Siri Opens the App Door, While AI's Infrastructure Bill Grows

Siri Gets More Ways Into Apps

An app's most useful feature can be buried three screens deep. Apple is giving developers a way to put those actions within reach of a spoken request, including requests that take several steps.

Our main story today is Apple's opening of Siri AI to custom app actions, and what happens when the assistant becomes another way to operate the software you already use.

On 3 November, Apple released a developer preview that expands third-party integration beyond the categories it previously supported. Developers can describe more of their app's own functions to Siri, along with content the assistant can retrieve. That gives an app maker more room to decide which useful parts of the app can be reached through conversation.

There are two pieces to the action side. App Intents exposes individual functions, such as checking a delivery or making a reservation. New skills describe how Siri should combine several actions into a more complex request. An individual function gives the assistant something it can do. A skill describes how those functions fit together.

That distinction is useful because everyday requests often contain more than one operation. Finding the relevant information and doing something with it can involve separate steps, even when a person expresses the whole job in one sentence. Developers now have a way to describe that relationship, giving Siri a route through the app's functions.

There's a content change too. Siri can retrieve app entities from Spotlight's semantic index without requiring them to fit an existing schema. An entity is an item of information the app makes available; a schema is the predefined structure it would otherwise have to match. Semantic search helps find content by meaning.

For an app whose information doesn't fit Apple's existing categories neatly, that broadens what can become discoverable. It also puts content retrieval alongside actions. The assistant can have access to both the thing someone is asking about and the function they want to use. That's a much more interesting integration surface for a developer than a collection of isolated shortcuts.

My read is that this could change where app design effort goes. A screen guides someone through choices visibly. A conversational request can compress those choices into a few words. The app maker has to make the available actions clear enough for that shorter interaction to work, including where permission or confirmation belongs. A delivery check and a reservation have different consequences, even if each begins with a casual request.

That makes the definition of an action part of the user experience. What does the action do? What information does it need? At what point is the person committing to something? Those are ordinary product-design questions, now expressed through an assistant as well as through buttons and screens.

The opportunity for developers is to make an app's existing usefulness easier to reach. People could spend less effort remembering where a feature lives, while the app still supplies the function behind the request. For now, this is developer testing; wider use depends on app makers implementing and shipping integrations. The strongest reason to explore it is a task that becomes genuinely simpler when someone can ask for it in their own words.

Tencent's Hy4 Takes on Visual Work

That's Apple's route into app actions. The shorter news starts with a model designed to work through software itself.

Tencent launched the full Hy4 model on 4 November, adding image understanding and computer use to its Hy family. Tencent says it can work through graphical interfaces and command-line tools, with targeted training involving professional software such as Blender and CAD applications.

That combination connects two kinds of interaction. A graphical interface presents controls and information visually. A command line gives software a text-based way to carry out operations. A model able to use both has more possible routes through a job, depending on how the application exposes its functions.

The inclusion of professional design software is interesting because these environments involve more than reading text from a screen. The user is working with a visual result, and the next action depends on what's already there. Image understanding can help connect that current state to the operation being requested. Computer use supplies the action side of that loop.

For developers, Hy4 is available through Tencent Cloud and Tencent applications, and OpenRouter lists a live API offering. That provides a hosted route for evaluating the model in a workflow. Downloadable weights and system cards are still promised for later, with the release date and licence unresolved. A self-hosted deployment therefore remains a future option.

That timing creates a practical split between two kinds of evaluation. A developer exploring whether Hy4 can handle a visual task can use the hosted offering. Someone choosing a model specifically to operate on their own infrastructure still needs the weight release and its terms before making that deployment choice.

I'd put the emphasis on a complete piece of work: whether the model can interpret the starting state, take the relevant actions and leave the software in the intended state. That's where the combination of vision and computer use becomes valuable. A tool that saves effort on a repeatable visual workflow could be useful even if the organisation has little interest in changing its general-purpose chat model.

Hy4 adds another hosted candidate for that kind of software operation. Its fit will come down to the particular application and task a developer needs it to handle.

An Agent Stop Button at the Gateway

Once software can take actions, the person running it needs a practical way to interrupt the work.

CrowdStrike announced immediate availability of an AI Kill Switch in Falcon Guardian on 3 November. An authorised administrator triggers the intervention, and the action is recorded in an audit trail.

