The Morning Paper — 15 to 19 October 2026

What changed in A I

Technology exports, electricity supply and access for safety evaluators are central to this week's developments in artificial intelligence. The lead story concerns China's proposed export rules and the distinction between transferring technology and using its services.

Power agreements in Indiana bring customer funding and construction conditions into view. Further stories examine the rights of outside evaluators, preparations for incidents during advanced model development, and research into quantum circuits that use fewer resources. A new software integration promises a more familiar route to Amazon's computing hardware.

Brief items cover industrial applications, a proposed benefit fund and reported exposure of access credentials. This is The Morning Paper for the fifteenth to the nineteenth of October, ending with the decisions and evidence to watch next.

China's technology-export proposal

China's commerce and science ministries have proposed additions to the country's technology-export catalogue, alongside an immediate registration requirement. That is according to the official notice reproduced by Science and Technology Daily. The proposed artificial intelligence entries exclude technology already made public through open sourcing or academic publication. For technology within the stated scope, the notice describes both an obligation applying now and categories being considered for the future.

The consultation runs from the sixteenth of October to the fifteenth of November. During the interval between publication of the notice and the revised catalogue formally taking effect, covered exports must be registered before export. This requirement also covers exports already in progress. The consultation timetable and the registration timetable therefore have different end points.

The proposed catalogue contains forty additions, with twenty-three concerning artificial intelligence. The remaining entries concern biomedicine and solar technology. Within artificial intelligence, the headings include large models, agents, computing clusters and information processing. These are headings for types of technology. Most of the proposed entries have technical thresholds, so the relevant specification matters when assessing a particular system or proposed transfer.

Registration covers several forms of transfer. An owner might transfer ownership of a technology, license it to a foreign recipient, or make a change that gives a foreign party ownership or actual control.

The registration information identifies the parties and destination, describes the technology and intended use, and gives the projected value of the transaction. The permanent treatment of the proposed additions remains part of the catalogue consultation and the process for bringing the revised catalogue into effect.

The ministry's explanation, reproduced by Beijing News, proposes fourteen prohibited categories and twenty-six restricted categories. Those totals cover all the sectors in the proposed additions. A prohibited category would prohibit the covered technology export. A restricted category would make licensing the relevant route. The technical thresholds help determine which technology falls within each proposed category.

The examples make the distinction more concrete. The ministry proposes prohibiting technology for generating malicious code. It proposes a licensing requirement for artificial intelligence technology that discovers software vulnerabilities.

For large models, the ministry also distinguishes a transfer of covered technology, or its deployment overseas, from foreign access to a service. Its explanation places access through an A P I, a website or an app outside the described large-model technology-export category. An A P I is an interface through which software requests access to a service. The explanation also includes software provided as a service and lightweight deployment on a user's device among these access routes.

The ministry describes the revision as routine and says its aims are preventing misuse and loss of control. It presents the proposal as generally applicable across countries and companies. It also describes licensing facilitation for eligible open-source releases, overseas facilities and research cooperation. Products manufactured using these technologies sit outside these catalogue changes, according to its explanation.

The practical inference is to examine registration, proposed restrictions and service access separately. Registration already applies within the notice's scope. The final catalogue remains proposed, and the ministry's service-access distinction remains part of its stated coverage. The next material development is the consultation's outcome and the technical criteria governing the resulting categories.

Power plans and financing risk

X-energy and American Electric Power have announced feasibility work for a nuclear project near Indiana's retired Breed Generating Station. Amazon will provide development capital as the project reaches agreed milestones. The agreement concerns development, with construction approval and financing through operation still to be established.

The proposal comprises twelve reactors with a combined capacity of nine hundred and sixty megawatts. Feasibility work will help determine whether the development can proceed. The companies identify further agreements, financing, licensing, supply chains and the performance of a first project of this kind among the conditions affecting delivery.

Amazon says it intends to pay its full costs, while American Electric Power describes financial safeguards as essential. These statements address responsibility for the expense. The work of securing permission, arranging finance and delivering the equipment remains part of the project. A commitment about who pays and evidence that a plant can enter service answer different practical questions.

