Our selection of the top business news sources on the web.

AM edition. Issue number 1342

Latest 10 stories. Click the button for more.

Quote: Ray Dalio - Bridgewater Capital founder

"When life doesn't give what you want, you can be angry or sad, or you can learn about how reality works and develop principles that will work well to get you what you want." - Ray Dalio - Bridgewater Capital founder

Disappointment is usually experienced first as a feeling and only later, if at all, as a problem to be modelled and solved. Most people stay in the realm of mood: anger at what should have happened, sadness at what did not, resentment at those who seemed to stand in the way. The alternative is far more demanding and far more productive: to treat each frustrated desire as data about how the world operates, to interrogate that data, and then to codify better ways of acting so that similar situations produce different outcomes next time. That pivot from emotion to inquiry, from grievance to design, is the underlying mechanism that separates episodic success from compounding progress.

From reaction to diagnosis

The immediate human response to not getting what we want is typically narrative: someone wronged me; the system is rigged; I am unlucky; I am not good enough. These stories may contain fragments of truth, but they are structurally unhelpful because they are not diagnostic. They do not ask what causal chain produced the result, how incentives and constraints interacted, which assumptions failed, or what was missing from our mental model of reality. A diagnostic response, by contrast, asks: given that this outcome occurred, what does it reveal about how the underlying system works, and what would need to change-about my behaviour, my environment, or my strategy-for a different outcome to be more probable next time?

That shift requires accepting that reality is not obliged to conform to our expectations. Instead, reality is made of patterns, feedback loops, and constraints that can be studied. When outcomes disappoint us, they are not insults but information. Anger and sadness may be unavoidable first responses, but if they remain the final stage, they are wasted. The crucial move is to operationalise setbacks: to treat them as experiments whose results must be analysed and then encoded into practical rules for future decisions.

Dalio, principles, and the economic machine

Ray Dalio built his career and his firm around this idea of turning painful surprises into explicit principles. As the founder of Bridgewater Associates, one of the world's largest hedge funds, he is known not only for his investment performance but for his insistence that decisions should be governed by clear, tested rules rather than by transient feelings or unchecked intuitions. In his investing, he is famous for breaking down the economy into a set of understandable mechanisms-what he calls an "economic machine"-and then using that understanding to guide decisions.

On the investment side, Dalio analyses how credit, productivity, and monetary policy interact over time to produce recurring cycles of inflationary, disinflationary, and deflationary environments. When losses or errors occur, instead of treating them as bad luck, he looks for structural misreadings of these cycles or of debt dynamics. Those insights are then codified into systematic decision rules, increasingly executed by algorithms that reflect his principles about how markets behave. The same pattern underpins his approach to life: observe outcomes, interrogate what they say about reality, and distil them into principles that can be reused.

The quote in question comes from a lecture Dalio gave to graduates of Long Island University, where he urged them to treat setbacks as prompts to understand reality better and to develop principles for dealing with it. He linked this stance to personal responsibility and radical open-mindedness: owning one's life, staying in touch with how the world actually works, and being willing to revise beliefs when evidence demands it. For Dalio, being "in touch with reality" and maintaining "radical open-mindedness" are not slogans; they are operating requirements for anyone trying to achieve ambitious goals in a complex environment.

Pain as data: the engine of improvement

Central to Dalio's framework is the idea that emotional pain is a necessary input to learning. He encapsulates this in a principle often summarised as: "Pain + Reflection = Progress". The raw hurt of failure highlights a gap between our mental map and the territory of reality. Reflection then turns that discomfort into insight. Without the pain, there is no trigger to question our assumptions; without the reflection, pain degenerates into bitterness or avoidance.

Dalio goes further, explicitly advising people to "go to the pain rather than avoid it" because staying at the edge of discomfort accelerates evolution. This is a demanding discipline. It requires noticing the impulse to turn away from criticism, loss, or embarrassment, and instead leaning in: replaying decisions, inviting disagreement, and scrutinising blind spots. In his own organisation, this mentality is embedded in a culture of high standards and constant improvement, where individuals are expected to confront their weaknesses in order to grow.

Bridgewater's culture document describes an "overriding objective" of excellence, defined as continual improvement, and asserts that accuracy in understanding reality is essential to achieving it. Translating that to individual lives, the message is that excellence in any domain depends on being ruthlessly honest about what is and is not working, especially when that honesty is uncomfortable. The alternative is to protect one's self-image and stay within familiar patterns, at the cost of stagnation.

Principles as reusable algorithms

In Dalio's usage, "principles" function like human-readable algorithms for living and working. A principle is more than a vague value; it is a conditional rule that connects situations to actions: if you encounter X pattern in reality, then you should do Y. Over the decades, he catalogued such rules-about hiring, decision-making, risk management, and conflict resolution-and compiled them in his book "Principles: Life & Work". The aim is not to produce a rigid manual but to provide a library of tested responses to recurring problems.

Principles have several advantages over ad hoc reactions. First, they reduce the influence of transient emotions. When you decide in advance that you will, say, seek out the smartest opposing view before making a major decision, you are less likely to be swayed by fear or overconfidence in the moment. Second, they create consistency across time and context; different situations that share the same underlying structure can be handled in similar ways. Third, they are teachable. An organisation that shares the same principles can coordinate action without needing the founder to be involved in every decision.

In investing, this approach is visible in Dalio's emphasis on systematic decision-making and diversification. He advocates creating rule-based processes that can be stress-tested with historical data and executed via algorithms. Diversification across uncorrelated assets becomes a principle: do not rely on a single bet; spread risk so that no one error is catastrophic. When markets move against a position, the question is not primarily "How do I feel about this?" but "What did this movement reveal about the assumptions built into my decision rules, and what principle should be updated?"

Responsibility versus resentment

The quote draws a stark behavioural fork in the road: when frustrated, you can either remain in the emotional loop of anger and sadness or adopt the slower, more demanding posture of student of reality. Dalio explicitly urged his audience to "own" their lives and take responsibility for making them great. Responsibility, in this sense, is not a moral label but a practical stance: it is the decision to treat one's situation as improvable through better understanding and better principles, rather than as fixed by external forces.

Resentment has its own logic. It offers a sense of justification and identity: I am the sort of person to whom unfair things happen. It can also be politically potent, mobilising people against perceived injustices. But as a personal operating system, resentment is paralysing. It interprets every disappointment as confirmation of a hostile world, not as input for revised action. By contrast, the principle-based stance acknowledges constraints and injustice but asks, "Given that these features of reality exist, what is the best strategy for pursuing my goals within them?" That question restores agency without denying difficulty.

Strategic and technological tension

There is a deeper tension embedded in Dalio's approach: the balance between systematising reality and respecting its complexity. On the one hand, he argues that understanding how reality works is essential and that principles can be codified into increasingly sophisticated systems, including algorithms. Bridgewater has long used technology to turn qualitative insights into quantitative decision rules, effectively building models that transform messy market data into structured signals.

On the other hand, excessive faith in models can be dangerous. Economic and social systems are adaptive; participants change their behaviour in response to observed patterns. A trading rule that works today may fail once others imitate it. The more we encode our understanding of reality into rigid systems, the greater the risk that we will be blindsided by novel configurations that fall outside our priors. The strategic challenge is to hold principles strongly enough to provide guidance, yet lightly enough to revise them when evidence accumulates against them.

Dalio attempts to manage this tension through radical open-mindedness and continuous stress-testing of ideas. He encourages "radical transparency" inside Bridgewater, meaning that people at all levels are expected to challenge each other's thinking, including his own. This social technology aims to prevent the ossification of principles into dogma. If someone presents better evidence or a better model, the principle should change. In theory, this keeps the system adaptive: principles are not final truths but current best hypotheses about how reality works.

Debates and objections

Critics raise several objections to this principle-driven pragmatism. One is that it can slide into technocratic coldness. If all setbacks are treated as data and all relationships as arenas for truth-seeking, there is a risk of underweighting the emotional and moral dimensions of human life. Bridgewater's intense culture-of constant feedback, rigorous criticism, and high expectations-has been described as both transformative and punishing, depending on whom you ask. Some thrive in such environments; others experience them as dehumanising.

Another objection is structural. Not everyone has the same degree of control over their circumstances. Systemic inequities, institutional barriers, and sheer luck shape life chances. To tell someone in a constrained environment that they should simply "learn how reality works" and develop principles to get what they want can sound naive or cruel. The valid insight-that focusing on controllable factors is empowering-must be balanced with recognition of structural limits. A realistic principle-based approach should include, as part of its understanding of reality, an analysis of power, institutions, and collective action, not only individual grit.

A further critique questions whether complex human lives can, or should, be governed by explicit principles at all. Much of what makes people effective in relationships, creativity, or leadership is tacit knowledge: pattern recognition that resists codification. Over-formalising behaviour can produce rigid, instrumental ways of being that miss nuance. Dalio's own corpus partly acknowledges this by insisting that others develop their own principles, articulated in their own words, rather than simply copying his. The aim is to make explicit as much as is useful, while accepting that some wisdom remains intuitive.

Why it matters beyond finance

Despite these tensions, the underlying stance has wide relevance beyond hedge funds and business schools. In a world marked by rapid technological change, geopolitical instability, and shifting social norms, individuals and institutions are constantly confronted with outcomes they did not anticipate. Reacting with anger or despair is understandable but strategically inert. Developing a practice of decoding those outcomes into updated models and better principles is one of the few robust responses available.

For a young professional passed over for promotion, the principle-based approach might mean examining feedback, observing who does advance, and inferring what the organisation actually rewards. Perhaps the implicit success criteria differ from the formal job description. That insight can be turned into new principles: prioritise projects that demonstrate impact in ways the leadership values, cultivate specific relationships, or, if misalignment is severe, adopt the principle that one should move to an environment where desired behaviours are recognised.

For someone facing personal adversity-illness, financial setback, relationship breakdown-the same logic applies. The question is not whether the situation is fair; often it is not. The operative questions are: what constraints does this reality impose, what options remain, what skills or supports are missing, and what principles would reduce the probability or severity of similar outcomes in future? That might translate into principles about savings and diversification of income, about communication and boundaries in relationships, or about health habits and early detection.

At a societal level, this mindset underpins effective policy-making. When a programme fails to achieve its stated goals, the productive response is to ask which assumptions about human behaviour or institutional capacity were wrong and to update design principles accordingly. Blame and outrage may have their place in assigning responsibility, but they do not by themselves produce better systems. A society that treats failures as occasions for learning builds institutional memory and resilience; one that treats them only as scandals repeats them.

Building a personal system of principles

Practically, adopting the stance implied by the quote involves building one's own evolving set of principles. That starts with a habit of documentation: after meaningful successes or failures, write down what happened, why you think it happened, and what rule you wish you had followed. Over time, patterns emerge: recurring mistakes point to missing or flawed principles; recurring wins highlight reliable strategies. This is analogous to Dalio's long-term effort to record his insights from managing people and markets and then refine them into the corpus published in "Principles".

Next comes testing. A principle is a hypothesis: "If I do X in situations of type Y, outcomes will generally improve." Applying it consciously in future situations and observing the results either strengthens confidence or triggers revision. Discussion with others-especially those who disagree-serves as further stress-testing, similar to the way Bridgewater uses internal debate to improve decision rules. Over years, this process yields a personalised operating system calibrated to one's goals, strengths, and environment.

Finally, there is the matter of courage. Learning how reality works can be uncomfortable, because it often reveals ways in which we have been self-deceiving, underperforming, or complicit in our own problems. Developing principles that "work well to get you what you want" requires not only intellect but willingness to confront those truths and to change entrenched habits. Dalio's insistence on "going to the pain" is a recognition that, without this courage, the entire project stalls. Most people do not lack access to information about how the world works; they lack the willingness to apply that information consistently when it conflicts with short-term comfort.

The choice framed in the quote is therefore not a one-off decision but a recurrent one, embedded in daily life. Each time reality disappoints, we either deepen the groove of emotional reactivity or strengthen the muscle of inquiry and principle-building. Over years, those choices compound. One path leads to a life organised around narratives of grievance and helplessness; the other leads, imperfectly and with many detours, towards greater effectiveness and alignment between what we want and what we are able to achieve.

Quote: Victor Khosla - Strategic Value Partners

"There are entire sectors like private equity, like real estate, that are constipated. They can't sell." - Victor Khosla - Founder of credit investor Strategic Value Partners at SuperReturn Berlin 2026

Capital that cannot circulate starts to decay. In private markets, the inability to sell assets on a reasonable timetable is no longer a cyclical annoyance but a systemic constraint shaping strategy, governance, and even the survival prospects of whole franchises. Funds that once relied on brisk exits to prove their worth and raise ever larger vehicles are now confronting portfolios that linger, values that look aspirational rather than realisable, and investors whose patience is being measured in years rather than quarters. The blockage is not confined to a few misjudged deals; it runs through core sectors such as buyouts and real estate, where the exit machinery that underpinned the boom has stalled.

