Microsoft plans multi-gigawatt Azure Maia deployments starting in 2027
Microsoft says Azure Maia will become its primary AI inference system, with Maia 200 already serving production workloads and Maia 300 deployments planned for 2027.
• Microsoft plans multi-gigawatt Azure Maia deployments from 2027 as its primary AI inference system.
• Maia 200 now serves Microsoft Copilot; GitHub Copilot and Microsoft Foundry are due by the end of 2026.
• Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems when it deploys in 2027.
Read more
Microsoft says Azure Maia will become its primary AI inference system, with Maia 200 already serving production workloads and Maia 300 deployments planned for 2027.
• Microsoft plans multi-gigawatt Azure Maia deployments from 2027 as its primary AI inference system.
• Maia 200 now serves Microsoft Copilot; GitHub Copilot and Microsoft Foundry are due by the end of 2026.
• Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems when it deploys in 2027.
Read more
AMD unveils Helion rack design for AI agents
Helion gives AMD partners a rack design for agent execution and, in the enterprise configuration, GPU inference in the same rack.
• AMD unveiled Helion, an OCP-compliant rack design for running AI agents.
• It supports CPU-only or mixed CPU/GPU configurations, with air or liquid cooling.
• Partner-built Helion configurations are scheduled for November 2026.
Read more
Helion gives AMD partners a rack design for agent execution and, in the enterprise configuration, GPU inference in the same rack.
• AMD unveiled Helion, an OCP-compliant rack design for running AI agents.
• It supports CPU-only or mixed CPU/GPU configurations, with air or liquid cooling.
• Partner-built Helion configurations are scheduled for November 2026.
Read more
Clearing chat history helps AI research agents find better solutions
In these tests, resetting chat context helped research agents explore more solutions and achieve higher, more consistent scores on the same compute budget.
• In a Microsoft Research study, Fork-and-Flush scored 0.78 across 13 tasks, versus 0.47 for single runs on the same compute budget.
• The method starts parallel branches with fresh chat contexts while keeping their workspace files and results.
• Across three Qwen3.5 model sizes, run-to-run variation peaked on MAX-3-SAT tasks near each model’s ability limit.
Read more
In these tests, resetting chat context helped research agents explore more solutions and achieve higher, more consistent scores on the same compute budget.
• In a Microsoft Research study, Fork-and-Flush scored 0.78 across 13 tasks, versus 0.47 for single runs on the same compute budget.
• The method starts parallel branches with fresh chat contexts while keeping their workspace files and results.
• Across three Qwen3.5 model sizes, run-to-run variation peaked on MAX-3-SAT tasks near each model’s ability limit.
Read more
Meta says Muse Wonder ranks near leading AI models
Meta’s reported benchmark results suggest its models are more competitive with leading AI systems, though benchmarks may not reflect real-world performance.
• Meta says its largest Muse Wonder model ranks below Claude Opus 5.5 and above Claude Fable 5.1 on an AI index.
• Meta says the models add Active Memory and Active Focus to help agents stay on task.
• Meta says it plans to release larger open-weight models in the future.
Read more
Meta’s reported benchmark results suggest its models are more competitive with leading AI systems, though benchmarks may not reflect real-world performance.
• Meta says its largest Muse Wonder model ranks below Claude Opus 5.5 and above Claude Fable 5.1 on an AI index.
• Meta says the models add Active Memory and Active Focus to help agents stay on task.
• Meta says it plans to release larger open-weight models in the future.
Read more
Amazon introduces priming to build hybrid AI models from pretrained Transformers
According to Amazon, priming lets developers test hybrid model designs using less than 0.5% of a source Transformer’s pretraining tokens.
• Amazon says priming builds hybrid AI models from pretrained Transformers at over 99% lower pretraining cost.
• In AWS tests, a 32-billion-parameter hybrid matched its source on reasoning benchmarks and decoded up to twice as fast.
• Amazon has released eight-billion and 32-billion parameter hybrid models, code and a vLLM plugin under Apache 2.0.
Read more
According to Amazon, priming lets developers test hybrid model designs using less than 0.5% of a source Transformer’s pretraining tokens.
• Amazon says priming builds hybrid AI models from pretrained Transformers at over 99% lower pretraining cost.
• In AWS tests, a 32-billion-parameter hybrid matched its source on reasoning benchmarks and decoded up to twice as fast.
• Amazon has released eight-billion and 32-billion parameter hybrid models, code and a vLLM plugin under Apache 2.0.
Read more
Anthropic adds defenses against Claude distillation attacks
Anthropic says the defenses make it harder to use Claude’s reasoning to train other models, though some legitimate users face integration changes or blocked requests.
• Anthropic says new Claude defenses have reduced reasoning data harvested by distillers by about 90 to 99 percent.
• The defenses restrict reasoning access, screen extraction attempts and limit accounts that repeatedly make them.
• Some developers have had to update integrations, and Anthropic now charges for certain blocked requests.
Read more
Anthropic says the defenses make it harder to use Claude’s reasoning to train other models, though some legitimate users face integration changes or blocked requests.
• Anthropic says new Claude defenses have reduced reasoning data harvested by distillers by about 90 to 99 percent.
• The defenses restrict reasoning access, screen extraction attempts and limit accounts that repeatedly make them.
• Some developers have had to update integrations, and Anthropic now charges for certain blocked requests.
