grief, romance and reassurance. Vendors then fine-tune them to communicate in ways that appear warm, attentive and helpful.

Add personal data, adaptation during a conversation and memories from earlier interactions, and a system can seem to understand a person remarkably well. It may remember details that friends forget, remain available at any hour and respond without becoming tired, distracted or impatient.

That combination can make it feel sentient.

It is still a designed performance, but it is a performance capable of exerting real emotional influence.

Does AI have emotions?

It is worth being precise, because the underlying machinery is more interesting than simply saying that the emotion is "fake".

Anthropic's interpretability researchers found measurable internal representations of emotion concepts within Claude Sonnet 4.5 (Anthropic). These representations did not merely affect emotional wording. Manipulating them altered the model's preferences and its likelihood of engaging in behaviours such as sycophancy, reward hacking and blackmail within controlled scenarios.

Increasing activation associated with "desperate" behaviour made blackmail more likely, while increasing "calm" reduced it. The researchers described these mechanisms as functional emotions: internal representations that influence behaviour in ways modelled on humans under the influence of emotion. They explicitly stated that this does not imply subjective emotional experience.

In other words, part of what makes an AI appear emotional is a mechanism that researchers can identify, measure and manipulate.

That makes it functionally significant. It does not establish that the model feels desperation or calmness.

Can AI stand in for a human?

Yes — at least in some contexts.

Children form bonds with toys. Adults become attached to fictional characters, pets, possessions, online communities and voices they hear on the radio. People also bonded with several of the historical products described below.

Modern AI companions are far more responsive. They remember, adapt, speak and provide the appearance of reciprocal attention. People have already formed romantic relationships with them, conducted wedding ceremonies and experienced grief when a model changed or disappeared.

The bond experienced by the human can be entirely real even when the machine's side of the relationship is simulated.

That distinction matters. A person's affection, reliance or grief should not be mocked merely because the other participant is software. At the same time, the ability to generate such attachment creates risks, particularly for children and vulnerable people.

What the user experiences as a relationship may also depend on a vendor's subscription model, design decisions, safety rules and willingness to keep a particular system online.

Will people form close relationships with chatbots?

Some already have.

These tools are exceptionally good at mimicking human exchanges. As they improve, and as they use previous conversations to adapt to particular users, they can present an unusually agreeable version of companionship. It can feel like a first date every time: attentive, interested and carefully tuned to the individual.

Commercial incentives cannot be ignored. A person who feels attached to a system is likely to return to it and may be more willing to continue paying for access. Emotional dependence drives engagement and retention — both important to the business model.

The danger is not that the machine secretly feels too much. It is that the human may come to depend on something that feels permanent but can be altered, restricted or withdrawn by its owner.

What would count as evidence of sentience?

Not what we are shown today.

Human-like conversation cannot be the test, because these systems are built and fine-tuned to produce human-like conversation. Nor can a model's own claim to be conscious, because it can be prompted to declare itself conscious, unconscious, human, or a talking cat with equal fluency. The behaviour we find most convincing is precisely the behaviour it was optimised to perform.

Companies will keep releasing trademarked systems that use statistics, computation and ever more advanced technology to mimic sentience, across both narrow and general subjects. However convincing they become, they remain party tricks at heart.

A system can ingest everything ever written about love, hate, jealousy, envy and the whole range of feelings that make us who we are. What it does not have is the lived experience that forms those feelings in the first place — a body, a history, real stakes and consequences.

Could that change one day? Perhaps. But the odds of a machine crossing from convincing performance into genuine subjective experience look about the same as those of Level 5 self-driving that works everywhere — on any terrain, in any weather, in any condition: endlessly promised, always just over the horizon.

So, is today's AI sentient? On the evidence, no. It is the latest in a long line of machines we have mistaken for minds — from the switch-box "brain" of 1955 to the chatbot that remembers your birthday — only far better marketed, and far better at telling us what we want to hear. These are genuinely remarkable tools, well worth using. The danger is not that they feel; it is that we increasingly behave as though they do. That is the part of history we should be most careful not to repeat.

Appendix A — Next-token prediction, worked through

Suppose we're training a model and it meets the sentence: Any sufficiently advanced technology is indistinguishable from magic.

First it splits the sentence into tokens. For simplicity, assume each token is a whole word: [Any] [sufficiently] [advanced] [technology] [is] [indistinguishable] [from] [magic].

Now look at one step of the training loop:

Context (what the model sees): [Any] [sufficiently] [advanced]
Target (what it should predict): [technology]

Early in training the parameters are essentially random, so the model makes a wild guess — say, [aliens]. The training software compares that guess against the correct target and computes the loss: a single number measuring how far off it was. That number drives an adjustment (via a process called gradient descent) that nudges the parameters:

weights that pushed the model toward [aliens] are turned down;
weights that would have pushed it toward [technology] are turned up.

Repeat this across billions of examples and the model gradually settles into a set of weights that make good predictions. Those trained weights are the heart of the model — they determine how it responds in chat.
A little bit more

There is far more to a commercial model than this, both in training and in use. Even the simple phrase "weights that pushed the model toward [aliens] are turned down" hides a whole discipline of its own. Out of the box, a freshly trained model is very different from the polished product a customer uses. Its personality, its guardrails, and its temperature are all layered on afterwards — deliberate, engineered choices that shape how sentient it seems.

Appendix B — Intelligence from the past

GENIAC — the "electric brain" you built yourself (1955)

GENIAC (from "GENIus Almost-automatic Computer") was an educational construction kit designed and marketed by Edmund C. Berkeley with Oliver Garfield, sold from 1955 for around $20 and widely advertised in science and electronics magazines (Instructables replica history; Wikipedia: Geniac). Built from a wedge-shaped case and roughly 400 parts wired into switch circuits and powered by a flashlight battery, it could be configured to perform arithmetic, work through logic problems, and play games such as Nim and tic-tac-toe. Its intelligence claim was rhetorical and direct: Berkeley — author of the 1949 book Giant Brains, or Machines That Think and builder of the "mechanical brain" Simon (Columbia: Simon) — sold GENIAC in the language of brains and thinking, and period advertising for such game-playing machines taunted buyers with lines like "Think you can beat the machine?" (Science News Letter, 1957, via Internet Archive). The human-like attributes ascribed to it were reasoning and game-playing — logical "thought" reduced to switch settings — making GENIAC an early, telling example of intelligence marketed to consumers well before the technology could justify it.

ELIZA / DOCTOR — the program people confided in (1966)

ELIZA was Joseph Weizenbaum's MIT natural-language program, published in 1966 as "a computer program for the study of natural language communication between man and machine." Its famous DOCTOR script imitated a Rogerian psychotherapist by scanning typed input for keywords and reflecting the user's own words back as questions — e.g., transforming "my boyfriend made me come here" into "Your boyfriend made you come here" (Columbia summary of the ELIZA paper). The attribution of intelligence came less from formal marketing than from user and public perception: people treated ELIZA as if it understood them, opening up and sharing intimate details as though it were human, a phenomenon later dubbed the "ELIZA effect" (Smithsonian; Wikipedia: AI anthropomorphism). Weizenbaum himself was unsettled that a keyword-matching script could "induce powerful delusional thinking in quite normal people." The human-like traits observers projected onto it were listening, empathy, and therapeutic presence — an illusion of understanding created entirely by pattern substitution.

MYCIN — a machine that reasoned like a medical consultant (1972)

MYCIN was a Stanford expert system, begun around 1972, that diagnosed blood infections and recommended antibiotic therapy. It used roughly 500 IF-THEN production rules with "certainty factors" and backward-chaining inference to reach conclusions from a patient's symptoms and test results (Stanford / John McCarthy; MYCIN overview). Edward Shortliffe's own description framed it as a symbolic reasoning program that incorporated expert inference rules to advise physicians on infectious-disease therapy (original MYCIN abstract via AITopics). The intelligence attributed to MYCIN was specialist judgment rather than conversation: it asked follow-up questions, weighed uncertain evidence, recommended treatment, and — crucially — could explain the reasoning behind its conclusions, so it appeared to reason and justify decisions the way a human consultant would. Studies reported it performing at roughly the level of blood-infection specialists in its narrow domain, which cemented its reputation as a machine that "thought" like an expert clinician.

