Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

UpNext AI

UpNext AI artwork

ARTEN’s Crop-Disease Images, Moonshot’s Nvidia Chip Deal, and Smart-Grid Security

Nov 19, 20268 minAudio

The ARTEN-Enhanced Multi-Crop Disease Dataset, Moonshot AI’s canceled Nvidia Blackwell chip-access deal, Dean Ball’s essay on institutional control, and research on smart-grid attack paths lead this episode. We also cover Anthropic’s IPO pricing, Palo Alto Networks’ planned Applied Compute acquisition, OpenAI’s frontier-risk approach, Microsoft’s MAI models, and Simon Willison’s Datasette Lite rewrite. **ARTEN crop-disease images:** A Scientific Data paper publishes enhanced images derived from an existing multi-crop collection and identifies an open question about preserving disease-specific visual features. https://www.nature.com/articles/s41597-026-08834-4 **Moonshot AI and Nvidia chips:** The Information reports that Moonshot scrapped a planned chip-access deal through an Indonesian data center after a U.S. remote-access rule. https://www.theinformation.com/articles/chinas-moonshot-ai-scraps-150-million-deal-access-nvidia-chips-indonesia **Dean Ball on human authority:** Ball’s essay examines the possibility of gradual delegation of institutional decisions to AI. https://openai.com/index/the-silent-consensus **Smart-grid security research:** Researchers compare language-model-guided attack-path rankings with cascading-failure-based references. https://arxiv.org/abs/2611.24765v1 **Anthropic IPO pricing:** The Guardian reports the company’s share price for its planned market debut. https://www.theguardian.com/technology/2026/nov/18/anthropic-share-price-stock-market-ipo **Palo Alto Networks and Applied Compute:** The Information reports the planned stock acquisition of the customized-model startup. https://www.theinformation.com/briefings/palo-alto-networks-acquire-ai-firm-applied-compute-4-5-billion **OpenAI frontier risk:** The company describes safeguards from training through deployment. https://openai.com/index/managing-frontier-ai-risk **Microsoft MAI models:** The Decoder reports on models unveiled at Ignite and Microsoft’s use of in-house alternatives. https://the-decoder.com/microsoft-is-replacing-outside-ai-models-in-its-own-products-with-cheaper-in-house-alternatives/ **Datasette Lite:** Simon Willison details a browser-based rewrite for sharing demos and reproducing bugs. https://simonwillison.net/2026/Nov/17/datasette-lite/

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show.

Welcome to the UpNext AI podcast. It's Thursday, November 19th, 2026, and here's what matters in AI today.

Plant-disease AI depends on the pictures used to build and test it. In Scientific Data, researchers present the ARTEN-Enhanced Multi-Crop Disease Dataset, a processed version of an existing collection of leaf photographs. It contains 21,875 enhanced images corresponding one-for-one to 21,875 original images. They cover five crops—banana, chilli, groundnut, radish, and cauliflower—across 30 healthy and diseased categories. No new field observations were collected for this release.

The researchers applied an image-enhancement system called ARTEN to address variations in lighting, blur, and contrast that can complicate visual analysis. Each output is standardized to 640 by 640 pixels, and the images are organized into training, validation, and test subsets. The enhancement workflow and code are public, giving computer-vision teams a way to reproduce the processing and compare experiments.

There is an important distinction for anyone using these pictures to develop a disease detector. The paper reports image-level similarity measures, but the researchers say further validation is needed to establish whether the enhancement preserves disease-specific details such as lesion boundaries, texture, and discoloration. The contribution here is a reusable, documented image resource; better-looking images alone do not establish better diagnoses.

From agricultural imagery to the hardware behind AI models: The Information reports that Moonshot AI has scrapped a deal worth $150 million to access Nvidia Blackwell chips through a data center in Indonesia. Moonshot is the Beijing-based developer of the Kimi models. According to two people with direct knowledge cited by the outlet, it had contracted with a Southeast Asian cloud provider to rent a sizable cluster of Blackwell Ultra B300 chips, with access planned to start as early as next year.

The cancellation follows a U.S. Commerce Department rule announced this week restricting Chinese companies’ remote access to advanced chips. The B300 chips are barred from sale to China, while overseas data centers had offered a route to use restricted chips without bringing them into the country. This deal gives the policy a concrete commercial consequence: a planned source of advanced computing for Moonshot is now off the table.

The question of access leads to a different question: who keeps authority as AI capabilities grow? In an essay published by OpenAI, its head of strategic futures, Dean Ball, argues that companies could gradually hand more decisions to AI because each delegation makes economic sense on its own. His concern is the cumulative result: people might lose effective control of businesses and institutions without anyone explicitly choosing that outcome.

