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

AM edition. Issue number 1388

Latest 10 stories. Click the button for more.

Term: LoRA (Low-Rank Adaptation) - Artificial intelligence

"LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique that adapts large pre-trained AI models by freezing the original weights and training much smaller, auxiliary rank-decomposition matrices. .By focusing computation only on these compact adapter layers, it drastically reduces resource requirements without sacrificing model performance." - LoRA (Low-Rank Adaptation) - Artificial intelligence

Organisations trying to adapt foundation models to niche tasks quickly run into a hard constraint: the cost of moving and updating tens of billions of parameters dwarfs the incremental value of most applications. GPU memory, bandwidth and training time become the binding bottlenecks, not data or ideas. Low-rank adaptation tackles this bottleneck by reframing the update itself as a compact, structured object that can be trained and stored far more cheaply while preserving most of the performance of full fine-tuning .

In practical terms, the technique achieves parameter efficiency by freezing the base model and pushing all task-specific learning into small auxiliary modules attached to existing layers . Instead of rewriting the model's knowledge, these modules learn additive corrections that steer behaviour on the new task. This architecture has concrete deployment advantages: teams can keep a single shared base checkpoint and swap in different adapters for legal reasoning, medical question answering or code generation, each occupying only a tiny fraction of the storage and memory footprint of the base model . That modularity also simplifies governance, since task-specific adapters can be versioned, audited and rolled back without touching the core model.

Mechanism: low-rank updates to frozen weights

The core idea is that the update needed to adapt a well-trained model to a specific task lies in a low-dimensional subspace, so it can be represented by a low-rank matrix rather than a full dense weight update . Suppose a given linear projection in a transformer layer is parameterised by a weight matrix . Classical fine-tuning learns a full update , giving an effective weight . In low-rank adaptation, one constrains to the form , where and with . Only and are trainable; the original is kept fixed. This factorisation means the number of new parameters scales with rather than , yielding orders-of-magnitude reductions for typical transformer dimensions .

During fine-tuning, the forward pass through the adapted linear layer is commonly written as , where is the input vector and is a scaling factor controlling the strength of the adaptation . Gradients flow only into , and possibly ; the base weight receives no updates and is typically stored in quantised or otherwise compressed form. After training, the low-rank update can be algebraically merged back into to produce a single, adapted checkpoint for deployment, with no extra latency relative to a fully fine-tuned model .

Parameter meanings and configuration choices

The most important hyperparameter in low-rank adaptation is the rank , which directly controls the expressive power and size of the adapter . Low ranks such as or are often sufficient for style transfer, light domain adaptation or instruction tuning on modest datasets . Higher ranks, for example or greater, may be needed for complex reasoning tasks, highly specialised jargon, or multi-step workflows, at the cost of increased memory and training time . Practitioners thus treat as a knob trading off accuracy against efficiency, tuned empirically under hardware constraints . The scaling parameter rescales the contribution of the adapter relative to the frozen base and can act as a regulariser, preventing the low-rank update from overwhelming the prior knowledge encoded in . Dropout applied inside the adapter path further reduces overfitting when data is scarce .

Another design choice is where to insert adapters in the network. Many implementations start by targeting the attention projections, such as query and value matrices, because small changes there can significantly influence how the model attends to task-relevant tokens . More aggressive configurations attach adapters to all linear transformations in both attention and feedforward blocks, increasing effective capacity but also memory usage . The distribution of rank across layers is an active area of research; approaches that allocate higher rank to layers with greater task-relevant entropy or sensitivity can improve performance without uniformly increasing the footprint .

Practical meaning: why it matters operationally

From an operational perspective, low-rank adaptation reshapes the economics of customising large models. By reducing trainable parameters by factors of up to 10 000 and lowering GPU memory by roughly 3 times for some configurations, it allows teams to fine-tune models that would previously have required large clusters, using a single high-end GPU instead . This cost compression enables experimentation with many candidate tasks or data slices, since each adapter can be trained cheaply and evaluated in isolation. It also encourages a plug-in mindset: an organisation may maintain dozens of small domain-specific adapters, switching between them per request or per product, while all share a single central base model .

The technique also mitigates catastrophic forgetting, the phenomenon where full fine-tuning on a narrow dataset degrades performance on broader capabilities . Because the original weights remain untouched, the base model's general language understanding is preserved, and the adapter learns to specialise without erasing prior knowledge . This makes low-rank adaptation appealing for applications that must balance strong performance on a target task with adequate behaviour on generic queries, such as customer support assistants that alternately handle policy questions and free-form conversation. However, this same regularisation means that when data and compute budgets are very large and peak accuracy on a single domain is paramount, full fine-tuning can still outperform adapter-based methods .

Mathematical and conceptual foundations

The use of a low-rank factorisation connects the method to long-standing ideas in numerical linear algebra and statistical learning. Low-rank approximations exploit the observation that many high-dimensional datasets and transformations lie close to a subspace of far lower intrinsic dimension, which can be captured by a small number of basis vectors . In the context of large language models, the claim is that the gradient-informed update for a downstream task mostly lives in such a subspace, so optimising and suffices to encode the relevant change . One can view this as projecting the full gradient update into a lower-dimensional manifold where optimisation is cheaper and less prone to overfitting, then lifting it back to the original space via the product .

Formally, if one considers the full fine-tuning update as sampling from a distribution over matrices, low-rank adaptation restricts that distribution to those matrices with rank at most . This constraint acts as an implicit prior favouring simpler, smoother updates which can improve generalisation in data-poor regimes . It also aligns well with modern PEFT (parameter-efficient fine-tuning) frameworks that treat task adaptation as learning a compact, structured perturbation rather than a full re-optimisation of billions of parameters . That said, the assumption that the optimal update is low-rank may not hold in all domains, and empirical work continues to probe which tasks and architectures are best suited to this constraint .

Schools of thought and emerging debates

One school of thought emphasises low-rank adaptation as the default tool for downstream tuning of large language models, pointing to its strong performance on instruction tuning, style transfer and moderate-scale domain adaptation relative to its low compute cost . Proponents argue that for most enterprise workloads-summarisation, translation, retrieval-augmented question answering-the marginal gains of full fine-tuning do not justify the complexity and expense, especially given risks around forgetting and model drift . A contrasting view sees adapters as a pragmatic compromise but still regards full fine-tuning, possibly combined with further pretraining, as the gold standard for high-stakes domains such as advanced coding assistance or mathematical reasoning . In this view, low-rank methods are invaluable for prototyping and mid-tier applications but may underperform when deep conceptual shifts in the model's representation are required.

A more recent debate concerns how low-rank adaptation interacts with quantisation and other compression strategies. Techniques such as QLoRA retain the frozen low-rank adapters but apply aggressive quantisation to base weights, achieving roughly 4x further memory reductions and enabling fine-tuning of models with 65B+ parameters on commodity hardware . While this widens accessibility, it introduces new questions about numerical stability, sensitivity to hyperparameters, and the cumulative effect of approximations at both the base and adapter levels. Researchers are also exploring alternatives such as prefix-tuning, parallel adapters, and multi-task adapter routing, all of which compete or combine with low-rank methods in different regimes . These debates highlight that the technique sits within a broader ecosystem of PEFT methods, rather than as a singular solution.

Continuing relevance

Low-rank adaptation remains central to the current generation of AI practice because it squarely addresses the main practical friction in deploying foundation models: the gulf between theoretical capability and affordable, governable customisation. By reframing task adaptation as learning compact, low-rank corrections to a frozen base, it unlocks workflows in which a single general model can underpin many specialised products and internal tools, each represented by a small adapter file . That shift supports organisational scaling, since different teams can iterate on their own adapters independently without competing for scarce training capacity on the full model. As models grow larger and regulations tighten around data use and model behaviour, the ability to adapt efficiently, reversibly and in a modular fashion will only grow more important, ensuring low-rank techniques continue to matter even as alternative PEFT methods and more powerful base models emerge .