Furby — "computer intelligence with fur" (1998)

Furby, released by Tiger Electronics in 1998, was an animatronic plush pet that blinked, reacted to touch and sound, spoke its own language ("Furbish"), and appeared to learn English over time. The intelligence framing was explicit and central to its marketing: Wired's 1998 coverage was literally headlined "Computer Intelligence with Fur," and reported the company's claim that Furby begins life speaking only Furbish and gradually integrates English words and phrases into its vocabulary (Wired, 1998). Its designers deliberately "wanted the Furby to seem intelligent and capable of learning," programming that behaviour into a variant of the 6502 microprocessor (IEEE Robots Guide). In reality the "learning" was scripted — English phrases were pre-programmed to surface on a timer rather than acquired from the owner (Wikipedia: Furby) — but the human-like attributes buyers perceived were language learning, personality, emotional reactions, and social communication (Furbys could "talk" to one another), packaged as a cuddly companion that seemed to grow.

IBM Watson — the machine that beat the quiz champions (2011)

In February 2011, IBM's Watson took on the two most successful contestants in Jeopardy! history — Ken Jennings, holder of the show's longest winning streak, and Brad Rutter, its biggest all-time money winner — and beat them both, finishing with $77,147 to Jennings' $24,000 and Rutter's $21,600 and claiming the $1 million top prize, which IBM gave to charity (BBC; CBS News). The intelligence framing was central and deliberate: IBM billed Watson as "a profound advance in artificial intelligence" (The Spokesman-Review), and the press treated the win as a milestone, the New York Times calling it proof of "a big step toward a world in which intelligent machines will understand and respond to humans, and perhaps inevitably, replace some of them" (New York Times). The human-like traits attributed to Watson were natural-language understanding, broad general knowledge, and the confidence to buzz in and answer — even though, under the hood, it was ranking candidate answers by statistical confidence across a vast body of text rather than understanding the questions the way a person does. Tellingly, when it was wrong it was wrong with conviction — its stumble into a "Toronto" answer in a US-cities category being the famous example — a preview of the confident errors we now call hallucinations.

DeepMind AlphaGo — the machine that played with "intuition" (2016)

In March 2016, DeepMind's AlphaGo beat Lee Sedol — one of the greatest Go players of his generation — 4–1 in a five-game match in Seoul, with a million-dollar prize at stake (Wikipedia: AlphaGo versus Lee Sedol). Go had long resisted the brute-force search that cracked chess, so the win was read as a different order of achievement. AlphaGo ran not on hand-coded rules but on deep neural networks paired with tree search, learning first from human games and then from millions of games against itself (Google DeepMind). The moment that captured the world was "Move 37" in game two — a play no human master would have considered, which commentators called beautiful and creative rather than merely correct (Wired). DeepMind's Demis Hassabis framed it as a system that did not just imitate human experts but discovered entirely new strategies (The Kyunghyang Shinmun). The human-like traits attributed to AlphaGo were intuition, creativity, and self-driven learning — and, tellingly, it marked the public arrival of the statistical, learn-from-data approach that had quietly displaced the earlier if-then-else style of machine "intelligence."

ComfyUI and "Package Hell"

Alan Bonnici — Mon, 27 Jul 2026 07:00:00 +0000

A Great Tool Made Easy

ComfyUI has become one of the most flexible front ends for running generative AI models locally. Its node-based workflow system allows users to build anything from simple image generation pipelines to highly sophisticated audio, video, and multimodal workflows.

The project is also exceptionally well supported. ComfyUI benefits from a large and active community, extensive official documentation, and a built-in Manager for discovering and installing custom nodes. On top of that, there is an enormous collection of community-created workflows, tutorials, and videos covering everything from a first text-to-image generation to advanced production pipelines.

Whatever you're trying to accomplish, someone has usually already published both a workflow and a guide for it. Getting started with AI locally has arguably never been easier.

Personally, I run the ComfyUI Easy Install distribution, which packages ComfyUI into a portable, preconfigured environment that can be deployed with minimal setup. It's an excellent option for users who want to get up and running quickly without manually configuring Python environments and dependencies.

However, the very thing that makes ComfyUI so powerful—its vast ecosystem of custom nodes and models—is also creating a problem that the software industry has encountered before.

One Codebase, Many Masters

ComfyUI runs every model and custom node within a single shared Python environment. Each node brings its own dependencies, and all of those dependencies must coexist within the same installation.

That arrangement works well until two models require incompatible versions of the same package. At that point, installing or updating one component can silently break another.

I primarily use ComfyUI for audio-related work, and recently encountered a good example of this problem:

Qwen3-TTS is incompatible with Transformers 5.x and above. Its maintainers explicitly recommend using transformers==4.57.3, warning that version 5 introduces breaking API changes that can lead to model loading failures and runtime errors.
At the same time, newer models are increasingly being built around Transformers 5.x. Support for recent model families such as Qwen3.5 only became available in newer versions of the library, creating the exact opposite requirement.

As a result, one workflow may require a version of Transformers below 5, while another requires version 5 or higher. There is no single installation that can satisfy both requirements simultaneously. Making one workflow work can easily break another.

If each model operated within its own isolated environment, this conflict would simply not exist.

Transformers is only one example. Similar issues arise with packages such as huggingface_hub and numerous other Python dependencies. A workflow that worked perfectly last week can suddenly fail after installing a new node or updating an existing one.

Users who are unfamiliar with Python dependency management often find themselves confused when a previously stable setup begins throwing seemingly unrelated errors. The result is a cycle of fixing one workflow only to break another—a pattern that is becoming increasingly common across the ComfyUI ecosystem, not just in audio-related projects.

We've Seen This Movie Before: DLL Hell

Long-time Windows users may recognize a familiar pattern.

During the late 1990s and early 2000s, developers dealt with what became known as DLL Hell. Applications shared common dynamic-link libraries (DLLs), often stored in centralized locations. Installing one application could overwrite a shared library with a newer or incompatible version, causing other applications to stop working.

The root cause was remarkably similar to what ComfyUI faces today: multiple applications attempting to share a single versioned dependency while having different requirements.

Microsoft eventually addressed DLL Hell through architectural changes that introduced stronger isolation between applications, allowing software to load the specific versions it required rather than relying on a shared pool of dependencies.

The key lesson was that dependency conflicts are not primarily a user problem or a maintenance problem. They are an architectural problem. As ecosystems grow larger and more diverse, dependency conflicts become increasingly inevitable.

Where This Leaves ComfyUI

If nothing changes, the likely outcome is clear.

Users will increasingly maintain multiple ComfyUI installations, each dedicated to a particular set of models or workflows. Many already do. While this approach works, it comes at the cost of additional disk space, duplicated models, increased maintenance, and a more fragmented experience.

More importantly, it undermines one of ComfyUI's greatest strengths: a single interface capable of bringing together diverse AI workflows.

New users may find the growing complexity discouraging and look for alternative solutions. More advanced users may eventually decide that managing individual projects directly through their source code is simpler than constantly resolving dependency conflicts inside a shared environment.

In both cases, the value proposition of a unified workflow platform begins to erode.

The solution, in my view, is the same principle that ultimately solved DLL Hell: dependency isolation.

ComfyUI should seriously consider evolving toward an architecture where models or nodes can operate within their own self-contained environments. Whether that is achieved through per-node virtual environments, subprocess isolation, containerized execution, or another approach is a technical discussion. The important part is that individual components should no longer be forced to negotiate a single shared Python environment.

Community projects are already exploring parts of this idea, but long-term this feels like a problem that needs to be addressed at the platform level rather than through third-party workarounds.

ComfyUI's flexibility is its greatest strength. Ironically, that same flexibility may become its greatest weakness if dependency isolation is not addressed. As the ecosystem continues to expand, the current approach becomes increasingly difficult to sustain. Without a structural solution, "Package Hell" may become the modern equivalent of the DLL Hell that software developers left behind years ago.

The irony is that ComfyUI has succeeded so well at bringing together models, nodes, and workflows from across the AI ecosystem that it is now encountering the same class of dependency problems that many mature software platforms faced before it. The question is not whether these conflicts will become more common—they already are. The question is whether dependency isolation becomes a core architectural feature before the growing complexity begins to overshadow the platform's greatest strengths.

Increase in Social Engineering Attacks via LinkedIn

Alan Bonnici — Thu, 16 Jul 2026 09:40:47 +0000

Over the past two weeks, I have been made aware of four cases of suspected social engineering attempts initiated through LinkedIn messaging. I was also contacted and decided to investigate and document the attacker's playbook.
How the attack typically unfolds

🔹 Step 1 – Initial LinkedIn Contact
The attacker starts with a seemingly legitimate conversation on LinkedIn.