The current enforcement point is the gateway. That's the controlled route through which an agent's activity passes. Endpoint-level termination, which would stop execution on the machine itself, is still forthcoming. For a security team, that location is the central detail in understanding what the new control can interrupt.

An agent's job can involve activity in several places. Controlling its gateway can contain the activity passing through that gateway. Local execution sits at a different enforcement point. So the useful deployment question is which parts of the agent's work pass through the route Falcon Guardian controls.

That turns a dramatic product name into a fairly concrete architecture discussion. The administrator needs to know where the intervention takes effect and which access paths it covers. If the relevant action goes through the controlled gateway, the stop control is positioned to intervene there. A separate execution path needs its own coverage.

The human trigger also defines the operating model. An authorised person decides to intervene, and the audit trail records that intervention. That record can help the organisation establish when the action was taken as it reviews what happened. It gives the stop decision a place in the operational history of the agent.

CrowdStrike also signed agreements to acquire Fiddler and Dreadnode, for agent observability and security testing. Those acquisitions and their planned Falcon integrations are pending. The available product change is the gateway-level stop control.

For organisations already using Falcon Guardian, my takeaway is that this makes the agent's access route newly important to the security review. A stop button becomes useful when it sits in the path of the work you need to stop. Mapping that path is a much more specific exercise than asking whether an agent has a kill switch somewhere in the stack.

Google Connects Data Across Clouds

There's another practical obstacle for enterprise agents: the information they need often lives in more than one cloud.

On 4 November, Google announced general availability of cross-cloud interconnect and caching for AWS and Azure within its Lakehouse offering. The update sits within Google's wider Agentic Data Cloud announcements.

It allows Google Cloud to query supported external tables while the source files remain with the other cloud provider. An organisation can connect analytics and conversational agents to that information without first carrying out a wholesale migration.

The mechanism is useful to understand. A query reaches data stored elsewhere, and retrieved blocks can be cached inside Google Cloud. A cache keeps data available closer to where it's being used, which can reduce repeated retrieval for suitable workloads. The source stays in its existing location while some of the data used by queries crosses the cloud boundary.

That gives an organisation a more flexible way to work with distributed information. A migration can be a substantial project in its own right. Being able to query supported tables where they already live could let an AI project proceed without making relocation of all those files a prerequisite.

It also changes the work around permissions and location. Access to the external table and the presence of cached blocks in Google Cloud both belong in the design. An organisation evaluating this service has to account for the source and the cache when deciding where its information is available.

On price, Google advertises zero variable egress fees for its private offering. Egress is the charge for data leaving a cloud. The documentation also describes chargeable public-internet transfers and private transport costs that depend on configuration. The selected network path therefore belongs in the cost calculation alongside the query workload.

I can see this being attractive where developers are spending too much effort maintaining copies and pipelines purely to bring scattered data together. The benefit would be less of that supporting work, with agents able to reach useful information across existing cloud boundaries. The decision is an architectural one: compare the transport, cache and query costs with the copying and maintenance the organisation could avoid.

Schools Face a Linked AI and Search Choice

For schools, access to AI is becoming tangled up with a service students already know very well.

Google announced on 3 November that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will default to off for school student accounts from kindergarten through year twelve, until administrators enable them.

These new account controls are planned. Google says schools will receive details in the coming months before rollout, and it hasn't established an Australian rollout date. That gives administrators an upcoming policy decision rather than an immediate change they can assume has already reached their accounts.

Search is the particularly consequential inclusion. Chalkbeat reports that Google's education chief confirmed schools rejecting AI entirely would also have to leave Search disabled, because AI is integrated into it. A school wanting an AI-free setting would therefore be giving up access to a familiar search service through those student accounts as well.

That makes the decision more difficult than choosing whether to enable a new classroom application. Administrators are weighing AI access alongside an established way students find information. The educational policy and the practical access choice become connected.

There are separate changes to Guided Learning. It became globally available in Google Classroom on 3 November. Its Search counterpart begins with selected US schools later in November. Those are different availability tracks: a global Classroom release, a limited US Search introduction, and the account controls still to come.

For an Australian school, that distinction is useful for planning. Classroom availability can be considered now. The Search pilot's limited scope and the pending details of the new controls leave other decisions for later. The exact control granularity is still unspecified, so administrators don't yet have the full menu of settings behind the announcement.

My concern here is the cost of the policy choice. Default-off access puts the initial decision with the school. Linking an AI-free approach to disabling Search makes that approach more disruptive. A school may have sound reasons for its position on student AI use; the available controls determine how much else has to change to put that position into practice.