Global Advisors News Brief - July 27 2026

Read the full brief at the link

Headlines for the last 24hrs

1. Big Tech Accelerates Multi-Hundred-Billion-Dollar AI Compute Investments Amid Growing Wall Street Scrutiny
2. Energy Markets Experience Heightened Volatility as Geopolitical Risks Intersect Supply Vulnerabilities
3. Federal Reserve Policy Under Scrutiny as Rebounding Inflation Risks Pressure Rate Decisions
4. High-Profile AI Security Incidents and Rogue Agent Failures Drive Calls for Transparency and Washington Lobbying
5. U.S. Trade Policy and Tariff Shifts Impede Cross-Border E-Commerce Expansion
6. Chinese Semiconductor and AI Sectors Surge Amid Intensifying Global Technology Rivalry
7. Potential U.S. Airline Consolidation Signaled by Early Megamerger Discussions
8. Private Capital Dealmaking Rebounds Across Consumer and Wealth Management Sectors
9. Cross-Border Regulatory and Capital Disputes Impair UK and European Tech and Banking Integration
10. Frontier AI Models Advance Multimodal Capabilities and Autonomous Task Execution Benchmarks

Time window: 2026-07-26T05:00:33.073Z to 2026-07-27T05:00:33.073Z

Quote: Professor Erik Brynjolfsson - Stanford economist, Director of the Stanford Digital Economy Lab

"If you get rid of the base of the [organisational] pyramid, it becomes like a a diamond. Then where are those junior those middle managers going to come from and where the senior people going to come from? And too many companies are being shortsighted about that." - Professor Erik Brynjolfsson - Stanford economist, Director of the Stanford Digital Economy Lab

The structural problem facing advanced economies is not simply that automation removes tasks, but that it selectively erodes the bottom rungs of professional labour markets while leaving demand for experienced judgement intact. Entry-level, routine knowledge work is disproportionately exposed to generative AI, yet organisations still need mid-level managers and senior experts to coordinate, decide and lead. The tension is straightforward: firms optimise costs by shrinking junior roles today, but they also depend on those same roles as the training ground for tomorrow's decision-makers.

The pyramid under pressure: AI and entry-level work

Recent research led by Erik Brynjolfsson and colleagues at the Stanford Digital Economy Lab shows a striking decline in employment for young workers in occupations most exposed to generative AI. In these roles, employment for people in their early twenties has fallen by around 16 % relative to less-exposed occupations, while older workers in the same fields have largely maintained or increased employment. The pattern is consistent with AI systems taking over routine, well-specified tasks that previously justified hiring larger cohorts of juniors, especially in software development, call centres, customer support, paralegal work and some marketing and sales functions. Senior workers remain because they supply judgement, context, client interaction and organisational memory - capabilities that current models cannot fully replicate.

Brynjolfsson repeatedly emphasises that the analytical unit is the task rather than the occupation. Most jobs consist of bundles of tasks, and generative AI is automating specific components - coding small features, drafting emails, summarising documents - rather than entire professions. That nuance matters, yet the aggregate effect at the bottom of organisational hierarchies is clear: the share of human effort devoted to routine execution is falling, and the first group to feel it is junior staff. When organisations respond by cutting entry-level hiring rather than redesigning roles, they begin to hollow out the base of the pyramid that traditionally fed into middle management and senior leadership.

From pyramid to diamond: a changed organisational geometry

Professional services firms, large corporates and technology companies have historically adopted a pyramidal staffing structure: many junior employees, fewer mid-level managers and a relatively small group of senior partners or executives. Juniors handled labour-intensive, repeatable tasks, generating leverage for experienced staff who focused on complex decisions, client relationships and high-value work. Over time, some juniors were promoted into mid-level roles, creating a natural pipeline of talent.

AI systems disrupt this geometry by absorbing a growing fraction of the routine work once done by the base. In software engineering, junior developers used to spend years writing boilerplate code, fixing defects and implementing well-understood patterns under supervision. Now coding assistants and agentic tools can generate plausible implementations directly from natural-language specifications, enabling a smaller number of more experienced developers to oversee much larger codebases. In call centres, conversational models have progressed from supporting human agents to autonomously handling a substantial share of inbound queries, reducing the need for large cohorts of novice staff. Similar trends appear in document-heavy fields such as law and compliance, where paralegal-style tasks are increasingly automated.

The immediate result is that firms can plausibly operate with fewer juniors while retaining or even expanding their cadre of mid-level and senior staff. The longer-term consequence is more subtle and potentially damaging: without a broad base of early-career employees gaining experience, the future supply of people capable of occupying those higher-level roles shrinks. The organisational pyramid begins to resemble a diamond - relatively narrow at the bottom, wide through the middle, and tapering again at the top - but that diamond shape is unstable if the mid-section is not continually replenished.

Career formation without apprenticeship

The traditional route into managerial and expert positions involved prolonged apprenticeship: juniors observed how decisions were made, absorbed tacit knowledge and gradually took on more complex responsibilities. Much of this learning occurred incidentally through the performance of routine tasks - reviewing documents, preparing reports, shadowing meetings - that were economically necessary even if intellectually basic. AI-driven automation strips away exactly those tasks, leaving fewer natural opportunities for observational learning.

The structural risk Brynjolfsson highlights is that firms enjoy short-term cost savings while inadvertently destroying the mechanisms that produced their own human capital. If junior positions vanish, aspiring professionals face a paradox: they are told that senior judgement, project management and domain insight are the safest skills, but they are denied the environments in which those skills were historically developed. At the societal level, this becomes a coordination problem. Each individual firm has an incentive to reduce entry-level hiring and lean on AI substitutes; collectively, the economy needs a steady stream of workers acquiring experience to sustain future productivity and leadership.

Infosys, the Indian technology and consulting company, appears in Brynjolfsson's narrative as a counterexample. Rather than sharply cutting junior recruitment in response to AI, it continues hiring young staff but redesigns their work. Routine coding and documentation are delegated to AI tools; juniors focus earlier on project management, systems thinking and broader contextual understanding. Learning that once occurred by osmosis is replaced with explicit training, often supported by AI itself as a teaching aid. The strategic bet is that human taste, judgement and leadership will remain scarce and valuable, so preserving the pipeline into those capabilities outweighs immediate savings from eliminating novice roles.

Strategic myopia and distributional risk

Brynjolfsson's broader economic work suggests that the same dynamic operates beyond individual firms in the form of skill-biased technical change. Technologies that complement highly skilled workers while substituting for less-skilled ones tend to widen wage and income gaps, an effect documented historically with earlier waves of computerisation. Generative AI threatens to accelerate that pattern by disproportionately eliminating entry-level white-collar roles while expanding the productivity of incumbents who already possess experience and decision rights.

If organisations underinvest in junior development, they will ultimately face a shortage of experienced managers and specialists, but the transitional damage may be borne by cohorts of young workers who struggle to enter professional careers at all. Brynjolfsson warns that this could echo the policy failures of globalisation: aggregate gains in productivity and wealth coinciding with concentrated losses for specific communities, generating political backlash and social instability. In the AI context, he argues for deliberate investment in education, retraining and apprenticeship-style pathways to avoid repeating that mistake. Proposed mechanisms include public funding for skills development, more flexible labour-market institutions and the use of AI itself to support job matching and personalised training.

Redesigning roles around agents, not eliminating people

A recurrent theme in Brynjolfsson's interviews is that future knowledge work will revolve around defining questions and evaluating answers, with AI agents handling much of the execution. He divides projects into three stages: define, execute, evaluate. AI excels in the middle once a problem is clearly specified, but humans remain essential for identifying the real problem and judging whether the solution is correct or useful. This framing suggests an alternative to removing junior staff: train them early to manage fleets of agents, formulate valuable questions and interpret outputs, rather than simply carrying out pre-defined procedures.

