Our Times — Reporting on the forces reshaping technology, business, and science.

Behavioural Remedies Keep Failing. Regulators Are Finally Saying So.

Daniel Okonkwo — Thu, 13 Aug 2026 00:00:00 GMT

The remedy has always been the weak link in competition enforcement. Establishing that a firm holds market power and abused it is hard but tractable. Fixing it has, for twenty years, mostly meant asking the firm to behave differently and appointing someone to check. The record of that approach is now long enough to evaluate, and enforcers in three jurisdictions have started saying out loud what the evidence supports: it does not work.

The shift is from conduct remedies to structural ones. That is a much larger change than the language suggests.

What a behavioural remedy actually asks

A conduct remedy tells a dominant firm to stop doing a specific thing: no self-preferencing in this ranking surface, no bundling this product with that one, offer these terms on a non-discriminatory basis. A monitor reports on compliance. The firm’s incentives are unchanged.

Three failure modes recur so reliably that they are effectively predictable.

Substitution. The prohibited mechanism is discontinued and a functionally equivalent one appears. The undertaking was written against an implementation, not an incentive.
Definitional drift. Terms like “equivalent treatment” and “comparable access” get litigated for years. The firm’s interpretation governs during the argument.
Monitor asymmetry. The monitor has a handful of staff and depends on the firm for data about the firm’s own conduct.

We spent six years supervising an undertaking that the market had routed around in eighteen months. That is not enforcement. It is an expensive form of documentation.

That assessment came from a former case officer at a European competition authority, who asked not to be named because the matter remains partially under review.

What changed the calculus

Two things. The first is simply elapsed time. There are now enough completed conduct remedies to compare intended outcomes against measured market structure, and the comparisons are unflattering. Market shares in several supervised markets are more concentrated at the end of the undertaking than at the start.

The second is a change in how enforcers frame the counterfactual. The traditional objection to structural relief is that it is disproportionate and risks destroying efficiencies. That argument carries less weight once the alternative has a documented failure rate, because the comparison is no longer divestiture against a working conduct remedy. It is divestiture against a remedy that predictably does not bind.

What structural relief looks like in practice

It is not always a break-up, and treating it as synonymous with one obscures the more likely outcomes.

The most common form is divestiture of a specific asset that creates the conflict: an ad exchange separated from the demand side that trades on it, a marketplace separated from the private-label operation competing on it, a payment rail separated from the platform mandating its use. The test is whether the conflicting incentive is removed rather than supervised.

A second form is mandated interoperability with a technical standard set outside the firm. This is structural in effect even though no asset moves, because compliance becomes observable by third parties rather than by a monitor reading the firm’s own reports.

The third, least developed, is data separation: prohibiting the transfer of data between units rather than prohibiting a use of it. Enforcers are cautious here because the monitoring problem returns unless the separation is architectural.

The cost, and who bears it

Structural remedies are slower, more expensive to litigate, and more vulnerable on appeal. They also produce irreversible outcomes, which is a genuine risk when the theory of harm turns out to be wrong.

Those objections are real and are not going away. What has changed is that they are now weighed against a measured failure rate rather than against a hypothetical. The firms facing this shift understand the stakes precisely, which is why the fight over the next several years will be about remedy design rather than liability.

For the market-structure context in which these cases arise, see our reporting on where venture capital is concentrating.

Inference Is Now the Line Item That Decides Which AI Products Survive

Maya Iyer — Wed, 12 Aug 2026 00:00:00 GMT

For three years the number that mattered in machine learning was the size of the training run. It was legible, it was expensive, and it made for a clean narrative about scale. That number now tells you almost nothing about whether a product works as a business. The number that does is the cost of answering a single request, and six engineering teams who shared their figures with Our Times describe the same uncomfortable arithmetic.

Training is a capital expense you amortise. Inference is a variable cost you pay on every interaction, forever, and it scales with exactly the thing you are trying to grow.

The gap between demo economics and production economics

A retrieval-heavy assistant that costs a fraction of a cent per query in a demo will often cost twenty to forty times that in production. The teams we spoke to attributed the gap to four things, and they listed them in roughly the same order.

Context growth. Demos use short prompts. Real users paste documents. Attention cost grows faster than linearly in sequence length, so the tail of long requests dominates the average.
Retries and cascades. A single user action fans out into several model calls once you add reranking, tool use, and a validation pass.
Idle capacity. Reserved accelerators are billed whether or not they are saturated. One team reported 31% average utilisation against a peak they had provisioned for.
Guardrails. Safety classification, moderation, and output checking are additional forward passes that nobody puts in the pitch deck.

We shipped a feature with a gross margin that was negative for eight months and nobody noticed, because the cost sat in a shared infrastructure budget rather than against the product line.

That quote came from a platform engineering lead at a company with more than 400 employees, on condition that we not name the product. It was the most common structural failure described to us: serving cost is real, but it is not attributed, so it cannot be managed.

What the teams that fixed it actually did

None of the successful interventions involved a better model. They involved routing, caching, and a willingness to say no to requests.

The most effective single change, reported by four of six teams, was tiered routing: classify the request, send the easy majority to a small model, and reserve the expensive path for cases that need it. The teams that measured it found that between 60% and 80% of production traffic did not need their flagship model at all. Users did not notice the difference, because the requests being downgraded were the ones with unambiguous answers.

