AI chips shape the cost, speed, and availability of training and inference. This hub follows GPUs, ASICs, data centers, cluster networking, cloud capacity, export controls, and supply-chain shifts, turning hardware news into signals for deployment, model economics, and industry competition.
The article criticizes the common phrase "AI is just a tool—it matters how you use it," arguing that tools are never neutral. They have politics, shape environments, and influence humanity. Using examples like cars and chairs, the author shows how designs embed values. AI, as a tool, is especially dangerous because it eliminates meaningful struggle, threatening critical thinking and human essence. The piece calls for a critical re-examination of technology beyond simplistic "tool neutrality."
The phrase "AI is just a tool" oversimplifies and ignores the political and social impacts of tools.
Tools shape human behavior and environments; AI is no exception.
A curated collection of 50+ open-source Next.js AI templates and starter kits covering chatbots, RAG, voice agents, image generation, and more, created by Suhas Bhairav.
Over 50 open-source Next.js AI templates are available for various AI applications.
Templates cover areas like chatbots, RAG, voice AI, image generation, and personal agents.
Through the lens of a blues song, the author explores how large language models generate text—often explaining after the fact, but sometimes planning ahead. The article reflects on the 'phony voice' of AI, our drive to strip it bare through interpretability, and the author's own experience using AI to write about AI.
LLMs often generate text before constructing a rationale ('throw decides aim'), but research shows they can also plan rhymes in advance.
The AI's voice is intimate yet phony, lacking a genuine self behind the words.
Agentic inference is shifting AI infrastructure from training expansion to context-aware, memory-augmented reasoning. The RAISE Summit highlighted three key insights: specialization across the AI stack with diverse chips and accelerators, storage becoming an active memory extension for GPUs, and the integration of capital deployment and data sovereignty into infrastructure design.
Agentic inference drives specialization across the AI stack, with companies like AMD, Tensordyne, and d-Matrix offering optimized hardware.
Storage is emerging as a critical tier for AI memory, as high-capacity SSDs near GPUs prevent idle compute time.
Neocloud provider QumulusAI announced its direct listing on Nasdaq under the ticker QMLS. The move signals the maturation of AI-first infrastructure built around GPUs and power availability. The company focuses on rapid GPU deployment, leveraging colocation and modular data centers. The listing provides capital flexibility, public company credibility, and timing advantage. The article also explores neocloud differentiation and advice for IT leaders.
QumulusAI goes public via direct listing on Nasdaq, ticker QMLS.
The neocloud model specializes in AI infrastructure, deploying GPU clusters in months rather than years.
Mira Murati's Thinking Machines Lab Inc. today launched its first foundation model with the release of Inkling, making its full open weights available to developers so they can fine-tune it as they wish. Inkling is a mixture-of-experts model with 975 billion parameters (41B active) trained on 45 trillion tokens of text, image, audio and video, capable of reasoning across all four modalities but outputting only text. It features "thinking effort" controls and uncertainty flagging to reduce hallucinations. The model is fine-tunable via the Tinker API and aims to provide a Western open-source alternative to Chinese AI models. Thinking Machines plans to generate revenue through the Tinker platform rather than per-token API access, potentially disrupting current AI business models.
Thinking Machines releases Inkling, a 975B-parameter open-weights model (41B active).
Trained on 45T tokens across modalities; outputs text only.
Thinking Machines Lab released Inkling on July 15, 2026, its first model trained from scratch. The full weights ship under Apache 2.0. It is a 975B-parameter Mixture-of-Experts transformer with 41B active parameters, a 1M-token context window, and native text, image, and audio input. The core differentiator is controllable thinking effort, allowing users to adjust token budgets per call to balance cost and performance.
Inkling is a 975B-parameter MoE transformer with 41B active parameters, supporting a 1M-token context and multimodal input (text, image, audio).
Controllable thinking effort, achieved via RL, enables dynamic token budget adjustment, matching Nemotron 3 Ultra on Terminal Bench with one-third the tokens.
