Today's 12 Stories — Monday, August 10, 2026

Industry Dynamics Story 1 of 12

OpenAI Buys Its Way Into the Office Suite With NextSlide

OpenAI has acquired NextSlide, a startup whose product turned prompts, notes, documents and research into finished, editable presentations. The company confirmed the deal publicly on 8 August, though founder Ahmed Beshry acknowledged the transaction had actually closed earlier in the year. The entire NextSlide team is moving to OpenAI to work on ChatGPT. Financial terms were not disclosed.

The delay between closing and disclosure is itself informative. Acquisitions get announced when the acquirer wants the market to draw a conclusion, and the conclusion available here is that ChatGPT is no longer positioned as a chat product with features bolted on. It is being assembled into a productivity surface.

Beshry is not a first time founder. He was a cofounder of Caper AI, the smart shopping cart company that Instacart bought in 2021. That pattern matters for how executives should read this deal. OpenAI did not buy a slide generator. It bought a team that has already taken an AI interface from research demo to deployed retail hardware once before, and it placed that team directly on its consumer product.

For enterprise buyers the strategic question is uncomfortable and immediate. Presentation software is one of the most defensible categories in enterprise software, protected less by features than by file formats, template libraries, brand compliance and two decades of institutional muscle memory. None of those moats were built to withstand a system that produces the finished artifact from the source material directly, skipping the authoring step that the incumbent tools exist to support.

The near term risk is not that ChatGPT replaces the presentation tools already deployed across an organization. It is that the work moves upstream without anyone deciding it should. When the analyst who was going to build the deck asks a model to build it instead, the governance apparatus attached to the sanctioned tool, meaning brand review, version control, data classification and retention, does not travel with the output. The deck still ships. The controls do not.

There is also a procurement dimension worth naming early. Organizations that negotiated enterprise agreements for productivity suites priced on seat counts assumed those seats would keep doing the same work. If a meaningful share of document and presentation production migrates to a general assistant, the value delivered per seat changes on both sides of the contract, and the renewal conversation changes with it.

The immediate action for technology leaders is not a policy on presentations. It is an inventory question. Find out where finished business artifacts are already being produced outside sanctioned tooling, because the honest answer in most organizations is that this migration started well before any vendor announced it, and the acquisition simply confirms which vendor intends to serve it.

Generative AI Story 2 of 12

Meta Ships Muse Code and Turns the Terminal Into Contested Ground

Meta has launched Muse Code, an AI coding agent built on its Muse Spark 1.2 model and released in beta in early August. The tool is designed to write code, fix bugs, verify its own results automatically and manage complex software projects rather than answer isolated questions about them.

The significant detail is where it runs. Muse Code is a terminal agent, which places Meta in direct competition with the command line coding tools that Anthropic and OpenAI have been building out, and away from the editor plugin model that defined the first generation of AI assisted development. That choice is strategic rather than cosmetic. An agent living in the terminal sits next to the build system, the test runner, the version control client and the deployment tooling. An agent living in the editor sits next to the text.

Automatic verification is the capability executives should interrogate most closely. The bottleneck in AI assisted engineering has not been generating plausible code for some time. It has been establishing whether the generated code is correct, and the review burden that question creates has quietly absorbed much of the productivity that these tools were purchased to deliver. An agent that runs its own tests and iterates against failures is attacking the actual constraint. Whether it attacks it successfully is an empirical question that no vendor benchmark will settle for any particular codebase.

Meta arriving here also changes the competitive shape of the category. Until now the serious terminal agents came from labs selling model access as their primary business, which meant their commercial incentive pointed toward token consumption. Meta does not monetize developers the way those companies do, and a competitor with a different revenue model can price a category in ways its incumbents cannot comfortably match.

For engineering leadership the governance questions arrive before the productivity ones. An agent with permission to modify files, execute commands and manage projects has a materially larger blast radius than an autocomplete suggestion, and the controls most organizations wrote for AI coding assistance were written for autocomplete. Credential scope, repository permissions, branch protection and audit logging all need to be revisited on the assumption that the agent acts rather than suggests.

The practical recommendation is to treat any terminal agent as a new class of privileged automation rather than as a new developer tool. That means running it first against low sensitivity repositories, requiring human approval at the merge boundary regardless of how confident the automatic verification appears, and instrumenting what the agent actually did rather than what it reported doing. The distinction between those two records is where the real operational risk lives.