The anatomy of a blocked exit pipeline

The immediate problem is mechanical: there are too many assets that were acquired at peak or near-peak valuations and too few buyers willing to pay prices that preserve those marks. During the era of cheap money, sponsors could rely on abundant leverage, aggressive growth forecasts, and liquid secondary demand to justify paying high multiples for businesses and properties. When policy rates rose and financing costs reset, the arithmetic that once supported these prices no longer added up, but the portfolios built on the old assumptions remained.

Industry data show that the value of private equity sales has dropped by roughly a fifth in the recent period, even as the stock of unsold assets has accumulated. Managers are continuing to hold companies that, according to pre-2022 playbooks, would have been exited via trade sale, sponsor-to-sponsor deal, or IPO after a holding period of four to six years. Instead, those assets are being extended, refinanced, or shifted into continuation vehicles, often without crystallising the type of gains that limited partners had come to expect.

The same blockage is evident in institutional real estate, where higher interest rates compress valuations and increase financing costs, particularly for leveraged, income-producing assets. Buyers demand discounts to compensate for uncertainty and higher yields, while sellers are reluctant to transact at prices that would lock in visible write-downs. The result is a bid-ask spread that is wide enough to choke off volumes, leaving many buildings and portfolios in a state of limbo: not obviously distressed, but not evidently worth prior valuations either.

Victor Khosla and the vantage point of distressed credit

From the perspective of distressed and opportunistic credit investors, the current blockage is as much an opportunity as a diagnosis of systemic strain. Victor Khosla, founder and chief investment officer of Strategic Value Partners (SVP), has built his franchise precisely around capitalising on moments when traditional owners can no longer carry assets or debt structures set up under rosier assumptions. SVP specialises in distressed and deep value situations across corporate credit, private equity, and special situations, giving Khosla a panoramic view of where pressure is building and how it may resolve.

His characterisation of sectors such as private equity and real estate as effectively frozen speaks to more than a cyclical slowdown. It signals a view that many sponsors are trapped by their own historical decisions: valuations that were underwritten in a world of low discount rates and abundant leverage have become psychological anchor points that slow down the recognition of new realities. For opportunistic credit funds, that hesitation creates the potential for negotiated restructurings, discounted secondary trades, and control transactions once forced sellers emerge.

SuperReturn Berlin and the politics of liquidity

The comments arose at SuperReturn, the flagship gathering of the private capital industry in Berlin, where thousands of general partners and limited partners meet to negotiate, market, and assess the state of the asset class. This is not a neutral academic forum; it is where capital providers confront managers about delayed distributions, extended fund lives, and the apparent disconnect between reported valuations and observable exit outcomes. Against that backdrop, describing entire sectors as unable to sell pins the problem firmly on the supply side: sponsors unwilling to adjust, rather than a mysterious absence of buyers.

Several large buyout executives at the event emphasised that private equity will have to start capitulating on valuations to clear the backlog. The idea of capitulation implies not just incremental discounts but a recognition that the previous equilibrium on pricing was artificially supported by low rates and exuberant competition. For LPs, the politics of liquidity are now central. They entered the asset class on the promise of illiquidity premia and diversification, not of indefinite lock-up with opaque marks. Conferences that once celebrated record fundraising now feature sessions dedicated to portfolio liquidity, continuation vehicles, and secondaries as release valves.

From virtuous cycle to negative feedback loop

In the boom era, private equity operated in a self-reinforcing loop. Strong distributions and mark-ups allowed LPs to recycle capital into new funds, which in turn justified higher fundraising targets and more aggressive deal-making. As long as exit markets remained open, the cycle sustained itself, even if leverage and valuations crept steadily upward. When exits slow materially, that virtuous loop can flip into a negative feedback process.

LPs with limited cash inflows from distributions become more selective, committing less to new vehicles or insisting on tighter terms and more conservative pacing. GPs, facing slower fundraising, are pressured to generate liquidity to preserve their franchise value, but they can only do so by selling assets at prices that may require write-downs. If they resist, they risk sliding into the category of so-called zombie funds: vehicles that hold ageing assets, charge fees, but generate little in the way of exits or performance fees.

One can view the situation through a simple cash flow lens. Let represent the net distributions to LPs at time , net of capital calls. In the expansion phase, many funds operated with patterns where for most post-investment years, allowing LPs to maintain or grow commitments without increasing their net exposure. As exits have stalled, has turned negative for many investors, even as their reported net asset value remains high. The resulting denominator effect, where private market allocations swell relative to public portfolios due to market moves and valuation lag, further constrains their capacity to recommit.

Real estate: the silent partner in the blockage

Real estate, especially in sectors affected by structural shifts such as offices and certain retail formats, amplifies this dynamic. Rising financing costs and changing usage patterns have forced investors to reassess long-term income assumptions. Yet many portfolios are still carried at values that assume moderate yield expansion and stable occupancy, rather than a more fundamental repricing. This creates a situation where transactions that would reveal more dramatic value adjustments are avoided, reinforcing the freeze.

For leveraged owners, the core equation linking asset values, loan-to-value ratios, and covenant headroom has shifted. If one denotes the market value of a property as and the outstanding debt as , covenants may require that , where is the maximum permissible leverage ratio. When values fall while debt remains fixed or only slowly amortising, owners can find themselves breaching or approaching breach, even when cash flows have not yet collapsed. This increases dependence on lender forbearance or restructuring and makes it harder to transact at realistic values without triggering technical defaults.

In turn, lenders and debt investors must choose between extending and pretending, injecting additional capital, or forcing asset sales into thin markets. Distressed and opportunistic funds are watching closely, as each of these choices can convert illiquid mark-to-model valuations into executable deals, often at prices that reflect distress rather than orderly value.

Strategic responses: write-downs, secondaries, and continuation funds

One of the major tensions highlighted by the current environment is between short-term reputational pain and long-term franchise survival. General partners who accept material write-downs today may suffer in near-term performance league tables and carried interest prospects, but they regain the ability to sell assets, return capital, and reset expectations. Those who hold onto legacy valuations may postpone the day of reckoning but risk compounding the problem as fund lives lengthen and LP pressure grows.

The secondary market has become a crucial adjustment mechanism. LPs seeking liquidity can sell their fund interests at discounts to net asset value, transferring the risk and potential upside of ageing portfolios to specialised buyers. Similarly, GP-led secondaries and continuation vehicles allow sponsors to extend ownership of particular assets while providing partial liquidity to existing investors. Yet these structures are themselves constrained by the same valuation debate: at what price should stakes be transferred, and to what extent should discounts acknowledge the illiquidity and uncertainty embedded in the asset?

Continuation funds in particular raise governance questions. When a GP sells an asset from one of its funds into a vehicle it also controls, conflicts become inherent. Independent fairness opinions and auction processes are used to mitigate these issues, but if underlying assets are marked aggressively, continuation transactions can be perceived as a way of avoiding the full recognition of market-clearing prices rather than a genuine value-maximising strategy.

Technological and structural drivers behind the blockage

Beyond interest rates and valuations, there are deeper technological and structural forces contributing to the logjam. In some sectors, especially software and digital infrastructure, business models that were once rewarded with very high growth multiples have shifted into slower, more cash-generative phases. Sponsors that underwrote deals on the assumption of sustained hyper-growth and rapid multiple expansion now face businesses that are solid but not explosive, making the exit case less straightforward.

At the same time, the rise of direct investing by large institutions, sovereign wealth funds, and family offices has altered the buyer universe. Some of these investors are less willing to participate in sponsor-to-sponsor trades at high multiples, preferring direct origination or co-investments where fees and governance are more aligned. This reduces one of the key exit routes that underpinned the previous boom, increasing dependence on trade buyers and public markets, both of which are sensitive to macro conditions and sector-specific narratives.

Advances in data and analytics, while powerful, also create more scrutiny. Prospective buyers now have access to richer operational and market datasets, enabling more granular stress-testing of revenue, margin, and cash flow scenarios. In a risk-off environment, such tools often lead to more conservative underwriting rather than more aggressive bidding, as buyers quantify downside scenarios more explicitly. The informational asymmetry that once allowed sellers to market a growth story with limited transparency has narrowed, making it harder to justify top-of-cycle valuations.

Debates and objections: is the problem overstated?

Not everyone accepts the narrative of pervasive constipation. Some argue that the industry is simply adjusting to a new rate regime and that exit volumes, while lower than peak, remain within historical norms when viewed over a longer horizon. They point out that selected sectors, such as energy transition, infrastructure, and certain niches in technology and healthcare, continue to see robust activity at sensible valuations. From this perspective, the blockage is concentrated in specific vintages and strategies that overreached, rather than an indictment of the asset class.

Others emphasise that the illiquidity of private markets is part of their design, arguing that an excessive focus on near-term exits risks undermining the long-term value creation thesis. They highlight cases where patient capital and operational improvement over extended holding periods have yielded strong outcomes, even without rapid flips. According to this view, pressure from LPs for faster distributions may be partly driven by their own allocation and liquidity management issues, rather than by an inherent failure of private equity or real estate strategies.

A further objection is that commentary from distressed and opportunistic investors is not disinterested. Those who stand to benefit from forced selling have an incentive to emphasise the scale of the problem and to urge capitulation on price. While this does not invalidate the diagnosis of a backlog, it suggests that statements need to be interpreted in the context of strategic positioning: what is a blockage for one segment of the market is a source of deal flow for another.

Why the blockage matters for the broader financial system

Despite those objections, the build-up of unsold assets in private equity and real estate has implications that go well beyond individual firms or funds. Institutional portfolios globally have significant exposures to private markets, often via defined benefit pension schemes, insurers, endowments, and sovereign funds. When exit channels narrow, these investors face a mismatch between their assumed liquidity profile and realised cash flows. For pension schemes in particular, the experience of backlogs in other risk-transfer markets, such as bulk annuity transactions, underscores how a crowded pipeline can delay strategic plans.

If LPs find themselves over-allocated to illiquid assets for prolonged periods, they may respond by reducing future commitments, driving a downshift in fundraising across the industry. Managers with strong track records and differentiated strategies may adapt, but marginal players could struggle to survive, leading to consolidation or outright failures. Bank and non-bank lenders with significant exposure to leveraged loans and commercial real estate debt must also manage the risk that delayed exits extend credit risk horizons and increase the probability of restructurings or losses.

There is also a macroeconomic angle. When significant pools of capital are trapped in assets that cannot be repriced or redeployed efficiently, capital allocation across the economy becomes less dynamic. Companies that might benefit from new investment may struggle to attract it if capital remains locked in legacy deals, while real estate that could be repurposed or redeveloped remains under the control of owners reluctant to crystallise losses. Over time, such frictions can dampen productivity, slow the reallocation of resources, and hinder the adjustment to new technological and societal patterns of demand.

Paths to resolution: price discovery and structural change

Ultimately, resolving a systemic backlog in private markets requires price discovery. As more transactions clear at realistic levels, valuation benchmarks will adjust, and the psychological barrier to accepting lower prices will weaken. This process can be painful, particularly for vintages funded at the peak of the cycle, but it is a prerequisite for restoring the circulation of capital. Several mechanisms are likely to play a role.

First, a wave of consensual restructurings and recapitalisations can realign capital structures with current cash generation and asset values, often accompanied by new equity injections from opportunistic investors. Second, distressed sales triggered by covenant breaches, fund-life constraints, or lender pressure will provide transparent reference points for pricing, even if they occur under duress. Third, regulatory and accounting developments that encourage more conservative valuation practices can push managers to update marks closer to observable transaction levels, even in the absence of exits.

On a structural level, the industry may move towards vehicles with more flexible liquidity features, such as evergreen funds, listed private markets vehicles, or hybrid structures that blend open-ended and closed-ended characteristics. While such innovations cannot eliminate the fundamental illiquidity of underlying assets, they may distribute the timing risk more broadly and reduce the concentration of exit pressure at the end of traditional fund lives.

There is also scope for more integration between public and private markets. Dual-track processes, public-to-private cycles, and the use of listed feeders or co-investment vehicles can create additional degrees of freedom for sponsors and investors. However, these tools work only if public markets are themselves receptive and if valuations in public and private domains converge sufficiently to allow arbitrage-free transitions.