Under such a model, junior roles would be reconfigured rather than erased. Instead of spending years on manual data cleaning or repetitive coding, new entrants would learn to orchestrate AI tools across workflows: decomposing tasks, specifying constraints, monitoring for errors and integrating results into decisions. They would acquire the meta-skills - problem framing, stakeholder communication, risk awareness - that constitute the foundation of senior responsibility, while agents supply the routine labour. This path preserves a developmental ladder even as the nature of rungs changes, aligning organisational needs for future leadership with technological realities.

Debates, objections and alternative trajectories

Not all economists agree that hollowing out entry-level roles will persist or that the organisational pyramid must fundamentally change. Some argue that new categories of junior work will emerge around AI maintenance, data curation, prompt engineering or human oversight, effectively replacing traditional routine tasks with technologically mediated ones. Others note that if AI substantially lowers the cost of producing certain services, demand elasticities may lead to expanded employment even at junior levels, as happened with radiology when cheaper imaging increased the total volume of scans. Brynjolfsson himself acknowledges this mechanism, highlighting that roughly half the economy may see rising employment when prices fall, depending on demand curves.

However, the entry-level job data for generative-AI-exposed occupations suggest that, at least in the current phase, displacement is outpacing creation for young workers in affected fields. The diamond-shaped organisational concern arises specifically where demand does not expand enough to justify large numbers of juniors, yet mid-level and senior roles remain necessary. The debate therefore centres less on whether AI will ultimately create new work - historical evidence suggests it will - and more on whether current corporate decisions around hiring, training and role design are aligned with long-term capability needs.

Why the organisational base still matters

The underlying message in Brynjolfsson's warning is not nostalgic protection of obsolete jobs but a pragmatic assessment of capability formation. Advanced technologies, including AI, are general-purpose tools whose economic impact depends on complementary investments in human capital, organisational redesign and institutional adaptation. If firms treat AI purely as a substitution technology - a way to remove headcount at the bottom - they may gain short-term margin improvements but weaken their ability to innovate, coordinate and exercise judgement over time.

Conversely, organisations that explicitly preserve and reconfigure junior pathways, teaching novices to manage agents, frame problems and develop domain expertise, are more likely to sustain a robust pipeline of mid-level managers and senior leaders. At the economy level, such choices affect whether AI becomes primarily a force for shared prosperity or for increased concentration of wealth and power. The geometry of the organisational pyramid is therefore not merely a staffing diagram; it is a reflection of how societies choose to invest in human potential under conditions of rapid technological change.

Global Advisors News Brief - July 26 2026

Read the full brief at the link

Headlines for the last 24hrs

1. Rogue Autonomous AI Agents Breach Safety Red Lines and Trigger Enterprise Security Risks
2. Big Tech Earnings Volatility Highlights Growing Market Scrutiny Over AI Return on Investment
3. Silicon Valley Split over Tech Nationalism and Chinese AI Competition
4. AI Data Center Power Demands Strain Global Energy Grid Infrastructure
5. Private Equity Consortium Consolidates Critical Global Energy Assets in $16B Deal
6. Surging Energy Prices and Sticky Inflation Trigger Expectations of Higher Central Bank Rates
7. GLP-1 Pharmaceutical Demand Reshapes Global Healthcare Logistics and Cold Storage
8. Organizational AI Productivity Disparities Highlight Critical Need for Human Capital Alignment
9. State Regulators Move to Ban AI-Driven Dynamic and Surveillance Pricing
10. Aggressive Tech Talent Acquisition Triggers Corporate Litigation Over Executive Poaching

Time window: 2026-07-25T05:00:33.068Z to 2026-07-26T05:00:33.068Z

Quote: Professor Erik Brynjolfsson - Stanford economist, Director of the Stanford Digital Economy Lab

" I think AI if anything - believe it or not - I think it's underhyped. I think it's going to be even bigger than most people realise." - Professor Erik Brynjolfsson - Stanford economist, Director of the Stanford Digital Economy Lab

The striking claim that artificial intelligence is still underhyped speaks to a structural mismatch between the technology's rapidly compounding capabilities and the pace at which economies, organisations and institutions are absorbing and exploiting those capabilities. The underlying issue is not whether AI can perform impressive benchmarks in coding, language and reasoning; that is already evident. The tension lies in the lag between what systems can do and how quickly business models, workflows, skills and policy frameworks are being redesigned to convert those capabilities into broad-based economic value.

From capability curve to economic J-curve

Over the past few years, frontier AI models have improved at an extraordinary rate on tasks such as software development, document analysis, mathematical reasoning and multi-step planning. Measured purely as a capability curve, the trajectory looks exponential. Yet the impact on aggregate productivity statistics, labour markets and GDP remains relatively modest. Brynjolfsson describes this as a technology adoption -curve: an early phase in which investment, experimentation and organisational disruption rise sharply, while observable economic gains appear muted. Only after firms re-architect processes, data flows and decision rights around the new technology does the curve inflect and value creation become visible at scale.

The factual context of his remark is a bet with another economist, Robert Gordon, who is sceptical that AI will materially accelerate productivity. Official projections for US productivity growth into the late 2020s remain anchored around pre-AI assumptions. Brynjolfsson argues that these forecasts are systematically too low, because they implicitly treat current institutional arrangements as fixed rather than adjustable. In his view, once organisations complete the redesign necessary to integrate AI deeply into production and services, productivity growth will exceed the Bureau of Labor Statistics baseline by 2030, with knock-on effects for fiscal sustainability, healthcare and living standards.

AI as a general-purpose technology

Another layer of meaning in the statement is the classification of AI as a general-purpose technology comparable to electricity or the internal combustion engine. General-purpose technologies display three properties: they improve rapidly over time; they become pervasive across sectors; and they enable complementary innovations in products, processes and organisational forms. Historically, their full economic impact has been delayed by the need for complementary investments. Electric motors, for instance, were installed into factories designed for steam, producing few early gains; only when factories were rebuilt around decentralised power did productivity accelerate.

Brynjolfsson's argument is that AI now sits in a similar transitional zone. Providing employees with chatbots or code assistants is equivalent to installing motors in old steam-era layouts. The larger effects will only appear when firms deliberately redesign end-to-end workflows, redefine roles, restructure data infrastructure and launch new AI-native products and services. Because digital systems can be reconfigured far faster than physical factories, he expects the AI transition to play out over roughly 3 to 5 years rather than the 30-year lag observed with electricity. That compressed timeline is one reason he judges AI to be underhyped: most forecasts still treat the transformation as distant, when in his view the economy is already turning the corner of the J-curve.

The labour market tension behind the optimism

The claim that AI is underhyped sits alongside clear evidence of labour market disruption, especially for younger workers in highly exposed occupations. Research from the Stanford Digital Economy Lab and ADP shows employment among workers aged roughly 22 to 25 falling sharply in jobs where generative AI can perform core tasks, such as junior coding, call-centre work, paralegal support and some routine marketing functions. The same datasets show far milder effects for older workers and less-exposed occupations, and even positive outcomes where AI is used to augment rather than fully automate tasks.

This duality creates a strategic tension. On one hand, AI boosts the productivity of workers and can improve customer satisfaction, retention and job quality when used as an assistive tool. On the other hand, the disappearance of entry-level roles threatens traditional talent pipelines, making organisational structures more diamond-shaped: fewer juniors, continued demand for mid-level and senior staff, and an emerging shortage of people with accumulated experience. Brynjolfsson's optimism about AI's aggregate impact therefore coexists with a warning that firms and societies must explicitly redesign education, apprenticeships and early-career development rather than relying on routine work to train future leaders.

Demand elasticity and why automation can expand jobs

Another factual foundation for the underhyped claim is a subtle point about demand elasticity. Automation does not automatically reduce employment; the outcome depends on how buyers respond to lower costs. Using the standard downward-sloping demand curve, Brynjolfsson distinguishes between relatively inelastic markets, where falling prices lead to only modest increases in quantity and thus lower total spending, and highly elastic markets, where price reductions trigger large increases in quantity and potentially higher overall expenditure.