Ball calls this possible future the “silent consensus,” and rejects the idea that human authority must inevitably give way to more capable machines. He also acknowledges uncertainty about how strongly human institutions would hold up under competitive pressure. For leaders deciding where to deploy autonomous systems, the essay poses a governance question worth making explicit: which decisions should remain under meaningful human authority as the systems improve? It is Ball’s argument, adapted from a lecture, rather than an announced OpenAI policy or a forecast established by new research. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast.

For the research note, consider a power grid where a cyber intrusion could trigger physical failures. Defenders need to identify attack paths that threaten consequential parts of the network, and they need to do it quickly enough to assess changing scenarios. A paper released earlier this week tests whether a large language model can help rank those paths using network connections and structured threat descriptions. The researchers compare its output with impact rankings derived from cascading-failure simulations.

In their cyber-layer evaluation, the overlap between the model’s top three critical nodes and the simulation-based reference rose from about one-third to as high as all three, depending on the approach. In a standard test grid called the IEEE 14-bus system, the proposed framework also ran about 75 to 85 percent faster than the researchers’ fusion-based comparison method. The result suggests a useful way to prioritize security analysis in that evaluated grid setting; agreement with simulated paths is the measure here, rather than a demonstrated defense against real-world attacks.

...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ...

A few other developments before we close. The Guardian reports that Anthropic priced shares for its planned stock-market debut at $83 per share, above its earlier indicated range of $70 to $78 per share. Trading is slated to begin Thursday.

The Information reports that Palo Alto Networks plans to acquire Applied Compute for $4.5 billion in stock. Founded by three former OpenAI researchers, Applied Compute helps businesses train and run customized versions of open-source AI models.

OpenAI has outlined its approach to frontier-model risk from training through deployment, including model alignment, sandboxing, blocking safeguards, and monitoring. It is a description of the company’s safeguards, not a report of newly measured results.

The Decoder reports that Microsoft’s AI division unveiled seven models at Ignite, covering work such as coding, security, and image generation. The outlet says Microsoft is shifting some of its own products toward cheaper in-house alternatives to outside models.

And Simon Willison has rewritten Datasette Lite to run the Datasette Python application in a browser using Pyodide and service workers. Developers can share links to demos with selected versions and plug-ins, making work-in-progress features and bug reproductions easier to show.

Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers.

And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. US AI Chip Access Rules, OpenAI’s GPT-6.2 Models, and Nvidia’s Inception Talks

    Nov 20, 20268 minAudio

    The Trump administration’s proposed restrictions on China’s remote access to AI chips lead this episode, alongside OpenAI’s GPT-6.2 Astra and Sol pricing, Nvidia’s reported Inception acquisition talks, and the...

    Credit the people on this episode

  2. OpenAI’s GPT-6.2 Astra and Sol, ArisGlobal’s Bedrock Costs, and Nvidia’s Scientific Agents

    Nov 18, 20269 minAudio

    OpenAI’s GPT-6.2 Astra and Sol launch across ChatGPT Work, Codex and the API; Latent Space examines independent results and early user reactions. ArisGlobal’s NavaX on Amazon Bedrock, the 3CBench infrastructure...

    Credit the people on this episode

  3. OpenAI’s NATO Cyber Tools, Researcher Access, and the U.S. AI Chip Access Rule

    Nov 17, 20268 minAudio

    OpenAI’s Daybreak work with NATO’s Communications and Information Agency, ENISA, and CERT Polska leads today’s episode. We also cover ChatGPT for Academic Researchers, a Nature Computational Science commentary on AI in...

    Credit the people on this episode

  4. OpenAI and LVMH’s Journalism Deal, Nvidia’s AI Day Seoul, and XR Blocks

    Nov 16, 20268 minAudio

    OpenAI and LVMH’s Media Division are bringing ChatGPT Enterprise into publication workflows and linking ChatGPT users to journalism; Nvidia’s AI Day Seoul highlighted robotics and agents, while executives at Meta, Jack...

    Credit the people on this episode

  5. Amplitude’s Wave, Nvidia’s Hugging Face Deal, and Metal Sandbox

    Nov 13, 20268 minAudio

    Amplitude’s Wave connects OpenAI-powered product analysis with coding agents and impact measurement; Hugging Face CEO Clément Delangue responds to Jensen Huang’s departure amid Nvidia’s pending acquisition. Also...

    Credit the people on this episode