computational resources Archives

Winsage

April 27, 2026

“The work we’re doing together remains ambitious”: OpenAI ends Microsoft exclusivity in a tone that doesn’t match last week’s leaked memo

Microsoft and OpenAI have redefined their partnership, ending Microsoft's exclusivity with OpenAI. Microsoft will continue as OpenAI's primary cloud partner, with products launching on Azure first, but OpenAI can now collaborate with other cloud providers if necessary. Key points include: - Microsoft retains its primary cloud partner status, with OpenAI products debuting on Azure unless Microsoft cannot meet capabilities. - OpenAI can distribute products across any cloud platform. - Microsoft's license to OpenAI's intellectual property is non-exclusive and extends through 2032. - Microsoft will no longer share revenue with OpenAI, while OpenAI will continue payments to Microsoft until 2030, capped at a predetermined total. - Microsoft remains a major shareholder in OpenAI. The partnership has faced challenges, including CEO Sam Altman's brief dismissal in late 2023, which displeased Microsoft CEO Satya Nadella. OpenAI's revenue chief acknowledged Microsoft's foundational support while indicating constraints on OpenAI's engagement with enterprise demands. Both companies are committed to ongoing collaboration amidst growing competition in the AI landscape.

Winsage

April 14, 2026

Veteran Windows dev shows off AI running on 47-year-old PDP11 with 6 MHz CPU and 64KB of RAM — ‘gloriously absurd’ project runs transformer model written in PDP-11 assembly language

Veteran Windows developer Dave Plummer is using a 47-year-old PDP-11 system with a 6 MHz CPU and 64KB of RAM to train a neural network called ‘Attention 11,’ developed in PDP-11 assembly language. The model is tasked with reversing a sequence of eight digits, requiring it to understand structural rules similar to those used in contemporary large language models like ChatGPT. Plummer emphasizes the importance of optimizing for the system's constraints, stating that constraints drive creative engineering. The model operates with 1,216 parameters and uses fixed-point math with 8-bit precision. After approximately 350 training steps, the model achieved 100% accuracy in the number-reversing task in about 3.5 minutes. Plummer argues that modern AI operates on the same mechanical principles as this vintage machine, just at a much larger scale, and suggests that companies focusing on efficiency and optimization may gain an advantage in the AI landscape.

Tech Optimizer

December 16, 2025

EnterpriseDB updates WarehousePG with features that fuel AI

EnterpriseDB has introduced new features for EDB Postgres AI for WarehousePG, including per-core pricing and real-time data streaming. WarehousePG is an open-source PostgreSQL data warehouse derived from the Greenplum Database project, part of the EDB Postgres AI platform launched in May 2024. The enhancements include: - Real-time streaming data capabilities for immediate access to data for AI applications. - A predictable per-core pricing model to help manage AI development costs. - Improved data observability and sovereignty features with flexible deployment options across cloud and on-premises environments. These updates aim to address financial pressures associated with AI development, as enterprises increasingly invest in AI technologies. The per-core pricing model stabilizes costs compared to consumption-based models, while the data streaming capabilities support modern AI workflows. The update also includes an AI-ready architecture, flexible deployment options, and data governance features.

TrendTechie

November 12, 2025

AI-торрент: децентрализованная сеть, где твои GPU станут валютой для супер-AI

A decentralized P2P network called "AI Torrent" is designed for AI model inference, based on principles such as a BitTorrent economy where nodes exchange computational resources, a Smart Swarm Architecture using specialized models, and Self-Organizing Intelligence that allows popular models to migrate to active nodes. Users without computational resources can interact through a standard API or chat interface, while those with resources can register as nodes and earn utility tokens by performing inference tasks. AI model creators can upload models and receive royalties through smart contracts. The economy operates on utility tokens (AIT), with revenue distribution of 70% to seeders, 20% to model developers, and 10% to a DAO fund. The network aims to be cheaper than centralized alternatives and has mechanisms for liquidity and stability, including trading on DEX platforms and staking. Existing projects in decentralized AI demonstrate the viability of P2P inference, and the "AI Torrent" seeks to integrate their best features while focusing on making inference accessible and affordable. Challenges include latency in P2P systems, which the network aims to address through geo-DHT and edge caching.

Tech Optimizer

November 1, 2025

Amazon Aurora makes PostgreSQL Serverless generally available

The PostgreSQL-compatible edition of Aurora Serverless is now generally available, allowing users to create database instances that operate only when necessary and automatically adjust capacity based on demand. Customers are billed on a per-second basis for active database capacity and standard Aurora storage costs. Users can specify minimum and maximum capacity when creating a database, and the architecture separates storage from computational resources, allowing automatic scaling from a minimum of 10GB to a maximum of 64 TB. Aurora Serverless PostgreSQL is accessible in regions including US East (N. Virginia and Ohio), US West (Oregon), EU (Ireland), and Asia Pacific (Tokyo).

AppWizard

August 1, 2025

Google releases Gemini 2.5 Deep Think for AI Ultra subscribers

Google is rolling out its latest AI model, Gemini 2.5 Deep Think, which is available exclusively through the Gemini app and requires a subscription to Google's premium AI Ultra plan. Deep Think enhances the capabilities of Gemini 2.5 Pro by increasing "thinking time" through greater parallel analysis, allowing it to explore various problem-solving methods. It requires several minutes to formulate a response, contributing to its proficiency in design aesthetics, scientific reasoning, and coding. Deep Think has been benchmarked against other models, showing superiority over Gemini 2.5 Pro and competing models like OpenAI's o3 and Grok 4. It scored 34.8 percent on Humanity's Last Exam, outperforming other models that typically score between 20 to 25 percent.