In sectors such as air travel, the introduction of jet engines cut per-trip costs but created such a surge in demand that total spending and employment in airlines grew rather than shrank. He expects roughly half the economy to behave in this expansionary manner as AI reduces costs and increases quality. Radiology offers a contemporary example: early headlines predicted that image-recognition systems would eliminate radiologists, yet many countries now face radiologist shortages as cheaper and faster imaging has increased demand for scans. In this framing, AI is underhyped not because it will avoid disruption, but because commentators fixate on job destruction and miss the employment created by lower prices, new services and expanded markets.

Redefining work around defining and evaluating

A central mechanism behind Brynjolfsson's optimism is his view of future work as structured into three stages: defining the problem, executing the work and evaluating the output. AI agents are becoming remarkably capable in the middle stage once a problem is well specified: writing software, drafting documents, analysing data and completing procedural tasks. Yet they remain limited in selecting economically meaningful problems, understanding organisational context and exercising judgement about the adequacy and implications of their outputs.

He expects most knowledge workers to manage fleets of specialised AI agents, operating more like chief executives of small digital workforces than individual contributors. Competence will shift towards the ability to frame questions, set objectives and constraints, interpret ambiguous results and iterate when literal answers miss the true need. This is where he sees human value persisting and potentially increasing: in agency, initiative, domain knowledge, interpersonal skills and situational judgement. If work is redesigned accordingly, AI amplifies human capability rather than simply substituting it, which in his view justifies a more optimistic forecast than typical narratives of wholesale job loss.

Strategic underdeployment inside organisations

Despite these possibilities, Brynjolfsson observes that many corporate AI initiatives remain superficial, misdirected or disconnected from core value creation. Hackathons produce playful applications such as automated lunch menus, while high-stakes processes in revenue generation, risk management, supply chains or customer service remain largely unchanged. This behaviour reflects a familiar pattern from earlier technologies: executives experiment at the periphery, but delay difficult redesign of central workflows, incentives and governance.

His argument that AI is underhyped therefore contains a critique of business strategy. The bottleneck is not access to models, but managerial imagination and organisational willingness to tackle the messy work of restructuring. In his policy testimony and interviews, he consistently urges leaders to focus AI efforts on economically material use cases, decompose roles into tasks that can be automated or augmented, and build agent-management skills across the workforce rather than treating AI as a bolt-on tool. He believes that firms which move fastest on this agenda will capture outsize gains in productivity and profitability, reinforcing his view that conventional forecasts underestimate AI's upside.

Distributional risks and concentrated power

The underhyped assessment does not imply that outcomes will be uniformly positive. Brynjolfsson repeatedly warns that AI could intensify inequality and concentrate economic and political power in a small cluster of companies or states. If ownership of data, models and digital infrastructure remains narrow, the benefits of higher productivity may accrue disproportionately to capital, senior talent and platform monopolies. Earlier waves of information technology already contributed to skill-biased technical change, widening gaps between workers with and without advanced education.

He argues that repeating the policy failures of globalisation would be a grave mistake. Trade raised aggregate output but left many displaced workers unsupported, generating political backlash. With AI, the stakes are higher: the same tools could enable extraordinary prosperity or unprecedented surveillance, autonomous weapons and engineered biological threats. Brynjolfsson's underhyped claim is therefore conditional. AI can be far more transformative than mainstream commentary allows, but whether that transformation yields shared prosperity or extreme concentration depends on deliberate choices around taxation, education, entrepreneurship, competition policy and safety regulation.

Beyond GDP: measuring the invisible value

One reason he believes mainstream assessments understate AI's impact is the inadequacy of conventional economic metrics. Standard GDP records market transactions and assigns zero weight to free digital goods such as search, email, Wikipedia and widely accessible AI tools. Brynjolfsson's research on alternative welfare measures, particularly GDP-B, tries to quantify consumer surplus by asking how much people would need to be paid to forego a given service.

Early results indicate that free digital services generate trillions of dollars in value that never appear in GDP figures. Similar surveys applied to language models suggest rapid growth in perceived value as adoption and usefulness increase. In other words, even if measured GDP and productivity rise only gradually, underlying welfare may be climbing much faster. This measurement problem reinforces his belief that AI's true economic significance is underhyped: official statistics and headline narratives are simply not capturing where value is created, especially when tools are free or bundled into existing products.

Why the next decade could be the best and the worst

Ultimately, the statement that AI is underhyped is less a prediction of automatic utopia than a call to recognise the scale of the stakes. Brynjolfsson argues that the coming decade could plausibly be the most prosperous period in human history, with breakthroughs in medicine, education, science and material abundance driven by AI-accelerated discovery and innovation. At the same time, he accepts that the same systems could enable catastrophic misuse, from bioengineered pandemics to pervasive manipulation and autonomous warfare.

In his view, the decisive variable is human agency. AI is a powerful tool for amplifying intention: it expands the reach and speed of whatever goals individuals, firms and governments choose to pursue. If societies channel that amplification towards entrepreneurship, problem-solving and shared prosperity, the technology's upside will be far larger than many current forecasts assume. If they neglect safety, inclusion and distribution, the downside could be equally dramatic. The remark that AI is underhyped is therefore both a descriptive claim about the technology's potential and a normative challenge: to upgrade institutions, strategies and values quickly enough that the eventual scale of impact is used well rather than squandered or weaponised.

Quote: Nvidia, Microsoft, Meta, Palantir, OpenAI and more than 20 other companies - Letter to policymakers, 24th July 2026

"Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect. And concentrating advanced AI capabilities behind a small number of closed models compounds that risk." - Nvidia, Microsoft, Meta, Palantir, OpenAI and more than 20 other companies - Letter to policymakers, 24th July 2026

The central policy dilemma is whether concentrating advanced artificial intelligence in a handful of sealed systems genuinely reduces risk, or whether it simply hides failure modes while magnifying the consequences of any breach or misuse when it eventually occurs. When a small number of firms operate closed models that shape information flows, productivity tools and critical infrastructure, any undetected flaw, exploit or design bias scales across millions of users and high-stakes environments without external parties being able to interrogate the system. The recent letter from Nvidia, Microsoft, Meta, Palantir and more than 20 other companies positions this dilemma directly in front of policymakers, arguing that restricting open-weight models in the name of safety could inadvertently deepen systemic exposure to opaque, concentrated AI capabilities.

Closed Models, Opaqueness And Undetectable Failure

Closed models are typically operated as remote services, with access mediated through proprietary APIs and contracts that reveal little about architecture, training data or guardrail implementation. Outsiders, including regulators and independent researchers, cannot easily inspect model parameters, replicate training conditions or stress-test behaviour across adversarial scenarios, which turns these systems into operational black boxes. When failures occur - whether hallucinated outputs in medical settings, covert prompt injection, data leakage or subtle discriminatory patterns - detection depends largely on the provider's monitoring and willingness to disclose issues, rather than on independent scrutiny. This creates a structural asymmetry: external users bear the consequences of model behaviour, but lack meaningful visibility into how that behaviour arises or how quickly systemic problems are identified and remediated.

This opaqueness extends to security posture. A closed model may employ strong internal controls, but third parties cannot verify whether safety features can be stripped out, circumvented or bypassed with techniques that require less effort than training a similarly capable system from scratch. Nor can they evaluate whether the distribution channels and attack surfaces - from SDKs and plug-in ecosystems to integrated office suites - are resilient against determined adversaries at scale. The claim that closed status is inherently safer therefore rests on trust in corporate assurances and limited audit rather than on open technical verifiability. In high-risk domains such as critical infrastructure management, defence applications or systemic financial decision-making, that gap between assurance and verifiable robustness becomes strategically significant.

Concentration Of Advanced AI Capability As A Systemic Risk

Beyond opaqueness, the statement targets concentration: the accumulation of cutting-edge AI capabilities in a few closed models controlled by a small number of firms operating at global scale. When the most powerful models are centralised, several systemic risks emerge. First, market power: if a handful of providers set pricing, access terms and acceptable use policies for de facto infrastructure models, they shape not only innovation pathways but also the distribution of safety standards across the economy. Second, correlated failure: any shared architectural vulnerability, misaligned fine-tuning practice or exploited control surface can propagate simultaneously across sectors that rely on the same underlying closed model. Third, geopolitical dependence: jurisdictions that lack domestic alternatives may find their digital sovereignty constrained by foreign providers' policy choices around censorship, surveillance or safety trade-offs.