The wider collaboration has an ambition exceeding five gigawatts across projects by twenty thirty-nine. That is a goal for the collaboration as a whole. The Indiana proposal has its own development requirements and would contribute capacity if it proceeds through them.

A separate agreement between NiSource and Microsoft also concerns electricity in Indiana. NiSource expects service to begin in twenty twenty-eight under a twenty-year term. The company says Microsoft will fund the generation and transmission it requires. The agreement retains execution and regulatory risks, so the expected service date depends on the work and approvals needed to deliver it.

The arrangement would add approximately one thousand one hundred megawatts to GenCo's shared portfolio. A related battery contract covers six hundred megawatts for the shared pool. The battery announcement gives a power figure, while storage duration remains an open detail. That duration would be needed to establish how long the stated output could be sustained.

NiSource forecasts one point seven five billion dollars in savings for existing N I P S C O customers across its data-centre contracts. Its previous forecast was one point four billion dollars. The new figure is an expectation for the combined contracts.

Financing conditions provide another part of the picture. Reuters reported that AirTrunk's planned real estate investment trust listing, Firmus's listing and S B Energy's listing had been pulled or paused as capital became harder to secure on expected terms. Its account uses the preceding week as context, so the report's publication date and the dates of individual financing decisions have different scopes.

The same account described continuing large transactions and strong regional debt financing. Interviewees distinguished established operators with revenue already under contract from businesses depending on future construction and projected revenue. The distinction concerns how much of the expected business is already secured and how much still depends on delivering a project. Their concern centred on financing terms and execution risk.

An earlier essay by Dwarkesh Patel offers a conditional explanation for possible pressure on computing prices. He argues that computing could become more expensive if economically valuable demand for artificial intelligence grows faster than capacity. His argument depends on uncertain assumptions about capabilities and growth. It is background reasoning about a possible future, with the outcome conditional on those assumptions.

The inference from these examples is that growing demand can coexist with difficult financing. Allocating costs to a technology customer clarifies who is expected to pay. Permitting, construction and the customer's performance still help determine whether capacity arrives. The evidence to watch is therefore both financial and physical: further agreements, approvals, financing and progress towards service under the conditions the companies have described.

Inspection rights and incident response

Anthropic has announced access for six staff from the evaluation organisation M E T R to company hardware, offices and internal tools. It describes their access as broadly comparable to that of its internal risk teams. The arrangement gives an outside evaluation team access to examine the company and its systems, subject to specific boundaries on access and disclosure.

Those rights belong to the participating team. Customer data and material protected by attorney-client privilege remain outside the arrangement. That privilege protects confidential exchanges with lawyers. Access to model architectures, the designs of the models themselves, requires explicit additional approval.

Publication rights have their own conditions. Anthropic says M E T R can publish its analyses on its own authority. Material in predefined sensitive categories can be redacted, meaning removed from the published account, and Anthropic ultimately decides whether information falls within those categories. M E T R may also publish a statement describing its level of access. Independence in deciding to publish and authority over sensitive disclosures are separate parts of the arrangement.

A separate development concerns preparations for failures during advanced model training and evaluation. The government agencies N I S T, N S A and C I S A have published ten voluntary practices for incident readiness. They organise the work into preparation, detection, response, recovery and reporting. These stages cover what operators arrange beforehand, how they recognise an incident and how they act through recovery.

The guidance gives a specific meaning to a control incident. It involves significant actions that a system initiates independently and that compromise the boundaries of its task or human oversight. The source of the action matters to this definition. An externally directed attack has an adversary directing the activity; a control incident concerns actions initiated by the system itself. Control incidents and ordinary cyber incidents can overlap.

Preparation includes limiting the permissions available to an agent. Monitoring and logs should be protected from interference by the agent being monitored. The guidance also recommends testing monitors and preserving evidence. These practices address both the ability to detect an incident and the ability to establish what happened afterwards.

During an incident, the document calls for stopping the affected agent's activity. Depending on the severity, other runs may also need to pause. Before activity restarts, credentials and potentially compromised environments need attention. Recovery therefore includes the conditions under which the system resumes work, as well as the immediate response to the behaviour that triggered concern.