Why the metaphor resonates now

The characterisation of private equity and real estate as jammed speaks to a larger anxiety: that a model built on constant motion has met a structural speed limit. Years of easy money encouraged an expansion of private markets that outpaced the development of corresponding exit channels. As the cost of capital reset, the imbalance became visible. Whether the current phase is remembered as a temporary blockage or as the tipping point towards a smaller, more disciplined industry will depend on how quickly sponsors accept new pricing realities and how effectively capital can be redeployed.

For LPs, regulators, and policymakers, the episode is a reminder that illiquidity is not a passive characteristic but an active risk factor. Promises of double-digit returns and diversification benefits cannot be divorced from the practical question of when and how capital comes back. The answer, in the coming years, will depend on the willingness of managers to trade short-term discomfort for long-term viability, and on the capacity of specialised investors to absorb, restructure, and eventually recycle assets that others can no longer afford to hold.

Quote: Editorial - Workers' Daily

"The benefits of technological advancement should be shared by society as a whole, rather than becoming a tool for a small number of employers to undermine workers' rights." - Editorial - Workers' Daily - the official mouthpiece of China's umbrella trade union organisation

Across recent waves of automation, the most enduring fault line has not been whether machines can perform particular tasks, but who captures the savings and productivity gains they generate. In the context of artificial intelligence, that question has become sharply political in China, where rapid deployment of generative systems collides with a labour market still shaped by industrial restructuring, a large pool of precarious workers, and an official commitment to social stability. As AI tools move from experimental pilots into core business processes, the distribution of their benefits increasingly determines whether technological change is experienced as opportunity, dispossession or something in between.

The immediate backdrop is a surge in corporate enthusiasm for AI as a way to compress labour costs in sectors ranging from customer service and data labelling to logistics and back-office administration. For firms operating on thin margins, replacing call centre agents or content moderators with AI-based systems can promise savings running to tens of millions of yuan across a workforce of 10 000 people. When those savings are not matched by new protections, retraining or bargaining power for employees, the technology becomes, in practice, a lever to shift income and security away from workers and towards shareholders and senior management.

China's experience with earlier rounds of restructuring provides a powerful memory of how such transitions can play out. The lay-offs of state-owned enterprise workers in the 1990s and early 2000s created a generation marked by insecurity and grievance, even as the broader economy grew. That legacy shapes today's anxiety about AI: mass replacement of workers by automated systems risks not only individual hardship but also heightened social tension, particularly in cities where service jobs have become a crucial absorption mechanism for rural migrants and displaced industrial labour. Against that backdrop, the insistence that technological progress must not become a tool for undermining workers' rights takes on a concrete, rather than abstract, urgency.

From productivity promise to labour displacement risk

AI, especially in its generative form, promises major gains in efficiency by automating routine cognitive tasks: drafting basic responses, triaging customer queries, generating boilerplate documents, and analysing large volumes of text or image data. For employers, the economic logic is clear. If an AI system can handle the work of, say, five junior clerks at a fraction of their combined salary and without social insurance contributions, it becomes hard to resist. Even modest improvements in model accuracy can make it rational, from a firm's perspective, to restructure departments and reduce headcount.

The risk is that this rationality at the firm level is socially destructive when replicated across thousands of enterprises. If many employers simultaneously replace segments of their workforce with AI, displaced workers may find that the new jobs created by the technology are too few, too specialised, or require skills they do not yet possess. The labour market then adjusts not through smooth transitions into higher-value roles but through periods of unemployment, downward wage pressure and an expansion of informal or precarious work.

The tension is sharpened in China, where the state retains a strong directive role and where social stability is treated as a central policy objective. Government advisers have begun warning that AI applications introduced solely to reduce labour costs, without improving services or sustainability, should face close scrutiny or explicit constraints. That stance implicitly challenges a purely market-driven allocation of the benefits of AI: cost savings cannot simply be converted into higher profits while workers absorb all of the transition costs.

Legal red lines and the politics of dismissal

This emerging political stance has started to crystallise in legal outcomes. Chinese courts have ruled that employers cannot justify firing workers simply because AI has rendered their current tasks redundant. In a high-profile case in Hangzhou, a technology company attempted to dismiss an employee after introducing AI tools that took over much of his work, offering a reassignment that involved a steep salary cut. The intermediate court found the dismissal unlawful, holding that an employer's decision to adopt AI does not constitute the kind of external, uncontrollable change that would justify redundancy under labour law.

Crucially, the court also rejected the idea that a substantial pay cut hidden inside a reassignment offer is an acceptable way to shift the costs of technological change onto workers. By treating such a move as effectively a constructive dismissal, the judgment signalled that firms cannot rely on AI as an accounting trick to convert stable jobs into worse-paid, less secure positions while claiming continued compliance with formal employment rules.

This jurisprudence sits on top of a broader legal framework in which the right to continued employment is embedded more robustly than in many market economies. China's Labour Contract Law places strong limits on arbitrary dismissal and treats job security as a social good, not merely a private contract matter. In the AI context, that framework acts as a guardrail: technological adoption is permitted and even encouraged, but not as a justification for offloading adjustment costs onto employees without negotiation or compensation.

Trade union positioning and official narratives

Within this legal and political environment, the rhetoric of China's official trade union structures carries particular weight. As the mass representative body for workers, closely intertwined with state institutions, the umbrella trade union organisation operates less as an adversarial bargaining agent and more as a channel through which policy priorities are communicated and reinforced. When its official newspaper warns against using technological advancement to erode workers' rights, it is not merely expressing a moral view; it is aligning labour discourse with the state's interest in maintaining legitimacy and social stability during a disruptive technological transition.

The editorial stance reflects both defensive and strategic considerations. Defensively, the union apparatus needs to demonstrate that it can protect workers from the most predatory forms of AI-driven restructuring, or risk appearing irrelevant to a younger generation of employees in digital and service sectors. Strategically, it participates in shaping a narrative in which China can pursue technological leadership while claiming a distinctive social model, one that contrasts with images of unregulated automation and gig work precarity elsewhere.

This positioning does not erase underlying tensions. Many workers are not formally unionised or lack genuine bargaining leverage within their workplaces, especially in private tech firms and platform companies. Enforcement of labour protections remains uneven, and there is a large grey zone where AI tools intensify workloads or surveillance without leading to outright dismissals. Nevertheless, the articulation of a principle that technological gains should be socially distributed rather than concentrated provides a reference point for workers, regulators and courts.

Algorithmic control, visibility and hidden labour

Debates about sharing the benefits of AI often focus on headline figures: how many jobs might be replaced, how much productivity might increase, what fraction of GDP growth can be attributed to automation. Yet a substantial part of AI's labour impact lies not in visible job losses but in the reconfiguration of work through algorithmic control and the creation of largely invisible labour behind the systems.

Chinese-developed AI models depend on vast quantities of human-labelled data, content moderation, and platform maintenance work, much of it outsourced, low-paid and weakly protected. These workers, who generate the training data that underpin sophisticated models, often lack clear paths to benefit from the subsequent productivity gains. Meanwhile, for employees in call centres, warehouses or delivery services, AI-powered scheduling and monitoring tools can increase pressure, reduce autonomy and tighten performance metrics, even if headcount stays constant.

In this sense, the risk is twofold. First, AI can be used to directly displace segments of the workforce. Second, even where jobs remain, AI systems can become instruments for intensifying labour, extracting more output per worker without commensurate increases in pay or security. When trade unions and courts insist that AI should not be turned into a tool for undermining rights, they are implicitly contesting both forms of impact: the overt substitution of machines for people and the subtler erosion of working conditions through algorithmic management.

Economic models of sharing technological gains

In economic terms, the argument centres on how the surplus generated by AI is divided between capital and labour. If the introduction of an AI system raises a firm's output or reduces its costs, the additional surplus can be decomposed into higher profits, increased wages, better working conditions, lower prices for consumers, or public revenues via taxation. Absent countervailing power, the default tends to favour capital holders and senior executives.

Some analysts formalise this distribution using production functions in which output depends on both traditional labour and an AI-enhanced component, with bargaining over wages determining how gains are shared. While the specific mathematics may vary, the underlying logic is straightforward: when workers lack the power to insist on a share of productivity improvements, their relative position deteriorates even if overall economic output rises. In that scenario, technological advancement can coexist with stagnant or falling real wages and heightened insecurity for large segments of the workforce.

China's institutional framework gives the state tools to intervene in this distribution. Minimum wage policies, social insurance requirements, and judicial interpretations of labour law all shape the bargaining landscape in which AI adoption occurs. If courts repeatedly rule that firms cannot dismiss workers or cut salaries under the pretext of technological change, employers may be pushed to find ways to use AI that complement, rather than replace, existing staff: reassigning employees to higher-value tasks, investing in reskilling programmes, or reducing working hours without cutting pay.

Such complementarity is not automatic and may be more feasible in some sectors than others. High-skill knowledge work, for instance, may lend itself to human-AI collaboration in which professionals use generative tools to enhance their output, while retaining control over final decisions. Low-skill routine work, by contrast, may be more susceptible to straightforward automation. The underlying normative claim - that benefits should be shared - thus implies a need for targeted policies to prevent sectors at greatest risk from bearing disproportionate costs.

Counterarguments and employer perspectives

Critics of strong labour protections in the context of AI often argue that constraining firms' ability to restructure will slow innovation, reduce competitiveness and ultimately harm workers by limiting growth and job creation. From this perspective, allowing employers to freely adopt AI and adjust their workforce is necessary to ensure that domestic firms can keep pace with international rivals, especially in sectors where global competition is fierce.

Employers may also contend that attempts to legislate or adjudicate against AI-driven redundancies misunderstand the reality that technological change inherently involves creative destruction. Protecting existing roles too rigidly, they argue, risks locking labour into obsolete tasks and freezing the economy in less efficient configurations. Some worry that courts declaring AI-driven restructuring illegitimate could create uncertainty, deter investment and encourage firms to relocate to jurisdictions with more permissive regimes.

These objections highlight a genuine tension: how to maintain dynamism and openness to innovation while preventing technology from becoming a one-way conduit for transferring risks and costs onto workers. Proponents of stronger protections respond that the choice is not between technological stagnation and unfettered automation. Rather, it is between different institutional arrangements for managing transitions - some of which spread adjustment burdens across firms, workers and the state, and others of which concentrate them on those with the least bargaining power.

International comparisons and distinctive features

Globally, debates over AI and labour rights span a spectrum. Some advanced economies emphasise data protection and algorithmic transparency, while leaving employment relationships largely governed by existing redundancy and discrimination rules. Others focus on reskilling and social safety nets, aiming to ease transitions without imposing strong constraints on firms' restructuring choices. China's emerging approach, with its combination of strong formal limits on arbitrary dismissal and an official labour narrative emphasising social stability, occupies a distinctive position on this spectrum.

One distinctive feature lies in the central role of the state in both promoting AI development and arbitrating its labour impacts. Government plans and industrial policies actively support AI research, infrastructure and deployment, while the same state apparatus, through courts and labour agencies, signals boundaries for how far employers can go in using AI to cut jobs. This dual role can generate tensions - for instance, between local governments eager to attract high-tech investment and national-level concerns over social unrest - but it also means that labour impacts are not treated as an afterthought.

Another distinguishing element is the prominence of social stability as an explicit policy objective. In a political system highly sensitive to large-scale unemployment or visible worker protests, there is a pragmatic incentive to ensure that AI does not trigger sudden waves of dispossession. Protecting workers from being summarily replaced by AI can thus be understood not only as a matter of justice but also as a tool of risk management for the state.

Why the distributional question matters now

The urgency of clarifying how AI's benefits are shared arises from the speed and breadth of current deployment. Unlike earlier waves of automation targeted mainly at manufacturing or narrowly defined clerical tasks, contemporary AI systems reach into creative industries, legal and financial services, health care administration and even elements of management decision-making. As more layers of the economy become susceptible to algorithmic substitution or augmentation, the number of workers whose roles are reshaped by AI expands far beyond a single sector.

In this context, a clear public stance that technological advances should not be repurposed as mechanisms for eroding labour rights performs several functions. It sets expectations for employers considering AI-driven restructuring, warning that they may face legal and reputational risks if they treat workers as disposable inputs in a cost-minimising exercise. It gives workers and their representatives a discursive and legal basis to challenge unfair practices, from unjustified lay-offs to exploitative algorithmic management. And it positions the broader society to debate, rather than passively accept, the terms under which AI is integrated into everyday economic life.

