Canva launches Craft as Block introduces speech-only Dessa
Canva brings its own image model to Dream Lab and an API, while Block demonstrates Dessa, trained entirely on human speech. Independent testing of Kimi K3.1 Pro adds agentic results and task-level costs.
GitHub updates PR exploration, enterprise apps and exposure alerts; AWS and Bayer describe multi-turn training, operational assistants and deployment workflows. Other developments concern cross-service advertising data, publisher objections to scraping terms and the timetable for new compute capacity.
01
Models and creative tools
3 stories
2026-10-23Canva
Canva launches Craft, an image model that plans before it renders
Canva has launched Craft, its first in-house image generation model, through Dream Lab for Canva users worldwide and a public API. It plans element placement, scale, visual hierarchy and text before rendering an image at up to 2K resolution. Canva says it supports more than 30 aspect ratios and text in 55 languages.
Its training data includes millions of human-made design templates, spanning posters, presentations and infographics. Magic Layers can turn generated images into editable elements. Canva MCP access and the faster Craft-fast and Craft-ultrafast variants remain on the roadmap.
2026-10-23Block
Block introduces Dessa, trained on speech alone for full-duplex conversation
Block has introduced Dessa, a conversational model it says was trained from scratch entirely on human speech, without text or pretrained models. It takes audio in and produces audio out, listening and speaking simultaneously. The demonstrations include interruptions, backchannel responses, pauses and changes in intonation.
The team has published live conversation recordings and says inference can run locally on a MacBook Pro with an M4 Max or better. Early access is limited to a small group of AI safety researchers. The expressive conversations are demonstrations from the team; this is not yet a generally available consumer product.
2026-10-23Artificial Analysis
Artificial Analysis measures Kimi K3.1 Pro at 55, with stronger terminal performance
Artificial Analysis reports an Intelligence Index score of 55 for Kimi K3.1 Pro, up 11 points from Kimi K3’s 44 and fourth in that evaluation’s ranking. Terminal-Bench 4.0 rises from 13% to 67%, while AutomationBench-AA, which tests agentic SaaS workflows, improves from 58% to 73%.
Lower token prices do not translate directly into equally large task savings. The evaluation records about 55,000 output tokens per Intelligence Index task, 14% more than the predecessor. With first-party token prices 33% lower, cost per task remains roughly similar at $1.98. The model’s weights had not been published when the evaluation appeared.
02
Developer tools and security administration
3 stories
2026-10-22GitHub
Visual Studio adds PR exploration and moves custom models to the new Agent mode
GitHub’s September roundup for Visual Studio 2026 adds pull-request exploration to the Git agent. Attach a PR to Copilot Chat to receive a summary with links to comments and files; this requires the GitHub or Azure DevOps MCP servers. NuGet Audit warnings also gain a fixer action that asks Copilot to address vulnerable packages.
Bring Your Own Model remains in preview, supports Microsoft Foundry, OpenAI, Anthropic and Ollama, and is enabled by default in all three Visual Studio editions. It now requires the new Agent (Preview); the previous Ask and Agent modes no longer support it. Existing Ollama models must be added again after updating. NuGet fixes separately require signing in to Copilot and enabling the NuGet MCP server.
2026-10-23GitHub
Enterprise installations for GitHub Apps become generally available
GitHub has made enterprise-level GitHub App installations generally available, expanding API coverage for governance, security and administration. More workflows that relied on classic personal access tokens can move to GitHub Apps, with permissions selected for the automation being built.
The added coverage includes repository policies and rulesets, reading bypass requests, two-factor authentication requirements, domain verification, IP allowlists and access restrictions. Some organization-level gaps are also closed. The updates are planned for GitHub Enterprise Server 3.24; existing apps still need the relevant permissions configured to use these APIs.
2026-10-23GitHub
GitHub public monitoring adds credential checks and incident-response integrations
GitHub’s public monitoring now supports closing and reopening alerts, activity timelines and filtering, alongside ongoing validity checks for supported credentials. Security teams can distinguish Active, Inactive and Unknown states, and connect findings to incident-response workflows through a REST API and enterprise webhooks.
The feature remains in public preview for GitHub Enterprise Cloud, including its data-residency offering. It is included with GitHub Advanced Security or Secret Protection at no extra cost. Closing an alert changes its state for the enterprise; it does not revoke the credential or remove it from the public repository.