Policy & Regulation Story 3 of 12

The EU AI Act Stopped Being Theoretical on 2 August

From 2 August 2026 the European Commission's AI Office, working with national authorities, began enforcing the Artificial Intelligence Act. New transparency rules started to apply on the same date. The regime that companies have been preparing for since the Act was adopted now has an enforcement apparatus behind it.

The obligations that bite hardest fall on providers of general purpose AI models designated as carrying systemic risk. Those providers must perform model evaluations including documented adversarial testing, assess and mitigate systemic risks, report serious incidents to the AI Office without undue delay, and maintain an adequate level of cybersecurity protection. Each of those is a continuing operational duty rather than a one time filing, which is the detail most compliance programmes have underweighted.

The designation threshold is compute based. The Act presumes systemic risk for general purpose models trained using more than ten to the twenty fifth floating point operations. A numerical trigger of this kind has a property that legal thresholds usually lack: it is knowable in advance and it is not negotiable after the fact. A company knows before it starts a training run which side of the line the resulting model will land on, and it cannot argue its way across afterward.

For executives outside the handful of labs that train frontier models, the temptation is to read this as somebody else's regulation. That reading is wrong in a specific way. The transparency obligations propagate downstream. Organizations that deploy general purpose models into products sold in Europe inherit disclosure duties, and the evidence needed to satisfy them has to come from the provider. Any contract with a model vendor that does not oblige the vendor to supply that evidence has left a compliance gap that the deploying company will be asked to close on its own.

The incident reporting duty deserves separate attention because it changes internal timelines. Reporting without undue delay is not compatible with an escalation path that routes model failures through a quarterly review. Organizations need a defined trigger for what counts as a serious incident, a named owner, and a path from the engineer who notices to the regulator that does not pass through six approval layers.

The realistic near term expectation is not a wave of headline fines. Regulators with new powers typically spend their first enforcement period establishing what compliance looks like in practice and testing their own procedures. That period is the window in which a company's documentation either exists or does not, and the organizations that treat this month as preparation rather than exposure are the ones that will be able to answer when asked.

AI Safety Story 4 of 12

Washington Chooses Voluntary, and Draws a Line Around Open Weight Models

The White House finalised a voluntary safety testing framework for frontier AI models in early August and convened the largest AI companies, including OpenAI, Anthropic, Google and Meta, to review it. The framework is administered by the Center for AI Standards and Innovation, known as CAISI, which sits inside NIST.

The mechanism at the centre of the framework is early access. It gives the United States government the ability to review frontier models for up to thirty days before public release. In practice that is a pre deployment evaluation window, and it represents a meaningful shift in the relationship between American AI labs and the federal government even though participation is not compulsory.

Two design decisions define what this framework is and is not. The first is that it is voluntary, which means its authority rests on the willingness of five or six companies to keep participating. Voluntary frameworks are durable when participants believe the alternative is worse and fragile when a participant decides the cost has exceeded the benefit. The second decision is that the framework applies to closed frontier models and excludes open and open weight models from federal security review.

That exclusion is the part that deserves executive attention, because it creates an asymmetry that is structural rather than incidental. The models subject to government pre release evaluation are precisely the ones distributed through controlled APIs where the provider retains the ability to monitor and revoke access. The models exempt from it are the ones whose weights can be downloaded, modified and run privately, where no provider retains any control at all. The review burden and the containment capability are pointed in the same direction, which is the opposite of how a risk based regime would allocate them.

Placing the function inside NIST rather than a new regulator is also a deliberate signal. NIST produces standards and measurement methodology, not enforcement actions. An organization reading this framework as the beginning of American AI licensing is reading it wrong. It is closer to the beginning of an American AI testing methodology, and methodology tends to outlive the administration that commissioned it in a way that enforcement postures do not.

For enterprises the practical consequence is a documentation asset they did not previously have. A model that has been through CAISI evaluation carries evidence that a model outside the process does not, and procurement functions should start asking whether a given model participated. The answer will not settle a purchasing decision on its own, but for regulated buyers it is quickly becoming the kind of question that is awkward to have never asked.