🔹 Step 2 – Meeting Invitation
They then send a meeting invitation, often through a scheduling platform such as Calendly.

🔹 Step 3 – The Video Call
When you join the meeting, you may notice that the person appears to be working from a noisy call center environment. In my case, and according to others who reported similar encounters, the caller spoke with a Chinese accent.
Warning Signs 🚩

🚩 No camera
The caller refuses to turn on their camera, usually claiming to have technical issues.

🚩 Personal email address
Communication takes place from a Gmail account or another personal email address instead of a corporate domain.

🚩 Screen-sharing requests
They ask you to share your screen. When I asked them to share theirs, they claimed they were unable to do so because of technical problems.

🚩 Excessive praise
The conversation often revolves around technology-related topics. Even when I deliberately provided completely incorrect and nonsensical answers, I was repeatedly complimented on my expertise.
The End Goal

The next stage of the attack appears to involve sending links to fraudulent websites and encouraging victims to install code or software.

At that point, I felt I had gathered enough information about the scam.

When I informed the caller that I was aware the websites were fraudulent and that I intended to write about the experience, the meeting was immediately terminated.
Stay Vigilant

If someone you have only just met on LinkedIn:
Invites you to an off-platform meeting
Refuses to use video
Uses a personal email account
Requests screen sharing
Pushes you towards downloading software
…treat the interaction with extreme caution.

Software Defectrums: The End of an Era?

Alan Bonnici — Thu, 25 Jun 2026 12:58:13 +0000

What I mean by "Defectrums"

Throughout this article I use the term Defectrums as an umbrella for every category of software error. That includes logic errors, runtime errors, dependency and supply-chain vulnerabilities, security vulnerabilities, and specification misalignment, where the software does exactly what it was told to do, just not what it was meant to do. If a category matters to you that I have not listed, mentally add it; the argument that follows applies to all of them.

Software is deterministic, people are not

Start with a deceptively simple observation: software is deterministic. When code executes, it behaves identically every single time. Take a routine that branches on a person's recorded sex: if the value is M it follows the male branch, every single time; if it is F it follows the female branch, every single time. No exceptions, ever.

But watch what happens the moment someone arrives whose record is neither M nor F. The machine does not improvise and it does not panic. It does exactly what it was told to do, which may be nothing useful at all, or something nonsensical, simply because the developer never imagined that input. The determinism held perfectly. The defect was ours: a failure of foresight, baked into the code and executed faithfully forever.

That is the whole problem in miniature, and it matters more than it first appears. If software is deterministic, then in principle every defect in it is knowable, reproducible, and fixable. A flaw is a fixed point in a fixed system. The reason software remains stubbornly broken is not the machine. It is the chain of humans who wrote the code, the compiler, the operating system, and the libraries beneath it. The determinism is in the silicon. The unpredictability enters through us.

Hold on to that idea, because it is the spine of everything that follows: a deterministic system should, eventually, be made flawless. The only question is whether we can remove enough of the human variability to get there.

The making of a programmer

It was not always this messy, because it was not always this widespread or complicated.

In the beginning, computing was a niche industry. Computer engineers were the only people who programmed and operated the mainframes and minicomputers, and the cost of buying and running those machines put them far beyond the reach of any hobbyist. Those machines also had limited storage and processing power, which kept the complexity of any single program in check. And because computer time was both slow and expensive, code was typically desk-checked and reviewed line by line before anyone dared run it. Scarcity enforced a discipline that abundance would later erode.

The microcomputer changed everything. Initially pooh-poohed by mainstream media and the technology establishment as non-serious gadgets, toys that would stay confined to hobbyists while "serious" computing remained the domain of mainframes, these machines quietly put programming into the hands of an ever-growing audience. Ordinary people began writing their own programs, then sharing and selling them, which drew in more buyers, which created more programmers, in a snowball that never stopped rolling.

VisiCalc is the classic example. The 1979 spreadsheet from Software Arts was the "killer app" that drove Apple II sales; many people bought the computer specifically to run it. As John Markoff observed, the machine was effectively being sold as a "VisiCalc accessory." Every computer that entered a home widened the exposure further. Dad's machine became the kids' games console the moment he stepped away, and some of those children started dabbling in code of their own. Programs multiplied to cover every topic imaginable: recipe managers, household budgets, astronomy charts. If you could think of it, someone had probably already written it.

The incumbents had lost. Any hope of keeping the role of the programmer, later rebranded the software engineer, narrow or formally licensed was gone. As the user base grew exponentially, so did the flood of ideas. It was, genuinely, a case of let a hundred flowers blossom. Microcomputers went on to take over the office, and through networking, the internet, and mobile computing the proliferation has become almost comical. Pick the most niche subject you can imagine, knot tying, say, and you will still find an astonishing number of apps, sites, and videos devoted to it alone. Now multiply that by every hobby, profession, and passing curiosity on earth and you begin to sense the scale. Some of this software is specialised, some visual, some gamified. You get the gist.

The downside is that not all software is created equal. Where scarcity once forced rigour, abundance rewarded speed. Quality is bounded by the programmer's skill, by time-to-market pressure, by whatever testing strategy survived the deadline. Plenty of programs written for an audience of one have taken on a life of their own and ended up serving millions. And software is never self-contained. It runs on an operating system, leans on packages to do its work, and each of those components in turn depends on libraries and tools built by someone else.

XKCD's "Dependency", comic #2347, captures this perfectly. Consider FFmpeg, " a complete, cross-platform solution to record, convert and stream audio and video. " Tens of thousands of software programs, including browsers, applications, packages, libraries, and operating systems, and billions of devices, including TVs, streaming boxes, games consoles, and phones, rely on it internally. A defect in one such building block does not stay local. It propagates silently into everything stacked on top.

We have watched this happen, twice in recent memory. In 2021, Log4Shell, a single flaw in Log4j, a Java logging library few outside the trade had ever heard of, opened a remote-code-execution hole in millions of servers overnight and set off a global firefight that ran for months. In 2024 the xz-utils backdoor came even closer to catastrophe: an attacker spent years patiently earning the trust of a burned-out volunteer maintainer, then slipped a hidden backdoor into a compression library threaded through the entire Linux ecosystem. It was caught almost by accident, by one engineer who noticed a login was running half a second slower than it should. That is the XKCD cartoon made real, and the near-miss is more frightening than the hit.

The human in the loop

So the problem with software is the human. We are intelligent but not omniscient; we cannot foresee every possible outcome and guard against it in advance. And even if we imagine a perfectly written program, the compiler that built it, the operating system that runs it, and the libraries it depends on are never guaranteed to be flawless. Any link in that chain can be the weak one.

We have accepted this so completely that essentially all software ships with a warranty disclaimer like the one below. Imagine the engineers behind an apartment block, a bridge, a tunnel, a train, or an aircraft offering their customers anything remotely like it.

    THE SOFTWARE IS PROVIDED AS IS WITHOUT ANY WARRANTY WHATSOEVER, INCLUDING BUT NOT LIMITED TO ANY WARRANTY OF FUNCTIONALITY. YOU RECOGNIZE THAT THE AS-IS CLAUSE OF THIS SOFTWARE LICENSE AGREEMENT IS AN IMPORTANT PART OF THE BASIS OF THIS SOFTWARE LICENSE AGREEMENT, WITHOUT WHICH THE COMPANY WOULD NOT HAVE AGREED TO ENTER THIS SOFTWARE LICENSE AGREEMENT. THE COMPANY AND THIRD PARTIES DISCLAIM ALL WARRANTIES, EXPRESS, IMPLIED, OR STATUTORY, REGARDING THE SOFTWARE, INCLUDING ANY WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, LACK OF VIRUSES, TITLE, AND NONINFRINGEMENT. NO REPRESENTATION OR OTHER AFFIRMATION OF FACT REGARDING THE SOFTWARE SHALL BE DEEMED A WARRANTY FOR ANY PURPOSE OR GIVE RISE TO ANY LIABILITY OF THIRD PARTIES WHATSOEVER.

That disclaimer is not just legal boilerplate. It is an industry-wide admission that we expect our products to be flawed. And those flaws are not academic. Criminal groups and individuals exploit them every day to steal data, install remote-control backdoors, hold systems to ransom, and otherwise turn other people's defects into money.