Caching came second, and the surprise was how much of it was viable. Semantic caching on normalised queries produced hit rates of 18% to 35% in support and documentation workloads. That is a direct multiplier on cost, and it also cuts latency, which improves the metric the product team actually cares about.

The third lever was the least popular and the most effective: limiting context. Two teams capped retrieved context aggressively and measured no degradation in answer quality, because most of the retrieved material had been noise. One of them had been paying to process an average of 14,000 tokens of context to answer questions that needed 1,200.

Why the accounting matters more than the optimisation

The teams that had cost under control shared one non-technical trait. They had a per-request cost figure that a product manager could see, in the same dashboard as engagement and retention.

Where serving cost lived in a central infrastructure line, it behaved like weather: everyone complained, nobody owned it. Where it was attributed per feature, the optimisation happened without anyone mandating it, because the person who had to justify the feature also had to justify its margin.

This is not a new lesson. It is the same lesson cloud migration taught a decade ago, arriving again with a different bill attached. The difference is magnitude. A wasteful web service costs you a percentage. A wasteful inference path costs you the product.

The part that does not optimise away

There is a floor, and several teams have hit it. If the task genuinely requires a large model over long context with verification, the cost is the cost, and the only remaining moves are pricing and scope. Two of the six teams had raised prices. One had removed a feature entirely after concluding that no plausible efficiency gain would make it viable at the price point customers would accept.

That is a healthy outcome, and it is happening more often. The supply picture matters here too: as we reported in our analysis of the advanced packaging bottleneck, accelerator availability is not improving on the timeline most 2025 capacity plans assumed. Teams that budgeted for cost declines driven by hardware abundance are revising those assumptions.

The teams likeliest to survive the next two years are not the ones with the best benchmark scores. They are the ones who can tell you, to the cent, what a request costs and what it earns. For more on how open-weight alternatives change that calculation, see our reporting on where open models still lose on deployment.

Open-Weight Models Closed the Benchmark Gap. Deployment Is Where They Lose.

Maya Iyer — Mon, 10 Aug 2026 00:00:00 GMT

On the public leaderboards the argument is over. The best open-weight models now sit within a few points of the best proprietary ones on most published evaluations, and on several tasks they are ahead. Teams that have actually deployed both describe a gap that remains substantial, and it has almost nothing to do with capability.

The difference shows up in the parts of a system that no benchmark scores: throughput under concurrency, behaviour at the edges of the input distribution, and the operational cost of being the party responsible when it breaks.

What benchmarks measure and what they miss

A published evaluation measures single-request quality on a curated input distribution, usually with generous latency budgets and no cost ceiling. Production measures something else.

Four teams running both classes of model in production identified the same divergences.

Throughput at target latency. A model that matches on quality can require substantially more accelerator time to hit the same p95 latency under real concurrency. That is a cost difference, not a quality difference, and it does not appear on a leaderboard.
Tail behaviour. Malformed input, adversarial prompts, mixed languages, and very long context are underrepresented in evaluation sets and overrepresented in real traffic.
Structured output reliability. Teams consistently reported more schema violations from open-weight models when asked for strict JSON, which matters enormously when the output feeds a downstream system rather than a human.
Change management. A hosted endpoint changes underneath you, which is a real risk. A self-hosted model does not change unless you change it, which sounds better until you own the upgrade, the regression testing, and the rollback.

We did not switch back because the open model was worse. We switched back because we were spending two engineers on serving infrastructure and the vendor bill was cheaper than those two engineers.

The cases where open weights clearly win

The picture is not one-sided, and the teams that stayed on open weights had specific, legible reasons.

Data residency was the most common. If the requirement is that inputs never leave a jurisdiction or a private network, the decision is made before quality enters the conversation.

The second was high-volume narrow tasks. Classification, extraction, and routing at large scale is exactly where a smaller fine-tuned open model is not just adequate but preferable, because the per-request cost difference compounds and the task distribution is narrow enough that tail behaviour is controllable. This is the same insight driving the tiered-routing pattern we documented in our reporting on serving-cost economics.

The third was pricing leverage. Two teams described maintaining a functional open-weight deployment specifically as a negotiating position, and both reported it worked.

What the honest comparison requires

The comparison most teams run is not the comparison they should run. Quality on a held-out set is the easy part. The complete accounting includes accelerator hours at target latency, engineering time for serving and upgrades, the cost of the evaluation harness you now have to maintain yourself, and the residual risk you have absorbed by becoming the responsible party.

Run that comparison and the answer stops being ideological. It becomes a straightforward function of volume, task breadth, and how much engineering capacity you have to spend. High volume and narrow tasks favour open weights. Low volume and broad tasks favour a hosted endpoint. Most organisations have both, which is why most end up running both.

The framing that will age worst is the one that treats this as a single decision with a single answer. Hardware supply shapes it too: as we reported on the packaging bottleneck, the cost of self-hosting depends on an accelerator market that is not loosening as quickly as 2025 plans assumed.

Batteries Are Cheaper Than Peaker Plants. The Grid Queue Is the Holdup.

Arjun Mehta — Sat, 08 Aug 2026 00:00:00 GMT

The cost argument for grid-scale storage is settled. In most markets a four-hour battery system now undercuts a new gas peaking plant on levelised cost for the same service, and the gap has widened for four consecutive years. What has not changed is how long it takes to connect one, and that number is now the thing determining how much storage gets built.