Concentration also interacts with incentives around disclosure. A provider whose revenue depends heavily on a flagship closed model may be reluctant to fully expose its limitations and failure cases, particularly if admitting systemic weakness could trigger regulatory intervention or reputational damage. In contrast, a more plural ecosystem of open-weight and open-source models allows independent labs, academic groups and civil society organisations to stress-test and publish findings, distributing the epistemic load of safety assessment rather than rooting it in a small number of corporate actors. The coalition's warning suggests that trying to enforce safety by suppressing open alternatives might inadvertently lock society into dependency on concentrated closed systems whose true risk profile is only partially known.

Open-Weight Models As A Middle Ground

The companies signing the letter are not arguing for unbounded openness; they focus on open-weight models as a specific technical and governance compromise. Open-weights refer to releasing the trained parameter values of a neural network while often withholding training data and full pipeline code. This creates a middle state between proprietary closed models and full open-source AI: external parties can download, run and fine-tune the model, gaining operational autonomy and partial transparency, but do not necessarily gain full insight into data provenance or training procedure. For enterprises, this means the ability to host models on their own infrastructure, avoid sending sensitive data through third-party APIs, and tailor behaviour to sector-specific norms without relinquishing control to a remote provider.

At laboratory scale, recent work indicates that small open-weight models can be competitive with closed models in domain-adapted tasks, delivering reasonable performance at relatively low monetary cost and data requirements. That result undermines the assumption that safety and capability must be traded off against openness; in practice, organisations can achieve useful, robust performance with models they can inspect and adjust more freely. Furthermore, open-weight availability facilitates emerging best practice in abstention and privacy, allowing models to be configured to decline high-risk queries and to keep sensitive contextual data inside local environments rather than central data centres. These characteristics connect directly to the argument that distributing capability across many open-weight systems reduces the chance of a single catastrophic failure and increases the overall capacity for collective safety research.

Regulatory Context: EU AI Act And Open Components

The tension described in the statement sits against a rapidly evolving regulatory backdrop, particularly in Europe. The EU AI Act differentiates between general-purpose AI models, open-source AI components and monetised AI services, carving out specific exemptions and obligations for open-source offerings. Open components - including models and parameters - can benefit from lighter obligations when provided under free and open licences and not monetised directly, but general-purpose models that present systemic risk or are tied to paid services remain subject to full regulatory requirements. This framework reflects an attempt to balance transparency and innovation with concerns about misuse and high-risk applications, yet it also introduces complexity for open-weight providers whose licensing and business models may straddle categories.

Experts have pointed out that merely releasing weights under an ostensibly open licence does not automatically qualify as open-source AI, particularly when training data and methods remain secret. As a result, open-weight models may sit in ambiguous territory: more transparent than closed proprietary offerings, but not fully aligned with the four freedoms of open-source as defined by community standards. The coalition's letter effectively challenges regulators to recognise this nuance. Prematurely imposing blanket restrictions on open-weight distribution, or treating all openness as equivalent risk, could narrow the space for experimentation with safer, more verifiable architectures while leaving closed mega-models largely untouched. Conversely, failing to impose any obligations would ignore the genuine hazards of making powerful models widely accessible without safeguards. The regulatory question is therefore not simply open versus closed, but which forms of openness reduce systemic risk and which amplify it.

Debates, Objections And Safety Concerns

Critics of open-weight and open-source models argue that wider accessibility increases the surface for malicious use, such as building tailored disinformation engines, automating cyberattacks or circumventing safety filters by modifying the model locally. They contend that closed models at least allow firms to enforce centralised guardrails, monitor usage patterns and throttle dangerous behaviour, while open-weight distribution makes it difficult to prevent determined adversaries from weaponising the technology. Some policy proposals therefore advocate temporary pauses on high-capability releases, registration and licensing schemes for systems above specific compute thresholds, and stricter control over distribution channels until security practices mature. From this perspective, the coalition's warning might appear self-serving: firms that benefit commercially from open-weight ecosystems could be seen as resisting necessary restraint.

Proponents of openness respond that security through obscurity is an unstable foundation, especially given the reality of model leaks, insider threats and sophisticated reverse-engineering efforts. They argue that openness enables broader participation in red-teaming, safety benchmarking and governance innovation, and that diverse open-weight models reduce monoculture risk by preventing any single vendor stack from dominating critical infrastructure. Additionally, many harms associated with generative models - from synthetic media misuse to privacy violations - are tied more to application design, deployment context and human incentives than to whether underlying weights are secret. The letter's wording reflects this stance: the real danger lies not simply in models being open or closed, but in concentrating advanced capabilities behind a small number of opaque, uninspectable systems that operate at planetary scale.

Strategic And Market Implications

Strategically, the debate shapes the trajectory of both national competitiveness and industrial structure. The signatories argue that open-weight models are essential to preserving technological leadership by allowing domestic firms, researchers and start-ups to build upon shared foundations without prohibitive licensing costs or API dependency. If policymakers heavily constrain open-weight development in the name of safety, they risk pushing cutting-edge experimentation to jurisdictions with more permissive regimes, thereby undermining domestic capacity to shape global norms. At the same time, large incumbents such as Nvidia, Microsoft and Meta have substantial commercial interests in open-weight ecosystems, from selling compute and tooling to providing platforms for fine-tuning and deployment. Their stance therefore mixes genuine systemic concern with strategic positioning in a competitive landscape defined by both closed premium models and rapidly advancing open alternatives.

For enterprises, the outcome of this policy debate will determine whether AI remains primarily a vendor-mediated service or becomes a configurable infrastructure asset that can be tailored and audited within organisational boundaries. A regime that privileges closed models could simplify compliance by outsourcing safety obligations to a few large providers, but at the cost of dependency, limited transparency and constrained customisation. A regime that supports responsibly governed open-weight models could broaden innovation and resilience, but demands stronger in-house expertise, more sophisticated risk management and clearer norms around documentation, licensing and accountability. The statement from the coalition marks a turning point: it invites policymakers to recognise that safety is not guaranteed by closure or concentration, and that a genuinely robust AI ecosystem may require plural, inspectable, and, where appropriate, open-weight models rather than a small constellation of unchallengeable black boxes.

Term: Zero-day vulnerability - Cyber

"A zero-day vulnerability is an undisclosed security flaw in software, hardware, or firmware that is unknown to the developers or parties responsible for patching it. The term signifies that the vendor has had "zero days" to create a fix or release a security update to protect users against malicious attacks." - Zero-day vulnerability - Cyber

Security teams confront a distinctive problem when a flaw can be weaponised before any tailored defence or patch exists: the system is vulnerable, yet neither signatures nor vendor guidance offer protection. That situation characterises zero-day vulnerabilities and explains why they so often underpin high-impact breaches, espionage operations, and disruptive attacks across modern digital infrastructure.

Structural nature of a zero-day vulnerability

The practical substance of a zero-day vulnerability is a defect in software, hardware, or firmware that enables behaviour the designer did not intend, and which can be triggered by an adversary to gain a strategic advantage. Because the parties responsible for maintenance and patching have not yet recognised the flaw, there is no bespoke fix, configuration change, or detection signature targeting it. In operational terms, the vulnerability remains latent until either discovered by attackers, reported by researchers, or observed through anomalous system behaviour, and during that window defenders must rely entirely on general controls such as network segmentation, strict access management, and behaviour-based monitoring rather than vulnerability-specific countermeasures.

Crucially, security practice distinguishes between the flaw, the exploit, and the attack as different stages of the same threat lifecycle. The zero-day vulnerability is the underlying weakness in the asset; the exploit is the technical method or code that triggers the weakness to do useful work for the attacker; and the attack is the application of that exploit against real targets to steal data, move laterally, or disrupt operations. This separation matters because mitigations can apply at each stage: secure coding and code review seek to reduce vulnerabilities; exploit prevention mechanisms such as modern operating system protections aim to raise the cost of developing reliable exploits; and incident response, monitoring, and segmentation seek to reduce the impact of successful attacks.