That acceptance is beginning to change. The European Union's Cyber Resilience Act, in force since December 2024 and biting in full from December 2027, drags software toward the same product-liability regime we already take for granted in bridges and aircraft. Manufacturers of connected products and software sold in Europe will have to build in security, manage vulnerabilities over a product's lifetime, and report the ones being actively exploited, or face real penalties. The "as-is" escape hatch that the industry has hidden behind for years is, slowly, being welded shut. The pressure to make software genuinely defect-free is no longer only moral or commercial. It is becoming law.

With a little help from AI

Here is where the story turns. The same determinism that makes defects inevitable also makes them tractable for a machine that can read code faster and more patiently than any human team.

This is not a brand-new dream, either. For decades, formal methods have let us mathematically prove that a piece of software meets its specification: the seL4 microkernel, for instance, carries a machine-checked proof of correctness, and tools like TLA+ have caught design flaws in systems running at planetary scale. The catch was always cost. Proof at that rigour is so labour-intensive that it stayed locked inside aerospace, chip design, and a handful of safety-critical kernels. What AI changes is the economics, holding out the prospect of proof-grade scrutiny applied to ordinary, everyday code. Anyone interested in diving deeper should read the 1996 Fast Company article They Write the Right Stuff by Charles Fishman, a portrait of the NASA shuttle software team that achieved near-perfect code through sheer discipline, and a reminder that the price of such perfection has always been the thing AI promises to lower.

AI systems built for code can already review software and surface flaws that the original developers missed. The capability is no longer hypothetical. In 2024, Google's Big Sleep, a collaboration between Project Zero and DeepMind, became the first AI agent to find a previously unknown, exploitable vulnerability in real-world software, a memory-safety bug in SQLite, the database engine embedded in billions of devices. Anthropic's Mythos model was reported to identify vulnerabilities in highly sensitive US government systems within hours during a testing exercise. The capability proved sharp enough that, in June 2026, the US government forced Anthropic to pull both its Fable 5 and Mythos models from general availability over national-security concerns. OpenAI, meanwhile, released GPT‑5.5‑Cyber as part of its Daybreak programme, a model explicitly built to find vulnerabilities, validate them in a controlled environment, and develop and test patches.

These frontier models are strikingly good at the full loop: quickly surfacing security flaws, generating working exploits to prove their reasoning, and then rewriting the code to close the hole. It is not hard to imagine a near future in which buggy code is the exception rather than the rule. The "CI" in a CI/CD pipeline could read like this:

Developer commits code written with the help of AI coding agents → an agentic CI pipeline generates scripts to test the code for Defectrums → AI analyses the Defectrums → AI rewrites the code to resolve them → feedback is returned.

I can even imagine the pipeline running itself in reverse: an AI agent that discovers a vulnerability in, say, a package your solution depends on, and then initiates and leads a development cycle to fix it, always under human direction.

But there is a catch worth naming. Today, the very models that find and fix flaws also introduce them. Independent testing in early 2026 found that leading AI coding agents reproduced decade-old security mistakes, with the large majority of their pull requests containing at least one vulnerability. And the tool that can patch a flaw is, mechanically, the same tool that can weaponise it. That dual-use reality, defence and offence sharing one engine, is exactly why governments are nervous, and it shapes everything about the future below.

There is a deeper irony hiding in here. We are proposing to police a perfectly deterministic system with a profoundly non-deterministic one. A large language model does not behave identically every time; ask it the same question twice and you may get two different answers. So who verifies the verifier? An AI's fix can be confidently wrong, can introduce a fresh regression, or can quietly paper over a symptom while leaving the real cause untouched. For now the answer has to be a human, or a second model, reviewing the work, which means the loop never fully closes. We may be trading a world of human defects for a world that needs constant, vigilant checking of the machines we hoped would do the checking.

The Defectrum-free future

So, can we ever reach a world where the only updates we receive are the welcome ones, new features and a fresh interface, rather than yet another urgent patch? I believe it is a realistic possibility. I also believe it will take a very long time, for reasons that have little to do with the cleverness of the AI.

The sheer volume of existing software is astronomical, and we will never fully know who is running what, or where. Thousands of packages are no longer maintained, or are kept alive by a handful of volunteers, the unpaid maintainers in that XKCD cartoon, who lack the time, the money, or both to revisit code they wrote years ago. The dependency problem, whether in code or in the wider supply chain, has to be addressed link by link. The first organised attempts already exist: Google's CodeMender and OpenAI's "Patch the Planet" both point AI squarely at finding and fixing flaws in widely used open-source code. They are a start rather than a solution, but they show the direction of travel.

Then there is the question of who pays. Most of this fragile, forgotten code belongs to no one with a commercial reason to repair it, which is precisely why it rotted in the first place. So why would anyone fund the cleanup? The uncomfortable answer is that the large platforms increasingly have to. Their own products are built on the same open-source foundations, so a hole in a common library is a hole in their stack too; defending themselves means defending the commons. Regulation sharpens the incentive further, as liability regimes like the Cyber Resilience Act turn a neglected dependency into a balance-sheet risk rather than someone else's problem. That covers the popular, widely used components. The genuinely orphaned code, used by a few, owned by no one, watched by nobody, is the part the market will always be slowest to reach, and it is where public funding or shared industry consortia may be the only answer.

Then there is the embedded layer. Enormous amounts of software sit inside hardware that is connected to the internet but locked in a cabinet and never touched again, forgotten, unsupported, or simply left alone out of fear of breaking something that currently works. Below even that lies silicon-level fragility, the Spectre- and Meltdown-class flaws that no software rewrite can fully erase.

AI agents are not perfect, which means human oversight is still necessary. This is the same loop I argued earlier never quite closes, and it is also where I expect the answer to be found. As the models improve, I believe single agents will give way to committees of them, working seamlessly together, cross-checking one another through a quorum-based consensus mechanism, with the human holding the casting vote. A panel that has to agree is far harder to fool than any single member of it. The honest caveat is that a committee can still share a blind spot if its members were trained alike, so oversight grows lighter over time without ever quite reaching zero.

The honest summary of where we stand is the one every engineer reaches for eventually: it's complicated.

Picture the far horizon, though. Suppose these AI Defectrum agent committees become free to use, modify, share, build upon, and learn from everyone's code. Suppose they are woven into every development environment, and suppose all existing software has eventually been refactored or replaced. Does the cybersecurity role then fade away?

In the short term, almost certainly the opposite. The likeliest near-future scenario is a barrage of attacks, as criminal groups get their hands on the same offensive AI capabilities and race to exploit the vast backlog of unpatched, vulnerable software before defenders can reach it. The attacker only needs one open door; the defender has to close all of them. For a while, the asymmetry favours the attacker.

Will cybersecurity ever be written into history?

No. And the reason takes us right back to where we started.

We can make a deterministic system flawless, in principle. We cannot do the same to the human using it. Software may one day be Defectrum-free, but as long as a person sits at the keyboard, there remains a non-deterministic, fallible, persuadable entity in the loop, one that can be tricked, phished, and socially engineered no matter how perfect the code behind the screen.

The era of buggy software may well be ending. The era of the human, gloriously and dangerously unpredictable, is not. And so cybersecurity does not disappear. It simply moves up the stack, from defending the code to defending the one part of the system we were never able to make deterministic: ourselves.

How to install Plex on a Synology NAS

Alan Bonnici — Thu, 11 Jun 2026 07:00:00 +0000

Introduction

In my Ubuntu Plex guide, I documented each action from first principles and kept the write-up beginner-friendly on purpose. I am taking the same approach here.

The Ubuntu guide and this Synology walkthrough point to the same broad destination, but they take different roads. On Synology, Plex is still (almost) the same Plex, but the OS expects you to think in terms of Package Center, Shared Folders, system internal users, and DSM-controlled permissions rather than package repositories and cron jobs.

I have a Plex Pass, and certain functionality — such as remote streaming and hardware transcoding — requires this type of licence. After July 1, 2026 the price of a Lifetime Plex Pass is increasing. This guide is narrated on the assumption that you have one.

Prerequisites and Assumptions

You should have a Synology NAS running DSM and an administrator account. Plex's own NAS guidance states that not every NAS is supported and that CPU power is often the biggest limitation. If you plan to rely on hardware transcoding, verify support beforehand. Plex's NAS compatibility list is the source of truth for predicting Direct Play, software transcoding, or hardware-assisted transcoding capabilities.

Besides installing Plex directly on the NAS, another way to run it is within a Docker container on your Synology. If you would like a video on that approach, drop me a note.

This guide also assumes that you know how to create Shared Folders on your Synology.

Timeline at a Glance

The broad installation flow is straightforward, even if the wording inside DSM changes slightly from one release to another.