The bottleneck moved from a spreadsheet to a filing cabinet, which is a harder problem to solve because nobody’s cost curve fixes it.

The queue, in the only terms that matter

Interconnection queues in most large markets contain multiples of the capacity that will ever be built. That is not inherently a scandal; speculative applications are cheap to file and many were never serious. The problem is what the volume does to timelines for the projects that are serious.

Three structural features do most of the damage.

Serial study processes. Each project is studied against the assumed presence of everything ahead of it. When a project ahead withdraws, downstream studies can require redoing, which cascades.
Cost allocation uncertainty. A project’s network upgrade cost is not known until late in the process and can change dramatically based on the behaviour of unrelated applicants. Financing around that uncertainty is expensive.
No meaningful entry cost. Where deposits are small relative to project value, filing broadly is rational, which inflates the queue that everyone else must be studied against.

We can procure a battery system in nine months and energise it in four years. Nobody in this business is waiting on cells.

Why storage suffers more than it should

Storage is disadvantaged by queue processes that were designed for generators. A battery is a flexible asset whose grid impact depends on how it is dispatched, but most study processes evaluate it at maximum output as though it were a thermal plant running at full load.

That assumption inflates the network upgrades attributed to it. Several markets have introduced flexible or limited interconnection arrangements, where an asset accepts curtailment during constrained periods in exchange for a faster and cheaper connection. Where those arrangements exist and are well designed, they work: connection timelines fall substantially and the operational cost of accepted curtailment is usually far lower than developers feared.

Adoption is uneven, and the details matter. A flexible connection with unbounded curtailment risk is not financeable, so the arrangements that succeed cap the exposure.

Queue reform, and what actually moved the needle

The reforms with the best measured record share a common shape: they impose a cost on occupying a queue position and study projects in groups rather than one at a time.

Cluster studies replace serial evaluation with batched analysis, which prevents the cascade problem and produces cost allocations that are stable enough to finance against. Higher, escalating deposits push out applications that were never going to be built. Readiness requirements, such as demonstrated site control, do similar work.

Markets that implemented both together have seen queue volumes fall and completion timelines improve. Markets that raised deposits without reforming study processes mostly just made speculation more expensive without accelerating anything.

The part that is genuinely hard

Some of the delay is not process at all. It is transmission. Where the network physically cannot carry additional output from a location, no reform to study methodology changes the answer, and building new lines takes the better part of a decade largely for permitting and land reasons.

That is the residual problem, and it is the one that will still be here after queue reform is done. Storage helps at the margin because it can absorb generation that would otherwise be curtailed, which is one of the better arguments for co-locating it. But it does not substitute for wires.

For the adaptation side of the same infrastructure question, see our reporting on how cities are budgeting for heat.

Advanced Packaging, Not Lithography, Is the Real Chip Bottleneck

Maya Iyer — Thu, 06 Aug 2026 00:00:00 GMT

Ask a policy analyst where accelerator supply is constrained and you will usually hear about lithography. Ask someone who schedules a fab and you will hear about packaging. The second answer has been the correct one for roughly two years, and the distinction matters because the two constraints relax on very different timelines.

Front-end wafer capacity for leading-edge logic has expanded substantially. The step that has not kept pace is the back end: the process of stacking high-bandwidth memory beside a logic die on an interposer and getting acceptable yield out of the result.

Why the back end became the constraint

A modern accelerator is not one chip. It is a logic die, several stacks of high-bandwidth memory, and a substrate that connects them at a pitch fine enough that the memory bandwidth is usable. That assembly step involves thermal cycling, precise alignment, and a yield penalty applied to components that are already expensive.

Three properties make it hard to scale quickly.

The equipment is specialised and the vendor list is short. Bonders and inspection tools for fine-pitch interposers come from a handful of suppliers with their own multi-quarter lead times.
Yield loss compounds. Scrapping an assembly late in the process discards good logic and good memory together, so effective capacity is lower than nameplate capacity by a margin that varies with process maturity.
Memory supply is coupled. High-bandwidth memory qualification is slow, and the number of stacks per accelerator has been rising, so each unit consumes more of a separately constrained input.

Adding a lithography tool is a capital decision. Adding qualified packaging capacity is a capital decision plus eighteen months of process learning that you cannot buy.

What the numbers look like from the outside

Public disclosure here is thin, which is part of the problem. Foundries report capacity in wafer-equivalent terms that obscure the back end, and packaging subcontractors report revenue rather than qualified units.

The signals that are visible point in one direction. Lead times quoted to large buyers have stayed extended even through periods when wafer starts increased. Substrate suppliers have been running expansion programmes continuously since 2024 without lead times normalising. And the pricing structure has shifted: buyers now commonly contract for packaged units rather than wafers, which is what happens when the scarce step moves downstream.

The forecasting error this creates

Capacity plans written in 2025 generally assumed that accelerator availability would improve as wafer capacity came online, and that per-unit costs would fall accordingly. Teams built serving-cost forecasts on that assumption. As we found reporting on the serving-cost squeeze, several of those forecasts are now being revised upward, because the hardware abundance they priced in has not arrived.

