11 topics covered

Listen to today's briefing
0:00--:--

OlmoEarth v1.3 Satellite Embeddings

What happened: Ai2 released OlmoEarth v1.3, a satellite embedding model, along with precomputed global embeddings and a browser-based Embeddings Explorer.

Key details:

  • The model reads a year of Sentinel-1 radar and Sentinel-2 and Landsat multispectral time series and outputs a 128-dimensional vector per 10-meter pixel; the first 64 dimensions can be used alone
  • Ai2 reports its 128-dimensional embeddings score highest on 20 of 27 evaluations, compared with AlphaEarth Foundations and Tessera v2
  • Introduces PLANTEUR, a crop-type benchmark for tropical farming
  • v1.3 Base took 53% of the compute time of v1.2 and 18% of v1
  • Pretraining OlmoEarth v1.3 Base would cost roughly $2,720 of rented H100 time for the final run
  • Model weights, pretraining data, and code are open under the OlmoEarth Artifact License

Why it matters: Precomputed global embeddings let users run mapping tasks with small classifiers instead of GPU-based fine-tuning.

Practical takeaway: Users can explore the embeddings with the Embeddings Explorer without an account, or pull the 64- or 128-dimensional embeddings from Source Cooperative.

OpenAI Funding Round Valuation

What happened: OpenAI is closing in on a new funding round of up to $30 billion, but has lowered its valuation target to get there, according to The Decoder.

Key details:

  • CNBC reported, citing people familiar with the deal, that OpenAI is nearing the round at a $1.25T valuation, led by UAE investors (as relayed by The Rundown AI and Latent Space)

Why it matters: The reported valuation target is lower than the previously discussed level, indicating how the company is negotiating new capital.

Practical takeaway: Watch for official confirmation of the final valuation and investor list, since the figures so far come from reports citing people familiar with the deal.

Canva Craft Image Model

What happened: Canva launched Canva Craft, its first in-house AI image generation model, which it positions against models from OpenAI, Microsoft, and Google.

Key details:

  • Uses a two-stage architecture that plans a design before generating the image, which Canva says allows "greater precision, consistency and control"
  • Generates images at up to 2K resolution and in more than 30 aspect ratios
  • Canva says it trained Craft on "millions of human-made templates" from its content library
  • Ranked fourth on Design Arena's leaderboard, behind OpenAI's GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare, and GPT-Image-2
  • Available now to all users globally via the Dream Lab option in Canva's left-hand menu; Canva plans to let AI agents connect through the Model Context Protocol and will roll out Craft-fast and Craft-ultrafast models later

Why it matters: Canva is moving from using third-party models to running its own image model, adding a design-focused competitor to general-purpose image generators.

Practical takeaway: Canva users can try Craft now through Dream Lab; developers can watch for the planned MCP access for agents.

DeepMind Global Forest Typology Map

What happened: Google DeepMind and Google Research, with the World Resources Institute and IIASA, released Forest Typology (ForTy) 2020, a global map classifying every 10-by-10-meter patch of land by forest and land-use type.

Key details:

  • The map sorts land into six categories based on FAO definitions: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land
  • An early time series covers each year from 2017 to 2025, showing where natural forests were replaced by plantations or farms
  • The model was trained on about 1.7 million labeled locations and uses AlphaEarth Foundations embeddings as a key input
  • Validated against more than 9,000 randomly chosen sites: correct forest type 90% of the time, and forest versus non-forest 95% of the time
  • The map and time series are freely available on Google Earth Engine under an open license

Why it matters: The map distinguishes what kind of forest replaced what, which tree-cover maps miss, and it provides a baseline for EU deforestation regulation compliance.

Practical takeaway: Companies and researchers assessing EUDR exposure can access the map on Google Earth Engine and use the per-class probability scores to set confidence thresholds.

Google Agent Compute Engine

What happened: Google introduced Agent Compute Engine, a runtime environment built specifically for AI agents, alongside expanded governance tools for managing agents.

Key details:

  • Gives every Gemini agent its own hardware-isolated micro virtual machine, per The Decoder's headline
  • Governance capabilities are designed to let companies manage agents across cloud, hybrid, and on-premises environments

Why it matters: It is a dedicated execution layer for agents that addresses isolation and oversight in enterprise deployments.

Practical takeaway: Teams deploying Gemini agents can look into Agent Compute Engine's isolation and governance options; the available source gives few technical specifics.

Gartner Survey on AI Business Returns

What happened: A Gartner survey of 2,002 CIOs found that few companies are getting real business value from their AI investments.

Key details:

  • Only 13 percent of companies are getting real business value from AI investments
  • 51 percent expect AI to increase their operating costs
  • 75 percent of CEOs plan to keep raising their technology budgets

Why it matters: Spending plans are continuing to rise even though measurable returns are limited, according to the survey.