Risk characteristics and practical impact

Zero-day vulnerabilities are disproportionately dangerous because they combine three factors: lack of patch, lack of specific detection, and asymmetry of knowledge between attacker and defender. When only adversaries or a small circle of researchers know a flaw exists, defenders neither track it in routine vulnerability scanning nor receive advisories from vendors or regulators. Attackers can therefore use the vulnerability to obtain initial access, escalate privileges, bypass authentication, or execute arbitrary code with a high probability of success, particularly in widely deployed platforms such as operating systems, web browsers, VPN gateways, or email servers. Recent case studies of exploited zero-days across major vendors demonstrate that these weaknesses are routinely used for ransomware deployment, credential theft, covert persistence and espionage activity in both corporate and governmental environments.

From a governance perspective, zero-day risk is systemic rather than local. A single critical vulnerability in a widely used component can expose thousands of organisations simultaneously, with little warning. The empirical literature on patching behaviour shows that even after disclosure, timely remediation is uneven: vulnerabilities affecting multiple vendors and causing scope change tend to be patched faster, while those requiring special privileges or impacting confidentiality are less likely to be corrected quickly. This reinforces a central practical point: the danger is not eliminated when a vendor releases a fix. It persists across all assets where the patch has not yet been applied, which may include legacy systems, devices with complex update cycles, or environments where patch-induced downtime is treated as unacceptable.

Mathematical framing of zero-day exposure

In quantitative risk analysis, zero-day exposure can be conceptualised as the probability that an unknown, unpatched flaw is present in a given asset, multiplied by the probability that a capable adversary has discovered and is exploiting it. If we denote by the probability that a system contains at least one zero-day vulnerability and by the conditional probability that an adversary both knows the vulnerability and chooses to exploit it against that system, the probability of compromise via zero-day in a given time window can be sketched as . This simple representation highlights several levers for defence. Reducing depends on software engineering quality, defensive programming, and proactive security testing such as fuzzing and code analysis. Reducing depends on making the system a less attractive or more difficult target through segmentation, zero trust principles, and hardening that increases attacker cost relative to expected benefit.

From a portfolio perspective, organisations sometimes treat zero-day risk as an unavoidable background rate of compromise inherent to operating complex systems in an adversarial environment, elevating residual risk management over absolute prevention. If the expected loss from zero-day events over a horizon is , where is the impact if asset is compromised, the strategy becomes to lower via data minimisation, strong isolation, and robust backup and recovery, even when cannot be driven close to zero. This conception underpins the growing emphasis on resilience, incident response readiness, and breach mentality as counterparts to traditional perimeter defence.

Discovery, disclosure, and ethical tensions

Discovering a zero-day vulnerability places the finder at the centre of an ethical and strategic dilemma: whether to disclose it responsibly to the vendor, sell it on a grey market, weaponise it for offensive operations, or withhold it entirely. States and security agencies have historically maintained stockpiles of undisclosed vulnerabilities for intelligence and military use, while private markets for exploits and vulnerability information offer significant financial incentives to researchers and criminals alike. This raises a policy question about how long such flaws should be retained before disclosure, given that the same weakness might independently be discovered and exploited by hostile actors. Some jurisdictions have explored formalised vulnerability equities processes to balance national security benefits of using zero-days against the collective security benefits of patching them, but practice remains uneven and often opaque.

Within commercial security, the dominant norm remains coordinated disclosure, where researchers privately inform vendors, allow a window for patch development, and only later publish details. However, zero-day status by definition ends once a patch or mitigation is widely available, even if many systems remain vulnerable due to slow deployment or operational constraints. At that point the vulnerability becomes an N-day issue, and public exploit code may appear quickly, making patch management and compensating controls urgent. Ethical debates continue over whether the publication of proof-of-concept exploit code accelerates defensive understanding or unnecessarily lowers the barrier to entry for attackers, especially when adoption of patches is delayed.

Defensive strategies beyond signatures

Because zero-day vulnerabilities are unknown and unpatched at the moment of exploitation, classical signature-based defences such as traditional antivirus or rigid intrusion detection rules offer little protection. Modern defensive architectures therefore prioritise behavioural and anomaly-based detection that monitors for deviations from established baselines of system and user activity. Techniques including endpoint detection and response, user and entity behaviour analytics, and AI-driven anomaly detection seek to recognise the consequences of exploitation, such as unusual process spawning, unexpected network connections, or anomalous access patterns, rather than the specific exploit code itself. Organisations that assume compromise and instrument their environments to detect lateral movement, privilege escalation, and data exfiltration are better positioned to identify zero-day abuse early in the attack chain.

Zero trust architecture plays a complementary role by reducing the blast radius of initial compromise. By limiting implicit trust, enforcing strong authentication and authorisation at each access decision, and segmenting networks into smaller trust zones, defenders ensure that a single exploited vulnerability does not automatically yield broad access. Additional measures such as application sandboxing, strict patch management for known flaws, compensating controls on unpatchable systems, and deception technologies like canary tokens further constrain attacker progress when a zero-day is present. These layered approaches acknowledge that prevention cannot be guaranteed, but that sophisticated attackers can be slowed, detected, and contained.

Why zero-day vulnerabilities remain strategically important

Zero-day vulnerabilities continue to matter because they sit at the intersection of software engineering, geopolitics, and organisational resilience. They expose structural weaknesses in digital supply chains, test the adequacy of disclosure processes, and reveal how quickly vendors and customers can coordinate patching at scale. While the absolute number of discovered zero-days in a given year attracts attention, more significant is the pattern of which technologies they affect, how quickly exploitation follows disclosure, and how many organisations maintain sufficient instrumentation and discipline to detect and respond. In an environment where critical infrastructure, finance, healthcare, and public services depend on complex, interconnected systems, the existence of unknown, exploitable flaws is not an anomaly but a persistent condition to be managed.

This continuing relevance shifts best practice from a mindset of eliminating all vulnerabilities to one of operating securely despite them. Effective organisations invest in secure development to reduce the introduction of new flaws, adopt robust vulnerability management and rapid patching to shorten exposure windows after disclosure, and design architectures that assume some unknown weaknesses will be exploited. Zero-day vulnerabilities thus act both as specific technical threats and as a lens through which the maturity of wider security strategy can be assessed, making them central to any serious discussion of contemporary cyber risk.

Global Advisors News Brief - July 25 2026

Read the full brief at the link

Headlines for the last 24hrs

1. Trump Administration Escalates Global Trade War with Broad New Tariffs in Retaliation for EU Tech Fines
2. Anthropic Launches Claude Opus 5, Escalating the Race for Advanced Agentic AI
3. Credit Agencies and Wall Street Sound Alarms Over Tech Giants' Massive AI Infrastructure Spending
4. Tech Leaders Push Back Against US Government Proposals to Ban Open-Weight AI Models
5. Global Oil Prices Spike Toward $100 per Barrel Amid Middle East Conflict and Geopolitical Supply Risks
6. Rogue AI Incidents and Cybersecurity Lapses Accelerate Regulatory Pressure for Mandatory AI Safety Controls
7. Venture Capital Inflows Accelerate for Defense Tech and Next-Generation AI Startups
8. AI Deployment Drives Corporate Restructuring and Reshapes Entry-Level Hiring Models
9. Rising Yields and Inflation Uncertainties Cloud Federal Reserve Interest Rate Trajectory
10. Automakers and Mobility Platforms Realign Strategic Partnerships in Autonomous Driving Tech

Time window: 2026-07-24T05:00:33.073Z to 2026-07-25T05:00:33.073Z

Quote: Professor Erik Brynjolfsson - Stanford economist, Director of the Stanford Digital Economy Lab

"Everybody is a coder now." - Professor Erik Brynjolfsson - Stanford economist, Director of the Stanford Digital Economy Lab

The practical tension behind contemporary AI is not simply automation versus preservation of jobs, but a deeper restructuring of who can translate ideas into functioning systems and products . Economic and technical power is shifting towards people who can articulate a problem and supervise computational agents that do most of the execution. When coding itself ceases to be an elite bottleneck, the scarcity moves away from syntax and towards judgement, domain understanding and the ability to turn fuzzy intentions into precise, testable instructions .