AI Models Story 5 of 12

Alibaba Puts a 2.4 Trillion Parameter Model in Play

Alibaba has released Qwen3.8-Max, a Mixture of Experts model with 2.4 trillion total parameters, and made it broadly available in August. The model supports a context window of up to 1 million tokens and activates roughly 95 billion of its parameters during inference.

That activation ratio is the number that matters commercially, and it is easy to skim past. A model can hold 2.4 trillion parameters worth of learned capability while paying inference costs closer to a 95 billion parameter model, because only the relevant experts are engaged on any given forward pass. The headline parameter count describes what the model knows. The activated count describes what each query costs to serve. Sparse architectures have decoupled those two figures, and any procurement conversation that still treats total parameters as a proxy for cost is working from a mental model that the architecture retired.

The million token context window changes a different calculation. At that scale the engineering pattern of chunking a corpus, embedding the chunks, retrieving the relevant ones and assembling a prompt becomes optional for a large class of problems rather than mandatory. Retrieval pipelines were built to work around a constraint. Where the constraint no longer binds, the pipeline is now a cost centre and a source of failure modes rather than an enabling technology. That does not mean retrieval disappears, because cost, latency and precision still favour it in many workloads, but the default is no longer obvious.

The strategic dimension is where this model sits in the competitive map. Alibaba is moving its largest model toward open availability at a moment when the most capable openly distributed systems increasingly come out of Chinese labs. For enterprises the practical effect is that the performance argument for paying frontier API prices is narrowing, and the remaining arguments are about support, indemnification, data governance and jurisdictional risk rather than raw capability.

Those remaining arguments are not trivial and executives should resist the framing that they are. Deploying a model whose weights you control means accepting responsibility for evaluation, safety tuning, serving infrastructure and security that a vendor otherwise absorbs. The organizations that have moved successfully to open weight models have generally done so because they had a specific workload with predictable shape and enough volume to justify the operational investment.

The recommended posture is to benchmark rather than to migrate. Take the three highest volume model workloads currently running on a frontier API, measure them against an openly available alternative on the actual task, and price the total cost including the engineering time the switch would consume. That number, not a benchmark leaderboard, is the one that should drive the decision.

AI Infrastructure Story 6 of 12

Hyperscaler Capex Keeps Climbing, and Microsoft Just Changed How It Is Counted

The 2026 infrastructure numbers from the largest cloud operators have now landed, and they describe a buildout still accelerating. Alphabet raised its capital expenditure guidance for the year to a range of 195 billion dollars to 205 billion dollars. Meta narrowed its own range to between 130 billion dollars and 145 billion dollars. Amazon told investors in its February earnings release that it expects to invest about 200 billion dollars across the company during 2026.

Microsoft supplies the most instructive figure, and not because of its size. In its fiscal 2026 third quarter materials the company said it expects to invest roughly 190 billion dollars in capital expenditures for calendar 2026, including approximately 25 billion dollars attributable to higher component pricing. Microsoft has separately extended the estimated useful life of its datacenters and office buildings to twenty five years from fifteen.

That accounting change deserves more executive attention than the headline totals. Extending useful life spreads the cost of an asset across a longer depreciation schedule, and it also shifts how future datacenter leases are classified. Finance leases count inside capital expenditure. Operating leases do not. A company can therefore report a lower capital expenditure figure while its underlying investment plans remain unchanged, and readers who track the reported number as a proxy for physical buildout will register a slowdown that did not occur.

There is a second implication that runs the other way. Depreciation schedules encode a judgement about how long an asset produces value. Extending the assumed life of AI datacenter equipment from fifteen years to twenty five is a statement that this infrastructure will remain productive far longer than the current hardware refresh cycle suggests. If that judgement proves optimistic, the correction arrives as an impairment rather than as a missed quarter, and impairments are lumpy and unpleasant in a way that ordinary depreciation is not.

The component pricing disclosure is the other detail worth isolating. When roughly 25 billion dollars of a single company's annual capital spending is attributed to price increases rather than additional capacity, that is a supply constraint expressing itself financially. Enterprises negotiating compute contracts should expect that pressure to arrive in their own pricing, because the cost is being absorbed upstream and it does not stay absorbed indefinitely.

For technology leaders the useful discipline is to stop reading capital expenditure headlines as capacity signals. Ask instead what physical capacity a given figure buys at current component prices, and what accounting treatment it reflects. The two questions have diverged this year, and the gap between them is now wide enough to mislead.

