All episodes

AI & Tech Daily

ChatGPT reaches checkout, and open AI gets more practical

17:46

ChatGPT Wallet adds user-approved purchases for eligible US subscribers. OpenShell moves into community governance, llama.cpp establishes compatibility boundaries, Runway opens its robot-control model, and World Labs makes generated scenes editable. Plus AWS cross-account secret replication, Atom Computing's quantum R&D award, and Salesforce's self-hostable transcription model.

Full transcript

Read the episode.

I'm Jesse Owen. This is AI and Tech Daily.

ChatGPT reaches checkout

An AI agent can now take an eligible ChatGPT user's purchase through checkout without receiving their actual card details. The customer still approves the transaction.

Our main story today is ChatGPT Wallet, and the question of how you give an agent the ability to pay while keeping control of the purchase. OpenAI began rolling it out through its dots agents on 13 November. It adds a concrete step to AI-assisted shopping: completing a payment with the user's authorisation.

The launch flow separates the agent's work from the customer's decision to spend. You review the purchase and approve it through a security check. The payment network then supplies substitute payment credentials, so the agent can proceed without being given your actual card details.

That separation is the interesting piece. Finding an item and paying for it involve different permissions. A system can help narrow down a choice without needing the ability to charge a card. Wallet creates a way to cross that boundary for a particular purchase, with a human decision built into the flow.

For someone using an agent to shop, the potential convenience is fairly easy to picture. The agent's assistance can extend into checkout, while the customer retains the point at which they say yes to spending. Payment becomes part of the task the agent can carry through, instead of a separate manual step left entirely to the user.

The initial audience is narrow: US ChatGPT Pro subscribers aged eighteen or older, using eligible US-issued cards. Australian availability hasn't been announced. So for listeners here, this is a look at a new purchasing arrangement entering use overseas.

There's also a practical dependency on the shop receiving that payment. OpenAI says websites and fraud systems can block or decline agent purchases. An authorised transaction still has to make it through the merchant's systems. That creates an interesting problem for retailers: recognising that a customer has approved an agent's purchase, even though the interaction may look different from someone doing everything themselves.

A merchant could therefore encounter customers whose payment approval is valid but whose automated shopping activity gets stopped along the way. The commercial question becomes whether the checkout can accommodate that authorised behaviour. From the customer's side, the experience depends on both ends working together. Having Wallet available is one part; getting through a particular shop's checkout is another.

OpenAI also describes a future in which people delegate spending within preset limits. Today's launch uses approval for each purchase. The Wallet terms, effective 12 November, already allow for advance authorisation, but that broader spending mode remains future functionality in the launch announcement.

Those two arrangements would feel quite different in use. Approving a purchase means deciding on a transaction presented to you. Setting a budget in advance would let an agent make spending decisions within that permission. The launch gives us the first arrangement, and a useful view of where OpenAI wants to take it.

My read is that this makes agent shopping more practical by giving payment its own explicit approval step. It also gives merchants something specific to work on: accepting a transaction the customer has authorised through an agent. The next useful measure of progress is how consistently those purchases can get through checkout, because that's where the promised convenience reaches the customer.

OpenShell gets shared governance

That's the checkout question. A shorter development now concerns the permissions behind an agent's work.

NVIDIA contributed OpenShell to the Cloud Native Computing Foundation on 10 November. It's joining as a Sandbox project, with development moving to the ai-openshell organisation and NVIDIA continuing its engineering work.

OpenShell is a runtime for agents: the software environment in which they operate. It enforces boundaries around access to files, networks, credentials and inference providers, which are the services or systems running the models. These are concrete permissions governing what an agent can reach while carrying out a task.

That gives platform engineers a place to express controls around an agent's activity. A task may involve reading files, contacting a service and using a model. Each of those interactions reaches a resource outside the agent itself. Controlling those connections is part of making agents usable inside an organisation with existing access rules.

CNCF says OpenShell is adopting Cedar as its policy language. A policy language provides a structured way to state access rules. For developers integrating a runtime, having those rules expressed explicitly makes the permission model available to inspect and work with alongside the rest of the infrastructure.

The governance change matters for a different reason. A project housed under a community foundation offers other organisations a place to contribute to the software they may depend on. It opens the possibility of shared maintenance and integrations across suppliers. NVIDIA's continuing involvement gives that transition an engineering contributor already familiar with the project.

Sandbox is CNCF's early maturity stage. The practical question from here is whether a wider group of maintainers and integrations develops around OpenShell. That's what would turn the governance move into a stronger shared foundation for organisations running agents.

For platform leaders, I see value in being able to inspect and help shape the controls their agent infrastructure relies on. Permissions touch the organisation's own files and services, so the ability to understand how they're enforced is useful in its own right. Shared governance could also reduce dependence on one supplier as that community grows.

A clearer contract for local AI

If you build AI into an application, a library update can create a surprisingly large amount of work.