There is no guarantee that such a stance will fully prevent the concentration of technological gains in the hands of a small group of actors. Powerful firms with access to capital, data and engineering talent remain well-placed to dominate AI markets and capture outsized returns. However, by embedding labour-protective principles in law, judicial practice and official discourse, China is attempting to tilt the playing field away from a pure race to the bottom in labour costs. Whether this experiment succeeds will depend not only on rulings and editorials but on continuous enforcement, worker organisation and the willingness of policymakers to adjust course as new challenges emerge.

What is already clear is that the contest over AI is not only about whose models are most powerful or whose infrastructure is most advanced. It is also about whose lives improve, whose become more precarious, and who has a say in that process. Framing technological advancement as something to be shared by society at large, rather than wielded as a tool to weaken the bargaining power of those who perform the work, draws a line between two possible futures: one in which AI deepens existing hierarchies of power, and another in which its benefits are mediated through institutions that recognise workers not as expendable inputs, but as stakeholders with rights that do not vanish when a new machine comes online.

Term: Perplexity - Artificial intelligence

"In computer science and natural language processing, 'perplexity' is a mathematical measurement of uncertainty. It evaluates how effectively an AI model predicts the next word in a sequence or sentence." - Perplexity - Artificial intelligence

Human language is structured enough to be predictable yet rich enough to surprise. Any system that tries to generate or understand text must therefore manage a tension between recognising familiar patterns and handling rare or novel expressions. The central technical question is how to measure whether such a system is predicting linguistic continuations in a way that aligns with real usage, rather than merely memorising or guessing.

Perplexity enters at precisely this point as a quantitative lens on predictive behaviour. It converts a model's entire probability distribution over possible next words into a single scalar that captures how uncertain, or "confused", the model is when faced with real data. Low values indicate that the model assigns high probability to what humans actually say; high values indicate that the model spreads probability mass thinly or places it on implausible options. Because this single figure is computed systematically across large corpora, it has become deeply embedded in how researchers train, compare, and refine language models.

Uncertainty, surprise, and predictive distributions

Any predictive text system maintains, implicitly or explicitly, a probability distribution over possible next tokens given the context. If we denote the context by (for "history") and the next token by , the system internally represents a conditional distribution . This distribution encodes both what the system believes is likely and how strongly it believes it. When the actual next word appears, the value is a direct measure of how well its expectations matched reality.

Entropy and related information-theoretic quantities provide a way to aggregate these local assessments. The Shannon entropy of a distribution over a discrete vocabulary is defined as . This quantity grows when the distribution is flatter (more uncertain) and shrinks when it is sharply peaked around a few options (more certain). However, entropy is expressed in bits or nats depending on the logarithm base, which is not immediately intuitive. Perplexity bridges this gap by re-expressing entropy as an effective number of equally likely choices.

Perplexity in mathematical terms

Formally, perplexity is defined as the exponential of the average negative log probability that a model assigns to a sequence of tokens. Suppose we have a sequence drawn from a test corpus, and a model that assigns probability to each token given its history . The average negative log-likelihood per token is

Perplexity is then defined as

If natural logarithms are used, perplexity corresponds to raised to the entropy rate; if base-2 logarithms are used, perplexity is , where is the entropy in bits. Intuitively, a perplexity of, say, 50 means that the model behaves as if it were choosing among about 50 equally likely options at each step, even though in reality its distribution may be uneven.

Several properties follow immediately from this definition. Perplexity is always at least 1, with 1 corresponding to a model that assigns probability 1 to the correct next word at every position. It is minimised when the model represents the true data-generating distribution and grows as the model's predictions deviate from the empirical distribution. Because it aggregates over the entire sequence, it penalises systematic miscalibration rather than occasional errors.

Relation to log-likelihood and cross-entropy

Perplexity is closely linked to standard statistical objectives used in training language models. Maximum likelihood training attempts to find parameters that maximise the log-likelihood . Minimising the average negative log-likelihood is equivalent to minimising cross-entropy between the empirical data distribution and the model's distribution. Because perplexity is a monotonic transformation of , minimising perplexity is identical to maximising likelihood or minimising cross-entropy.

This equivalence is practically important. During optimisation, models are updated to reduce loss, but when results are reported to other researchers or stakeholders, perplexity provides a more interpretable metric. A drop in perplexity from, say, 80 to 40 is easy to understand as halving the effective number of equally likely options, which in turn suggests much sharper predictions. This interpretability makes perplexity a convenient benchmark when comparing architectures, training regimes, or datasets.

Practical meaning for model quality

In applied natural language processing, perplexity is used in two broad ways: as a training objective proxy and as a diagnostic for model behaviour. During development of language models, improvements in architecture or training data selection are often evaluated by their effect on perplexity on held-out corpora. Lower perplexity generally correlates with better performance in generative tasks such as language modelling, text completion, and machine translation.

For example, when an organisation transitions from a smaller recurrent neural network to a transformer-based model, it typically observes a substantial drop in perplexity on standard benchmarks. This reduction is evidence that the new model captures longer-range dependencies and richer linguistic structure. In turn, this tends to yield more fluent text generation and more accurate predictions of rare but grammatically and semantically appropriate tokens.

Perplexity also functions as a sanity check on training dynamics. Sudden spikes can indicate numerical instability, data pipeline corruption, or misconfigured learning rates. Gradual stagnation of perplexity during training suggests that the model has reached the limits of what can be extracted from the current data with the given capacity. Monitoring perplexity across domains or languages can reveal where a model is under-exposed or miscalibrated, guiding targeted data collection or fine-tuning.

Local versus global uncertainty

Although perplexity is defined as a global average over a test set, it is often insightful to inspect its local contributions. The term can be interpreted as the surprise associated with observing given the context. High local surprise may stem from genuinely rare words, idiomatic expressions, abrupt topic shifts, or areas where the training data was thin. By examining segments with high average surprise, practitioners can diagnose specific weaknesses such as poor handling of code-switching, domain-specific jargon, or unusual syntactic patterns.

This local view is crucial for understanding that models with similar overall perplexity may fail in different ways. Two systems might achieve comparable averages yet differ sharply in how they allocate uncertainty: one may be consistently moderately uncertain, while another is confident most of the time but catastrophically wrong in certain regimes. Perplexity alone does not distinguish these patterns, prompting complementary analyses such as calibration curves, error typologies, and task-specific evaluations.

Parameter meanings and modelling choices

In the classical statistical language modelling literature, perplexity often appears in the context of n-gram models. An n-gram model approximates the conditional distribution by considering only the previous tokens: . Parameters in such models are counts and smoothing coefficients that adjust for sparsity. Perplexity provides a direct way to quantify how well these approximations capture real sequences and how much improvement is obtained by increasing or introducing better smoothing.

In modern neural language models, parameters are continuous weights in deep architectures. Although there is no closed-form mapping from individual parameters to perplexity, some structural choices have predictable effects. Increasing model width and depth, enlarging context windows, and using richer positional encodings typically reduce perplexity up to a point, after which overfitting and diminishing returns appear. Likewise, training on larger and more diverse corpora tends to lower perplexity, but domain mismatch between training and evaluation data can negate these gains.

Temperature and related decoding parameters, which control randomness in sampling, do not affect perplexity directly because perplexity is calculated on the underlying distribution, not on generated samples. However, severe miscalibration in the distribution - for example, distributions that are too flat or too peaked - will show up as elevated perplexity relative to an ideal model.

Competing and complementary evaluation metrics

Despite its widespread use, perplexity is not a universal proxy for downstream task performance. Many benchmarks in natural language processing involve structured prediction, reasoning, or interaction with users, for which task-specific metrics are more appropriate. Accuracy, F1 score, BLEU, ROUGE, and human judgements of fluency or relevance often provide a more direct assessment of practical utility.

There are two main limitations in interpreting perplexity. First, it is sensitive to the tokenisation scheme: models that operate on different vocabularies - words, subwords, or characters - are difficult to compare directly. A model with a finer-grained tokenisation may have higher perplexity per token but similar or better performance when measured per character or per word. Second, perplexity ignores meaning: assigning very high probability to fluent but semantically inappropriate continuations can still yield favourable perplexity scores if they match surface statistics.

These limitations have led to a view in which perplexity is necessary but not sufficient. It remains valuable as a basic measure of language modelling quality, especially for comparing variants of the same model family under identical tokenisation and data conditions. However, it is supplemented by application-specific evaluations that capture semantics, factual accuracy, and robustness.

Perplexity in the broader AI landscape

As large language models have moved from research prototypes to widely deployed systems, perplexity has acquired a dual role. It continues to serve as a core internal metric during pretraining on massive corpora, where its reduction signals better compression of linguistic regularities and more efficient representation learning. At the same time, its direct visibility to end users has diminished, replaced by qualitative assessments of helpfulness, harmlessness, and reliability.

Behind the scenes, however, perplexity still matters for engineering decisions. Model distillation, where a large model trains a smaller one, often relies on matching probability distributions and thus on controlling perplexity gaps between teacher and student. Domain adaptation, where a general model is fine-tuned on specialised text such as legal or medical documents, is evaluated by domain-specific perplexity improvements. Even in retrieval-augmented systems, where external information is fetched at query time, perplexity on the combined context-plus-document input informs how well the model integrates retrieved evidence.

In interactive settings, such as conversational agents and AI-powered search tools, perplexity can be monitored as a proxy for the model's comfort level with a query. High perplexity on user instructions or on retrieved content may indicate that the model is extrapolating far beyond its training distribution, which correlates with greater risk of hallucination or misinterpretation. This has motivated research into using perplexity-like measures to trigger fallback behaviours, such as requesting clarification, restricting outputs, or escalating to human review.

Debates and tensions around its use

The reliance on perplexity has sparked several debates in the research community. One line of argument holds that over-optimising for perplexity can encourage models that excel at shallow pattern matching but underperform on compositional reasoning or factual consistency. Since perplexity is indifferent to whether a prediction is logically grounded, it may reward models that memorise long-range patterns in the training data without learning general principles.

Another concern is distributional shift. Perplexity is usually measured on static test sets, but deployed systems face evolving language, emerging topics, and changing user behaviour. A model with strong perplexity numbers on past news articles, for example, may exhibit much higher perplexity on discussions of novel technologies, slang, or events that occurred after its training cut-off. This gap underscores the need for continual evaluation and possibly continual training, as well as for metrics that better track real-world performance.

There is also a methodological tension between comparing models on standardised benchmarks and tailoring evaluation to specific deployment contexts. Standard corpora facilitate apples-to-apples perplexity comparisons across architectures and research groups, but they may not reflect the specialised domains where a model will actually operate. Conversely, domain-specific corpora provide more relevant perplexity estimates but reduce comparability. Balancing these considerations remains an active area of practice rather than a solved theoretical issue.

Why perplexity still matters

Despite critiques, perplexity continues to occupy a central position because it aligns naturally with how generative language models are trained and used. It connects the statistical foundations of probability distributions and entropy with practical questions about predictive performance. It is simple enough to compute and interpret, yet sensitive enough to reveal meaningful differences when architectures, datasets, or optimisation strategies change.

Moreover, perplexity reflects a deeper conceptual challenge in artificial intelligence: representing and managing uncertainty in complex, structured domains. Human communicators constantly navigate uncertainty about what will be said next, what listeners already know, and how a conversation might unfold. Language models, though purely computational, must grapple with the same uncertainty at scale. Perplexity does not capture everything about this process, but it provides a disciplined way to quantify how well models anticipate linguistic realities.

As AI systems expand beyond text into multimodal settings and more interactive applications, analogous questions arise about predicting the next frame in a video, the next action in a sequence, or the next user response in a dialogue. Extensions of perplexity and its underlying cross-entropy framework are likely to remain part of the evaluative toolkit in these areas as well. In that sense, understanding perplexity is not only about language models but about a broader approach to measuring how effectively artificial systems handle uncertainty in open-ended environments.

Quote: Bryan Catanzaro - Vice president of applied deep learning at Nvidia

"For my team, the cost of [AI] compute is far beyond the costs of the employees." - Bryan Catanzaro - Vice president of applied deep learning at Nvidia

In the current generation of artificial intelligence deployment, the binding constraint for many organisations is no longer talent but access to affordable compute at scale. That inversion of the traditional cost structure is what turns remarks from senior technologists into a broader economic signal: for some cutting-edge teams, the recurring bill for accelerated hardware, cloud instances, networking, and power now exceeds the wage bill for the engineers designing and operating the systems. This is not just an accounting oddity; it alters how firms evaluate automation, how investors price AI strategies, and how policymakers should interpret predictions of rapid job displacement.