Read more
SpaceXSI bundles Grok, Cursor and X Premium in one subscription
X Subscription puts SpaceXSI's AI, coding and X features under one membership, with access and usage capacity varying by tier.
• SpaceXSI announced X Subscription, combining access to Grok, Cursor and X Premium across paid tiers.
• Individual plans cost $8, $30, $100 or $200 a month, with different features and access at each tier.
• Premium, Super and Ultra are available for teams, while enterprise customers can get customized plans.
Read more
X Subscription puts SpaceXSI's AI, coding and X features under one membership, with access and usage capacity varying by tier.
• SpaceXSI announced X Subscription, combining access to Grok, Cursor and X Premium across paid tiers.
• Individual plans cost $8, $30, $100 or $200 a month, with different features and access at each tier.
• Premium, Super and Ultra are available for teams, while enterprise customers can get customized plans.
Read more
Google launches Gemini 4 Argon globally with API access and new subscription availability
Developers and eligible subscribers can now use Argon globally, while Google says its early users have applied it to engineering, security and research tasks.
Read more
Developers and eligible subscribers can now use Argon globally, while Google says its early users have applied it to engineering, security and research tasks.
Read more
Apple launches HomeView smart display for $299
HomeView brings Apple’s personal apps and smart home controls to a shared display, though personal features depend on user recognition and a compatible device.
• Apple introduced the $299 HomeView, its first smart display, with Siri AI and Apple Home controls.
• Tabletop and wall-mounted versions support FaceTime, Matter and up to six user profiles.
• Preorders are open in eight countries, with devices shipping November 13th.
Read more
HomeView brings Apple’s personal apps and smart home controls to a shared display, though personal features depend on user recognition and a compatible device.
• Apple introduced the $299 HomeView, its first smart display, with Siri AI and Apple Home controls.
• Tabletop and wall-mounted versions support FaceTime, Matter and up to six user profiles.
• Preorders are open in eight countries, with devices shipping November 13th.
Read more
Nvidia prepares RTX Spark laptops with local AI tools for October 16 launch
Nvidia says RTX Spark owners will be able to run supported AI agents locally and use optimized creative apps and games when laptops launch October 16.
• Nvidia says RTX Spark laptops will be available October 16 with tools for running AI agents locally.
• Its Nemotron 3.5 Bolt 74B-A9B 24GB model is due October 15, and the PAIR network tool is out of beta.
• Adobe optimizations are in beta, while Nvidia says DLSS 5 will reach RTX Spark at launch.
Read more
Nvidia says RTX Spark owners will be able to run supported AI agents locally and use optimized creative apps and games when laptops launch October 16.
• Nvidia says RTX Spark laptops will be available October 16 with tools for running AI agents locally.
• Its Nemotron 3.5 Bolt 74B-A9B 24GB model is due October 15, and the PAIR network tool is out of beta.
• Adobe optimizations are in beta, while Nvidia says DLSS 5 will reach RTX Spark at launch.
Read more
Shopify to open its catalog to more AI shopping agents
Shopify’s planned expansion gives more AI agents access to its catalog while requiring buyers to approve every purchase.
• Shopify will soon let more AI agents browse and buy from over a billion listings in its catalog.
• Shoppers can connect an agent to Shop and check out with Shop Pay inside the conversation.
• Buyers must approve each purchase and can set rules such as budget limits or fast shipping.
Read more
Shopify’s planned expansion gives more AI agents access to its catalog while requiring buyers to approve every purchase.
• Shopify will soon let more AI agents browse and buy from over a billion listings in its catalog.
• Shoppers can connect an agent to Shop and check out with Shop Pay inside the conversation.
• Buyers must approve each purchase and can set rules such as budget limits or fast shipping.
Read more
Nvidia pledges $250 million in compute for UK AI safety research
Nvidia says the compute will support UK research on secure AI agents, open-model security and safety assessment benchmarks.
• Nvidia pledged compute access valued at $250 million over five years for UK research on AI agent safety and security.
• The company says $50 million in compute access will be provided annually, mostly for the wider UK research community.
• UK-based Nvidia Cloud Partners will provide access, but Nvidia has not named providers or given an access date.
Read more
Nvidia says the compute will support UK research on secure AI agents, open-model security and safety assessment benchmarks.
• Nvidia pledged compute access valued at $250 million over five years for UK research on AI agent safety and security.
• The company says $50 million in compute access will be provided annually, mostly for the wider UK research community.
• UK-based Nvidia Cloud Partners will provide access, but Nvidia has not named providers or given an access date.
Read more
SolarWinds unveils Avior for enterprise IT operations
SolarWinds says Avior can take on first-line IT work while keeping actions subject to permissions, logging and risk-based approval.
• SolarWinds unveiled Avior to handle first-line enterprise IT operations.
• Avior can group alerts, address configuration changes and retire stale device records.
• SolarWinds says higher-risk actions require human approval; Avior is available through guided onboarding.
Read more
SolarWinds says Avior can take on first-line IT work while keeping actions subject to permissions, logging and risk-based approval.
• SolarWinds unveiled Avior to handle first-line enterprise IT operations.
• Avior can group alerts, address configuration changes and retire stale device records.
• SolarWinds says higher-risk actions require human approval; Avior is available through guided onboarding.
Read more