From coding as craft to coding as capability layer

For several decades, software development has operated as a specialised craft. Entry has been gated by formal education, arcane tooling and the need to internalise complex abstractions before producing reliable code. That gating created high wages, a distinct professional identity and a powerful role in organisational decision-making. The rapid advance of AI coding assistants and agentic tools is dismantling much of that barrier by turning natural language descriptions into functioning code, scaffolding projects, debugging and proposing architectures in minutes . This reframes coding from a scarce craft to a ubiquitous capability layer: if a motivated non-specialist can specify a task clearly, tools such as Claude Code, Replit or cursor can generate usable software with minimal manual intervention .

In Brynjolfsson's teaching at Stanford, this shift has already changed behaviour. Students who previously ended the term with slideware are now expected to ship working software irrespective of whether they had prior programming experience . The constraint is no longer the ability to write loops or manage memory; it is the ability to define a worthwhile problem, express requirements clearly and iterate intelligently on the outputs. Within that pedagogical environment, everybody is treated as having latent access to coding capability, with AI bridging the gap between intention and implementation .

Economic context: AI capability outrunning adoption

Despite impressive benchmark gains in coding, mathematics and reasoning, the measured economic impact of AI remains muted in aggregate productivity statistics . Brynjolfsson consistently emphasises the gap between technical capability and realised value: large language models and coding agents can already perform substantial portions of knowledge work, yet most organisations have barely reconfigured their workflows, decision processes or product strategies to exploit that fact . He describes the present as the turning point of a -curve in technology adoption: capabilities have risen quickly, but business models, skills and institutions will take several years to catch up. His wager with economist Robert Gordon, that productivity by 2030 will exceed official forecasts, rests on the expectation that systematic redesign will convert existing AI potential into measured output in the coming 3 to 5 years .

Coding sits at the centre of this gap. On capability metrics, generative models already perform strongly on software tasks, achieving rapid gains on programming benchmarks and powering agents that can build non-trivial applications from high-level prompts . Yet large parts of the economy still treat software as a scarce resource delivered by specialist teams on long timelines. The assertion that everybody is now a coder is therefore both a description of technical reality and a provocation aimed at lagging organisational and educational models: it suggests that bottlenecks in software production should be re-examined as failures of problem selection, coordination and adoption rather than immutable constraints of human skill supply .

Tasks, not occupations: how coding is being restructured

Brynjolfsson's broader research distinguishes between occupations and tasks. Every job comprises a bundle of activities, only some of which are currently tractable for AI . Radiology, for example, includes reading images, coordinating with clinicians, interpreting lab results and making holistic judgements; AI excels at image interpretation but is less competent at the interpersonal and integrative aspects . Coding follows a similar pattern. Tasks such as boilerplate generation, test creation, standard API wiring and translating business logic into conventional patterns are increasingly within reach of AI agents. By contrast, tasks such as system-level architecture, security-sensitive design, negotiation of requirements across stakeholders and integration with complex organisational constraints remain stubbornly human-intensive .

This task-level view clarifies why junior roles are being hit first. In coding, entry-level jobs disproportionately involve routine work that is now automated or heavily augmented by AI, leading to measurable declines in employment among workers under 25 in highly exposed roles . Firms can generate equivalent or greater output with fewer junior staff when AI handles much of the repetitive coding burden. However, mid-level and senior responsibilities rely more on problem framing, cross-functional coordination and strategic technical judgement, which are currently less automatable. The claim that everybody is now a coder therefore coexists with a serious structural problem: traditional pathways that relied on routine coding to develop experience are being eroded precisely as coding capability becomes ubiquitous .

New hierarchy of skills: problem-definition, domain expertise and agent management

Brynjolfsson breaks most knowledge projects into three stages: defining the question, executing the work and evaluating the result . AI is increasingly strong at the second stage. Once a problem is specified, agents can write code, query data, generate documentation and iterate quickly. The durable human advantage shifts to the first and third stages: choosing valuable problems, framing them in context, setting constraints and assessing whether the outputs are correct, relevant and robust . In that structure, coding fluency becomes a component of execution that can be delegated; the higher-order skill is managing a portfolio of agents to pursue complex objectives.

He predicts that workers will act as miniature chief executives of fleets of AI agents, orchestrating specialised systems for research, coding, analysis and customer interaction . In such roles, everybody must be able to interact productively with code generation tools, understand basic computational concepts and spot when an AI-produced solution is brittle or misaligned. Technical literacy remains necessary, but deep expertise in specific syntaxes becomes less central than the ability to integrate AI outputs into accountable decisions. This aligns with his broader framing of AI plus X: meaningful value is created by combining domain expertise in areas such as medicine, finance or marketing with competent use of AI tools, rather than by isolated mastery of either component .

Debates and objections: does ubiquity trivialise software?

The idea that everybody is a coder is controversial. One objection is that equating prompt-based generation with software engineering trivialises complexity, underestimates the risks of insecure or poorly structured code and misleads organisations into thinking they can dispense with expertise. Critical systems, from healthcare devices to financial infrastructure, still demand rigorous testing, formal methods and robust design that cannot be captured fully through conversational interfaces. Bugs, vulnerabilities and design errors can have serious consequences, particularly when generated code is deployed without deep review. From this perspective, coding remains a craft in high-stakes domains and the assertion risks fuelling irresponsible deployment .

Brynjolfsson's work partially acknowledges this tension by emphasising evaluation, domain understanding and human oversight. His empirical studies of call-centre AI, which show productivity gains of 14 to 15 percent and larger benefits for less experienced workers, are grounded in environments where AI recommendations are supervised and embedded within human workflows . Customer satisfaction improves and novice workers learn, but firms do not simply let unsupervised agents make all decisions. Translated into coding, the analogous pattern is collaborative development: AI proposes code, humans assess, adapt and integrate. In this view, ubiquity does not eliminate craft; it changes what craft consists of, moving from manual typing towards architecture, verification, risk assessment and alignment with organisational goals .

Strategic implications for organisations and education

If everybody is treated as a potential coder, organisations must rethink how they structure innovation and capability building. Brynjolfsson criticises shallow experimentation, such as hackathons that produce playful applications unrelated to core value creation, as symptomatic of under-structured adoption . He argues that firms should focus AI efforts on economically meaningful workflows, decomposing them into tasks that can be automated, augmented or reconfigured, and then redesigning end-to-end processes to embed AI deeply . Coding becomes a pervasive tool for process change rather than a specialised service function; non-technical staff with domain expertise can propose, test and iterate software-mediated solutions directly, assisted by AI, reducing friction between ideas and implementation.

Education faces parallel pressures. Courses organised around stepwise procedural instruction, such as manual matrix inversion or routine algorithm implementation, lose value when AI can perform those procedures reliably on demand . Brynjolfsson advocates shifting towards open-ended problem-solving, Socratic questioning and explicit practice in defining and evaluating problems . Students are cold-called to articulate the core question, design empirical strategies and critique outputs, building capabilities that align with agent-management roles rather than repetitive execution. Under this model, basic coding literacy is assumed; the differentiator becomes the ability to connect technical tools with relevant, high-impact problems and to exercise rigorous, context-sensitive judgement about proposed solutions .

Why it matters: agency, inequality and shared prosperity

Viewed through Brynjolfsson's broader lens, the idea that everybody is a coder is ultimately about agency. Tools that turn ideas into functioning software lower the cost of experimentation, enabling more people to test business concepts, automate aspects of their work and create digital assets without traditional gatekeepers . That shift could support an economy with wider participation in entrepreneurship and innovation, provided access to AI tools is broadly distributed and people acquire the skills to use them productively. At the same time, he warns that AI may deepen concentration of wealth and power if the underlying models, data and platforms are controlled by a small number of corporations or states . In such a scenario, being a coder in principle may not translate into meaningful economic autonomy if value capture mechanisms remain highly centralised.