The guidance warns that recovery training can make adverse behaviour harder to detect. It recommends independent review and appropriate public sharing. Existing legal reporting duties continue to apply, while reporting control incidents and near misses under this guidance is encouraged voluntarily.

Earlier conversations published by The Cognitive Revolution add a distinction about software proofs. Everett Hildenbrandt separates proving that translated code preserves its behaviour from proving that the behaviour is correct. The first establishes a relationship between versions of a program. Correctness is a further property to establish. Nathan Labenz questions whether reliable tools alone establish that an agent acts in line with intended goals. These are attributed arguments about the scope of assurance.

The inference from the access agreement and the guidance is that inspection and response procedures are becoming more concrete. Their effectiveness still depends on useful evidence, working monitors and disclosure in practice.

Quantum circuits with fewer resources

Anthropic describes research in which Claude Mythos five repeatedly wrote programs that generate quantum circuits. Humans supplied the tasks and scoring machinery. An automated evaluator tested candidate solutions and returned feedback, and the model used that feedback in further attempts. Occasional human prompts broadened the search. The resulting designs were then assessed through tests, analysis and proofs for particular components.

A quantum circuit has working resources called qubits and operations called gates. Reversible computation needs to clear temporary work while preserving information that the calculation still requires. That clearing process adds overhead. Reducing the amount of temporary work, or the operations needed to clear it, can reduce an algorithm's resource requirements. The research searches for ways to make that work more economical.

The score combines particular kinds of gate with the largest number of logical qubits needed at any point. Specifically, it multiplies the count of Toffoli or C C Z gates by peak logical-qubit use. The result is a resource estimate for a circuit design. Interpreting it requires keeping both factors in view, since a change in operations can accompany a change in working space.

The preprint reports improvements in both resource measures for seven of the eight tasks. Modular inversion, an arithmetic task, involves a trade-off between time and space. The comparison for a Hermitian decoder uses the authors' own basic reversible implementation as its baseline.

One arithmetic method reconstructs earlier choices from the values currently available. A calculation can follow different branches, and retaining a record of those choices takes space. Recovering the choices from current values reduces the history that needs to be stored.

A decoder illustrates another mechanism. The design omits extra storage that would be needed only rarely. It accepts a permitted probability of failure in exchange for the lower storage requirement.

For the Reed–Solomon decoder benchmark, the product of gate count and qubit use falls by approximately a factor of twelve. The allowed failure probability is at most one in a thousand.

Another example concerns elliptic-curve point addition with two hundred and fifty-six bit values. The relevant comparison circuit is the earlier design with the smallest qubit requirement. This choice of baseline matters because the result trades a reduction in operations against a larger requirement for working resources.

The new design uses about a seventeenth as many gates and around twelve per cent more qubits than that comparison circuit.

Verification also has a defined scope. The work combines tests, mathematical analysis and component proofs, with bounded failure permitted in some tasks. Those forms of evidence support particular parts of the results. The measurement is a calculated gate and qubit requirement, with the allowed failure conditions attached to the tasks where they apply.

The inference is that a model can contribute to algorithm search when proposed designs receive structured feedback and checks. The demonstrated progress concerns circuit design and its resource requirements.

A new way to run on Trainium

TorchNeuron is a new connection between the PyTorch software framework and Amazon's Trainium computing hardware. The technical announcement schedules its public beta, a release for public testing, for the twentieth of October. At the close of this edition's reporting period, that release remains scheduled. The announcement describes a software integration with a defined set of supported features. The release is planned under the Apache two point zero software licence.

The integration is a native backend: the part of the software that connects model operations to the computing device. It supports eager execution for interactive work, in which operations can be run directly, and compilation for optimisation. The aim is to make the hardware available through the framework developers already use for their models and training work.

Hugging Face describes the practical change from its previous Optimum Neuron approach. That earlier route required custom model code and training machinery around a separate backend. The new integration exposes Trainium as a native device to familiar training and inference libraries. Training develops a model; inference runs it to produce outputs. The company presents reduced implementation work as a practical benefit of the integration.

