Exclusive-The AI founders who walked away from Bezos-backed Prometheus to model the universe

thestar.com.my Aug 25, 2026

Exclusive-The AI founders who walked away from Bezos-backed Prometheus to model the universe

Curated by Shubham · TechCapsules · Aug 25, 2026

AI-summarised brief · reviewed before publication

Accelerated Understanding Inc, co-founded by Anima Anandkumar and Benedikt Jenik, unveiled a new AI model designed to predict physical phenomena rather than process language. The system, which abandoned the standard Transformer architecture for neural operators, successfully handled 5 trillion data points in a single prompt during tests. This capacity is approximately 5 million times larger than what flagship models from Anthropic and Google can typically consume. The founders, previously pitched to lead Jeff Bezos-backed Project Prometheus, developed this technology independently. They argue that a nature-centric view of intelligence, focused on physics, offers superior capabilities for enterprise applications. Potential use cases include optimizing chip design, advancing robotics, predicting extreme weather, and analyzing geological data for energy firms. By replacing brittle, bespoke math models with a unified AI, the company aims to reduce trial and error in scientific research. Accelerated Understanding is prioritizing enterprise deals over consumer offerings, positioning its technology as a versatile tool for businesses requiring precise physical simulations and predictions across various industries.

💡 Why It Matters

· Shifting AI development from linguistic prediction to physical simulation addresses a critical bottleneck in hardware and scientific research.
· This approach promises to drastically reduce the time and cost required for complex engineering tasks like semiconductor design.

AI Capsules Entrepreneurs Startups

Read full article on thestar.com.my

The AI experts behind this adorable robot say it’s not a pet…

In May 2026, former iRobot CEO Colin Angle unveiled the Familiar, a plush, freckled robot developed by engineers from Disney, Boston Dynamics, and iRobot. The device is marketed as a physically embodied AI that learns and adapts to its owners, developing a distinct personality over time. Unlike typical robo‑pets, the Familiar tracks household routines, moods, and emotions, offering comfort, gentle coaching, and companionship. The creators claim it represents a shift toward robots that build and sustain human connection rather than merely performing tasks. The prototype debuted at the Future of Everything conference in New York.

💡 The Familiar’s ability to learn personal habits and respond emotionally could redefine how households integrate technology, moving beyond functional gadgets to emotionally responsive companions.

iphoneincanada.ca · Aug 26

OpenAI Releases First Benchmark Results for its Jalapeño AI Chip

OpenAI unveiled benchmark results for Jalapeño, its first custom AI inference chip, showing substantial gains in speed and energy efficiency. Tests using the SemiAnalysis InferenceX suite compared the processor against leading commercial systems on open‑weight models such as GPT‑OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. Jalapeño delivered 1.7‑3.6× lower latency, 1.5‑1.9× more work per watt, and 2.1‑4.1× faster performance on real‑time tasks. On the trillion‑parameter Kimi model it responded 3.4× faster while drawing under 550 watts, below its 700‑watt design limit. OpenAI accelerated development by employing its own AI tools, shrinking design time to nine months. The company plans to integrate Jalapeño into production data centers by year‑end, supplementing existing Nvidia hardware for handling the growing demand of users.

💡 Lower inference costs let OpenAI expand user capacity while reducing dependence on external chip makers, securing a hardware edge in the AI services market.

thenewstack.io · Aug 26

Anthropic gives chat and Cowork one memory

Anthropic announced that its AI models, Chat and Cowork, will now retain a limited memory of prior interactions, enabling the systems to reference earlier conversation context within a session. The company says the feature will improve continuity and relevance in user experiences, allowing the models to recall user preferences, previous queries, and ongoing tasks without requiring repeated prompts. Anthropic highlighted that the memory is bounded and designed to respect privacy, automatically clearing after a defined period. The update is part of Anthropic’s broader effort to enhance the usability of its conversational agents for both consumer and enterprise applications, positioning the firm against competitors that offer similar context‑aware capabilities.