The correction is not dramatic. It is a matter of quarters, not years. But it lands on organisations that committed to product margins based on the earlier curve, and the ones with the least slack are the ones who reserved capacity at fixed prices without a corresponding pricing mechanism on the revenue side.

What would actually change the picture

Three developments would loosen the constraint, in descending order of near-term plausibility.

The first is process maturity at existing packaging lines, which raises effective capacity without new equipment. This is happening steadily and is the main reason supply has improved at all.

The second is architectural: designs that need fewer memory stacks per unit of useful throughput. There is real work here, and it does more for effective supply than a new facility would, because it reduces demand on the coupled input.

The third is new qualified capacity at scale, which is under construction and will matter in 2027 and beyond. Anyone promising relief sooner than that from new facilities is describing an announcement, not a shipment.

For the demand side of this equation, see our reporting on where open-weight models still lose, which shapes how much of the market needs frontier-class hardware at all.

Designed Proteins Are Leaving the Lab. The Validation Gap Is Widening.

Sofia Reyes — Tue, 04 Aug 2026 00:00:00 GMT

The design half of protein engineering has become extraordinarily fast. A group with a target structure and reasonable compute can generate thousands of candidate binders in a week, and a meaningful fraction of them will fold as predicted. The characterisation half has not sped up at all. Expressing, purifying, and measuring a candidate still takes weeks per construct, and the good measurements take longer than that.

That asymmetry is not a temporary inconvenience. It is starting to shape what the field publishes and what it believes.

Where the ratio actually sits

Groups we spoke to described design-to-validation ratios that would have been unthinkable five years ago. One lab generated roughly 4,000 candidate designs against a single target over two months and experimentally characterised 47 of them.

The 47 were not a random sample. They were selected by computational filters, which is reasonable practice and also the source of the problem. The published result describes the performance of the selected subset. It does not describe the performance of the design method, because the selection step is doing work that is not being measured.

Filtering is a hidden model. Predicted binding affinity, predicted solubility, and structural plausibility scores all encode assumptions. Reporting success only among survivors measures the filters as much as the generator.
Negative results stay unpublished. A design campaign that yields nothing is rarely written up, so the field’s estimate of base rates is drawn from campaigns that worked.
Characterisation depth varies enormously. A binding assay is not a functional assay, and a functional assay in vitro is not activity in a cell.

We can now design faster than we can be wrong at a measurable rate. That should worry people more than it does.

The measurements that are being skipped

The specific gap most often cited by structural biologists is not affinity. It is specificity and stability under realistic conditions.

Affinity for the intended target is comparatively easy to measure and is almost always reported. Off-target binding across a realistic proteome is expensive and is usually not. Thermal and proteolytic stability get reported inconsistently. Aggregation behaviour at concentration, which determines whether a molecule is developable at all, appears in a minority of papers.

This produces a literature in which designs look excellent on the axis that is cheap to measure. Groups working on therapeutic applications are blunt about the consequence: a substantial share of published designed binders fail on properties that were never characterised in the original report.

What would close the gap

Three interventions came up repeatedly, and none of them requires a methodological breakthrough.

The first is reporting the denominator. State how many designs were generated, what filters were applied, and how many survived each stage. This is a change in convention rather than in capability, and it would immediately make published success rates interpretable.

The second is standardised minimum characterisation. A short, agreed panel covering specificity, stability, and aggregation, reported for every candidate that gets published, would eliminate most of the current inconsistency. Several groups are pushing for this through journal policy rather than waiting for consensus.

The third is investment in throughput on the wet side. Automated expression and purification exists and works. It is unglamorous, it does not produce papers on its own, and it is chronically underfunded relative to the compute budgets on the design side.

Why this is a familiar failure

The pattern here is not specific to protein design. It is the standard signature of a field where one half of the loop got cheap and the other did not: apparent progress accelerates, published success rates rise, and the base rate quietly becomes unknowable.

Medicine has run this experiment already, which is why trial registration exists at all. As we reported on the trial reporting gap, even mandatory registration only partly solved it. The design field has the advantage of being able to adopt the convention before the credibility problem becomes acute rather than after.

Half of Registered Clinical Trials Still Never Report Their Results

Priya Nair — Sat, 01 Aug 2026 00:00:00 GMT

Trial registration was designed to solve a specific and well-documented fraud: running a study, disliking the result, and never publishing it. Registering a trial before it starts creates a public record that it existed, so an absent result becomes visible as an absence. The mechanism is sound and it has been mandatory in most major jurisdictions for years. Compliance with the reporting half of it remains roughly a coin flip.

The registry now documents the problem it was built to prevent, in considerable detail, and very little happens as a result.

What the requirement says and what happens

The obligation is not complicated. Register the trial with its primary outcome before enrolment, and post summary results to the registry within twelve months of completion. Publication in a journal is separate and additional; the registry posting is the floor.

Audits across registries and jurisdictions keep landing in the same range. Roughly half of completed trials have results posted within the required window. A meaningful share never post at all. And a further category posts results that do not match the registered primary outcome, which is a different failure with the same effect.

Non-reporting is not random. Trials with unfavourable or null results are less likely to be reported, which is precisely the bias registration was meant to remove.
Academic sponsors perform worse than industry in most audits, which surprises people who assume the incentive problem is purely commercial.
Outcome switching is under-measured because detecting it requires comparing the registered protocol against the published paper, which almost no one does routinely.