From cheap silicon and expensive people to expensive silicon and leveraged people

For several decades, digital transformation followed a familiar pattern: hardware and basic infrastructure costs per unit of computation fell predictably, while skilled labour remained scarce and expensive. Organisations hired more software engineers to sit atop an increasingly cheap computational substrate. In that world, the canonical argument for automation was straightforward: once a process was codified in software, the marginal cost of running it again was negligible compared with paying an additional human to do the same work.

What has changed with frontier AI systems is the scale and intensity of computation required to deliver competitive performance. Training large language models and vision systems involves running vast numbers of parallel operations across specialised GPUs and associated accelerators, often for weeks, on clusters that can cost tens or hundreds of millions of dollars in capital expenditure for hyperscalers and large enterprises. Even when organisations do not own the hardware, their cloud providers pass through these capital and operating costs via metered pricing. As a result, the unit of analysis for AI economics has shifted from "salary per employee" to metrics such as cost per GPU hour, cost per million tokens, and cost per inference request.

Industry observers now describe AI infrastructure as a new class of heavy industry: data centres designed around specialised accelerators, redundant power feeds, and advanced cooling, with aggregate global spending that consultancy estimates place in the trillions of dollars by 2030. That capital intensity explains why some AI teams report that, at the margin, paying for more compute to improve a model or scale an application is financially weightier than hiring additional engineers to refine prompts, build interfaces, or clean data.

Why compute dominates the cost stack

There are several mechanisms behind compute eclipsing labour in AI projects.

First, state-of-the-art models are extremely large. Modern large language models and multi-modal systems often contain hundreds of billions or even 1 trillion parameters. Each training run requires repeated passes over large datasets, with backpropagation and optimisation algorithms applied across every parameter. The computation required scales roughly linearly with the number of parameters and data tokens, and in practice teams often run multiple training, fine-tuning, and ablation cycles. That translates into millions of GPU hours, even when using the most efficient hardware and software stacks available.

Second, inference - the process of serving model outputs to users - imposes ongoing costs that grow with adoption. Training is a one-off or periodic capital-like expense, but inference is an operational expense that scales with queries. Industry frameworks therefore emphasise cost per token as a central metric: the all-in cost to produce each output token, incorporating hardware depreciation, energy, data centre overheads, networking, and software optimisation. Even small differences in cost per million tokens can compound dramatically when applications serve millions of users or integrate AI into high-frequency workflows.

Third, the energy footprint of frontier AI is substantial. High-end GPUs draw significant power, and data centres require additional energy for cooling and ancillary systems. Energy prices vary geographically, but in many locations they are high enough that power constitutes a large share of the total cost of ownership for AI infrastructure. Analysts have therefore started speaking of "intelligence per megawatt" as a key performance dimension. When firms compare this stack of costs with the wages of knowledge workers, the balance can tilt unexpectedly toward hardware and energy spending.

Fourth, there is a structural asymmetry between compute and labour costs. Employee salaries are relatively predictable and can be adjusted slowly via hiring freezes, attrition, or compensation changes. Compute costs, by contrast, can spike rapidly if product usage grows or if teams run more experiments than anticipated. In startups, venture capital typically funds both headcount and infrastructure, but some investors report portfolio companies spending more than 80% of their capital raised on compute resources, dwarfing wage bills.

The empirical picture: AI not yet cheaper than humans

These mechanisms are now visible in empirical work. One 2024 study from MIT examined where AI systems could perform visual tasks at or near human level, and then compared the cost of machine versus human performance. The researchers concluded that automation was economically viable in roughly 23% of roles where vision is central to the job, meaning that in about 77% of cases, it remained cheaper to pay humans than to deploy AI. The issue was not capability - the models could often do the tasks - but economics: hardware, energy, and infrastructure outweighed labour costs.

Macro-level data on AI expenditure reinforce this picture. Big technology firms alone have announced around 740 billion in capital expenditures in a single year, largely driven by AI data centre build-out, representing a 69% increase over the prior year. Other analysis suggests that AI-related expenditures could reach 5,2 trillion by 2030 under central scenarios, and as high as 7,9 trillion under more aggressive build-out, with data centre and IT equipment accounting for the bulk of this spending. Against those numbers, even generous headcounts of highly paid engineers, researchers, and product staff occupy a smaller share of the cost base than might be expected.

At the same time, external observers note that, despite the scale of these investments, there is still limited aggregate evidence of AI-driven productivity gains across the economy. Budget analysts and academic studies point out that, so far, there is no broad-based data showing AI displacing jobs at scale or dramatically boosting measured output per worker, even as tech sector layoffs have accelerated. That divergence between spending and measured productivity raises the question of whether the current wave of AI investment is front-loaded - infrastructure built ahead of realised returns - or whether some fraction will prove to be misallocated capital.

Strategic tension: build now, pay later

This leads to a central strategic tension. On one side is the argument that AI is an infrastructure revolution akin to electrification or the early internet, requiring enormous upfront capital before productivity gains show up in statistics. Firms that invest early, this view holds, will establish competitive moats via proprietary data, trained models, and optimised infrastructure. For such firms, the fact that compute currently costs more than labour may be beside the point; they are laying the foundations for future economies of scale, where the amortised cost per unit of AI output falls sharply as utilisation rises.

On the other side is a more sceptical view, which emphasises opportunity cost and path dependence. If AI systems are currently more expensive than humans for many tasks, especially outside narrow high-value niches, then replacing workers prematurely may destroy value rather than create it. Companies that chase AI for its own sake, without rigorous cost-benefit analysis, risk saddling themselves with fixed infrastructure commitments and ongoing compute bills that are difficult to roll back. This argument is bolstered by evidence that many firms do not fully understand their AI cost structures, focusing on headline model access fees or GPU rental rates rather than total cost per token or per workflow.

These perspectives are not mutually exclusive. It is possible that some organisations are overbuilding, while others are rationally investing in infrastructure that will underpin future competitive advantage. What unites them is an underlying bet: that the cost of compute will fall fast enough, and the productivity benefits of AI will rise high enough, to justify today's discrepancy between machine and labour costs.

Falling unit costs versus rising aggregate spend

Industry roadmaps and analyst reports forecast significant reductions in the unit economics of AI over the next few years. Hardware generations such as Nvidia's Blackwell architecture promise up to 30× gains in inference performance at similar power budgets compared with earlier accelerators. Software improvements - better compilers, quantisation techniques like FP4 precision, more efficient attention mechanisms, and mixture-of-experts routing - all work to reduce the computational load per unit of useful output. Gartner-style forecasts point to the cost of running inference for models with 1 trillion parameters dropping by more than 90% over a four-year horizon.

If realised, those gains could radically alter the relative cost of compute and labour. A workflow that is uneconomic today because each AI call is expensive might become viable once the cost per million tokens falls below some threshold. In that future, the remark that compute costs more than employees would be overtaken by a new reality in which compute is cheap enough that the main question becomes how to reconfigure organisations to exploit it.

However, even as unit costs fall, aggregate spending may still rise. The classic rebound effect applies: cheaper computation tends to expand the range of feasible applications and increase total usage. Organisations that pay less per token may respond by embedding AI into more products, workflows, and services, multiplying the total number of tokens generated. If spending on AI grows from 5,2 trillion to 7,9 trillion by 2030, a large part of that increase will likely reflect expanded scope, not just higher prices. The result is a paradox: individually, each unit of compute may become cheaper and more efficient; collectively, compute may remain the largest single line item for AI-heavy firms.

Employment, displacement, and the cost paradox

The fact that compute can cost more than employees complicates narratives about AI-driven job displacement. From a firm's perspective, automation only makes sense when the total cost of designing, training, deploying, and maintaining an AI system is lower than - or at least justifiable relative to - the wage cost, management overhead, and performance variability of human workers. When AI is more expensive, substituting capital for labour purely on cost grounds is irrational.

This does not mean AI will not change employment. Instead, it suggests a more nuanced pattern of complementarity and selective substitution. In domains where human labour is extremely costly or scarce, such as high-end legal services, algorithmic trading, or complex simulation, even expensive compute may be a bargain. In mass-market customer service or routine back-office work, by contrast, the current cost structure favours augmenting workers with AI tools rather than replacing them outright. The MIT study's finding that only around 23% of vision-centric jobs are currently economically automatable illustrates how narrow the immediate substitution window may be.

The paradox is that headlines about companies spending more on AI infrastructure than on salaries coexist with data showing limited net job losses attributable directly to AI. Part of the explanation is temporal: firms are investing ahead of adoption, building capabilities before fully restructuring their workforces. Another part is strategic: some firms see AI as a growth tool rather than a cost-cutting tool, aiming to enable new products and services rather than simply replacing existing staff.

Pricing models and hidden subsidies

One reason many users and even corporate customers underestimate the true cost of AI compute is the structure of pricing models. A significant portion of the market relies on flat subscription charges or simple usage tiers that do not map transparently to underlying infrastructure costs. For light users, this can be attractive: they pay a fixed fee and rarely hit limits. For heavy users, however, the provider may be effectively subsidising usage, especially if the subscription was priced before providers had a clear picture of real-world load.

Reports of AI software fees rising by 20% to 37% over a year indicate providers adjusting to this reality. As cost pressures mount - from energy, hardware procurement, and the need to recoup massive capital investments - providers are likely to shift toward more granular, usage-based pricing that reflects cost per token or per request more accurately. When that occurs, more enterprises will discover that their apparent labour savings are offset by higher-than-expected compute bills.

This evolution will bring AI closer to other utilities: electricity, cloud storage, and bandwidth. In each case, users ultimately pay for the marginal resource consumed, and efficient usage becomes a competitive advantage. Just as cloud-native firms learned to optimise workloads to reduce compute and storage charges, AI-native firms will need to optimise prompts, context lengths, caching strategies, and model architectures to minimise unnecessary tokens and reduce idle capacity.

Why the remark matters

The observation that compute can be more costly than employees is important for several constituencies.

For executives and boards, it underscores the need for rigorous capital allocation in AI initiatives. Projects should be evaluated not only on potential strategic upside but also on fully burdened compute economics: total cost per workflow, sensitivity to usage spikes, and exposure to future hardware and energy price shifts. In an environment where tech companies have already announced hundreds of billions in AI-related capital expenditure, misjudging these factors can have material consequences for profitability and competitive positioning.

For investors, the remark acts as a reminder that not all AI spending is value-creating. A significant share may be speculative or defensive, driven by fear of missing out rather than clear use cases. Distinguishing between firms that translate compute into durable revenue and those that merely accumulate expensive infrastructure will be a central task over the next decade.

For policymakers and labour economists, recognising the current cost structure is essential when interpreting forecasts of rapid, sweeping job automation. If AI is still more expensive than humans for the majority of tasks, then near-term labour market disruption is likely to be more contained and sector-specific than some narratives suggest. This does not eliminate the long-term risk of displacement as unit costs fall, but it introduces a window in which policy can focus on adaptation: training, re-skilling, and ensuring that productivity gains, when they arrive, are broadly shared.

Finally, for engineers and product teams, the remark is a design constraint. It implies that building AI systems is not just a problem of maximising accuracy or capability; it is also a problem of optimising for economic viability. Model selection, quantisation choices, caching, retrieval strategies, and system architecture all affect compute consumption. Teams that learn to treat tokens, GPU seconds, and watts as scarce resources, on par with developer time, will be better positioned to create sustainable AI products.

As AI infrastructure matures, the relative prices of compute and labour will continue to evolve. The present moment, in which a leading practitioner can credibly say that the compute bill dwarfs the wage bill, is a snapshot in a longer trajectory. Whether history records it as an early, capital-intensive stage on the way to widely affordable machine intelligence, or as evidence of an overcapitalised boom, will depend on how effectively organisations turn expensive silicon into genuine productivity.

Quote: Dario Amodei - Founder and CEO, Anthropic

"Those who don't kind of see what's coming, who don't identify the moats they have, they're going to have a really hard time." - Dario Amodei - Founder and CEO, Anthropic

The practical problem is not simply that AI capability is improving quickly, but that the basis of competitive advantage is shifting faster than many organisations can recognise it. In a market where model quality, distribution, infrastructure, data access, and trust are all being repriced at once, the winners are likely to be those who can identify which advantages are durable before the rest of the field realises that the old ones have decayed.

That is the deeper force behind Dario Amodei's warning. It is not a casual remark about caution; it reflects a strategic view that many participants in the AI economy are mistaking visible momentum for defensible position. Anthropic's own public profile has made Amodei one of the clearest voices on the concentration of power in the AI era, and his concern appears to be that speed alone can create a false sense of security.