Enterprise AI Story 7 of 12

Gartner Expects More Than 40 Percent of Agentic AI Projects to Be Canceled by 2027

Gartner projects that more than 40 percent of agentic AI projects will be canceled by the end of 2027. For boards that approved agent programmes on the strength of pilot demonstrations, that forecast is the most useful planning input available this year, and it deserves to be read carefully rather than defensively.

A cancellation rate above 40 percent is not evidence that the technology does not work. Analyst forecasts of this shape describe organizational failure more often than technical failure, and the pattern is familiar from every prior enterprise automation wave. Projects die because the business value was never specified in measurable terms, because the cost of running the system exceeded the cost of the work it replaced, or because nobody owned the governance question until an incident forced someone to.

The economics deserve the most scrutiny. An agent that completes a task through many model calls, tool invocations and retries consumes vastly more inference than a single prompt, and the unit cost of an agentic workflow is frequently discovered rather than designed. A pilot running a few hundred times a month reveals nothing about what the same workflow costs at production volume, and the shape of the curve between those two points is not linear when retries and failure handling are included.

Governance is the second common cause of death. An agent that reads systems and drafts output is an assistant. An agent that writes to systems and takes actions is an actor, and it needs the accountability structure an actor requires: a named owner, a defined scope of authority, an audit trail of what it actually did, and a documented path for reversing an action it should not have taken. Organizations that skipped those questions during the pilot find them raised for the first time by an auditor, and the answer arrives too late to save the project.

The forecast also carries a quieter implication that is easy to miss. If more than 40 percent are canceled, the remainder are not. Some meaningful share of these programmes will reach production and will produce durable advantage for the organizations that built them. The distinguishing factor in prior automation waves was rarely technical sophistication. It was whether the sponsoring organization had defined what success meant precisely enough to recognise it.

The practical recommendation is to run a portfolio review now rather than at the next budget cycle. For every agent project underway, require three artifacts: the measurable outcome it is meant to move, the fully loaded cost per completed task at projected production volume, and the named individual accountable for what the agent does. Projects that cannot produce all three are not necessarily bad ideas, but they are the ones this forecast is describing.

AI Models Story 8 of 12

Moonshot Ships a 2.8 Trillion Parameter Open Weight Model

Moonshot AI released Kimi K3 in late July, an open weight model carrying 2.8 trillion parameters. The scale is the headline, but the licensing is the story. A model of this size distributed with published weights is a different kind of object from a model of this size served through an API, and the difference is not primarily technical.

An open weight release transfers control. An organization that downloads these weights can run them inside its own perimeter, on its own hardware, in a jurisdiction of its choosing, with no vendor able to observe the traffic, change the model underneath it, deprecate it on a schedule set elsewhere or revoke access. For a regulated institution that has spent two years negotiating data residency and model change management clauses into AI vendor contracts, that combination resolves several problems simultaneously.

It also creates problems that the vendor was previously absorbing. Safety tuning, evaluation, red teaming, serving infrastructure, security patching and the ongoing question of whether the model is behaving correctly all become the deploying organization's responsibility. Enterprises that have moved workloads onto open weight models successfully have generally done so with a specific workload in mind and a team capable of owning it, not as a general migration.

The competitive picture deserves clear reading. The frontier of openly distributed models has been set repeatedly by Chinese laboratories over the past year, and Kimi K3 continues that pattern at a scale that would have been considered a closed frontier release very recently. The capability gap that once justified paying premium API prices has compressed, and the arguments for closed frontier models are increasingly about support, indemnification and institutional risk tolerance rather than what the model can do.

Those arguments are legitimate and executives should not let a benchmark table dismiss them. Indemnification in particular has real balance sheet value, and no downloaded weight file provides it. Jurisdiction cuts in more than one direction as well. A model whose provenance sits outside the deploying organization's regulatory environment raises procurement questions that are political as much as technical, and those questions are being asked more sharply this year than last.

The recommended action is narrow and specific. Identify one high volume, low sensitivity workload currently running against a commercial API, and price what it would cost to run against an open weight model on infrastructure the organization already operates. Include the engineering time honestly. The resulting number is the only reliable evidence about whether this shift is relevant to a given organization, and it is available for the cost of an experiment rather than a strategy.