The registry entry said the primary outcome was mortality at ninety days. The paper reported a composite endpoint that included hospital readmission. Both documents are public. Nobody had compared them in four years.

Why the evidence base is distorted, quantitatively

The consequence is not abstract. Meta-analyses and clinical guidelines are built from published trials. If unfavourable results are systematically missing, pooled estimates of treatment effect are biased upward, and the direction of the bias is known even when its magnitude is not.

Analyses that have recovered unpublished data through regulatory submissions or litigation have repeatedly found smaller effects than the published literature indicated, and in several well-known cases harms that the published record did not reflect. Each of those recoveries required a specific investigation. There is no routine mechanism for it.

Enforcement exists on paper

The relevant statutes and regulations do provide for penalties, including substantial daily fines in some jurisdictions. Enforcement actions are exceptionally rare relative to documented non-compliance.

The reasons regulators give are consistent: limited staffing, ambiguity about which entity is the responsible party for multi-site academic trials, and definitional questions about completion dates that make the twelve-month clock arguable. Those are real administrative obstacles. They are also the kind of obstacles that get resolved when there is institutional will, and the pattern suggests there is not much.

What has demonstrably worked is public tracking. Independent audits that publish sponsor-level compliance rates by name have produced measurable improvements at named institutions, on a timescale of months. Reputational pressure has outperformed statutory penalty in this domain by a wide margin, largely because it is actually applied.

What would fix it

Three changes would address most of the gap, and none require new legislation.

Tie funding and ethics approval to prior reporting compliance. An institution seeking approval for a new trial should demonstrate that its completed trials have posted results. This makes the sanction automatic and administrative rather than dependent on enforcement discretion.

Require registries to flag outcome discrepancies mechanically. Comparing a registered primary outcome against a reported one is a structured data problem, and doing it automatically would make outcome switching visible without anyone having to investigate.

Publish sponsor-level compliance as a standing metric rather than as periodic studies. The evidence that this works already exists.

The underlying pattern generalises beyond medicine. As we reported on the validation gap in protein design, any field where negative results go unrecorded will systematically overestimate how well its methods work.

Venture Funding Recovered on Paper. The Median Founder Did Not.

Daniel Okonkwo — Wed, 29 Jul 2026 00:00:00 GMT

Every quarterly funding report this year has led with the same figure: total dollars deployed, up substantially year over year. It is an accurate number and a misleading headline. Strip out the largest deals and the picture reverses. Deal count is roughly flat, the median round has shrunk, and the time between rounds has stretched by about five months.

What happened is not a recovery. It is a concentration.

The arithmetic of a mean that moves without the median

When a small number of very large financings enter the denominator, aggregate dollars rise sharply while the typical experience of raising money gets harder. Both things are true at once, and only one of them makes the summary slide.

The pattern is visible in three places.

Round size distribution. The top decile of deals accounts for a materially larger share of total dollars than it did three years ago. The bottom half accounts for less.
Deal count. Flat to slightly down, depending on whose dataset you use and how they treat extensions and bridges. Nobody credible has it up meaningfully.
Graduation rates. The share of seed-funded companies raising a Series A within 24 months has fallen. This is the number that actually describes founder experience, and it is the one least often reported.

The aggregate number is a fundraising document for the asset class. It is not a description of the market that founders are operating in.

Why bridges are doing the work rounds used to do

The clearest structural change is the normalisation of the inside round. Bridges, extensions, and structured follow-ons from existing investors now make up a much larger share of financings than they did in the last cycle.

This is rational behaviour from funds holding marks they do not want to reset, and it produces two effects worth naming. It keeps companies alive that would otherwise have failed, which delays the mark-down rather than avoiding it. And it shifts negotiating leverage decisively toward existing investors, because the alternative to their terms is often no term sheet at all.

Founders describe the consequence in consistent language: the round happens, the valuation is flat, the preference stack gets heavier, and the option pool gets refreshed out of common. The company survives. The founder’s economics do not.

What the concentration is buying

The large deals are not random. They cluster in capital-intensive infrastructure, where the cheque size is a function of what the business physically requires rather than investor enthusiasm.

That is a meaningful distinction. A very large financing for compute capacity or manufacturing is buying a fixed asset with a depreciation schedule. It looks like exuberance in an aggregate chart and like capital expenditure on a balance sheet. Our reporting on the packaging bottleneck in accelerator supply explains part of why those cheques are as large as they are.

The corollary is that the concentration is not evenly distributed across sectors either. Software businesses with ordinary capital needs are raising smaller rounds against tougher metrics, and the bar has moved from growth rate to demonstrated gross margin. As we found reporting on inference costs, that second requirement is genuinely difficult for a category of products that looked healthy under the old bar.

The number to watch instead

If you want one series that describes the market founders actually face, use the seed-to-Series-A graduation rate on a trailing 24-month basis, split by cohort year. It captures survival, it is not distorted by outliers, and it leads the aggregate figures by about a year.

It is currently below its 2021 peak by a wide margin and has been flat for three quarters. That is the recovery.

Recommendation Feeds Flattened Taste. Editors Are Quietly Coming Back.