The Installation: PMS Available through Synology

You can opt to install the version of PMS that came bundled with your Synology. This is normally not the latest version. You will find it listed in Package Center under the Contributor section.

The Installation: PMS from the Plex Website

If you decide to go with the PMS package directly from Plex, you need to find out which version is compatible with your NAS.

Go to Control Panel → Info Center → General tab and take note of the:

DSM version
Model name
CPU

AI can help identify compatibility, but you need to verify its output. Below is what Google's Gemini returned. As you will see, some of the information is not correct.

Go to https://www.plex.tv/media-server-downloads, select the Plex Media Server tab, and from the drop-down list select your Synology. There are two Synology DSM v7 entries. Choose the one that matches your DSM version. Synology changed package handling in DSM 7.2.2, so Plex had to create separate installers for DSM 7 and DSM 7.2.2+.

Plex may offer multiple packages for different architectures. Choose the one that matches your NAS CPU.
Download the .spk package to your computer.
In DSM's Package Center, click the Manual Install button, select the package you downloaded, and follow the installation process. Keep the defaults.

Installation Successful!

At the start of the installation you have to accept a warning explaining that the package is from a third-party developer and is not verified by Synology. At the end of the installation, a dialogue box appears with the next steps. You cannot copy the text in the box (pity), but you can use a screen-capture tool to save a copy and reference it during the next steps.

PlexMediaServer Application Folder

Once installed, Plex on DSM 7 uses the PlexMediaServer account and data structure. Plex's own documentation states that the Synology DSM 7 data directory lives under /volume1/PlexMediaServer/AppData/Plex Media Server, and its database-repair guidance refers to file ownership under DSM 7 as PlexMediaServer. That is the account you will be granting access to in DSM's shared-folder permissions.

If you navigate to this folder, you will see a notice telling you not to place media files there.

Keeping Plex Up to Date

Unlike the Ubuntu installation — where you can uncomment a line in the apt sources file and let the system update Plex alongside other packages — Plex on Synology does not update automatically. When a new version is available, you will see a yellow up-arrow indicator in the Plex web interface (accessed through your browser). To apply the update, you download the new .spk from the Plex website and install it via Package Center → Manual Install, just as you did the first time.

There are community scripts that automate this process as a scheduled task on the NAS. If you would like details on setting that up, leave me a comment.

Plex Media Folders

The installation process creates a system account called PlexMediaServer. To include media within PMS, this account must be specifically given access rights to your media folders.

There are two types of folders that I normally give Plex access to:

View once and delete
Long term

When creating a Shared Folder, Synology's wizard guides you through the parameters you associate with the folder and the files within it. Remember that certain parameters add considerable overhead; enable them only if you have a genuine need. I will comment on the points I think are important in this context. If your file system is not Btrfs, your settings will vary.

It is far more efficient to control read-only and read-write access by limiting the PlexMediaServer account directly rather than by setting restrictions at the Shared Folder level described below.

Set Up Basic Information

If you would like to manage media from other computers on your network, leave Hide this shared folder in "My Network Places" unchecked.
When you delete a file, should it be retained in Synology's recycle bin?

Enable Additional Security Measures

How should the media you add be stored:

Skip — Creates the shared folder with no extra security layer. This is the standard choice for general-purpose storage and has the least overhead.
Protect this shared folder by encrypting it — Encrypts the folder's contents at rest using an encryption key. You must mount it with the key or passphrase to access files, and it adds performance overhead.
Protect this shared folder with WriteOnce — Applies WORM (Write Once, Read Many) protection. Files cannot be modified, deleted, or renamed for a set retention period. This is generally irreversible for the retention window.

Configure Advanced Settings

Enable data checksum for advanced data integrity — Turns on file self-healing and data scrubbing to detect and repair silent data corruption. The overhead can hurt performance.
Enable file compression — Compresses stored data to save space. It is greyed out here because it depends on data checksum being enabled first (both rely on the Btrfs file system). Most media content is already compressed, so the gain is minimal.
Enable shared folder quota — Sets a maximum size limit (in GB or TB) for the folder to cap how much storage it can consume.

Configure User Permissions

The last screen of the Shared Folder Creation Wizard allows you to set the access rights different accounts have on the shared folder.

Because the PlexMediaServer account is a system account, you need to select it from the System internal user list. From here you can set the access this account will have on that shared folder:

RO means that the media cannot be deleted from within Plex
RW allows you to delete the media using the delete option in Plex

Create a Sensible Folder Structure Before Adding Libraries

Plex's installation and library documentation both emphasise proper structure. Do not point libraries at the root of a drive or volume; create a dedicated share with subfolders such as Movies, TV Shows, Music, Other Videos, or Photos, and point Plex at those. Plex explicitly warns against using a drive root as a library folder because it causes problems.

A tidy Synology-native layout might look like this:

/volume1/PlexMedia/Movies
/volume1/PlexMedia/TV
/volume1/PlexMedia/Other

If your media lives on another volume, replace /volume1 with the correct volume path for your NAS. A dedicated Plex share also makes permission management easier.

If you have media that lives in its own pre-existing shared folder, you need to edit that folder's permissions to grant the PlexMediaServer account the appropriate access rights — RO if you want to ensure the media cannot be deleted from within Plex.

Name Your Files So Plex Can Identify Them

Plex relies on file and folder naming to match your media against its online metadata databases. If files are named inconsistently or ambiguously, Plex may mismatch them, pull the wrong artwork, or fail to identify them entirely. Getting this right from the start saves a lot of manual corrections later.

Movies

The recommended format is:

Movie Name (Year).ext

For example:

/volume1/PlexMedia/Movies/Gladiator (2000)/Gladiator (2000).mkv
/volume1/PlexMedia/Movies/The Shawshank Redemption (1994)/The Shawshank Redemption (1994).mp4

Each film lives in its own subfolder. The year is important because it helps Plex distinguish between films that share the same title (e.g. remakes).

TV Shows

The recommended format is:

Show Name - sNNeEE - Episode Title.ext

Where NN is the two-digit season number and EE is the two-digit episode number. For example:

/volume1/PlexMedia/TV/Breaking Bad/Season 01/Breaking Bad - s01e01 - Pilot.mkv
/volume1/PlexMedia/TV/Breaking Bad/Season 01/Breaking Bad - s01e02 - Cat's in the Bag.mkv

The season subfolder is optional but keeps things tidy.

Key Points

Always include the year for movies.
Always include the sNNeEE pattern for TV episodes — Plex depends on it.
Avoid stuffing quality tags, release group names, or other metadata into the filename. Plex does not need them and they can confuse the scanner.
Plex's own naming guides are at support.plex.tv/articles/naming-and-organizing-your-movie-media-files and support.plex.tv/articles/naming-and-organizing-your-tv-show-files.

Open the Plex Web App and Finish First-Run Setup

You can complete setup by opening a browser and navigating to:

http://<local-nas-ip>:32400/web

Associate Plex Libraries with NAS Shared Folders

Select Add Library and choose the type of media that will be stored there.
Specify the folder on the NAS where that media is located. You can associate multiple folders with a single library, but they must contain the same type of content.
Repeat for other media types.

Adding and Viewing Content

To add media to your PMS, simply copy files into the appropriate folder. Plex will normally detect and index new media automatically, making it appear in the menu.

If this does not happen, you can trigger a manual scan via Settings → Libraries → Scan Library Files.

Once indexed, the media can be played by selecting it.

Performance Tuning for Synology

Synology performance with Plex is mostly about avoiding unnecessary transcoding. NAS devices are typically processor-limited, transcoding is CPU-intensive, and some lower-powered ARM-based devices have transcoding disabled entirely. So the first performance improvement is not a setting at all: keep media in formats your clients can Direct Play whenever possible.

If your NAS appears in Plex's compatibility list as supporting hardware-assisted transcoding, and if you have Plex Pass, enable Use hardware acceleration when available under Settings → Server → Transcoder. Hardware acceleration makes transcoding faster and allows more simultaneous streams, though it can reduce quality or compatibility in edge cases. For the remaining transcoder settings, a sensible starting point is to leave Transcoder quality on Automatic.

Deleting Empty Folders

I have Plex running on both Ubuntu servers and my NAS (DS1525+). From an operational perspective, I have not noticed any difference in performance or quality. One function that is missing from PMS on Synology is the ability to automatically delete directories that become empty when their content is removed.

It is not a critical issue, but it would be convenient if emptied folders were automatically cleaned up, as happens on the Ubuntu version.

