Apple Filing Describes Phones Swapping AI Models That Fail

An Apple patent filing describes a phone that checks its own AI models and switches to a different one when a model fails. The international application WO2024040586A1, filed on 26 August 2022 and published on 29 February 2024, claims a user equipment that can “detect a failure status of a first Artificial Intelligence(AI)model of the UE;and switch to a second AI model that is different from the first AI model”.

By our arithmetic, 20 of its 30 claims cover that failure detection and switching, and the other 10 cover how the phone and the network agree on the quality of the data used to train the models. Patently Apple reported on 30 October what it calls a “newly published patent application” describing the same system; its article gives no application number.

Business Pill · Vertical Integration Explained

An evergreen Business Engineer explainer on vertical integration. It teaches the general idea only and does not discuss this filing or Apple.

What the Filing Claims

The abstract says the application covers “AI/ML model quality monitoring and fast recovery under model failure detection”. The models in question are AI models of the phone used in wireless communication. The description says “AI/ML techniques may be applied to the wireless communication systems(for example,cellular communication systems)to improve system performance”, and names “Channel State Information(CSI)feedback enhancement” among the uses.

Claim 12 says the phone detects a failure when “values for one or more inference quality metrics of the first AI model do not match one or more first preconfigured threshold values”. Claim 14 lists three such metrics: “inference accuracy of the first AI model”, inference integrity and “inference latency of the first AI model”.

Claim 16 adds a time condition: the failure is detected only if the metrics miss the thresholds “for a preconfigured time duration”. The description calls this “a time-to-trigger mechanism”. Patently Apple reads it as a way to avoid switching on every short dip.

Phone or Network Decides

The claims set out two routes. In one, the phone detects the failure itself, sends “model failure information associated with the first AI model” to the network and receives “a switch indication to switch to the second AI model” (claim 17). Claim 18 says that report “indicates one or more candidate AI models that are preferred by the UE”.

In the other route, the phone reports its metric values to the network, which sends the switch instruction (claim 24). Claims 27 and 28 say the reports can be periodic or triggered by events.

Negotiating Training Data

The first 10 claims cover a separate step. The network sends a “training quality enquiry message” and the phone answers with a “training quality response message” (claim 1). Claim 2 lists five metrics, including “a minimum size of a dataset that is to be used for training the one or more AI models”, “a maximum age of a dataset that is allowed to be used for training the one or more AI models” and “a time budget for training the one or more AI models”. Claim 5 says the phone can return “a rejection indication or an acceptance indication”.

The Filing Record

The Google Patents record names Apple Inc as applicant and lists seven inventors: Peng Cheng, Ping-Heng Kuo, Alexander Sirotkin, Fangli Xu, Huaning Niu, Haijing Hu and Ralf Rossbach. It also lists a related Chinese publication, CN119790631A. By our arithmetic, the filing date is 50 months before Patently Apple’s report.

Apple has also been in our coverage on 30 October for other reasons: BofA cut its iPhone estimate on memory costs, and Apple has an EU-disclosed pact with Ascend.io.

The Structural Read

As we read the claims, the filing treats an on-device model as something to be monitored and replaced, not trusted: claim 16 requires the failure to persist for a preconfigured time before it counts.

As we read claims 17 and 24, control is shared: the phone can name the models it prefers, but in both routes the claims have the network send the switch indication.

As we read claims 1 to 10, the network also sets the terms of training, down to a minimum dataset size, a maximum dataset age and a time budget.

Three Implications

FOR PHONE MAKERS Claim 18 lets the phone list the candidate models it prefers when it reports a failure, so the device keeps a say in which model replaces the one that failed.

FOR NETWORKS Claim 13 says the thresholds are configured by a network device, and claim 25 says the reporting is configured by the network device.

FOR READERS A patent application describes an invention; the filing does not say which Apple devices or modems would use it, or when.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework says “Foundation models become relatively commoditized as orchestration layers capture differentiation.”

As we read it, the filing puts that orchestration on the phone and the network: which model runs is set by monitoring and switching rules, not by any one model. The framework does not discuss this filing or Apple.

What Is Not Established

Patently Apple does not give an application number, and the Google Patents record we read lists no US publication for this family. We matched Patently Apple’s description to WO2024040586A1 on the claims it describes; we did not find a US publication of this application in the records we could read. A patent application describes an invention; it does not show that Apple will ship it, and nothing in the filing says which devices or modems would use it. We did not contact Apple.

The Bottom Line

An Apple patent application claims a phone that detects when an AI model serving its cellular link misses accuracy, integrity or latency thresholds for a set time and switches to a different model, with the network able to direct the switch. Twenty of its 30 claims cover that recovery step, by our count.