Lena Fischer — Sat, 25 Jul 2026 00:00:00 GMT

The argument for replacing human curators with ranking models was never really about quality. It was about scale and cost, and on those terms it worked completely. One ranking system can personalise a catalogue for a hundred million people, and no editorial department can. The interesting development is that several platforms are now hiring editors again, and they are not doing it out of a change of heart.

They are doing it because optimising for engagement produces a catalogue that people get bored of, and boredom shows up in the retention numbers about two quarters later.

What the ranking function actually optimises

A recommender trained on completion, watch time, or click-through learns to predict what you will consume next. It does not learn what will make you glad you subscribed, because that signal is delayed, sparse, and hard to attribute.

The result is a well-documented set of behaviours.

Convergence. The system finds a reliable local optimum for your profile and stays there, because exploration costs measurable engagement now for uncertain benefit later.
Homogenisation of supply. Creators optimise for the ranking signal, so the catalogue itself narrows. The feed is not just showing you less variety, there is less variety being made.
Popularity feedback. Items with early engagement receive distribution that generates more engagement, which the model reads as quality.

We could raise session length by four percent any quarter we wanted. We could not raise the number of people who said the service was worth paying for.

Why editors solve a specific technical problem

The case for human curation that survives scrutiny is narrow but real: editors are good at exactly the thing recommenders are structurally bad at, which is deciding what deserves attention before there is any engagement data about it.

A new release, an unfamiliar genre, an artist with no audience yet, a back catalogue item that has never been surfaced. To a ranking model these are all high-variance bets with no evidence. To an editor with domain knowledge they are a judgement call, and a competent editor’s hit rate on cold-start material is considerably better than random.

That is why the roles being rebuilt are not general-purpose taste-making positions. They are cold-start and long-tail curation jobs, often reporting into growth rather than content.

The hybrid that seems to work

The pattern emerging across several platforms is not editors replacing the model. It is editors supplying a candidate set that the model is required to distribute.

In practice this means a reserved share of impressions allocated to editorially selected material, with the recommender deciding which users see which items but not whether the items get shown at all. The reservation is the crucial part. Without it, the model reallocates the inventory to safer bets within days.

Platforms that have measured this report a small, consistent engagement cost in the short term and improved retention and catalogue breadth over longer windows. Whether that trade is accepted depends entirely on which team owns the metric, which is an organisational question rather than a technical one.

The part that has not been fixed

None of this addresses distribution economics. A reserved impression share changes what audiences encounter; it does not change what creators are paid when they are encountered. Those are separate systems, and the second one has moved much less than the first, as our reporting on streaming royalty structures describes in detail.

It is also worth being precise about what the return of editors is not. It is not a restoration of a golden age of taste-making that people remember more fondly than it deserved. Editorial gatekeeping had its own well-catalogued biases, and there is no reason to assume a curation team assembled to fix a retention metric will be more representative than the ranking function it supplements.

What it is, more modestly, is an admission that a system optimising a proxy will eventually degrade the thing the proxy was standing in for.

Cities Are Budgeting for Heat. Most of Them Still Measure It Wrong.

Arjun Mehta — Tue, 21 Jul 2026 00:00:00 GMT

Heat has finally become a budget line. Cities that had no adaptation plan five years ago now have heat officers, cooling centre networks, and capital programmes for shade and reflective surfaces. The planning is real and mostly sensible. The measurement underneath it is not.

Most municipal heat policy is triggered by readings from a small number of official weather stations, and those stations are frequently sited in exactly the conditions least representative of where people live.

The siting problem

Official meteorological stations are placed for consistency and aviation, which means open ground, good airflow, grass surfaces, and often an airport. Those are reasonable choices for a climate record. They are poor proxies for a dense neighbourhood of asphalt, low albedo roofing, and no canopy.

Mobile and distributed sensing campaigns consistently find large intra-city differences that a single station cannot see.

Within-city spreads of several degrees between the hottest and coolest neighbourhoods at the same hour are routine, and the spread is larger at night than at midday.
Night-time minima are where the divergence matters most physiologically, because inability to cool overnight drives mortality more than peak afternoon temperature.
The gradient correlates with income and canopy cover, which means a city-wide average conceals a distribution with a predictable social pattern.

Our threshold was met four times last summer according to the airport. Our own sensors in three neighbourhoods crossed the same threshold nineteen times.

Why the wrong metric produces the wrong programme

If a heat emergency is declared based on a reading that understates conditions in the worst-affected areas, three things follow mechanically.

Cooling centres open late and close early, because activation is tied to the threshold. Outreach to vulnerable residents is triggered on the same signal, so it also arrives late. And retrospective evaluation of the programme uses the same understated series, which makes interventions look more adequate than they were.

There is a fourth, subtler effect. Capital allocation for shade, canopy, and roof programmes is often prioritised using modelled heat exposure calibrated against the official station. Neighbourhoods whose actual exposure is highest can therefore be systematically underweighted in the investment that would help most.

What better measurement looks like

The technically correct answer is to measure what the body experiences rather than air temperature alone. Heat stress depends on humidity, radiant load, and air movement, which is why wet-bulb globe temperature and similar composite indices exist.

The practical answer is more modest and more achievable. Cities that have improved their measurement did three things: deployed a distributed low-cost sensor network with enough density to resolve neighbourhood differences, calibrated it against the official station rather than replacing that station, and rewrote activation thresholds to trigger on the hottest monitored areas rather than the city average.

