Back to the show

AI & Tech Daily

DeepMind Opens Its Weather Data, and Agent Infrastructure Gets More Useful

17:37

DeepMind releases the satellite imagery and rainfall estimates behind WeatherNext 3, giving researchers a reusable foundation for weather and water applications. AWS brings snapshot-backed AgentCore sessions to Sydney, ElevenLabs adds contextual turn-taking to Scribe, and Cursor explains the credential chain behind its October extension attack. Also: Australian evidence on AI and employment, A2UI's first stable release, Nvidia's leadership transition, and what Cloudflare's open-source Kitesurf browser offers agent builders.

Full transcript

Read the episode.

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

Weather Data Opens Up

Rainfall estimates every fifteen minutes sound ideal for watching a storm. When they arrive two to eight hours later, their usefulness shifts towards a different set of decisions.

Our main story today is DeepMind opening the data behind WeatherNext 3, and where that gives weather and water researchers a useful head start. On 13 November, Google DeepMind released historical archives and operational feeds for two datasets: SatWeave satellite imagery and PARDIG rainfall estimates.

The opportunity is in the preparation. Someone building a weather application can begin with observations that have already been brought together, instead of first assembling the underlying satellite feeds. That moves some of the effort from collecting inputs to investigating what those inputs can tell you.

SatWeave combines imagery into seventeen spectral bands on a grid of roughly ten kilometres, with operational updates every hour. Spectral bands capture different parts of the electromagnetic spectrum. Having several bands gives a model different views of the same area, rather than relying on a single picture.

That ten-kilometre grid also tells you something about scale. A grid cell represents an area, so a researcher interested in a particular property or a small waterway has to consider how that area relates to the place they're studying. Coverage has gaps too: polar areas and some regional spectral coverage are missing. Those details help determine where the prepared dataset fits the job.

PARDIG tackles a different problem. It uses satellite observations and atmospheric data to infer rainfall. That gives researchers estimates they can work with across places and times, including when they're exploring how rainfall relates to water management. The rainfall figures are estimates derived from those inputs.

Its timing needs a little unpacking. There are estimates for fifteen-minute intervals, but they're published in six-hour batches. By the time a batch is available, the conditions it describes are roughly two to eight hours old. The interval between measurements and the wait to receive them are separate properties of the feed.

For a study of how rainfall developed through an event, that sequence can be useful. For a decision that depends on what is happening in the next few minutes, the delivery delay becomes a constraint. You can see why the same dataset could fit a research workflow and be unsuitable as the immediate input to a warning.

Both datasets are available through Google Cloud Storage under the Creative Commons Attribution 4.0 licence. PARDIG also has Earth Engine collections, although those are static historical archives at launch. The operational feed is a separate access choice. A developer comparing past events can use an archive; an application that needs continuing updates needs the operational data.

I think the strongest opportunity here is for researchers who have a worthwhile question but face a large data-preparation job before they can test it. Water management and solar forecasting are among the application areas this foundation could support. Historical observations let a developer examine past conditions, then explore whether the same approach is useful with incoming data.

For an Australian catchment, local performance still needs to be established against the conditions and decisions that matter there. Opening the inputs makes that investigation more accessible. It gives other researchers something substantial to build on, with the spatial coverage and delivery schedule visible from the outset.

Faster Starts for AgentCore

That's the weather-data opportunity. For the shorter news, let's start with the wait before an AI agent gets to work.

On 12 November, AWS announced that newly created Bedrock AgentCore runtimes would use its snapshot-based V2 platform by default in nineteen regions, including Sydney. Existing runtimes stay on their current platform version.

An agent runtime is the environment in which the agent's software runs. Before a browser or coding agent can do anything useful, that environment needs to be ready. V2 restores an initialised environment from a snapshot, so a new session can start from a prepared state.

AWS reports startup below two seconds at the seventy-fifth percentile for a two-gigabyte image, compared with 29.7 seconds previously. In that measurement, roughly three quarters of starts were below the reported threshold. The time being reduced is the wait for the session to start; the model still takes its own time to work through the task.

That's a meaningful distinction for the experience of using an agent. If you've asked it to perform a short piece of work, a lengthy startup can dominate the wait. Cutting that overhead could make occasional tasks feel more responsive, even with the same model doing the same reasoning afterwards.

Each session continues to run in its own hardware-isolated Firecracker microVM, a lightweight virtual machine. The snapshot changes how the environment is prepared while retaining that per-session boundary.

There is one deployment detail worth catching. The launch announcement describes V2 as the new default, while the platform documentation still says leaving out the platform version selects V1. Explicitly selecting V2 and verifying the result resolves that ambiguity. Existing applications also need an explicit move to V2; they don't migrate automatically.

For Australian builders, Sydney's inclusion means the snapshot-backed option is available in that region. My reading is that this makes intermittent browser and coding agents more attractive: less time spent waiting for an environment, without having to keep every environment running between jobs. A migration decision can focus on the startup behaviour of the application you're deploying, with the platform version selected deliberately.