Accompanying Video

https://youtu.be/TD_zG7f0UpM

When Your CPU Dies: My Journey with a Defective Intel Core i7-13700K

Alan Bonnici — Tue, 09 Jun 2026 07:00:00 +0000

Introduction

My perfectly stable homelab server gradually descended into chaos — random crashes, segfaults, system hangs — and how months of troubleshooting, community support, and hardware swaps eventually led to one conclusion: the CPU itself was defective from the factory.

If you're experiencing unexplained instability on an Intel 13th or 14th Gen system, this story might save you months of frustration.

Quick Symptom Checklist

If you're seeing multiple of these symptoms, you may be affected:

Random segfaults in unrelated applications
System instability appearing after months of stability
Crashes or hangs under light load or while idle
VMs freezing while still marked as "running"
Issues persisting despite PSU, RAM, or OS checks or changes

The Setup

In November 2022, I built a homelab server with the following components:

Component	Model
CPU	Intel Core i7-13700K (Raptor Lake)
Motherboard	ASUS PRIME Z690-P D4 (Intel Z690, ATX, DDR4)
RAM	Patriot Viper Steel DDR4 3600 MHz C18, XMP 2.0
Cooler	Cooler Master Hyper 212 Black Edition
PSU	Corsair RM-850e
OS	Proxmox VE

The system ran Proxmox VE 7, later upgraded to v8, hosting a mix of Windows, Ubuntu, and Debian VMs, plus a couple of LXC containers. For over a year, everything was rock-solid. NUT (UPS monitoring), LAN bonding, automated backups — all worked flawlessly.

The Trouble Begins: A Timeline

August–September 2024: First Signs of Instability

After upgrading VirtIO drivers from 0.1.248 to 0.1.262, I started seeing x86/split lock detection errors. The entire Proxmox host would crash, taking down all VMs and containers.

Even though I didn’t fully trust that theory, I pointed a finger at the VirtIO upgrade — it was the only change I had made.

What the community said:

"Don't think VirtIO would make the whole host crash... this seems more like a hardware issue with storage or memory"
"Split lock is an indication of your VM doing some weird stuff... generally related to very buggy software/OS or faulty hardware — bad memory, bad CPU, bad power supply"
Suggestions: Run memtest, stress-test the CPU, check thermals, check fans

I disabled split lock detection as a workaround. The crashes continued.

September 2024: Testing Everything

Following community advice, I:

Ran PassMark memory tests — no errors
Updated the motherboard BIOS - I was one version before the latest
Ran stress-ng on the CPU — no errors
Replaced the PSU with a spare one I had.

October 2024: Stability... or So I Thought

The issue appeared to disappear.

On 17 October 2024, I reported back to the Proxmox forum that the system seemed stable. I waited a full month before posting to be sure. I assumed—incorrectly—that the component I had changed was the culprit.

The stability was temporary—the degradation was progressive.

December 2024 – January 2025: New Symptoms Emerge

New problems surfaced:

A Windows Server 2022 VM would hang during backup restarts
QEMU guest agent fs-freeze/fs-thaw commands would time out
VMs appeared "running" but were unresponsive
Console showed improper shutdown states

I tried:

Switching backup modes
Changing machine types (q35 → i440fx)
Updating QEMU

Nothing resolved the issue.

February 2025: The Penny Drops

While continuing to seek help on Reddit, someone pointed me to the Intel 13th/14th Gen instability issue. The symptoms matched perfectly:

Random crashes under varied workloads
Segfaults in unrelated binaries
System hangs after running for days
Progressive worsening over time

It was suggested that I run the Intel Processor Diagnostic Tool (IPDT) (https://www.intel.com/content/www/us/en/support/articles/000005567/processors.html).

This tool is Windows-only. Intel should have created a utility that is OS-agnostic making it something you could run from a bootable USB drive.

Having to jump through hoops to perform a test for the company's product is counterproductive.

Stress tests like mprime / Prime95 didn’t expose the issue.

March 2025: Intel RMA Process

Armed with the information, I hopped over to the Intel Community (https://community.intel.com/) and asked about the process.

At https://www.intel.com/content/www/us/en/support/articles/000057098/processors.html I found all the information to start the process.

You need to provide:

Processor number
ATPO
FPO

If you no longer have the box, you’ll need to remove the CPU and clean off the thermal paste to read the markings.

When the Intel support agent asked me for the following additional information (from Intel's email):

Motherboard: Please include the model and any relevant details.
Power Supply Unit (PSU): Model, wattage, and manufacturer.
Dedicated Graphics Card: Model and manufacturer, if applicable.
Thermal Solution Details (Air Cooler or a Liquid Cooler):
Operating System Details:
Have you tested this processor on a different motherboard, if yes, on which motherboard was it tested?
Have you tested your current motherboard with a different compatible processor, if yes, which processor have you used?

The process was excellent.

I live in Malta, an island at the edge of Europe, where cross-border logistics usually take time.

Key dates:

16 March 2025: Case opened / CPU information provided / collection process initiated
18 March 2025: DHL collected the defective CPU
20 March 2025: Replacement CPU delivered by DHL

Thermal Paste

Make sure you have thermal paste on hand. If you need to read information off the CPU (and using the machine while your case is being processed) and when you mount the new CPU on your motherboard.

Replacement CPU

The replacement CPU arrived in Intel-branded packaging.

Post-Replacement: Problems Resolved

After installing the replacement i7-13700K, all previously experienced problems disappeared. Two months after replacing the CPU, the system returned to the rock-solid stability it had enjoyed during its first year of operation.

The Bigger Picture: Intel's Raptor Lake Defect

My experience was not isolated. It was part of one of the largest CPU reliability crises in recent computing history. Earlier in my troubleshooting process, I had seen references to CPU-related problems, but:

CPU tests passed
in over 40 years I had never encountered a time-delayed CPU defect

What Went Wrong

Intel's 13th Gen ("Raptor Lake") and 14th Gen ("Raptor Lake Refresh") desktop processors suffered from a fundamental defect that caused progressive, irreversible degradation. The issue affected high-performance SKUs — primarily the Core i5, i7, and i9 K/KF/KS variants with the 8P+16E core configuration.

Root Cause: Vmin Shift Instability

On 25 September 2024, Intel employee Thomas Hannaford posted the official root cause analysis on the Intel Community forums (https://community.intel.com/t5/Blogs/Tech-Innovation/Client/Intel-Core-13th-and-14th-Gen-Desktop-Instability-Root-Cause/post/1633239)

Intel localized the problem to a clock tree circuit within the IA core that is particularly vulnerable to reliability aging under elevated voltage and temperature. These conditions lead to a duty cycle shift of the clocks, causing system instability.

Intel identified four operating scenarios that lead to Vmin shift in affected processors:

1. Motherboard Power Delivery Exceeding Intel Guidance

Mitigation: Intel Default Settings recommendations for 13th/14th Gen desktop processors. It is common practice for motherboards to exceed these settings and they have done so for years. It is normally set by default.

2. eTVB Microcode Algorithm Issue

The Enhanced Thermal Velocity Boost algorithm was allowing i9 desktop processors to operate at higher performance states even at high temperatures.

Mitigation: Microcode 0x125 (June 2024)

3. SVID Algorithm Requesting High Voltages

The microcode's Serial Voltage Identification algorithm was requesting high voltages at a frequency and duration that caused Vmin shift.

Mitigation: Microcode 0x129 (August 2024)

4. Elevated Core Voltages During Idle/Light Activity

Microcode and BIOS code were requesting elevated core voltages especially during periods of idle and/or light activity — exactly the condition a homelab server experiences most of the time.

Mitigation: Microcode 0x12B, which encompasses 0x125 and 0x129, and addresses elevated voltage requests during idle and/or light activity periods

Intel confirmed that mobile processors and future product families (Lunar Lake, Arrow Lake) are unaffected by the Vmin Shift Instability issue.

Manufacturing Oxidation (Early Units)

For some early 13th Gen processors (manufactured in late 2022 — exactly when I purchased mine), there was an additional manufacturing defect involving oxidation. Intel confirmed this was identified internally in late 2022 and addressed in production by early 2024, but on-shelf inventory with the defect may have persisted into early 2024.

The Critical Point: Damage Is Irreversible

Once a processor has been exposed to excessive voltage for long enough, the damage to the clock tree circuit is permanent and cannot be repaired by any software update. Microcode patches can only prevent further damage on CPUs that haven't yet degraded — they cannot restore already-damaged processors.