The ggml team released llama.cpp and ggml version 1.0 on 13 November, with an explicit commitment to signal incompatible public-interface changes through major version numbers. That gives developers embedding local inference a clearer way to judge what an upgrade could disrupt.

There are two closely related interfaces involved. The public C API is the set of functions application code can call. The ABI, or application binary interface, concerns how compiled software fits together. An incompatibility in either can affect an application using the library, and the announcement covers both.

The versioning rules distinguish those breaking changes from compatible additions and fixes. That distinction gives a maintainer more useful information before they begin an upgrade. A major-version boundary flags a possible migration; compatible changes sit within a different category. It's a small piece of information with a potentially large effect on planning.

Version 1.0 itself removes deprecated functions. Existing applications calling those removed interfaces need to migrate. So there is some immediate work for affected developers as the project establishes the new boundary. The release gives them a defined point at which to deal with that older integration code.

For anyone maintaining a private or cross-platform AI application, dependency changes sit alongside the work users actually see. The developer may want to improve the product, yet first has to establish whether a library update still connects correctly to the application. Clearer compatibility rules can make that maintenance work easier to anticipate.

They also help separate two different questions during an upgrade. One is whether your software still connects to the library through supported interfaces. The other is how your chosen model behaves and performs on the hardware you're using. This commitment addresses the interface question; model and hardware evaluation remain specific to the application.

My takeaway for builders is that a clearer dependency contract makes local inference easier to maintain as part of a longer-lived product. The first migration may take effort, particularly if old calls remain in the code. After that, version numbers offer a more dependable signal about where the next compatibility boundary lies.

Runway opens its robot foundation

Let's step away from application plumbing for a moment and look at software that moves a robot.

Runway released the weights, inference code and fine-tuning code for Praxis-1 on 11 November, under Apache 2.0. The model was introduced in September. This release gives robotics developers access to the material needed to run and adapt it, including for commercial applications.

Praxis-1 draws on video pretraining and jointly predicts upcoming video and robot movements. That ties a prediction about what comes into view to a prediction about how the robot moves. It's a way of connecting visual understanding with action inside the model.

The joint prediction is the technical idea to hold onto. Robot control involves a changing scene, and the movements themselves help change that scene. Predicting both gives developers a foundation that connects those two aspects of the task, which they can then adapt for their own hardware and use case.

The release includes a base checkpoint and robot-specific checkpoints. A checkpoint is a saved set of learned model parameters. Having those starting points, together with fine-tuning code, means developers can build on the video foundation already trained into the system. They can spend their effort adapting that foundation to the task they want their robot to perform.

Runway's technical report identifies two limits that are especially relevant to that work: confusion between similar-looking objects, and difficulty chaining multiple steps from a single instruction. Those are concrete clues about where integration and evaluation effort belongs.

Object confusion affects which thing the robot acts on. Task sequencing affects whether it can carry out a longer instruction across several actions. A developer considering Praxis-1 can use those distinctions to think about the shape of a suitable initial task and the additional work a more complicated one would require.

The useful opening here is access to a reusable foundation, with the code needed to adapt it. In my view, that can make experimentation more accessible to robotics developers who want to investigate video-based control without first recreating the underlying training work. Hardware adaptation and reliable task sequencing remain substantial parts of the engineering, and now more of the starting material is available to work with.

Atlas makes generated scenes editable

There's a different use for visual AI when you want to work inside the scene it creates.

World Labs released its Atlas creation tools and developer access on 12 November. The tools connect image-based scene generation to editing, camera movement and export, giving creative and simulation developers environments they can continue working with in a 3D workflow.

Viewfinder builds explorable scenes from images. Chisel provides control over scene structure. Splats to Mesh converts the scene representation into geometry that conventional 3D tools can use. Together, those capabilities take the output through several practical stages, from an image-conditioned scene to something a developer or artist can edit and export.

The export step is particularly useful. A generated result becomes easier to incorporate into existing work when it can move into the tools people already use. Mesh geometry gives the scene a form those tools understand, so the initial generation can become a starting point for further construction and changes.

Camera movement also changes how you engage with the result. An explorable environment lets you consider the scene from different positions. For visual prototyping, that could help someone investigate an idea spatially, then adjust its structure before taking it further.

There's one essential property of image-based reconstruction to understand here. World Labs says Atlas invents plausible content where the input images don't show the scene. Additional views constrain that reconstruction by supplying more visual information. An area hidden in the source images therefore contains generated detail, rather than an observation of what was there.

That distinction helps define the useful applications. A creative prototype can benefit from the system filling in an environment around the visible material. A factual reconstruction of a real place needs evidence for the details it represents. Providing more views helps constrain the result, but measurement accuracy hasn't been established by this release.

For creative developers, the opportunity I see is a shorter path from visual reference to an environment they can reshape. The combination of structural editing and export makes that path more interesting than generation alone. It leaves room for the artist or developer to make decisions inside the resulting scene, using the rest of their existing toolchain.