💡 Persistent session memory lets users build on prior exchanges, turning one‑off interactions into more fluid, task‑oriented dialogues.

prnewswire.com · Aug 26

Centric Consulting Earns Microsoft Agentic Business Solutions Specialization

Centric Consulting has officially earned Microsoft’s Agentic Business Solutions Specialization, a prestigious designation recognizing partners with proven capabilities in building intelligent AI-enabled business solutions. Based in Dayton, Ohio, the firm achieved this status by demonstrating technical depth and customer success using Microsoft Power Apps, Power Automate, and Copilot Studio. This milestone underscores Centric’s ability to help organizations transition from AI planning to implementation and long-term value realization. Joe Hartsel, Vice President and Microsoft partnership lead, stated the validation confirms their team’s expertise in delivering measurable business outcomes. The specialization builds upon Centric’s broader role as a Microsoft Prioritized Managed Partner, complementing existing designations in Cloud & AI Platforms, Security, and Modern Work. Through Microsoft’s co-investment programs, Centric can now help qualifying clients lower project costs and accelerate implementation. The firm continues to leverage its 27-year history of deploying strategic advisors across the United States and India to solve complex technology challenges in AI, cyber risk, and data analytics for international clients.

💡 This certification validates a critical shift from theoretical AI planning to tangible, low-code automation execution for enterprises. It provides organizations with a vetted pathway to reduce implementation costs and accelerate productivity gains through proven partner expertise.

sea.mashable.com · Aug 26

AI mental healthcare is here. Can it help people — or cause…

Talkspace, a prominent virtual therapy platform, has launched Tee, an artificial intelligence mental health application designed to provide therapeutic support. Developed by clinicians and trained on anonymous Talkspace data, Tee utilizes a specialized large language model to assist users over the age of eighteen. The service costs $19.99 per month and includes safety protocols to identify extreme distress, automatically escalating such cases to human caregivers. Dr. Nikole Benders-Hadl, Talkspace’s Chief Medical Officer, helped develop the tool to address the growing trend of patients using general-purpose chatbots for mental health needs. Unlike generic models, Tee is informed by actual therapy sessions and clinical expertise. The launch follows a significant increase in telemedicine adoption during the pandemic, with Talkspace expanding its offerings beyond asynchronous messaging to include this AI-driven solution. The company aims to improve clinical quality and efficiency while providing accessible support to users who might otherwise turn to unregulated digital health tools.

💡 This launch marks a critical shift from experimental AI wellness tools to clinically supervised digital care, directly addressing the safety risks of unregulated chatbots. By embedding professional oversight into an accessible, low-cost format, Talkspace attempts to legitimize AI therapy while mitigating the potential for harmful misinformation.

microsoft.com · Aug 26

Moving beyond release waves with the AI at Work roadmap

Microsoft is consolidating the roadmaps for Dynamics 365, Power Platform, and Dataverse into the unified AI at Work roadmap, effective September 2026. This strategic shift replaces the traditional twice-yearly release wave model with continuous publishing, allowing customers to view upcoming capabilities across business applications, productivity tools, and AI portfolios in a single destination. The new system updates status in real-time as features progress from development to launch, ensuring stakeholders can track rollout progress and plan adoption more effectively. Organizations can filter content by environment, subscribe to RSS updates, or integrate data into existing workflows using the Release Communications MCP Server. While product development remains unchanged, this transition aims to align communication with how customers actually utilize interconnected Microsoft services. Historical release plans will remain on Microsoft Learn for reference, but new capabilities will exclusively appear on the AI at Work roadmap as they are committed.

💡 The shift from periodic waves to continuous publishing forces enterprises to abandon biannual planning cycles in favor of agile, ongoing integration strategies. By centralizing visibility across disparate business tools, Microsoft compels organizations to treat AI adoption as a persistent operational priority rather than a scheduled event.