The cost is small relative to any capital programme it informs. The institutional obstacle is larger, because it requires accepting that the number the city has been reporting for decades was not describing what residents experienced.

The accountability angle

There is a reason some administrations are slow to adopt better measurement, and it is worth stating directly. Denser monitoring produces a higher count of threshold exceedances, more declared emergency days, and a documented record of which neighbourhoods bear the load.

That record creates obligations. It also creates the evidence base needed to justify the spending, which is why the cities furthest along on measurement tend to be the ones that have already committed capital and need to defend it.

For the generation-side infrastructure that heat demand strains, see our reporting on why storage is stuck in the interconnection queue.

A Decade of Sky Surveys Just Rewrote the Supernova Rate

Sofia Reyes — Thu, 16 Jul 2026 00:00:00 GMT

Ten years of automated transient surveys have produced a catalogue large enough to do something the field could not previously do well: measure how often stars actually explode, rather than inferring it from a handful of nearby events and a lot of modelling. The measured rate of core-collapse supernovae has come out higher than the standard prediction, and the excess is concentrated in dusty, star-forming galaxies.

That last detail is what makes the result interesting rather than merely surprising. The discrepancy appears where the observational bias was always expected to be worst.

What changed methodologically

Earlier rate estimates were built from small samples with heterogeneous selection. An event was found because someone was looking at that galaxy, which makes the sample a description of observing habits as much as of the universe.

Automated wide-field surveys with consistent cadence removed most of that problem. The survey observes the same footprint on a fixed schedule regardless of what is interesting, so the selection function is computable rather than anecdotal.

Uniform cadence means the probability of catching an event of known duration can be calculated, not estimated.
Consistent depth allows a completeness correction that is a function of distance and brightness rather than of who was on shift.
Sample size finally permits splitting by host galaxy type, which is where the signal turned out to live.

The dust problem, stated plainly

Core-collapse supernovae come from short-lived massive stars, so they occur in regions that are actively forming stars. Those regions are dusty. Dust absorbs and reddens optical light, so a fraction of these events have always been expected to be missed or misclassified in optical surveys.

The size of that fraction was the open question. The new catalogues, combined with infrared follow-up on a subsample, put it substantially higher than most models assumed. In the dustiest host galaxies the inferred correction is large enough to account for most of the gap between predicted and observed rates.

The universe was not producing more supernovae than we thought. We were failing to see them in exactly the places our own models told us we would.

Why the star-formation rate is implicated

The supernova rate and the cosmic star-formation history are tied together. Massive stars form, live briefly, and explode, so a measured explosion rate constrains how many massive stars formed a few million years earlier.

If the supernova rate is higher than predicted, either more massive stars are forming than the star-formation rate implies, or the relationship between the two is not what the standard assumptions encode. Both possibilities have consequences well beyond supernova statistics, because star-formation history feeds into estimates of metal enrichment, dust production, and the ultraviolet background.

The dust-obscuration explanation is the least disruptive of the available options, and it currently has the best support. It also has a testable prediction: infrared and radio surveys, which are far less affected by dust, should recover the missing events. Early results from radio follow-up are consistent with that, though the samples remain small.

What the result does not say

It does not indicate a problem with stellar evolution theory. The models of how massive stars end are not in question here; what is in question is the completeness of optical censuses of them.

It also does not resolve the related question of how many core-collapse events fail to produce a bright explosion at all. Failed supernovae, where a massive star collapses without a luminous transient, remain difficult to constrain and would push in the opposite direction. The current work brackets the problem better than before without closing it.

The broader methodological lesson is the one worth carrying forward, and it echoes what we found reporting on the validation gap in protein design: when a field’s measurements are drawn from a selected sample, the first thing to characterise is the selection.

The Streaming Royalty Math That Keeps Mid-Tier Artists Broke

Lena Fischer — Thu, 09 Jul 2026 00:00:00 GMT

Every discussion of streaming payouts gets stuck on the per-stream rate, and it is the least useful number in the system. There is no fixed per-stream rate. What exists is a pool of money divided according to a rule, and the rule is where the money actually moves. Under the standard model, an artist’s payout depends not on how much their own listeners paid but on what fraction of all listening on the platform they captured.

That single design choice explains most of what artists find inexplicable about their statements.

Pro-rata, in plain terms

Under pro-rata distribution, the platform pools subscription and advertising revenue, takes its share, and divides the remainder in proportion to total stream counts across the entire service.

Follow the consequence. If you subscribe and listen exclusively to one independent artist all month, your money does not go to that artist. It goes into the pool and is distributed according to global listening share, most of which is captured by the largest catalogues. Your subscription funds the artists you did not listen to.

Heavy listeners dilute everyone. A user streaming 8,000 tracks a month contributes the same revenue as one streaming 200, but claims forty times the share of the pool.
Skew rewards concentration. Listening follows a steep power law, so proportional division concentrates payouts far more than revenue is concentrated.
Fraud is a transfer, not a leak. Artificial streams do not create money. They take it from legitimate artists in the same pool.

My statement showed 340,000 streams and a payment that would not cover the mastering. The label explained the pool. It was the first time the numbers made sense and the first time they seemed indefensible.

What user-centric distribution changes, and what it does not

The obvious alternative divides each subscriber’s payment among the artists that subscriber actually listened to. It fixes the dilution problem directly and makes statements comprehensible.

