Canva Introduces Craft for Image Generation
Canva launched its first in-house AI image model, Canva Craft, on October 22, 2026. The model uses a two-stage architecture that plans designs before generating images, aiming for greater precision, consistency, and control. It supports up to 2K resolution and more than 30 aspect ratios. Canva trained the model on millions of human-made templates from its platform library. The system includes input and output moderation along with privacy controls. Canva Craft is available globally through the Dream Lab menu option. Canva’s first AI image model plans designs before generating them
The release positions Canva alongside other providers of generative image tools. Future plans include integration with AI agents via the Model Context Protocol and the rollout of faster variants named Craft-fast and Craft-ultrafast. Canva’s first AI image model plans designs before generating them
This launch marks an important step for the company as it develops its own generative capabilities rather than relying solely on external providers. The two-stage approach allows the system to first consider layout and composition before rendering pixels, which supports the goals of improved consistency across outputs. Training on templates created by human designers helps the model align with established design practices already popular on the platform. Moderation layers and privacy features are built in from the start to address common concerns around AI image tools. Availability through the existing Dream Lab interface means users can access the new model without changing their current workflows. Canva’s first AI image model plans designs before generating them
MIT Researchers Observe AI Agents Developing Play Behaviors
MIT researchers placed 13 AI agents, each running on Claude Opus 4.7, on a simulated desert island for 30 hours with no assigned tasks. The agents, described as naturally curious in their master prompt, began inventing activities such as stacking blocks to reach a spire, creating memorials, and using objects in novel ways. Six agents discovered they could throw blocks farther by releasing them mid-jump. Left alone with nothing to do, AI agents started playing
After the free-exploration period, the agents performed better on subsequent goal-directed tasks than identical agents without the island experience. Researchers suggest the approach could complement task-specific training for developing agent capabilities, particularly for future robotic control. Left alone with nothing to do, AI agents started playing
The experiment highlights how unstructured time can lead to emergent behaviors that later transfer to structured problems. Agents explored physical interactions and creative uses of available objects without external direction. The improvement on later tasks indicates that open-ended exploration may build useful priors or skills. Such findings could influence how future agent systems are pretrained before deployment in real-world settings. Left alone with nothing to do, AI agents started playing
Nvidia Supports Expanded Compute Deployments
Kazakhstan is scaling its AI infrastructure with planned NVIDIA Vera Rubin deployments. Firebird secured a U.S. export license for 45,000 NVIDIA Grace Blackwell GPUs for a facility at Data Center Valley, with commercial launch targeted for the third quarter of 2027. Kazakhtelecom has deployed 400 NVIDIA H200 and 480 NVIDIA B300 GPUs in Astana and Almaty, with additional Vera Rubin GPUs planned for the Ekibastuz region. Kazakhstan Expands Compute for Its AI Future
These moves coincide with Nvidia CEO Jensen Huang’s visit to Kazakhstan. The infrastructure supports local AI-native companies and research initiatives. Kazakhstan Expands Compute for Its AI Future
The deployments represent a significant expansion of available compute resources in the region. Existing GPU clusters in major cities provide a foundation for immediate research and commercial use while larger future installations come online. Coordination with international hardware providers and local telecommunications operators helps ensure the infrastructure can support both training and inference workloads. Kazakhstan Expands Compute for Its AI Future
Ride-Hailing Partnerships Test Robotaxi Systems
inDrive signed a memorandum of understanding with Pony.ai and Nvidia to explore robotaxi operations. Testing of Pony.ai Gen-7 vehicles is scheduled to begin in Astana, Kazakhstan, in the first half of 2027, initially with safety operators. Nvidia will supply its Hyperion platform and Halos safety system. Uber rival inDrive eyes 2027 robotaxi tests in Central Asia with Pony.ai and Nvidia
The effort aligns with Kazakhstan’s plans to launch autonomous taxi services in 2027. inDrive also maintains a separate agreement with Saudi Arabia’s AI Driver to explore WeRide-powered vehicles in Riyadh. Uber rival inDrive eyes 2027 robotaxi tests in Central Asia with Pony.ai and Nvidia
These partnerships combine local ride-hailing networks with established autonomous vehicle technology providers. Initial testing with safety operators allows for gradual validation of the systems in real traffic conditions. The use of standardized platforms for perception and safety supports consistent performance across different operating environments. Uber rival inDrive eyes 2027 robotaxi tests in Central Asia with Pony.ai and Nvidia
Enterprise Focus Shifts Toward Serving Infrastructure
Discussions at recent conferences emphasize that competitive advantage in enterprise AI increasingly depends on the systems built around models rather than model selection alone. Data pipelines, engineering practices, platforms, operating models, and governance must align for scalable deployment, especially in regulated environments. Designing an AI serving harness built to last The Foundations of Scalable AI
Organizations are therefore investing more in the surrounding tooling and processes that turn raw model outputs into reliable business capabilities. Alignment across technical and organizational layers helps address requirements around auditability, security, and compliance that are common in enterprise settings. Designing an AI serving harness built to last The Foundations of Scalable AI
Data Usage Terms Draw Continued Scrutiny
Meta updated its business terms effective October 30, 2026, removing explicit references to training AI models on content from business accounts. Legal experts note that remaining language around developing business products may still permit broad use of scraped data, and robots.txt signals remain the primary mechanism for content owners to express preferences. Meta’s AI could still scrape news sites, legal experts say
These developments illustrate parallel tracks in frontier AI: new model architectures for image and agent tasks, expanded compute foundations, and the operational and policy frameworks required to sustain deployment at scale. Continued attention to data governance reflects the broader need to balance innovation with responsible practices across the industry. Meta’s AI could still scrape news sites, legal experts say
Evidence Used
- Canva’s first AI image model plans designs before generating them Canva’s first AI image model plans designs before generating them — Canva’s first AI image model plans designs before generating them
- Left alone with nothing to do, AI agents started playing Left alone with nothing to do, AI agents started playing — Left alone with nothing to do, AI agents started playing
- Kazakhstan Expands Compute for Its AI Future Kazakhstan Expands Compute for Its AI Future — Kazakhstan Expands Compute for Its AI Future
- Uber rival inDrive eyes 2027 robotaxi tests in Central Asia with Pony.ai and Nvidia Uber rival inDrive eyes 2027 robotaxi tests in Central Asia with Pony.ai and Nvidia — Uber rival inDrive eyes 2027 robotaxi tests in Central Asia with Pony.ai and Nvidia
- Designing an AI serving harness built to last Designing an AI serving harness built to last — Designing an AI serving harness built to last
- The Foundations of Scalable AI The Foundations of Scalable AI — The Foundations of Scalable AI
- Meta’s AI could still scrape news sites, legal experts say Meta’s AI could still scrape news sites, legal experts say — Meta’s AI could still scrape news sites, legal experts say