Anthropic's Long Infrastructure Commitment

The scale changes sharply with this one: Anthropic has put a much larger infrastructure picture into its IPO filing.

An amended registration filing dated 2 November discloses approximately eight hundred and ninety-five point four billion US dollars in aggregate purchase and lease commitments. Substantially all of that relates to compute and infrastructure.

The payment timetable gives that enormous number its meaning. About sixty-nine point eight billion US dollars is due in the twelve months following 30 September 2026. Substantial remaining obligations extend into the period from 2031 to 2045. The total describes commitments reaching many years into the future.

There's also a conditional portion. Roughly two hundred and thirteen point eight billion US dollars involves arrangements with termination rights or conditions before they commence. Those terms affect how much capacity ultimately comes into use and which obligations proceed.

The filing reports third-quarter revenue of sixteen point nine five six billion US dollars. That describes three months of business. The commitment schedule stretches across much longer periods, so understanding the financing means following when payments fall due alongside how the business grows. The immediate cash demands and the long-term capacity plan are related parts of that assessment.

For enterprise customers, this is a more detailed view of the infrastructure obligations behind Claude's expansion. Capacity is part of a model supplier's ability to support growing use. Reserving it ahead of time can give the supplier room to serve that demand, while committing the business to substantial future spending.

The hard commercial question is utilisation: how much of the secured capacity gets used productively, and whether the revenue it supports can sustain the associated costs. A long commitment can be helpful when demand grows into it. It can become expensive when the business is paying for more infrastructure than it needs.

I'd treat these disclosures as useful information for anyone making a long-term supplier decision. They show both the scale of Anthropic's intended expansion and the payment horizons behind it. For infrastructure suppliers, those horizons describe the prospective business stretching years ahead. For customers, they make financing and capacity use more visible parts of assessing the service they're building around.

AMD's Rack-Scale Hardware Reaches Workloads

Some of that infrastructure story is already showing up in chip revenue and systems doing work.

AMD reported third-quarter data-centre revenue of about eight point five billion US dollars on 3 November, up ninety-seven per cent from a year earlier. It attributed the growth to EPYC processors and the launch of Instinct MI450 GPUs. The segment combines CPUs and GPUs, so that sales total covers both kinds of hardware.

AMD says Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure are running workloads on its first Helios rack-scale systems. That puts the new systems into operational use with named major customers, a useful stage beyond announcing the hardware.

Rack-scale systems bring the discussion up from an individual accelerator to the larger system used to run workloads. For organisations buying capacity at that level, the ability to put a workload on the system is an important step in evaluating an alternative.

My reading is that AMD's position is becoming more tangible for large AI deployments. Stronger segment revenue shows a growing data-centre business, while the named Helios deployments give customers concrete examples of where the new hardware is being used. Organisations considering their own deployment still have a workload-specific comparison to make on performance and cost. The useful development is another hardware route reaching use at major operators.

What Changes for You

If you're building an Apple app or looking after a WordPress site, there are two immediate details to pin down.

Apple's Siri preview accompanies the iOS, iPadOS, macOS and visionOS twenty-seven point three betas. Testing requires compatible Apple Intelligence hardware and a supported language. That gives Apple-platform developers a current environment for prototyping the custom actions and skills we've discussed.

App Store submissions for these integrations are planned for Northern Hemisphere spring 2027, which falls during Australia's autumn. Apple hasn't supplied a firm customer launch date or complete usage terms. Siri AI also remains unavailable on iPhone and iPad in the European Union.

For a developer, the practical opportunity now is to explore how an existing app function works as a conversational request. Customer release planning still depends on those later submission and launch arrangements.

For WordPress operators, the immediate issue is an upgrade flaw that has moved onto the known-exploited list. Canada's Cyber Centre reports that CISA added it to the catalogue on 3 November. The vulnerability can permit remote code execution during the WordPress upgrade window, meaning an attacker can cause code to run on the affected system.

The relevant fix was released in WordPress seven point one point four on 28 October. Confirmed exploitation is the new development, making this a maintenance-priority change for anyone responsible for an affected site.

The useful check is whether that security update reached the site and completed successfully. An automatic-update setting tells you how an update is intended to happen; the installed version tells you what the site is running. For an operator juggling several sites, that gives a concrete way to close the gap between a patch being available and the affected installation being fixed.