Funding & Investment Story 9 of 12

Fireworks Raises 1.5 Billion Dollars on a Bet Against General Intelligence

Fireworks AI announced a 1.505 billion dollar Series D at a 17.5 billion dollar valuation, led by Atreides Management, Index Ventures and TCV. The company disclosed alongside the raise that it has surpassed 1 billion dollars in annualized revenue run rate and serves more than 40 trillion tokens every day.

One disclosed figure explains the thesis better than the valuation does. More than 95 percent of the tokens Fireworks serves come from models specialized on customers' proprietary data rather than from general purpose frontier models answering general purpose questions. That ratio is a measurement of where production AI workloads have actually settled, taken from infrastructure carrying them at scale.

The distinction matters because the industry conversation and the production reality have drifted apart. Public attention tracks frontier model capability, measured on general benchmarks that reward breadth. The revenue underneath one of the larger inference platforms is coming from narrow models tuned on data that belongs to the customer and optimized for one job. Those are different products serving different needs, and an organization budgeting for the first while operating the second will misjudge both its costs and its competitive position.

The economic logic is straightforward once stated. A general model priced to recover the cost of training on everything is an expensive way to perform one repetitive task well. A smaller model tuned on a company's own data frequently matches or exceeds it on that specific task at a fraction of the serving cost, and the proprietary data is the part a competitor cannot replicate. Specialization converts a company's accumulated operational history into a durable advantage in a way that prompting a shared frontier model does not.

The 40 trillion token daily figure carries its own signal about market maturity. Volume at that scale does not come from experimentation. It comes from systems embedded in production workflows where the token consumption is a recurring operational cost that someone has already justified. The market has moved past the phase where inference spending is discretionary innovation budget.

For executives the actionable question is about data readiness rather than model selection. Specialization requires labelled, accessible, reasonably clean records of how a company actually does its work, and most organizations discover that this asset is in worse condition than assumed only after committing to a project that depends on it. The organizations positioned to benefit from this shift are the ones that invested in data infrastructure for reasons that had nothing to do with AI, and that investment is not something that can be assembled quickly once the need becomes obvious.

Funding & Investment Story 10 of 12

Together AI Raises 800 Million Dollars and Books the Power to Use It

Together AI announced an 800 million dollar Series C on 1 July, with investors including Aramco Ventures, NVIDIA, Vista Equity, General Catalyst and Salesforce Ventures. Alongside the equity the company secured commitments for over 500 megawatts of compute capacity to be capitalized independently by its new investors.

The power commitment is the more interesting half of this announcement and it is structured in a way worth understanding. The capacity is being capitalized separately from the equity round, which means the investors are funding physical infrastructure through a different instrument than the one that bought them shares. That structure exists because datacenter capacity has become the constraint that money alone does not resolve quickly. Grid interconnection queues, power purchase agreements and construction timelines operate on schedules measured in years, and a company that raises capital without simultaneously securing the electricity to deploy it has bought a slower version of the same problem.

The investor list reinforces the reading. An energy company and a chip manufacturer appearing in the same round alongside conventional growth investors describes a capital structure organized around securing supply rather than around financial return alone. Strategic investors in this market are increasingly buying access as much as equity.

Together AI's positioning is explicitly a bet that inference economics, not model capability, will determine where production workloads run. The company's argument is that the cost structure of closed frontier models looks manageable in a prototype and becomes unsustainable in production, because inference bills compound as usage scales while budgets do not. Organizations respond by rationing intelligence precisely when demand for it is accelerating, which is a poor outcome for everyone including the frontier labs.

That argument matches what enterprises report privately. The pilot that cost a few thousand dollars a month becomes a production system consuming a materially larger sum, and the finance conversation that follows is rarely about capability. It is about unit economics on a workload that has already proven its value, which is the worst possible moment to discover the cost curve.

For technology leaders the transferable lesson is about how to plan capacity rather than which vendor to select. If sophisticated investors are treating megawatts as the scarce input worth securing years ahead, an enterprise planning significant inference growth should treat its own compute commitments with similar seriousness. Reserved capacity, multi provider arrangements and negotiated pricing tied to volume are the instruments available, and they are considerably cheaper to arrange before a workload becomes critical than after.