Hugging Face's compatibility comparison covers the hundred most-used text-generation architectures. Its earlier software in that comparison is Optimum Neuron version zero point four point five. An architecture describes the design of a model, while an individual workload also has its own configuration and settings.

The company reports coverage rising from forty-six per cent with the earlier software to eighty-nine per cent with the new integration. These figures measure compatibility: which architectures the implementation supports. They provide evidence about the range of models developers can attempt to use through this route. Performance and cost for a particular workload would require measurements under that workload's conditions.

Device-specific settings still matter. Hugging Face's example recommends a static generation cache, with storage for generation fixed in advance. Automatic dynamic shapes and asynchronous weight loading are outside the example's supported configuration. The first concerns automatic changes to data dimensions; the second concerns how the model's saved values are loaded. The example produces a standard portable Transformers checkpoint, a saved model in a familiar format.

The technical announcement describes tests using reduced model configurations and records C P U fallbacks, where work returns to the central processor. The first release also has limits on supported data representations and on handling changing or abstractly specified data dimensions. Some features for calculations using less precise numbers remain in private beta. The published coverage therefore sits alongside conditions on which features run through the new hardware integration.

Amazon identifies support for Trainium two and Trainium three. It says Cartesia ran voice inference with minimal PyTorch changes. It also describes the Lieber Institute exploring biological and chemical modelling.

The inference is that reducing the work of adapting a model could make alternative computing hardware easier to evaluate.

Industrial A I, benefit sharing and OpenCode

Germany's economy and research ministries have signed a declaration of intent with Nvidia on industrial artificial intelligence. The government identifies specialised open models using Nemotron and Cosmos. It also describes a refocused Plattform Industrie four point zero spanning infrastructure and applications. The platform is open to all companies, and the initiative is intended to involve industry, researchers and smaller businesses. The platform maintains cooperation with international partners. The declaration establishes an agreed direction for future development of industrial models and their applications.

Anthropic has announced a fund holding five per cent of its shares, subject to stockholder approval. The fund will be a wholly owned limited-liability company. Possible uses include donations and direct distributions, with governance and milestones still to be announced. The commitment takes the form of company shares, and the arrangements for turning that value into benefits remain to be developed.

Further commitments of two per cent of shares for each fifty per cent increase in valuation above the baseline depend on board discretion. The fund's equity is reserved for uses outside lobbying and political giving. Separate founder and employee pledges sit alongside the company's conditional commitment.

The Hacker News reports OpenCode's account of an intrusion into a legacy system on the night of the sixteenth of October, potentially exposing provider keys hosted in Console. Those keys are credentials used to access model providers. The company put the affected share at one point four per cent of Console users. OpenCode's Dax Raad distinguished these from keys stored locally by the application and command-line tool. This remains the company's account of the incident as reported by the publication, with the extent of exposure described as potential.

OpenCode reportedly notified affected users the following day and told them to revoke affected keys at the provider. Users alleged unauthorised charges. The company described the incident as related to distillation, collecting a model's answers to train another model; that motive remains a company allegation. The identity of the attacker, complete intrusion mechanism and aggregate losses remain uncertain. The reported exposure concerns hosted Console credentials, with the affected proportion referring specifically to Console users.

What to watch

China's consultation closes on the fifteenth of November. Watch the resulting catalogue and its technical thresholds for evidence of which transfers the final categories cover. The immediate registration requirement and the ministry's stated treatment of service access remain relevant in the meantime.

TorchNeuron's public beta is scheduled for the twentieth of October. Availability and experience with specific workloads will show how the announced compatibility translates into practical use under the stated feature limits.

For the evaluator arrangement, watch for published analyses and M E T R's account of its access. Those would help establish what the agreement enables evaluators to examine and disclose.

For the Indiana power projects, further approvals, agreements and financing would change the assessment of delivery. Customer funding commitments establish intended responsibility for costs. Progress through the remaining conditions will determine how the proposals move towards electricity supply.