03
Enterprise agents and engineering practice
4 stories
2026-10-23AWS
AWS demonstrates outcome-based multi-turn training with a Wordle agent
AWS has published a SageMaker AI tutorial on multi-turn reinforcement learning, using the outcome of an entire task rather than labels for each intermediate step. In its Wordle example, gpt-oss-20b’s solve rate on held-out words rises from 61% to 84% at default reasoning effort, and from 15% to 69% at low effort.
The agent, six-guess environment and reward function run on AgentCore, returning one score at the end of each game. AWS reports approximately $339 in training-token charges, plus about $79 for the AgentCore runtime. This is a Wordle teaching experiment, with code and the complete training and evaluation workflow available.
2026-10-23AWS / Bayer Crop Science
Bayer uses 22 specialist agents to explain its crop-science analytics workflows
Bayer Crop Science describes its Agentic Decision Engine, a natural-language layer over an existing machine-learning analytics platform. A supervisor routes requests to 22 specialist agents running in production. The AWS and Bayer case study reports that roughly 80% of user questions in the first phase are resolved without an engineer.
For example, a breeder can ask why yesterday’s phenotyping workflow failed. A troubleshooting agent reads operational logs and platform status, then returns a diagnosis; the case study puts this process at under ten minutes, down from over an hour. Tools are scoped to each specialist, while AgentCore Gateway handles access and authorization for backend APIs.
2026-10-23AWS
AWS shows how to restore autoscaling safely during SageMaker endpoint updates
AWS has shared a reusable CDK deployment pattern for conflicts between SageMaker real-time endpoint updates and registered autoscaling targets. Changes such as switching instance types require deregistration first. Manual deregistration can cause deployment failures or leave an apparently successful deployment without autoscaling restored.
Three CloudFormation custom resources enforce the sequence: deregister the target, update the endpoint, wait for it to return to service, register it again and attach the scaling policy. A configuration hash triggers orchestration when parameters change. Asynchronous readiness checks avoid placing the entire waiting period inside one function invocation.
2026-10-23Sean Goedecke
Sean Goedecke argues for profiling and redesign before AI-assisted rewrites
Engineer Sean Goedecke examines the enthusiasm for using AI to rewrite software in C, Rust or assembly. He argues that many business applications spend substantial time waiting on databases, networks or disks, which a language change cannot accelerate. Many Python numerical workloads already delegate computation to optimized lower-level libraries.
His examples of useful optimizations include caching slow calls, moving work into background jobs, parallel execution, combining repeated database queries and fetching only the data needed for the current page. He advocates profiling and reducing unnecessary work first, reserving lower-level rewrites for genuinely compute-bound hotspots that cannot be simplified.
04
Data policies and compute infrastructure
3 stories
2026-10-23The Verge / Google
Google adds a cross-service exception to its personalized-ad policy
The Verge reports that Google’s October 16 privacy-policy update adds an exception to its promise not to use Gmail, Drive or Google Photos content for personalized ads. If users choose to make that data available to another Google service that supports personalized advertising, it can enter that service’s personalization process.
Google’s current policy specifically says Search may process data users allow it to access to personalize content, including ads. It still excludes sensitive targeting categories such as health and religion, and says identifying details such as names or email addresses are not shared with advertisers unless users request it. Cross-service data access depends on the relevant settings.
2026-10-23The Verge
Meta postpones new business terms until November 20 amid publisher objections
Meta has confirmed to The Verge that its new business terms will take effect on November 20, instead of October 30. Publishers objected that the wording could permit scraping their websites for AI training. Meta says the delay provides more time to explain the changes to partners.
The report notes that Meta had already removed some language about training its products, but objections remained. The US News/Media Alliance has called for revisions, while groups in Europe and the UK have appealed to regulators. Meta says it will clarify crawler use and how content owners can express crawling preferences through robots.txt.
2026-10-22NVIDIA
Firebird targets a Q3 2027 commercial launch for its Kazakhstan AI factory
NVIDIA outlines new AI-compute plans in Kazakhstan: Firebird has secured a US export license for 45,000 NVIDIA Grace Blackwell GPUs and is building its first phase with Kazakhtelecom at Data Center Valley, targeting a commercial launch in Q3 2027. A planned second phase would bring potential total capacity to approximately 500 megawatts.
The license and construction targets are distinct from installed equipment. The article lists Kazakhtelecom’s existing deployments of 400 H200 and 480 B300 GPUs in Astana and Almaty. The operator also plans more than 1,000 additional GPUs at Data Center Valley next year and over 1,000 Vera Rubin GPUs in Ekibastuz.