AI Research Story 11 of 12

An AI Agent Is Now Flying a Real F-16

DARPA and the United States Air Force disclosed on 16 July that an F-16 modified under the VENOM programme is undergoing in air testing with an artificial intelligence agent autonomously controlling the aircraft at Eglin Air Force Base. A human pilot remains in the cockpit monitoring the agent throughout.

The engineering choice underneath this milestone is the part with implications beyond defence. Brig. Gen. James Valpiani, the DARPA programme manager, said the team automated flight controls and sensors on a standard F-16 without changing the jet's core software. The autonomy was added as a layer around an existing certified system rather than by rewriting it.

That is a genuinely difficult constraint and it is the same constraint most large enterprises face. The systems that run a bank, an insurer, a hospital network or a utility are certified, audited, deeply understood and extremely expensive to modify. The prevailing assumption in enterprise AI strategy has been that meaningful autonomy requires modernising those systems first, which is why so many programmes stall in a multi year platform replacement before delivering anything. VENOM demonstrates the alternative at the most demanding end of the reliability spectrum: instrument the interfaces, add the intelligence outside the certified core, and leave the core untouched.

The control design carries the second transferable lesson. The VENOM Autonomy Kit lets the pilot toggle between traditional human control and AI control with the flip of a switch. Reversion is immediate, physical and requires no diagnosis of what the agent was doing or why. Most enterprise agent deployments have nothing equivalent. Stopping an agent that is behaving badly typically means revoking credentials, halting a service or filing a ticket, and each of those paths runs through a decision about whether the situation warrants intervention. A switch removes the decision.

The programme is also a statement about test infrastructure rather than about any particular model. VENOM aircraft will serve as the cornerstone for the next phase of AI development under DARPA's Artificial Intelligence Reinforcements programme, which will evaluate multiple competing agents in live flight. The investment is in the capability to test agents rigorously and repeatedly, on the reasoning that the agents will keep improving while the evaluation apparatus is what determines whether anyone can trust the improvement.

That ordering is worth borrowing. Most organizations have built agent capability faster than they have built the means to evaluate it, and the resulting gap is where the eventual incident will originate. The lesson from an agency flying autonomous fighter aircraft is that the evaluation infrastructure is the deliverable, and the agent is what runs on it.

AI Business Models Story 12 of 12

Eli Lilly Commits Up to 2.75 Billion Dollars to AI Discovered Molecules

Eli Lilly and Insilico Medicine have signed an agreement worth up to 2.75 billion dollars granting Lilly exclusive worldwide rights to develop and commercialize preclinical drug candidates discovered using Insilico's Pharma.AI platform. The structure of that sentence contains the entire significance of the deal.

Lilly is not buying software, a licence to a model, or a research collaboration to explore what AI might contribute to discovery. It is buying molecules, with the AI platform named as the method by which those molecules were found. That is a commercial arrangement in which the output of a generative system is the product being transacted, priced at pharmaceutical asset values rather than at software values.

The distinction matters because it settles a question that has been argued in the abstract for several years. AI in drug discovery has produced an enormous volume of partnership announcements structured as option agreements, milestone ladders and research collaborations, all of which allow a large pharmaceutical company to participate without committing to the underlying premise. An exclusive worldwide licence on preclinical candidates is a different posture. It says the acquiring company has evaluated specific compounds and wants to own them.

Executives outside life sciences should read this as a template rather than as sector news. The pattern is a generative system producing candidate outputs in a domain where validation is expensive, slow and rigorous, and a buyer with the validation apparatus paying for the candidates rather than for the tool. That shape recurs wherever generation is cheap and verification is expensive, which describes materials science, chemical engineering, semiconductor design and a growing list of others.

The caution is equally important and rarely stated plainly enough. Preclinical means the hard part has not happened yet. The attrition rate between a promising preclinical candidate and an approved therapy is severe, and it is severe for reasons that have very little to do with how the molecule was originally identified. Biology remains the constraint. A deal valued at up to 2.75 billion dollars is a milestone structure, and the difference between the headline figure and what is actually paid will be decided in clinical trials over years.

What the deal does establish is that the discovery stage has changed enough for a company of Lilly's scale to pay meaningfully for its output. For leaders in any research intensive industry, the question this raises is not whether to adopt a generative tool. It is whether the organization owns validation capability that would let it be the buyer in an arrangement like this one, because in this structure the validation infrastructure is where the leverage sits, not the generation.