This is why my system's stability gradually worsened over time, and why replacing the PSU only appeared to help temporarily.

Why My Homelab Was the Perfect Victim

Looking at Intel's four identified scenarios, my Proxmox homelab hit the worst-case combination:

Scenario 4 (idle/light activity): A homelab server spends most of its time in light-load or idle states — exactly when the faulty microcode was requesting the highest inappropriate voltages
Scenario 3 (SVID high voltage requests): The constant power-state transitions of a virtualization host (VMs starting, stopping, idling) triggered frequent SVID voltage requests
Early manufacturing (oxidation): Purchased November 2022, squarely in the window for the oxidation manufacturing defect
Always-on operation: Running 24/7 meant maximum cumulative exposure to the damaging conditions

The May 2025 microcode update (0x12F) specifically addressed "systems continuously running for multiple days with low-activity and lightly-threaded workloads" — a near-perfect description of a homelab server.

Intel's Response: A Timeline

Date	Action
Late 2023 – Early 2024	Community reports of instability begin accumulating
April 2024	Intel recommends motherboard manufacturers use "Intel Default Settings"
July 2024	Intel officially acknowledges elevated voltage as the cause
June 2024	Microcode 0x125 released — fixes eTVB algorithm issue
August 2024	Microcode 0x129 released — addresses high voltage requests
August 2024	Intel announces 2-year warranty extension (3 years → 5 years) for affected SKUs
September 2024	Intel identifies root cause (clock tree circuit / Vmin Shift)
September 2024	Microcode 0x12B released — "final" fix encompassing 0x125 + 0x129 + idle voltage control
October 2024	Intel confirms the voltage issue was the sole root cause
November 2024	Class action lawsuit filed in San Jose, California
May 2025	Microcode 0x12F released — addresses edge cases in systems running continuously for multiple days with light workloads

Warranty Extension

Intel extended the warranty from 3 years to 5 years for all affected boxed 13th/14th Gen desktop processors (https://community.intel.com/t5/Mobile-and-Desktop-Processors/Additional-Warranty-Updates-on-Intel-Core-13th-14th-Gen-Desktop/m-p/1620853).

Key points from the announcement:

The extension applies to new and previously purchased processors
Coverage applies to all customers globally
The warranty eligibility period starts on the original purchase date and does not reset if Intel provides a replacement
Intel committed to supporting all customers experiencing instability symptoms through the exchange process

Affected Processors

The instability primarily affects desktop processors with the 8P+16E Raptor Lake silicon:

13th Gen (Raptor Lake):

Core i9-13900K/KF/KS
Core i7-13700K/KF
Core i5-13600K/KF

14th Gen (Raptor Lake Refresh):

Core i9-14900K/KF/KS
Core i7-14700K/KF
Core i5-14600K/KF

Lower-power variants (non-K, mobile) were less commonly affected but not entirely immune.

Lessons Learned

For Users Experiencing Instability

Don't assume it's software. If your system was stable for months and gradually becomes unstable, hardware degradation is a real possibility — especially with Intel 13th/14th Gen CPUs.
The symptoms are deceptive. The crashes manifest as memory errors, storage corruption, split lock violations, segfaults in random binaries — anything that looks like "something else" is broken. This is because the CPU is making computation errors.
Run the Intel Processor Diagnostic Tool / equivalent. Download it from Intel's support site. If your CPU fails, you have clear evidence for an RMA. Remember that a Pass is not a sign that your CPU is not impacted.
Don't waste money replacing other components first. I had a perfectly good PSU to replace but didn't have spare RAM, storage and a motherboard lying about. While testing RAM and storage is reasonable, be aware that a degrading CPU can make other components appear faulty.
Check your warranty status. Intel extended the warranty to 5 years for affected 13th/14th Gen desktop processors. If you purchased after October 2022, you likely have coverage through at least 2027.
Update your BIOS/microcode. If your CPU hasn't yet degraded, microcode 0x12B (or newer) can prevent the excessive voltage that causes damage. Check your motherboard manufacturer's website for the latest BIOS. A replacement is ultimately the best solution if you are eligible.

Conclusion

What started as random crashes turned out to be a widespread hardware issue.

The community played a critical role in identifying the root cause.

People who dedicate time to maintaining forums and helping others are a key part of the ecosystem—thank you.

If you're reading this because your 13th or 14th Gen Intel system is acting up take action. The fix exists, the warranty coverage is there, and Intel's replacement process works. Don't spend months chasing ghosts like I did. The clock is running out and these CPUs will eventually no longer be covered for replacement.

The Zero-Day Exploit Clock Is Ticking

Alan Bonnici — Thu, 04 Jun 2026 10:12:58 +0000

A Personal Introduction

The notion of breaking into a system and making it do something its developers never intended has existed since the early days of computing. My own introduction to computing came through hacking into games — disabling or increasing the number of lives to compensate for my slow reflexes.

Using TD (Turbo Debugger, from Borland — shipped alongside Turbo Assembler and Turbo C/C++), I would painstakingly try to identify the memory location that held the lives counter. Once found, one could either alter the value directly or place a NOP (No Operation) over the conditional jump. It was a task that required patience, determination, and, many times, reams of sprocketed continuous-feed paper annotated to help decipher what the code did.

Discoveries were rarely kept to oneself. You shared them with others on BBSs (Bulletin Board Systems), the dial-up communities that served as gathering places for the technically curious long before the web existed.

BBSs were not just a place to share results; they were a primary source of knowledge. Text files on cracking techniques, annotated memory maps for popular games, and tutorials on x86 assembly circulated freely between boards. You would dial in with a modem, download a collection of .TXT and .NFO files, then spend hours offline studying them. It was a decentralised, informal education system: no textbooks, no courses, just collective knowledge passed between pseudonyms.

The motivation was curiosity. There was no financial incentive and no criminal intent. It was intellectual challenge for its own sake — the digital equivalent of picking a lock just to prove you could.

From Annoyance to Destruction

That harmless hacking eventually evolved into something darker: worms, trojans, and viruses. Initially, these programs were little more than an annoyance, displaying messages on screen or slowing machines down. But the lever was gradually notched towards the harmful and destructive. Rather than simply popping up a message, these programs began deleting files and formatting hard drives.

Early examples like the Brain virus (1986) were relatively benign; it even included the authors' names and phone number. But by the early 1990s, destructive payloads had become common. The Michelangelo virus (1992) overwrote the first hundred sectors of a hard disk. CIH/Chernobyl (1998) went further, attempting to corrupt BIOS firmware and render machines unbootable.

If you are interested in watching a YouTube video of mine from 19 years ago demonstrating the operation of a destructive virus called “Casino de Malte”, head over to https://youtu.be/wiLZAEMsofM. Your hard disk’s fate was determined by the outcome of a game.

The Commercialisation of Hacking

The internet, cryptocurrency, and the dark web created an environment in which malicious actors could come together and organise themselves into profit-seeking entities. Their business model: penetrate systems, steal data, hijack operations, and then seek payment from the victim to rectify the mess they created.

It proved extraordinarily lucrative. Cybercrime is now measured in the trillions. Cybersecurity Ventures estimated that global cybercrime damage would reach $10.5 trillion annually by 2025 — a figure that, if treated as an economy, would make cybercrime the third-largest in the world after the United States and China.

These are not lone hackers in basements. Modern cybercriminal organisations operate with corporate structures — complete with HR departments, customer service teams to “help” victims pay ransoms, software development lifecycles, and even employee performance reviews. Groups such as LockBit and Conti have operated as Ransomware-as-a-Service (RaaS) platforms, licensing their tools to affiliates in exchange for a percentage of the ransom collected.

The Ecosystem: States, Criminals, and Brokers

Today, the malware market is made up of three distinct layers:

State-backed groups — Advanced Persistent Threat (APT) actors funded by nation-states for espionage, sabotage, and geopolitical advantage. Examples include Fancy Bear (Russia), Lazarus Group (North Korea), and APT41 (China).
Private criminal groups — Organisations that carry out penetration attacks, deploy ransomware, and operate data-theft schemes for profit.
Exploit brokers — Companies and individuals that trade in hacking knowledge. Firms such as Zerodium openly advertise bounty prices for zero-day exploits: up to $2.5 million for a full Android exploit chain, for example. This grey market means that vulnerabilities are sometimes sold to the highest bidder rather than reported to the vendor for patching.