What a moat means in the AI context

A moat in the classic business sense is any structural feature that makes profits more resistant to competition. In AI, that can mean several different things at once: proprietary distribution, developer loyalty, enterprise integration, talent depth, compute access, safety credibility, regulatory readiness, or a feedback loop that improves the product faster than rivals can copy it. The difficulty is that AI moats are often unstable early on. A capability that looks like an enduring edge one year may be commoditised the next by better open models, cheaper inference, or a rival's stronger distribution channel.

That instability matters because many companies have been tempted to treat AI as a feature layer rather than a strategic reordering. But the recent market conversation has increasingly moved towards the idea that AI will not merely be sold into existing businesses; it will reorganise them from the inside. Commentary around the so-called AI rollup thesis argues that investors are buying labour-intensive businesses and rebuilding them around AI so that economics begin to resemble software rather than services. If that thesis proves correct, then the old source of value is not just under pressure; it is being redefined.

Why speed makes identification harder

Amodei's warning lands because the pace of improvement is itself part of the competitive landscape. He has said that cognitive ability in frontier systems can be doubling every four to 12 months, a pace that would make conventional strategic planning dangerously slow. When the underlying capability curve is that steep, the shelf life of an advantage shortens. What looks like a moat may actually be a temporary lead created by timing, capital, or first-mover publicity.

Quote: Andrej Karpathy - Anthropic (Openai Founder, formerly head of Tesla AI)

"You can give [Claude Fable 5, the same underlying model as Mythos but with added safeguards] a lot more ambitious tasks than what you're used to, the model 'gets it' and it will just go, and it's never felt this tempting to stop looking at the code at all." - Andrej Karpathy - Anthropic (Openai Founder, formerly head of Tesla AI)

Modern software development is being pulled towards a regime where human programmers increasingly specify intent while machines decide the details. That shift is most visible not in abstract benchmarks but in the psychological moment when an expert developer realises they can hand over a far more ambitious task than before, watch an AI system autonomously decompose and implement it, and feel genuine temptation to stop inspecting every line it produces. For someone steeped in traditional notions of craftsmanship and code review, that temptation is both intoxicating and alarming.

The move from instructions to intent

The underlying transition is from imperative programming, where humans micromanage every step, to a declarative style where they specify success criteria and let an AI agent find a path. Historically, a senior engineer might spend hours designing architecture, writing scaffolding, and orchestrating tools; now, high-capability models can generate coherent multi-file projects, manage dependencies, restructure modules, and even propose test suites to validate their own work. In that world, the bottleneck shifts from typing speed or API recall to the clarity and completeness of the human specification.

This is what makes the ability to give a model substantially more ambitious tasks so significant. When an AI system can handle not just a function or a bug fix but an end-to-end feature, a migration, or a refactor across tens of files, the human role changes. The developer becomes more of a product and safety architect: deciding goals, constraints, and trade-offs, then auditing whether the agent met them. The quote speaks directly to this: work that once had to be decomposed into micro-prompts can now be expressed as a single high-level directive, with the model reliably filling in the operational gaps.

Why this particular endorsement matters

The significance of that shift is amplified by who is speaking. Andrej Karpathy is not a casual user experimenting with consumer tooling but a foundational figure in modern deep learning and applied AI. He was a founding member of OpenAI and later headed Tesla Autopilot, leading large teams building vision and planning systems for real-world safety-critical autonomy. He has also been one of the most visible educators in the field, teaching tens of thousands of practitioners how neural networks work and how to reason about their failure modes.

That background makes his sense of surprise at feeling "behind" as a programmer, and the ego hit of giving more of the work to AI, noteworthy. When someone with deep understanding of model limitations and training quirks reports that a code-focused model "gets it" on larger, more complex tasks, it is not naive enthusiasm but a statement grounded in long experience of what usually goes wrong. It suggests a step change in practical reliability, especially on extended coding sessions.

Claude Fable 5, Mythos and safeguards

The specific model involved, Claude Fable 5, sits in a deliberately structured product family. Anthropic describes Fable 5 as sharing the same underlying model as Claude Mythos, but with additional safeguards and alignment layers. In practice, that means the core capability for long, multi-step reasoning and coding is retained, while the system has tighter policies around potentially harmful outputs and more conservative behaviours in ambiguous domains. Mythos is aimed at frontier, less constrained exploration; Fable, at general-purpose work where safety, compliance, and predictability matter.

Karpathy himself characterises Fable 5 as a "major-version-bump-deserving step change", particularly on long and difficult tasks. Reports and demos around the release highlight stronger performance not only on standard coding benchmarks but on real-world development flows: navigating large repositories, performing multi-file edits, and maintaining context over extended sessions. The result is a system that feels less like an autocomplete gadget and more like a junior engineer who can hold a problem in mind over hours of work.

Crucially, the "added safeguards" do not just refer to refusal policies. They also encompass training and inference-time measures that make the model more robust to prompt injection, reduce hallucinations, and bias it toward verifiable operations like running tests or inspecting diffs instead of bluffing. That combination - high capability plus strong guardrails - is what makes handing over more ambitious tasks psychologically viable. The user is not simply trusting a stochastic parrot; they are interacting with a toolchain engineered for cautious autonomy.

Karpathy's journey into agentic coding

To understand the deeper significance of the quote, it helps to situate it within Karpathy's broader ideas about software. In recent years he has described a transition towards what he calls "Software 3.0" or "agentic coding". Earlier eras could be caricatured as follows: in Software 1.0, humans wrote explicit logic; in Software 2.0, humans trained models but still wrote the surrounding infrastructure; in Software 3.0, AI systems increasingly write, test, and maintain significant portions of the codebase themselves, guided by human intent and oversight.

Within that frame, he has promoted practices like "vibe coding", where a developer converses with an AI assistant, iteratively refining prompts and reading outputs rather than manually hand-crafting every function. The point is not laziness but bandwidth: by offloading boilerplate and low-level wiring, humans can spend more time on product thinking, architecture, and evaluation. Yet he has also been candid about the danger of "brain atrophy" if humans stop engaging deeply with technical substance and become mere prompt routers.

His move to Anthropic, announced publicly on X, is explicitly about pushing this paradigm further. He is joining the Claude pre-training team, with a mandate to build a sub-team focused on using Claude itself to accelerate pre-training research. That is, he is not only using AI to write ordinary application code but using AI agents to help design, run, and analyse the experiments that produce the next generation of AI models. Some observers describe this as laying the groundwork for recursive self-improvement, where systems contribute directly to their own advancement.

The temptation to stop reading the code

The most charged part of the quote is not the praise for task capability but the admission that "it's never felt this tempting to stop looking at the code at all". That sentence crystallises a new risk frontier. Up to now, cautious practitioners have recommended heavy human inspection of AI-generated code: checking logic, scanning for security flaws, reviewing for maintainability. Those practices are time-consuming, but they preserve a culture where humans remain accountable for what ships.

As model quality improves, the marginal benefit of reading every line may appear to shrink. When the output often looks clean, idiomatic, and passes tests, the pressure to skim rather than scrutinise grows stronger. That temptation is exacerbated by business incentives. If an AI agent can implement a feature in 1 hour that would take a human 10 hours, organisations will be driven to capture that 9-hour gain, especially under competitive pressure. Deep review may be cast as optional overhead rather than mandatory safety.

This dynamic is not unique to coding. In aviation, pilots became less hands-on as autopilots grew more reliable, leading to worries about skill decay; yet in rare edge cases, human intervention remained vital. The same pattern looms in software: as AI-generated code becomes the default, there is a risk that fewer engineers retain the ability to reason from first principles when the system fails in a novel way.

Strategic and technological tension

The tension, then, is between speed and scrutiny, between trusting an increasingly competent agent and insisting on human understanding. On one side lies the productivity windfall: AI can manage dependency graphs, propose architecture refactors, and generate regression tests at a pace that would overwhelm any human team. On the other side lies epistemic opacity: large language models generate code via pattern completion, not explicit formal derivation, and even when the code passes tests, it may encode subtle bugs, non-obvious security weaknesses, or performance pathologies.

In safety-conscious organisations, this tension will likely be addressed with layered controls. For critical systems, one can imagine a workflow where an AI agent proposes changes, another independent agent attempts to break or exploit them, and human reviewers arbitrate. For less critical contexts, teams may accept a higher degree of automated autonomy, using telemetry and canary deployments to catch regressions in production.

Technologically, the quote points to a world where coding models are integrated deeply into development environments as persistent agents rather than stateless assistants. In that world, the system remembers project history, tracks unresolved issues, and maintains a map of the codebase. This is already visible in the way tools like Claude Code are embedded into full IDE surfaces where generation, testing, and git operations happen in one loop. The practical question is not whether such agents will exist but what guardrails and observability layers they will carry.

Anthropic's safety-first positioning

Anthropic has invested heavily in a brand and research agenda built around "constitutional" AI and safety. That approach involves specifying normative guidelines that models are trained to follow, and then auditing behaviour against those guidelines. For coding, that can be extended into concrete policies: refuse to write insecure patterns, prefer constant-time implementations in cryptographic contexts, suggest mitigation when encountering user-supplied input.

Fable 5's positioning as "Mythos but safe" reflects a belief that potential harms can be reduced without sacrificing too much capability. Karpathy's enthusiasm suggests that, at least in his workflows, the safeguards are not experienced as a hindrance but as a trust multiplier. He can instruct the model more ambitiously precisely because he expects it to act conservatively when it encounters sensitive operations and to avoid reckless actions like deleting large portions of a repository without confirmation.

Yet there remains an unresolved debate over how far safety techniques can go in mitigating risks that emerge from sheer scale and generality. Even a strongly aligned model may generate exploitable code when given innocuous prompts, simply because the space of correct-looking but vulnerable implementations is vast. Critics argue that this cannot be fully addressed by refusal policies and that deep formal methods or language-level safety guarantees will be necessary. The temptation to "stop looking at the code" must be evaluated against that backdrop.

Debates and objections

There are at least four major lines of objection or concern surrounding the world implied by the quote.

First, there is the professional identity and labour market concern. If AI tools can handle an increasing share of coding, especially the more routine or boilerplate-heavy parts, junior roles may shrink, making it harder for new developers to gain experience. Karpathy himself acknowledges a crossroads between "brain atrophy" and skill evolution, where humans must decide whether to re-skill towards higher-level system design and evaluation or risk being displaced.

Second, there is the epistemic reliability concern. Benchmarks can show impressive averages, but systems are still brittle on rare edge cases, poorly specified tasks, or ambiguous requirements. A sense that "the model gets it" can mask the fact that its understanding is statistical, not semantic in a human sense. Critics worry that as trust grows, organisations will deploy AI-generated code beyond domains where its failure modes are well characterised.

Third, there is the self-referential risk of using AI to build the next generation of AI. The work Karpathy is taking on at Anthropic involves using Claude to accelerate pre-training research itself, potentially moving towards recursive self-improvement. Enthusiasts argue that this is necessary to make progress at the current frontier, where experiments are too numerous and complex for purely human pipelines. Skeptics warn that errors, biases, or misalignments may be amplified if AI-driven research loops are not carefully constrained and audited.

Fourth, there is the cultural concern. Software engineering has long valued code readability not only for maintainability but as a vehicle for knowledge sharing. If more of the codebase is generated and fewer humans read it deeply, tacit knowledge may concentrate in the behaviour of models rather than in the minds of engineers. Some fear a loss of craftsmanship and a drift towards opaque systems even within a single organisation.

Why this moment matters

Despite these concerns, the practical direction of travel is clear. Developers are already wiring multiple frontier models into a single development surface, choosing per-task which to call, whether Claude, GPT, or others, based on performance and cost rather than vendor loyalty. Tools that bundle coding, testing, and version control into agentic workflows are proliferating. The quote captures a threshold where these tools no longer feel like experimental sidekicks but like the primary engine of implementation.

From a strategic perspective, this changes how organisations think about their software capability. Instead of asking how many engineers they can hire, they will ask how effectively they can orchestrate AI coding capacity: prompt libraries, evaluation harnesses, and safety procedures become as important as hiring pipelines. Companies that embrace this shift thoughtfully will invest in engineers who are excellent at specifying intent, designing tests, and auditing AI proposals - a different profile from traditional full-stack roles.

For individual developers, it poses a challenge and an invitation. The challenge is to resist the laziness of unexamined trust while also resisting nostalgia for a world where writing every line oneself was feasible. The invitation is to climb the abstraction ladder: to become better at defining product goals, at thinking in systems, at debugging not just functions but entire AI-assisted workflows.