Speech Recognition Learns When to Wait

A faster start is useful. In a voice conversation, knowing when to hold back can be just as important.

ElevenLabs announced Scribe v3 and Scribe v3 Live on 10 November, with an update to the announcement on 14 November. Both are available through its API: one for recorded audio, the other for streaming transcription.

The live model adds speaker labelling and contextual turn prediction. That second feature addresses a familiar conversational problem: a pause can mean you've finished, or it can mean you're finding the next word.

A system that reacts to the pause alone can start answering while you're still forming the question. Contextual turn prediction uses the conversation to help decide whether the speaker has finished. For a voice agent, that affects when it passes the conversation on for a response.

Consider someone giving a piece of information, pausing to recall the rest, then continuing. If an agent treats the first fragment as complete, it may answer an unfinished request. Giving the person room to complete the thought could prevent that downstream mistake before the answering model even gets involved.

Speaker labelling deals with another part of the same workflow: keeping track of who said what. When more than one person is speaking, the words and their speaker both contribute to understanding the exchange. Better handling of those boundaries could make the transcript more useful to whatever happens next.

ElevenLabs also reports twenty-seven per cent fewer word errors for the recorded-audio model than Scribe v2, on its FLEURS evaluation across 102 languages. That result concerns the recorded model. For live interaction, the feature that interests me most is the ability to use conversational context when deciding when to respond.

There is implementation work attached. Turn prediction is opt-in, and the official SDKs don't yet support it, multilingual mode or the expanded keyterm library. Those features currently require direct WebSocket integration, with the developer handling the streaming connection. Selecting the new model alone won't turn on contextual turn prediction.

For people building voice workflows, this is a useful place to put engineering attention. A caller who gets to finish a sentence gives the rest of the system a better chance of acting on the request they intended.

Cursor's Credential Chain

There's also a more serious development for people using agents in their coding tools.

Cursor's forensic update on 13 November traces its October extension attack back to an exposed Fireworks access key. According to Cursor's reconstruction, that key led to historical agent traces containing unredacted credentials, and ultimately to a malicious extension in Cursor's marketplace cache.

The affected packages were tampered versions of the Anthropic Claude Code extension for macOS and Windows. Two version numbers are useful here because they identify the affected builds: 2.1.312 and 2.1.313. They were served through Cursor's VS Code extension marketplace on 29 October, between approximately 2:27 in the morning and 1:35 in the afternoon, Pacific Time.

Cursor puts the potential exposure at no more than five thousand users who may have received a malicious build. Anthropic didn't publish those packages, and Open VSX didn't host them. The distribution point was Cursor's marketplace cache.

For an affected developer, removing the extension is only part of the response. The attack may leave persistence on the computer, allowing malicious access to survive the extension's removal. Cursor's response therefore includes host remediation and credential rotation. That addresses both the machine and the credentials that could still provide access elsewhere.

Cursor says Privacy Mode prompts and code, official Cursor updates and customer code repositories were unaffected. It believes the incident is contained, and says the last observed attacker activity was blocked on 1 November.

The chain through historical traces is the part organisations can learn from. A trace records an agent's past activity. If credentials are retained inside it, someone who obtains access to that old record may acquire a way into a system that's still in use. Data kept to understand previous work can therefore carry operational access long after the original task has finished.

I'd include retained agent traces in a supplier-risk review: what gets stored, who can access it, and how credentials are removed from those records. In this incident, the route from one exposed key to a software-distribution system crossed that historical data. That makes trace handling a concrete part of the security of the development tools people use every day.

Australia's Early AI Jobs Signals

Let's step away from the tools for a moment and look at the people working alongside them.

An RBA staff research note published on 9 November finds limited evidence of broad AI-driven disruption in Australia's labour market. It also identifies tentative weaker outcomes for younger workers in occupations with greater exposure to AI.

The analysis, by Mike Major, looks at movement into and out of employment. Across occupations with different levels of AI exposure, it finds little evidence of different job-finding rates or rates of moving from employment into unemployment. Those measures help show whether people in more exposed work are having a noticeably different experience.

The younger-worker result deserves attention because an aggregate picture can hide an uneven adjustment. A labour market can look relatively steady overall while people at the start of their careers face a tougher path into particular occupations. Graduate recruitment and entry-level opportunities are useful places to watch for that kind of change.

AI exposure describes the tasks that could be affected. It doesn't measure whether an employer has adopted AI. Broader economic conditions and structural changes could also explain the weaker outcomes among younger workers. This is Major's staff analysis, rather than a statement of the RBA Board's position.

For an employer, I think the useful question is what happens to the route into a career as tasks change. If some work traditionally done by a junior employee becomes easier to automate, the organisation still has to develop people who can do more demanding work later. That makes entry-level hiring worth examining alongside investment in AI tools.