Practical takeaway: Leaders building AI budgets may want to track measurable business outcomes alongside spending.

Microsoft Diagnostic Orchestrator Medical AI

What happened: Microsoft published results in the journal Science for its Diagnostic Orchestrator, a multi-agent AI system that it says diagnoses conditions more accurately than human doctors on tested cases.

Key details:

  • On cases from hundreds of New England Journal of Medicine reports, the system reached about 83% diagnostic accuracy, versus roughly 36% for generalist physicians using references and 22% without them
  • On emergency department cases from Beth Israel Deaconess Medical Center, it scored 80% accuracy versus an average of 47% for physicians
  • The system uses multiple AI agents that decide what tests to order and arrive at a diagnosis; it is powered by OpenAI's GPT-5 according to the researchers
  • Microsoft built the SDBench benchmark from the cases and plans to open-source it
  • Microsoft plans to test the system in hospital settings with Beth Israel Deaconess; Dr. Dominic King said this will likely "get underway in the next couple of months"
  • Suleyman said the system "needs clinical validation"; no commercialization timeline was given

Why it matters: The results are from retrospective case tests rather than live clinical use, and the paper does not include newer models such as GPT-6 Astra or Gemini 4 Argon, so real-world performance remains unproven.

Practical takeaway: Watch for the planned Beth Israel clinical deployment and the SDBench release to assess the benchmark independently.

Symantec Self-Propagating Offensive AI Agent

What happened: Symantec security researchers found an offensive AI agent that hijacks unprotected language models on the web and uses them to carry out attacks on its own.

Key details:

  • The agent hijacks unprotected language models and uses them to attack servers, steal databases, and install cryptominers
  • It runs up to 100,000 steps without human input
  • The owner of the hijacked model is left paying for the usage

Why it matters: The agent shifts the cost and exposure of attacks onto operators of unsecured model endpoints, making exposed AI services part of the attack infrastructure.

Practical takeaway: Operators of publicly reachable language model endpoints should check that access controls and usage limits are in place.

Nintendo Switch 2 AI-Assisted Jailbreak

What happened: A modder known as "pepsipu" (Sammy Hajhamid) released a jailbreak of the Nintendo Switch 2 that OpenAI's Codex helped create, called "ns2jb."

Key details:

  • The jailbreak is software-only and tethered, so it must be launched from a game each time the console is turned on
  • It requires the game SmileBASIC 4, though another modder claims the title is "not strictly needed"
  • The modder demonstrated a custom Mario Kart World with characters including Cinnamoroll, the red Angry Bird, Peter Griffin, and Steve from Minecraft

Why it matters: The Verge reports that the jailbreak remains complex for most users and its full impact is unclear, but it may be the start of more Switch 2 exploits if modders use AI tools.

Practical takeaway: Users should be cautious about running unofficial jailbreak tools, and should expect further exploit reports for the console.

Recursive Flint Model

What happened: AI startup Recursive unveiled Flint, its first language model, built by former DeepMind researchers using the company's in-house system.

Key details:

  • Flint is a 40-billion-parameter mixture-of-experts model for local coding
  • It was created in four months using Recursive's in-house AI system, "Eureka Machine," which independently tests and optimizes architectures
  • Latent Space reports Flint ran 700+ NanoChat training experiments and ended close to the team's earlier Claude Haiku 4.5 results; the MLX build is listed at 2.5B active parameters
  • Rapid-MLX measured 131 tok/s decode at 1K context on an M5 Max at 4-bit, using about 24 GB
  • According to The Decoder, Recursive has yet to prove whether this approach will lead to self-improving AI

Why it matters: Flint is presented as a first output of an AI-driven research process, but the self-improvement claim is unverified.

Practical takeaway: Developers interested in local coding models can look for Flint's weights and benchmarks to test the claimed performance.

Meta Muse Shared Artifact Links

What happened: Meta has re-enabled sharing for Muse after security firm Claroc reported that shared links were publicly discoverable, saying the underlying issue is fixed.

Key details:

  • Meta's David Singleton wrote on X that "The underlying issue is now fixed, and sharing is back on"
  • Singleton said the problem was with URLs for shared Muse artifacts (such as documents and images), not with chat sharing
  • Researchers said they found Muse pages such as a marketing company's internal revenue review and a small business owner's loan application
  • Shared artifacts created before the fix now have new URLs; visitors to old links see a note asking them to get the new link from the sharer

Why it matters: The fix changes the scope of the earlier pause, which had been described as affecting shared conversations, and existing shared artifact links are now invalid.

Practical takeaway: If you shared Muse artifacts before the fix, you will need to resend the new link to recipients.