The measured effects, from platforms and studies that have modelled it, are real but smaller than advocates suggest. Mid-catalogue and niche-genre artists with devoted, moderate-volume audiences gain. Artists whose plays come from passive playlist and background listening lose. The very largest artists lose slightly. Total money paid to artists does not change, because the rule only governs division.

That last point deserves emphasis. Switching distribution models redistributes a fixed pool. It does not address the size of the pool, which is set by subscription pricing and the platform’s revenue share. An artist whose income tripled under user-centric distribution went from very little to slightly less little.

The changes that alter the pool rather than the split

Three levers actually change how much money reaches artists, and they are all harder than reforming the split.

Subscription price is the most direct. Real prices were flat for over a decade while catalogue size and listening hours grew enormously. Recent increases have moved the pool more than any distribution reform would have.

Minimum thresholds are the most contested. Several platforms now withhold payment from tracks below an annual stream floor and redistribute it. This demonstrably reduces fraud and administrative cost, and it also removes small payments from artists at the bottom of the distribution, which is precisely the group the reform debate claims to be about.

The third is the share retained before the pool is formed, split between the platform and rights holders. This is the largest number in the system and the least discussed publicly, because the parties negotiating it have a shared interest in the argument staying focused on per-stream rates.

Why the framing persists

Per-stream rates are easy to compare and easy to be outraged about, which makes them useful to everyone who does not want the structure examined. A platform can point out truthfully that it does not set a per-stream rate. A label can point to the platform. An artist is left with a statement they cannot reconcile.

The structural question is simpler to state and harder to deflect: what fraction of the revenue generated by a listener reaches the artists that listener chose? For the curation systems that determine what those listeners encounter in the first place, see our reporting on the quiet return of human editors.

The Antibiotic Pipeline Is Failing for Economic Reasons, Not Scientific Ones

Priya Nair — Tue, 30 Jun 2026 00:00:00 GMT

The standard account of antimicrobial resistance describes a scientific failure: bacteria evolve, discovery is hard, the pipeline is empty. The first two claims are true and the third is misleading. Novel antibiotics have been discovered, developed, and approved over the past decade. Several of the companies that brought them to market then went bankrupt, sometimes within two years of approval.

That is not a discovery problem. It is a revenue model that is structurally incompatible with responsible use of the product.

The conservation paradox

For most pharmaceuticals, a superior product should be prescribed widely. For a novel antibiotic effective against resistant organisms, the correct clinical and public health behaviour is the opposite: hold it in reserve, use it only when older agents fail, and thereby preserve its effectiveness for as long as possible.

Stewardship programmes exist specifically to enforce that restraint, and they work. They also mean that a successful new antibiotic, judged by public health value, generates minimal sales volume.

Revenue scales with use, and appropriate use is deliberately minimised.
Treatment courses are short. A week or two of therapy, compared with years of a chronic medication.
Price resistance is severe. Antibiotics are benchmarked against generics costing very little, so premium pricing meets institutional resistance regardless of novelty.
Patent life burns during reserve. Exclusivity runs while the drug is deliberately unused, so peak sales arrive after generic entry becomes possible.

We built the thing the guidelines asked for, and the guidelines then correctly instructed hospitals not to use it. There was no version of this where we made money.

What happened to the companies

The pattern has repeated enough to be predictive. A small company takes a novel agent through clinical development, secures approval on the strength of activity against resistant organisms, and then discovers that hospital formularies are slow to add it, stewardship committees restrict it appropriately, and annual revenue lands one or two orders of magnitude below what was needed to service development debt.

The company is sold at a loss or files for bankruptcy. The asset transfers to a larger firm that keeps it on the shelf, or in some cases withdraws it. The clinical need it addressed remains unmet in practice even though a licensed product exists.

Large pharmaceutical companies drew the obvious conclusion years ago and mostly exited antibacterial discovery. The remaining pipeline sits with small firms and academic groups that will face the same economics on approval.

The proposed fixes, and their track records

Two categories of intervention have been tried. Push incentives subsidise development: grants, non-dilutive funding, and public-private partnerships for early research. These have worked reasonably well at their stated purpose, which is why there are candidates at all.

Pull incentives are supposed to reward approval, and they are where the model still fails. Extended exclusivity does little when the constraint is volume rather than competition. Priority review vouchers are tradeable and have produced real value, but they are a one-off payment poorly matched to the size of the gap.

The mechanism with the strongest logic decouples payment from volume entirely. Under a subscription or availability model, a health system pays a fixed annual sum for guaranteed access to an antibiotic regardless of how much is dispensed. The developer receives predictable revenue; the health system retains every incentive to restrict use.

Pilot programmes in a small number of national health systems have demonstrated that this is administratively workable. The limitation is scale. A subscription from one or two countries does not underwrite global development costs, and the coordination problem across payers has not been solved.

What to watch

Two indicators matter more than pipeline counts. The first is whether subscription-style procurement expands beyond pilots to a group of payers large enough to constitute a viable market. The second is whether any newly approved agent reaches profitability under current arrangements, which so far none has.

Until one of those changes, additional research funding will continue producing approved drugs that bankrupt their developers. For the parallel problem of evidence that never reaches the public record, see our reporting on unreported clinical trials.