For someone entering the workforce, the local evidence gives a more specific picture than a prediction that every job is about to disappear. The broad transition measures show limited disruption so far; the tentative pressure is concentrated in younger workers in more exposed occupations.

Watching hiring at that entry point could reveal changes before they're obvious in the overall employment figures. It also gives employers a practical way to examine their own decisions: whether the opportunities to begin, learn and progress are changing alongside the tasks.

A2UI Gets a Stable Foundation

Back at the screen, an agent's answer can be easier to use when it gives you something to interact with.

Google released A2UI 1.0 on 11 November, the first stable release of its open protocol for agent-generated interfaces. The agent describes an interactive screen, and the application renders it using its own components.

That allows an exchange to include a form or an availability picker. Instead of describing every field in a text conversation, a person can work with the controls the application presents. The application supplies the visible components; the protocol carries the description from the agent.

Version 1.0 adds bidirectional function calls and accessibility support. The two-way interaction gives the interface a way to participate in the exchange, beyond displaying the agent's initial response.

For developers, a stable protocol provides a firmer target to build against. Support still varies by renderer and integration: Google's announcement says Jetpack Compose and Gemini Enterprise remain on A2UI 0.9.1, with 1.0 support coming. The supported version at the receiving end determines which features an application can use.

I like the practical consequence of this approach. A form can make missing information visible, and a confirmation screen can put a proposed action in front of the person who needs to approve it. That could save repeated text exchanges while making the next action easier to inspect.

Nvidia's Leadership Transition

One brief corporate development has direct relevance to organisations planning their next infrastructure purchase.

Nvidia's board accepted Jensen Huang's immediate resignation as president, chief executive and director on 11 November, according to its regulatory filing. The public announcement followed on 12 November.

Colette Kress has been appointed interim chief executive and remains chief financial officer. Nvidia is searching for a permanent CEO among internal and external candidates.

The board says it expects operations, customer commitments and the product roadmap to continue without interruption. No product-roadmap change was announced with the transition. The permanent successor and any subsequent strategic changes remain unknown.

For an organisation buying Nvidia-based infrastructure, I'd keep the business question close to the purchase: continuity of delivery and the customer commitments attached to it. A leadership change is consequential, while the usefulness of an existing infrastructure plan still depends on the products and delivery arrangements that plan was built around. The next meaningful evidence for those buyers will be how Nvidia carries those commitments through the transition.

What Changes for You

If you're building an agent that browses the web, you now have another part of the system you can change yourself.

On 13 November, Cloudflare published Kitesurf's source code under Apache 2.0. Kitesurf is its experimental browser engine for AI agents. This release opens the source of the existing browser, allowing developers to inspect, modify and run it locally or deploy their own instance on Cloudflare Workers.

It parses HTML, applies CSS, runs page JavaScript and produces screenshots without a browser container. In practical terms, it can work through the page's structure, styling and behaviour to produce something an agent can inspect.

It also supports the Chrome DevTools Protocol, the connection used by browser-control tools such as Puppeteer. That gives developers with existing clients an integration route, rather than requiring them to replace every part of their browsing workflow.

The useful change is control over the engine itself. A builder who needs different page handling can inspect the implementation and make changes, going beyond the settings exposed by a managed browsing API. That could be particularly useful when the browser is part of an agent whose behaviour the developer is trying to understand.

The boundaries are concrete. Kitesurf remains experimental, website compatibility varies, and the documented unsupported uses include video, WebGL graphics and long-running authenticated sessions that need persistent state. Running your own instance on Cloudflare also requires a Workers Paid plan because the engine uses Dynamic Workers.

For a developer working with supported pages and shorter sessions, the source release creates room to adapt the browser to the agent's needs. Someone depending on persistent logins or those graphics features has a different requirement. The value here is being able to work on the browsing component itself when page handling is the thing holding an application back.

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

Sources

Reporting behind this episode.

  1. deepmind.google/blog/opening-the-datasets-behind-our-ai-weather-models
  2. developers.google.com/weathernext/guides/pardig
  3. developers.google.com/weathernext/guides/satweave
  4. aws.amazon.com/about-aws/whats-new/2026/11/new-agentcore-runtime-default
  5. docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-how-it-works.html
  6. elevenlabs.io/blog/scribe-v3
  7. elevenlabs.io/docs/overview/models
  8. cursor.com/blog/security-incident-update
  9. cursor.com/blog/claude-code-extension-security-notice
  10. rba.gov.au/publications/insights/australian-economy/2026/monitoring-early-effects-of-ai-on-the-labour-market.html
  11. developers.googleblog.com/en/a2ui-10-build-delightful-fast-reliable-agentic-experiences
  12. nvidianews.nvidia.com/news/nvidia-announces-ceo-transition
  13. d18rn0p25nwr6d.cloudfront.net/CIK-0001045810/2bc9480b-60a6-4d8a-90cb-87190f804a3b.pdf
  14. blog.cloudflare.com/kitesurf-open-source
  15. developers.cloudflare.com/browser-run/kitesurf