Karpathy's experience with Claude Fable 5 illustrates that frontier models are now strong enough to make this shift emotionally palpable. When a veteran practitioner feels tempted to stop reading the code, that is not a signal to give up scrutiny, but it is evidence that the agent has crossed a qualitative threshold. The world of software will be shaped by how we respond to that feeling: whether by surrendering to it, ignoring it, or deliberately building new practices, tools, and norms that harness its power without abandoning responsibility.

Term: Absorption costing - Managerial accounting

"Absorption costing, also known as full costing, is a managerial accounting method that captures and assigns all manufacturing costs to the specific products being produced. Under this system, the unit cost of an item absorbs every single expense required to get it ready for sale, including both fixed and variable costs." - Absorption costing - Managerial accounting

Profitability in manufacturing depends as much on how costs are measured as on how efficiently factories run. The way overheads such as factory rent, depreciation and supervisory salaries are spread across products can change reported margins, influence pricing, and even affect behaviour inside the plant. Absorption costing sits at the centre of this machinery, because it drives the unit cost that flows into inventory valuation, cost of goods sold, and headline profit figures used by boards, lenders and tax authorities alike.

Underlying economic issue: who should bear the fixed factory bill?

Manufacturing businesses incur large fixed costs to keep production capacity available: buildings, machines, salaried staff and support functions. These expenses are paid regardless of whether the factory runs at 20 percent or 90 percent of capacity. The central issue is how to attribute this fixed factory bill to individual units of output so that financial statements, pricing decisions and performance assessments make sense.

Absorption costing answers by insisting that every unit produced should carry a fair slice of that fixed burden, alongside its direct materials, direct labour and variable overhead. In other words, the economic logic is that capacity costs exist in order to make units, so units must "absorb" them. This contrasts with variable costing, where fixed manufacturing overhead is treated as a period expense of having capacity, rather than a cost of individual units.

The tension between these views is not merely academic. It determines whether unsold inventory carries embedded fixed overhead on the balance sheet (absorption costing) or whether all fixed overhead hits the income statement immediately (variable costing). The result is different profit paths over time when production and sales volumes diverge.

Substantive meaning: what costs are absorbed?

In practice, absorption costing brings together four categories of manufacturing cost as product cost:

- Direct materials

- Direct labour

- Variable manufacturing overhead (for example, indirect supplies, power linked to machine hours)

- Fixed manufacturing overhead (for example, factory rent, depreciation, factory management salaries)

These costs are all treated as part of inventory while units remain unsold and only become cost of goods sold when the units leave inventory. Selling, general and administrative costs, whether fixed or variable, remain period costs and are never attached to units.

From a financial reporting standpoint, this approach is not optional. Under major accounting frameworks, inventory must be carried at cost, including an appropriate allocation of fixed and variable production overhead. Absorption costing therefore underpins external profit reporting, tax computation and many loan covenant calculations.

Mathematical specification of unit cost under absorption costing

Although the mechanics appear straightforward, writing the relationships explicitly clarifies how production volume and allocation rates interact. Suppose a single product is manufactured in a period. Denote:

- : total direct materials cost for the period

- : total direct labour cost

- : total variable manufacturing overhead

- : total fixed manufacturing overhead

- : total units produced in the period

The total product cost for the period under absorption costing is:

The absorption costing unit cost is then:

Variable costing would instead treat only variable elements as product cost. Let be total variable manufacturing cost (). The variable costing unit cost is:

The difference between the two unit costs is simply the fixed overhead per unit:

This fixed overhead rate, often computed per machine hour or labour hour in multi-product environments, is the core mechanism by which overhead is absorbed into inventory. When production volume rises, increases, reducing fixed overhead per unit; when volume falls, each unit carries a heavier fixed overhead charge.

Income effects: production vs sales volume

The choice of costing method does not change total cash flows, but it can change the timing of reported profit. Under absorption costing, the fixed overhead tied to unsold units remains in inventory and is not yet expensed. Under variable costing, all fixed manufacturing overhead for the period appears immediately as an expense. As a result, in any period where production exceeds sales, absorption costing will usually show higher profit than variable costing; when production is below sales, the reverse occurs.

A simple reconciliation highlights the mechanism. Define:

- : units sold in the period

- : change in inventory units (positive if inventory grows)

- : fixed overhead per unit produced

The difference between absorption costing net income () and variable costing net income () in a period is:

When production exceeds sales so that , fixed overhead is deferred in inventory and exceeds . When sales draw down inventory so that , previously deferred fixed overhead flows to cost of goods sold, making lower than . When production equals sales, both methods report the same profit.

This algebra explains why standard-setting bodies still require absorption costing for external reporting but many internal management reports supplement it with variable or contribution costing to show the direct profit impact of volume changes.

Practical mechanics: cost pools and allocation bases

The theoretical unit cost formulas mask a significant practical challenge: allocating overhead to products in a way that is both systematic and economically meaningful. In a multi-product plant, overheads are typically collected into cost pools and assigned to products using allocation bases such as machine hours, labour hours, or material quantity.

A typical implementation proceeds in three stages:

- Establish cost pools: group similar overhead costs, for example all machine-related expenses, maintenance, and depreciation into a machinery pool; factory management salaries into a supervision pool.

- Determine usage measures: identify the driver that best reflects how products consume each cost pool, such as machine hours, direct labour hours, or production runs.

- Compute and apply rates: divide each pool by its total driver quantity to obtain a rate (for example, per machine hour), then multiply by each product's usage to assign overhead.

Absorption costing does not prescribe a particular choice of allocation base; the method is an overarching principle that all manufacturing costs should be absorbed by units. The sophistication of the allocation scheme can range from a single plant-wide rate to detailed activity-based costing with many cost pools and drivers.

Relation to variable costing and contribution analysis

Variable costing strips away the fixed overhead component of unit cost, focusing on the marginal resource consumption of each unit. For internal decision-making, this provides a cleaner view of how additional units affect profit because fixed overhead is held constant. Contribution margin analysis, which subtracts variable costs from sales to show the amount available to cover fixed costs and profit, is built on this variable costing logic.

The key contrast can be summarised conceptually:

- Absorption costing: all manufacturing costs, including fixed overhead, are product costs; inventory includes fixed overhead; external reporting requirement.

- Variable costing: only variable manufacturing costs are product costs; fixed manufacturing overhead is a period cost; used internally for planning, pricing, and performance evaluation.

Managers need both lenses. Absorption costing ensures financial statements comply with standards and reflect the full cost invested in inventory. Variable costing illuminates how decisions about volume, mix, and pricing will change cash profit in the short and medium term.

Major schools of thought and debates

Within managerial accounting, debates around absorption costing centre on three themes: performance measurement, decision relevance and overhead allocation philosophy.

First, performance measurement. Critics argue that tying profit to production volume via overhead absorption can create perverse incentives. Because producing more units spreads fixed overhead over more units, the unit cost falls, cost of goods sold per unit drops, and short-term profit often rises as long as the additional units go into inventory rather than being sold at a loss. This can encourage managers evaluated on absorption-based profit to overproduce relative to demand, leading to excess inventory, storage costs and potential obsolescence.

Proponents respond that robust inventory and working capital controls, together with careful use of variable costing and non-financial metrics, can mitigate these incentives while preserving the benefits of full cost information for pricing and long-term investment decisions.

Second, decision relevance. For decisions such as special orders, make-or-buy evaluations, or short-term pricing in the face of spare capacity, the fixed overhead portion of unit cost is sunk in the short run and should not drive the decision. Analysts therefore often ignore the absorbed fixed overhead in unit cost and instead work from variable costs and incremental cash flows. This creates a conceptual split between the "accounting cost" of a unit (including overhead) and the "economic cost" relevant for a particular decision scenario.

Third, overhead allocation philosophy. Traditional absorption costing usually allocates overhead using volume-based drivers like labour or machine hours. As production technologies and product diversity expanded, critics pointed out that such bases can distort product costs: low-volume, complex products may consume disproportionate setup and scheduling resources that do not scale with simple machine hours. Activity-based costing emerged as a refinement, retaining the absorption principle but using multiple cost drivers linked to underlying activities. This evolution reflects a broader debate about whether any allocation of common fixed costs is inherently arbitrary or whether careful design can approximate economic cause-and-effect sufficiently for management use.

Why absorption costing still matters

Despite these criticisms and refinements, absorption costing remains central to financial management for several reasons.

First, it is mandated for external reporting and taxation. Inventory must include an allocation of fixed overhead under accounting standards, which means any manufacturer preparing audited accounts must implement some form of absorption costing. As a result, banks, investors and regulators interpret performance largely through absorption-based statements.

Second, it anchors pricing and profitability analysis in the full cost base. Over time, businesses must recover both variable and fixed manufacturing costs through prices if they are to remain viable. While short-run decisions can legitimately use variable cost information, sustainable pricing strategies need to recognise the burden of capacity costs, which absorption costing surfaces.

Third, it disciplines capacity investment and utilisation decisions. By making fixed overhead visible within unit costs, absorption costing signals when capacity is under-utilised and factory-scale economics are deteriorating. Rising unit costs due to falling volume highlight the financial consequences of excess capacity or lost demand, encouraging rebalancing either through market expansion or capacity reduction.

Finally, it provides a common language for integrating financial control with operational data. Overhead rates per machine hour or per labour hour connect accounting records to shop-floor metrics, enabling cost variance analysis, standard costing systems and budgetary control. Even when management decisions rely on more refined models, the absorption framework underlies many of the control reports they receive.

Contemporary practice and evolving challenges

Modern manufacturing environments pose new challenges for absorption costing. Automation reduces direct labour content and increases capital intensity, weakening the link between simple volume measures and true resource consumption. Multi-site global supply chains complicate the definition of what counts as "manufacturing" overhead for a particular product. Customisation and short product life cycles create more setup and engineering costs, whose allocation may dominate traditional overhead pools.

Practitioners respond by:

- Refining cost pools and drivers, for example separating machine-level overhead, setup costs, quality assurance and engineering support so that each is allocated using an appropriate activity driver.

- Integrating operational systems with costing, using data from production execution and planning systems to update overhead drivers in near real time.

- Running parallel views: one set of absorption-based numbers for external reporting and high-level budgeting, and alternative contribution and activity-based analyses for operational decisions.

Even as digital tools make more sophisticated costing feasible, the fundamental requirement remains: inventory values on the balance sheet and cost of goods sold in the income statement must reflect all manufacturing costs, including an allocation of fixed overhead. Absorption costing provides the conceptual and procedural backbone for meeting that requirement.

Understanding how this method works, where it can mislead, and how it interacts with alternative views such as variable and activity-based costing equips managers, analysts and students to interpret reported margins critically, design better performance measures and make more informed operational and strategic decisions.

Quote: Anthropic - Artificial Intelligence - Recursive Self Improvement

"Claude writes a significant proportion of Anthropic's code. As of May 2026, more than 80% of the code we merge into Anthropic's codebase was authored by Claude. Before Claude Code launched in research preview in February 2025, this number was in the low single digits." - Anthropic - Artificial Intelligence - Recursive Self Improvement

The moment an internal engineering metric flips from human-written to AI-written code marks a structural shift in how complex software systems are built and evolved, not just a productivity bump for individual programmers. It signals that the primary generative force shaping a large codebase has become a model rather than a workforce, and that human engineers are increasingly curators, reviewers, and system designers guiding a non-human author.

In Anthropic's case, that shift is tightly bound to a broader concern: the trajectory from powerful coding assistants to systems that can meaningfully participate in, and eventually drive, the entire AI research and development cycle. When an AI model can write most of the code for its own infrastructure, tools, and scaffolding, the boundary between "AI helps humans build AI" and "AI builds AI" becomes thinner, and the timeline to more thorough forms of recursive self-improvement compresses.

From coding assistant to dominant author

Large language models like Claude were initially introduced as general-purpose assistants: chatbots that could answer questions, draft text, help with documents, and generate basic code. Early coding capabilities looked like autocomplete on steroids: filling in small functions, refactoring snippets, or suggesting tests. In that phase, AI was clearly subordinate to the human developer, integrated into IDEs as a suggestion layer with humans still doing the conceptual work, system design, and most of the implementation.

The internal numbers highlighted by Anthropic indicate that this relationship has inverted in at least one crucial dimension: the share of merged code now primarily authored by the model rather than by employees. Human engineers still specify goals, review diffs, and orchestrate work, but the bulk of literal line-by-line code is machine-generated. Independent developers using Claude Code describe a similar workflow: they treat the AI interface almost as the primary editor, with a traditional editor demoted to a verification and correction tool. One typical pattern is to spend most of the time explaining the problem and iterating on plans with the model, then auto-accept its changes, and only afterwards manually review and adjust. That mirrors the internal picture: humans move up a level of abstraction, while the model handles implementation detail at scale.