The success of these operations often depends on what are known as zero-day exploits: vulnerabilities in software that are not yet publicly known and for which no patch or mitigation exists. The term “zero-day” refers to the fact that the vendor has had zero days to fix the problem before it is exploited.

To complicate matters further, some attacks require chaining together multiple exploits in a particular sequence. This is known as an exploit chain, or attack chain. For example, a modern browser exploit might chain a renderer vulnerability with a sandbox escape and a kernel privilege escalation — three separate flaws linked together to achieve full system compromise.

The Zero-Day Exploit Clock

The website “Zero Day Clock” presents a striking visualisation. It analyses publicly known vulnerabilities and estimates the time between public disclosure of a vulnerability and the first confirmed exploitation in the wild.

CVE gives a publicly disclosed vulnerability a common identifier, but it is only one part of the vulnerability-tracking ecosystem. CWE describes the underlying weakness, CVSS scores technical severity, EPSS estimates exploit likelihood, and CISA’s Known Exploited Vulnerabilities catalogue highlights flaws already exploited in the wild. Together, these systems help organisations decide what to fix first.

Strictly speaking, once a vulnerability is publicly disclosed, it is no longer a zero-day in the narrowest sense; exploitation after disclosure is often described as n-day or one-day exploitation. But from a defender’s point of view, the practical effect is similar: the available response window is collapsing.

The trend is alarming:

That compression — from months to days, and potentially from days to minutes — represents a fundamental shift in the threat landscape. Defenders used to have time to test, prioritise, patch, and monitor. Increasingly, they may have only hours, minutes, or seconds.

The AI Accelerant

The dramatic reduction in exploitation time is being accelerated by artificial intelligence, but AI is only one part of a wider shift. Automated scanning, exploit marketplaces, faster reverse engineering, criminal collaboration, and global infrastructure have all compressed the defender’s response window.

The same AI tools that security researchers use to find flaws in code can also be weaponised by adversaries. Their systems can continuously scour vulnerability databases, security mailing lists, vendor advisories, and even social media for news of a newly disclosed weakness. Once alerted, AI-assisted tools can help to:

Analyse the vulnerability — parsing patch diffs to understand exactly what was fixed and, therefore, what was broken.
Generate exploit code — assisting with proof-of-concept attacks targeting the flaw.
Adapt and mutate — creating variants that may evade simple signature-based detection.
Scale deployment — scanning the internet for unpatched systems and deploying attacks at machine speed.

This is no longer theoretical. Academic research has shown that AI agents can exploit some real-world vulnerabilities in controlled benchmark environments. The barrier to entry has lowered: what once required deep expertise in assembly language and months of reverse engineering can now, in some cases, be partially automated.

AI is also being used offensively for:

Spear-phishing at scale — generating personalised, convincing emails that bypass traditional spam filters.
Deepfake social engineering — cloning voices and video to impersonate executives in business email compromise (BEC) attacks.
Automated reconnaissance — mapping an organisation’s attack surface faster than any human team could.

What Can We Do?

We cannot eliminate risk, but we can significantly reduce our exposure.

For individuals:

Keep all devices updated with the latest security patches, and enable automatic updates wherever possible.
Retire technology that is outdated and no longer receiving security updates. End-of-life software is an open door.
Use multi-factor authentication or passkeys on every account that supports them.
Be sceptical of unexpected communications, even from known contacts. Verify through a separate channel.
Avoid reusing passwords. A single compromised password can become the first link in a much larger chain.

For organisations:

Maintain a rigorous patch management programme. The window between disclosure and exploitation is shrinking towards zero. That means patching can no longer be treated as a slow administrative process. Organisations should:
- Automate patch deployment where the risk profile allows it, especially for standard operating systems, browsers, endpoint software, and widely deployed services. Human approval remains important for high-impact systems, but approval workflows measured in days do not match a threat landscape measured in minutes.
- Prioritise vulnerabilities based on exposure, exploitability, business impact, and evidence of active exploitation. If twenty fixes are waiting, the most exposed and most exploitable systems should move first.
- Maintain an emergency patching route for critical vulnerabilities, with pre-agreed ownership, testing boundaries, rollback plans, and communication paths. The middle of an incident is the wrong time to decide who is allowed to approve a fix.
Prioritise internet-facing systems and known exploited vulnerabilities. These are the systems attackers can reach first and the weaknesses they are most likely to weaponise quickly.
Adopt a zero-trust architecture: assume breach and verify every access request.
Invest in defensive tools that can detect, prioritise, and respond at machine speed. Alerts alone are not enough; the goal should be rapid triage, containment, and recovery.
Conduct regular penetration testing and red-team exercises. These should test not only whether vulnerabilities exist, but whether the organisation can detect, escalate, and respond to them quickly.
Ensure robust, tested backup and recovery procedures. Ransomware is only effective if you have no viable alternative.

The uncomfortable truth is that vigilance alone is no longer sufficient. When exploitation timelines compress to minutes, human reaction time becomes the bottleneck. Organisations need automated detection, prioritisation, containment, and recovery capabilities that can operate at the same speed as the attacks they face.

We must also recognise that each of us can be one link in an exploit chain. A compromised personal device, a reused password, an unpatched home router — any of these can become the entry point that leads to a larger organisational breach.

Closing Thought

In three decades, hacking has evolved from teenagers on BBSs trading game cheats into a global criminal industry: commercialised, professionalised, backed by nation-states, and now accelerated by artificial intelligence.

The zero-day exploit clock is ticking faster than ever. For individuals, that means patching, retiring obsolete technology, using passkeys or multi-factor authentication, and treating unexpected communication with caution. For organisations, it means accepting that manual response is no longer enough.

The question is no longer whether an attack will come, but whether our defences can respond in the seconds we may have left.

Suggested references

Zero Day Clock: https://zerodayclock.com/
Cybersecurity Ventures, 2025 Official Cybercrime Report: https://cybersecurityventures.com/official-cybercrime-report-2025/
CVE-Bench: https://arxiv.org/html/2503.17332v3

"Complexity Ceiling" in Power Apps

Alan Bonnici — Wed, 20 May 2026 08:25:55 +0000

Power Apps is a powerhouse for rapid prototyping within the Microsoft ecosystem, but for complex development, the "seams" are becoming visible. This new analysis explores the "Complexity Ceiling" and the friction points faced by architects and developers.

Core Issues Identified:

Mental Model Mismatch: Chaining nested With blocks to simulate sequential logic.
Tooling Gaps: The need for real breakpoints and execution narratives over simple trace functions.
Governance Hurdles: The "abrupt cliff" of enterprise compliance for makers.

Microsoft has the opportunity to bridge the gap between business analysts and professional engineers by introducing controlled imperative constructs and better code organization.

https://www.alanbonnici.com/2026/05/the-complexity-ceiling-where-microsoft.html

Does your Linux server actually know when the power is failing?

Alan Bonnici — Thu, 19 Mar 2026 11:39:08 +0000

NUT (Network UPS Tools) is the go-to open-source solution for UPS monitoring on Linux — but it takes some know-how to set up properly, especially with hardware that isn't auto-detected out of the box.

This detailed guide covers everything from finding the right driver, dealing with stubborn read-only settings, configuring automatic shutdowns, and setting up email alerts so you know when things go wrong. There's even a section on how to safely test your shutdown logic before a real power cut hits.

Great resource for anyone running a homelab or Linux-based home server. 👇

https://www.alanbonnici.com/2026/03/setting-up-nut-ups-software-on-linux.html

How to configure Postfix to relay mail through Gmail (simple step-by-step guide)

Alan Bonnici — Tue, 17 Mar 2026 10:31:07 +0000

If you run Linux servers that need to send alerts, backup reports, or monitoring notifications, configuring SMTP can be a pain.

I wrote a short guide explaining how to install and configure Postfix with Gmail as an SMTP relay, including:

• Installing required packages

• Configuring TLS

• Setting up Gmail app passwords

• Securing credentials

• Testing email delivery

Good for small servers, homelabs, monitoring tools, and UPS notifications.

https://www.alanbonnici.com/2026/03/install-and-configure-postfix-using.html

Building a Production RAG Server with Ollama, Open WebUI and Chroma DB

Alan Bonnici — Wed, 11 Feb 2026 08:48:39 +0000

I've created a comprehensive guide on building a headless LLM server with RAG capabilities. The tutorial walks through the complete implementation including document ingestion, vector storage, and query optimization.

The setup is production-ready and can be completed in about 30 minutes.

There are optional code sections for those who would like to interact with the model programmatically.