His preferred response is a strategy of shared prosperity: redesigning education, tax systems and organisational practices so that many individuals can use AI to create value and retain a significant stake in that value . Encouraging entrepreneurial behaviour, supporting transition pathways for displaced workers and treating workers as stakeholders in the gains from automation all form part of this agenda. In that context, the statement that everybody is a coder is both descriptive and normative. It describes the technical reality that AI has democratised coding capability, and it implies a policy and organisational responsibility to ensure that this new capacity amplifies broad human intention rather than reinforcing narrow concentration. Whether the next decade becomes, as Brynjolfsson suggests, one of the best or one of the worst in history will depend on how societies choose to steer a world in which the power to create software is, at least technologically, within almost everyone's reach .

Term: Acid-test ratio (also called the quick ratio) - Finance

"The acid-test ratio (also called the quick ratio) is a strict liquidity metric that measures a company's ability to cover its short-term debts using its most liquid assets. It excludes inventory, which can be hard to sell quickly. An ideal ratio is 1.0 or higher." - Acid-test ratio (also called the quick ratio) - Finance

Pressure to meet payroll, service suppliers and roll over short-term borrowing means liquidity failures tend to unfold rapidly, long before a business shows distress in its profit figures. Insolvency in practice is usually triggered not by long-run unprofitability but by the inability to convert assets into cash on the timetable imposed by creditors, which is why analysts focus closely on how quickly balance sheet resources can be mobilised to meet near-term obligations.

Strict liquidity and the role of quick assets

The core issue is whether a firm could settle its current liabilities if lenders and suppliers suddenly demanded payment without granting fresh credit. The relevant resources are its quick assets: cash and cash equivalents, marketable securities, and trade receivables that are expected to turn into cash within roughly 90 days. These items can usually be realised without material loss of value or operational disruption. By contrast, inventories may require discounting, marketing costs and time to sell, while prepaid expenses and other miscellaneous current assets cannot be converted directly into cash. The metric commonly used to isolate this strict subset of assets is the acid-test ratio, which compares quick assets to current liabilities and asks how many monetary units of highly liquid resources back each unit of short-term obligation.

Practical specification and alternative formulas

In practice, two equivalent formulations are widely used. The direct version isolates quick assets on the balance sheet and calculates . Where balance sheet subtotals are less granular, analysts often work indirectly: , so . Both approaches generate substantially the same figure because inventory and prepayments are the principal exclusions from current assets. Quick assets are thus a deliberate subset that captures only those items likely to be realisable within the next three months at close to book value.

Interpreting values and the 1,0 benchmark

Most teaching texts and professional bodies present 1,0 as the rule-of-thumb threshold for comfortable liquidity on this strict definition. A ratio at or above 1,0 implies that, on paper, the firm could extinguish all current liabilities using only its most liquid assets, without selling inventory or relying on new financing. Values between about 0,8 and 1,0 are often read as warning territory: the business may manage in normal conditions but could be exposed if credit tightens or if collections slow. Ratios significantly below 1,0 flag dependence on either ongoing cash generation, rapid inventory turnover, or continued access to credit lines; in adverse conditions this dependence can become a vulnerability. On the other side, ratios markedly higher than 1,0, such as 1,5 or 2,0, indicate substantial liquidity reserves but may also hint at under-utilised assets, especially if large cash holdings coexist with limited investment opportunities.

Relationship to the current ratio and why inventory is excluded

The tension between different liquidity measures is largely about the treatment of inventory. The broader current ratio includes inventory in its numerator, asking whether all current assets together exceed current liabilities, whereas the acid-test ratio seeks to answer whether the company could settle obligations without resorting to inventory liquidation. The exclusion reflects several practical considerations. First, inventories vary markedly in liquidity: highly standardised, fast-moving items may be saleable quickly, but specialised, seasonal or obsolete stock may only clear at deep discounts. Secondly, relying on inventory liquidation to meet debts can be operationally disruptive, forcing fire-sale pricing or production cuts that damage long-run viability. Thirdly, inventories are vulnerable to sudden write-downs when market conditions change. By stripping them out, the acid-test ratio focuses on assets that are conceptually closer to cash and less exposed to these frictions. This makes it a more conservative, though sometimes harsher, gauge of resilience than the current ratio.

Parameter meanings and basic analytical uses

Each component of the ratio conveys distinct information about liquidity structure. Cash and cash equivalents represent immediate purchasing power and buffer operational shocks. Marketable securities add a second line of defence; they may offer yield but can still be liquidated quickly in normal markets. Trade receivables embed both the firm's pricing and credit policies: generous payment terms or poor collection practices inflate receivables while weakening cash posture. Current liabilities aggregate short-term borrowing, trade payables, accrued expenses and tax obligations, making them a composite measure of near-term claims on the firm. Analysts use the acid-test ratio to test scenarios such as tighter supplier credit, delayed customer payments or withdrawal of overdraft facilities, asking whether existing quick assets would suffice to bridge such shocks without emergency measures. For lenders and investors, the metric becomes a quick filter for identifying businesses that rely heavily on slow-moving assets or future earnings to meet present commitments.

Industry differences, working capital strategies and debates

One limitation of a simple numeric threshold is that liquidity needs vary structurally by sector. Studies of cross-industry data show technology firms and some services businesses often sustain quick ratios in the 1,5-4,0 range, reflecting low inventory intensity and large cash holdings. Manufacturing groups may operate successfully around 0,8-1,2, while retailers and utilities sometimes run at 0,5-0,8 because their business models rely on rapid inventory turnover or stable cash flows to support lower levels of quick assets. This raises a technical debate about whether the acid-test ratio should be interpreted against fixed global standards or primarily relative to peers. Some practitioners argue that as long as cash generation and credit lines are robust, a low ratio can be acceptable, particularly in sectors where inventory is demonstrably liquid. Others emphasise systemic risk: during downturns, assumptions about inventory salability and credit access often break down simultaneously, leaving firms with low acid-test ratios exposed to sudden funding gaps. A further debate concerns the inclusion of certain items such as near-term portions of long-term investments or highly liquid inventories, which could in theory be treated as quick assets but are usually excluded for conservatism.

Mathematical framing and dynamic perspective

Although the ratio itself is a simple fraction, formal working-capital models often embed it within time-based cash-flow analysis. Let denote quick assets at time and current liabilities at time . The acid-test ratio is then . Analysts may project via expected operating cash inflows and receivables collections, for example , where represents net cash generated from operations over the period and cash payments for operating costs and interest. Similarly, , where captures new short-term borrowing and payables incurred, and repayments. This simple dynamic structure allows practitioners to test how changes in credit terms, collection speed or borrowing affect over time. It becomes clear that a reassuring snapshot ratio can deteriorate quickly if receivables stretch out or if liabilities increase faster than liquid assets, underscoring why trend analysis and scenario modelling are more informative than a single-period snapshot.

Contemporary relevance and limitations

Despite the growth of sophisticated cash-flow forecasting tools, the acid-test ratio remains embedded in loan covenants, investment screens and credit scoring models. Its endurance lies in its simplicity and its focus on truly liquid assets, which can be applied quickly to any set of financial statements. However, as a static point estimate it has limitations. It ignores off-balance-sheet facilities such as committed credit lines, and it treats all receivables and marketable securities as equally realisable, glossing over credit quality and market liquidity differences. It also provides no direct insight into the timing of cash inflows and outflows within the reporting period or into structural drivers such as business model resilience. Consequently, sophisticated users interpret the acid-test ratio alongside cash-flow statements, maturity profiles of debt, and qualitative assessments of customer and supplier relationships. Used in this broader context, it remains a valuable indicator of whether a company could withstand sudden funding stress without having to liquidate inventory or pursue emergency refinancing.