Secrets across AWS accounts

A smaller cloud change could simplify an awkward part of recovery planning.

AWS Secrets Manager added cross-account replication on 13 November. It can now maintain a copy of a secret in another AWS account, with separate ownership and encryption controls. The replica uses a customer-managed encryption key belonging to the receiving account, and it can be promoted to a standalone secret.

For cloud operators, that creates a managed way to keep credentials available under separate account control. AWS handles synchronising the values, reducing the custom replication work involved in maintaining that copy. Promotion lets the replica become independent when that's required.

The boundaries are specific: one cross-account replica per primary secret, in the same AWS Region. Updates are asynchronous and eventually consistent. In plain language, a changed value takes time to reach the copy.

That timing becomes relevant during credential rotation or failover. An operator planning to use the replica needs to account for the possibility that the most recent update hasn't arrived yet. Same-Region replication also serves account separation, rather than providing a copy in a different region.

I see this as useful account-level recovery plumbing. It supplies a separately controlled copy of the current credential, with managed synchronisation. Because changes flow through to that copy, recovering an older value still calls for a separate historical-backup approach.

Funding quantum control hardware

One funding announcement is worth a brief stop for the engineering work it supports.

On 13 November, the US Department of Commerce announced a final CHIPS research and development award of up to one hundred million US dollars for Atom Computing, which works on neutral-atom quantum computers.

The award targets semiconductors, control hardware and systems integration needed to manipulate and validate tens of thousands of qubits, the basic units of quantum information. That's a substantial development objective, with attention on the machinery needed to control and check a much larger system.

The signed award gives Atom support for further engineering work. Its value is a funding ceiling; the notice doesn't specify disbursement timing or a commercial-delivery date.

For organisations watching quantum computing, the interesting signal is where this investment goes. Increasing the number of qubits also creates work in the electronics and system integration required to operate them. Control and validation are part of making that scale useful.

My read is that this award puts resources behind a demanding systems-engineering problem. The milestone to watch is progress in manipulating and validating those larger numbers of qubits, which is what the funded work is intended to enable.

What Changes for You

If your work involves recorded speech, there's a new option for controlling how it becomes text.

Salesforce released SOV-Transcribe-1 on 12 November, with model weights under Apache 2.0. It's a speech-recognition model that builders can host themselves, and its transcription behaviour can be directed with instructions.

You can request verbatim text or a cleaned-up transcript, control formatting, and supply names or specialist vocabulary as recognition context. That gives developers a way to specify the kind of transcript their application needs, as well as the terms the model is likely to encounter.

Those choices have practical consequences. A verbatim transcript preserves the spoken wording, while a cleaned-up result serves a different reading need. Recognition context lets a builder provide vocabulary relevant to the recording instead of leaving the model to infer every unfamiliar name. The controls make transcription more adaptable to a particular workflow.

The model supports seven languages, including English. It processes recordings rather than producing a live, incremental transcript, and inputs of up to thirty minutes have been tested. So its immediate fit is work on recorded audio.

Operating it requires suitable compute and integration skills. The complete model has 26.6 billion parameters, including the audio encoder. Native framework support is still being upstreamed, and the released serving plugin targets vLLM 0.32.0. That's a specific compatibility requirement for developers considering deployment.

Self-hosting lets an organisation keep audio on infrastructure it controls. In exchange, it takes on the cost and work of running the model, and establishing how well it handles its own accents, vocabulary and recordings.

For builders working with sensitive recordings or specialised language, that trade-off is newly worth considering. They can now choose a transcription system with control over both where the audio is processed and how the resulting text is shaped.

You'll find the sources and full transcript at owenonthenet.com. Thanks for listening.

Sources

Reporting behind this episode.

  1. openai.com/index/introducing-chatgpt-wallet
  2. investing.com/news/stock-market-news/openai-launches-chatgpt-wallet-letting-ai-agents-pay-for-purchases-5026114
  3. openai.com/policies/chatgpt-wallet-terms
  4. cncf.io/blog/2026/11/10/welcoming-openshell-to-the-cncf-open-governance-to-secure-agentic-ai
  5. huggingface.co/blog/ggml-org/llama-cpp-v1
  6. github.com/ggml-org/llama.cpp/releases/tag/v1.0.0
  7. runway.com/news/company-news/praxis-1-is-now-available
  8. runway.com/research/praxis-1-video-pretraining-to-robot-control
  9. worldlabs.ai
  10. worldlabs.ai/atlas
  11. worldlabs.ai/blog/atlas
  12. aws.amazon.com/about-aws/whats-new/2026/11/secrets-manager-cross-account
  13. docs.aws.amazon.com/secretsmanager/latest/userguide/cross-account-replicate-secrets.html
  14. nist.gov/news-events/news/2026/11/department-commerce-announces-finalization-chips-rd-award-atom-computing
  15. huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1
  16. huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1