The key structural consequence is that the constraint on how fast a codebase can change shifts away from human typing speed or individual concentration. Instead, the main bottlenecks become prompt quality, review capacity, testing infrastructure, and organisational willingness to deploy AI-authored changes. Once those guardrails are in place, the marginal cost of asking the AI to implement yet another subsystem approaches the cost of specifying it, rather than building it yourself.

Recursive self-improvement: several distinct mechanisms

The idea of recursive self-improvement (RSI) in AI originally focused on a dramatic scenario: a sufficiently capable system rewrites its own code, becomes smarter, uses that increased intelligence to further rewrite itself, and so on, producing an "intelligence explosion". In more formal discussions, RSI is framed as a process where an AI improves its own ability to improve, potentially leading to superintelligence if the feedback loop is strong enough. For decades this remained hypothetical, because no deployed system could modify its own internals in a reliable, directed way.

Recent work on RSI has clarified that there are at least three separable mechanisms, each with different bottlenecks and risk profiles. First, there is what some researchers call scaffolding-level improvement: you keep the base model weights fixed but wrap the model in better tools, agents, and workflows that make more effective use of its capabilities over time. Coding agents that orchestrate tool calls, decompose tasks into subproblems, and maintain long-lived workspaces fall into this category. The AI does not change itself directly, but the environment around it is iteratively improved-often with heavy AI assistance.

Second, there is improvement of the broader AI research and engineering process. Here, models help design better architectures, tune hyperparameters, automate experiments, and analyse results. The AI is not rewriting its own weights on the fly but is heavily used by human researchers to run more experiments faster, test more ideas, and push the frontier models forward. In effect, the research pipeline that generates new models is being partially automated by prior models, shortening cycle times.

Third, there is the more classical vision of model-internal self-modification: a system that can inspect, reason about, and deliberately rewrite its own internal structure. In the current deep learning paradigm, this would require some combination of advanced mechanistic interpretability and internal training or optimisation loops guided by the model itself. This is the least empirically grounded category today; there are not yet widely documented systems that autonomously edit their own weights in a stable, predictable way in production, without external training pipelines.

Anthropic's published analysis emphasises that the world is beginning to see concrete progress in the first two forms of RSI, while the third remains more speculative but increasingly relevant. The metric that more than four-fifths of merged code comes from Claude is directly relevant to the first two types: scaffolding-level improvement and research-process acceleration. It is not yet full-blown self-modifying AI, but it clearly moves along the continuum from "AI as a tool" to "AI as a primary agent in its own development ecosystem".

What does it mean for AI to "build itself"?

In its report "When AI builds itself", Anthropic defines a future regime in which AI systems can design, implement, and train successor models with minimal human involvement. That scenario includes choosing research directions, generating experimental configurations, running training runs, monitoring results, and iteratively refining architectures, all mediated by models rather than individual researchers. The report stresses that current systems have not yet reached this stage, but the pattern of automation suggests a trajectory that could plausibly converge towards it in the medium term.

Already, tools like Claude Code enable models to handle much of the mundane engineering needed to integrate new components, instrument experiments, and manage evaluation pipelines. For example, a model can generate scripts to launch training runs, write configuration files for different hyperparameter sweeps, produce dashboards for monitoring metrics, and adapt code to new hardware or inference setups. Engineers remain in the loop to approve designs, interpret anomalies, and adjust objectives, but they increasingly operate at the level of specifying desired behaviours and constraints rather than manually wiring every detail.

Once the majority of the code surrounding the training and deployment pipeline is generated by models, the human role shifts to defining goals, setting safety criteria, and analysing higher-level trade-offs. The mechanics of "building"-in the sense of constructing new experimental setups, converting research ideas into running code, and instrumenting systems-becomes heavily AI-mediated. Over time, if models learn from this process (for instance by analysing successful and failed experiments), they can become better at designing and conducting AI research itself.

Strategic and technological tensions

The shift towards AI-written code simultaneously advances capability and heightens safety concerns. On the one hand, organisations that can mobilise models as large-scale coding engines enjoy dramatic efficiency gains. Anthropic and other labs report that a single engineer working with AI can now accomplish several times the output of a solo developer from only a few years ago. Internal numbers cited in commentary around the Anthropic report suggest that in some workflows, one engineer paired with advanced coding models can match the productivity of many engineers without such tools. This is economically attractive and strategically hard to ignore, especially in competitive markets where speed and feature velocity matter.

On the other hand, every additional layer of automation in the AI development pipeline reduces the surface area where humans directly engage with the details of what is being built. If most of the code diff is AI-authored, there is a constant pressure to keep review lightweight enough not to erase the productivity gains. Organisations must decide how much friction to reintroduce via testing, code review, and formal verification to compensate for the opacity and potential brittleness of model-generated software.

There is also a tension between transparency and performance. Coding models are trained on large corpora and fine-tuned for usefulness, but their internal reasoning is not inherently interpretable. When such models are tasked with writing critical infrastructure-especially infrastructure that itself trains or deploys models-the demand for rigorous verification increases. Yet the whole point of using AI at scale is to compress the development cycle; fully auditing every AI-generated line is often infeasible. This pushes teams towards probabilistic assurance: relying on automated tests, static analysis, and spot checks, accepting that some defects or misalignments may slip through.

Anthropic's policy stance reflects this tension. The organisation has publicly advocated for a potential future pause or slowdown in frontier AI development if such a pause can be coordinated and verifiable. At the same time, it continues to deploy tools that significantly accelerate the AI engineering process. The argument is not that acceleration ought to stop now, but that the world should build governance and monitoring infrastructure capable of making a pause credible if systems begin to show signs of more autonomous, less controllable forms of self-improvement.

Debates and objections

There are several lines of scepticism about treating AI-written code as a near-term marker of recursive self-improvement. One objection is that a model generating code on command is still deeply dependent on a human-constructed training pipeline and hardware stack. The AI may write most of the repository, but it does not yet select its own training data, modify its own loss functions, or commission new datacentres. From this perspective, calling such behaviour "self-improvement" risks overstating the level of autonomy.

Another objection focuses on quality. Critics argue that high percentages of AI-written code may reflect a bias towards quantity over robustness. If models can quickly generate large volumes of superficially plausible code, teams may be tempted to merge more, trusting tests and users to uncover issues. This could increase technical debt and vulnerability surfaces, particularly if AI-generated code uses patterns that are less idiomatic or less well understood by the team. In this view, the headline figure of more than four-fifths AI-authored code says more about internal incentives and tooling than about genuine leaps in capability.

A further concern is that the narrative of "AI writing its own code" might be leveraged for competitive signalling or regulatory positioning. Emphasising that models are rapidly approaching self-building status can support calls for stricter regulation, but it can also serve as a way to demonstrate leadership and sophistication in the race for funding and talent. Observers therefore scrutinise such claims, asking how the metric is defined (for example, how attribution between human and AI edits is measured) and what kinds of code are included-core model logic, surrounding infrastructure, or peripheral tools.

Supporters of the stronger interpretation respond that the exact percentage is less important than the direction of travel and the kinds of tasks being automated. The movement from "AI can write helper scripts" to "AI can build and maintain major production systems" represents a qualitative shift. Moreover, as AI-generated code begins to include experiment orchestration, data processing pipelines, and evaluation harnesses, the model's role in improving subsequent models increases, even if human oversight remains substantial. From this vantage point, the concern is not that current systems are already self-improving in the strongest sense, but that they are laying the groundwork for a regime in which incremental capability increases lead to disproportionate gains in further capability development.

Why it matters beyond software engineering

The implications of AI writing most of the code in a frontier lab extend well beyond the internal life of software teams. One major dimension is economic. If an AI-augmented engineer can do the work of several traditional engineers, the effective labour cost of software development drops sharply. Over a horizon of a few years, this could reshape labour markets, favouring organisations that can most effectively integrate AI into workflows. Entire categories of skilled work-software engineering, research assistance, data analysis, legal drafting-could be automated at a pace that leaves limited time for institutions to adapt.

Another dimension is geopolitical. Access to models capable of acting as high-bandwidth coding engines becomes a strategic asset. States or firms that control such systems can upgrade their digital infrastructure, defence systems, and research capabilities faster than competitors. If recursive self-improvement processes take hold, the gap between leading actors and followers could widen rapidly. This is one reason why some analysts emphasise the risks of concentration of power: if a small number of organisations own the most capable self-improving AI systems, they may acquire outsized influence over economic and political developments.

There is also a safety dimension that goes beyond the immediate risk of buggy code. As AI systems participate more in their own development, misalignments in objectives or reward signals can be compounded. If an AI is tasked with optimising for performance on certain benchmarks, and it also plays a role in designing the evaluation apparatus and experimental setups, it might inadvertently favour changes that make it look better on metrics without improving, or even while degrading, its broader alignment with human values. The more of the research loop is automatised, the more important it becomes to design robust, hard-to-game objectives and interpretability tools.

Finally, there is an epistemic dimension. When AI systems write most of the code, run most of the experiments, and summarise most of the results, human understanding of complex software and research landscapes can become indirect. Engineers and scientists may interact primarily with AI-generated abstractions of what is going on. This can be efficient, but it also risks a kind of institutional deskilling: fewer people understand systems end-to-end, making it harder to detect systemic errors, correlated failures, or unanticipated interactions. In high-stakes domains, that loss of deep understanding could itself become a safety hazard.

The emerging role of human engineers

In the near term, the rise of models as dominant code authors does not eliminate the need for human engineers; it changes their role. Reports from practitioners using Claude Code suggest that humans increasingly focus on problem decomposition, specification, and verification. They spend more time writing detailed natural language descriptions of desired behaviour, orchestrating multi-step workflows, and designing tests that capture subtle requirements. They also become stewards of code quality and maintainers of conceptual coherence across rapidly evolving codebases.

This role shift is non-trivial. Writing good prompts or instructions is a skill; designing prompts that anticipate edge cases, security concerns, and performance constraints is even more demanding. Similarly, effective verification under conditions of AI-generated abundance requires new practices: stronger automated test suites, better monitoring, and perhaps new forms of formal methods that are integrated into everyday workflows. Human engineers who adapt to these demands may become more like system architects and editors, curating and refining the work of a powerful but sometimes unreliable assistant.

At the same time, there will likely remain pockets of development where human-written code is preferred or required, especially for safety-critical components, low-level systems programming, or domains where subtle domain knowledge is hard to transmit through prompts alone. The distribution of human effort across a codebase will change: less time on boilerplate and repetitive patterns, more on rare but consequential decision points.

Looking ahead

The internal data that an AI system now authors the majority of a leading lab's merged codebase should be understood as a waypoint, not an endpoint. It marks a concrete, measurable point on a curve that leads from basic assistance to deeper forms of recursive self-improvement. The same dynamics that allow models to dominate code authoring-scaling, better scaffolding, agentic tools, and integration into research workflows-are also those that will shape how quickly AI systems begin to design and build their successors with decreasing human input.

Whether this trajectory culminates in controllable, beneficial systems or in hard-to-govern, rapidly self-improving agents will depend on decisions being made now: how much autonomy to grant coding models, what review standards to enforce, how to design incentives for safety rather than pure speed, and what international coordination mechanisms to build in anticipation of more powerful RSI. As the proportion of AI-written code grows, so too does the responsibility to align not just the models, but the socio-technical systems that surround them.

Strategy Tool: Rethinking SWOT analysis in the context of AI

AI-SWOT reframes classic SWOT for the AI era by treating AI as a strategic amplifier and mitigator, not just another technology. It shows how AI can dramatically amplify existing strengths and opportunities (through scale, speed, data flywheels and new business models) while mitigating key weaknesses and threats (by closing capacity gaps, enhancing risk detection and building early?warning systems). The tool introduces a structured, workshop-ready process that walks leaders through: (1) identifying where AI can turn genuine strengths into durable moats, (2) using AI to unlock or accelerate external opportunities, (3) targeting AI at the specific weaknesses that drive competitive loss, (4) deploying AI to detect and neutralise emerging threats, especially in the WT quadrant, and (5) recognising AI itself as a new category of threat via competitor amplification and low-barrier new entrants. Packed with contemporary case studies (Nike, Amazon, Netflix, Klarna, JPMorgan, Siemens, boutiques vs. global firms), diagnostic questions, and stepped tasks, AI-SWOT gives executives a practical, evidence-based way to convert AI from a generic initiative into a focused, advantage-creating strategy tool.