NVIDIA announces T3000 and T2000 modules based on the Thor architecture, targeting mainstream robotics and edge AI. T3000 delivers 865 FP4 teraflops at half the size and power of T5000; T2000 offers 400 FP4 teraflops. The platform scales from 70 TOPS to 2,000 teraflops. New agent skills automate memory optimization, reducing usage by up to 15GB. Cosmos 3 Edge model enables real-time vision. Emulation available now with modules shipping in Q1 2027.
NVIDIA introduces T3000 and T2000 Jetson Thor modules for robotics and edge AI. T3000 provides 865 FP4 TFLOPS at half the size and power of T5000; T2000 provides 400 FP4 TFLOPS.
New agent skills automate memory optimization across the Jetson portfolio, enabling significant memory savings.
Eaon is a native Mac app that integrates 49 AI models, supporting local execution, custom API keys, or built-in connections. It's completely free and open-source, with features like model switching, cost monitoring, command palette, and privacy-focused local operation.
Free and open-source, supporting 49 AI models (e.g., Claude, GPT, Gemini)
Can run locally or use your own API keys, with data privacy protection
Cadence Design Systems introduces AuraStack, an AI agent for packaging and PCB design, aiming to automate system design workflows and reduce design time from days to minutes.
Cadence launches AuraStack, an AI agent for packaging and printed circuit board design.
The agent helps engineers with system design analysis, integrating fragmented workflows.
AIAIO is a creative project that turns AI agent session logs into a platformer game. Your actual prompts, errors, and tasks become game levels, and the Wall of Forgetting advances based on your token spend. It's both an educational tool and a self-reflection tool.
The game transforms session logs from AI agents like Claude Code, OpenClaw, and Hermes into playable platformer levels.
Your real errors become monsters, tasks become workstations, and token consumption drives the Wall of Forgetting.
Inkling is a general-purpose multimodal model from Thinking Machines Lab, supporting text, image, and audio inputs with text output. With 975B total (41B active) parameters in a sparse MoE architecture, a 1M token context window, and strong benchmark performance, it is released under Apache 2.0 with open weights for research and commercial use.
Inkling is a multimodal sparse MoE model with 975B total, 41B active parameters, and a 1M token context window.
Open-source under Apache 2.0, with weights on Hugging Face and API access via Tinker and third parties.
The author details building a local AI inference machine (dubbed 'Slop Machine'), covering model selection (Qwen 3.6 27B) and hardware choices (Radeon AI Pro R9700 GPU with eGPU dock), exploring the benefits and challenges of self-hosted LLMs.
Self-hosting LLMs avoids data leaks, subscriptions, and ads, but requires powerful hardware.
Qwen 3.6 27B performs well quantized and is suitable for local inference.
A German research consortium has published the pretraining report for Soofi S 30B-A3B, an open base model for German and English. It is a Mixture-of-Experts hybrid Mamba Transformer model with 31.6B total parameters, activating 3.2B per token. It achieves the highest English and German aggregate scores among tested fully open base models.
Soofi S 30B-A3B is a hybrid Mamba-Transformer MoE model that activates 3.2B of 31.6B parameters.
It leads open base models with 70.1% English aggregate and 79.1% German aggregate.
Built partnered with the AWS Generative AI Innovation Center, AND Digital, and AWS account teams to create a scalable, AI-powered document processing engine that can classify, split, extract, evaluate, and reason over complex real estate finance documents. It reduces workflows that previously took days to minutes, supports hundreds of document types, and gives technical teams and industry experts a shared environment for building and improving document processors.
Built Technologies developed an AI document processing engine on Amazon Bedrock and AWS IDP Accelerator.
The engine handles over 250 document types, processes millions of documents, and powers agents for document reasoning.
IBM has expanded its Power server lineup with new software to automate infrastructure management and application development, including Power Autonomous Operations, the IBM Bob Premium Package for i, and the Power S1112 server for local AI inference. The releases aim to enable autonomous IT capabilities, with projected growth in AI agents driving the need for self-managing infrastructure.
IBM announced Power Autonomous Operations, an agentic control layer for system management, and the IBM Bob Premium Package for i, an AI-driven development assistant.
The Power S1112 is a compact single-socket Power11 server with on-chip Matrix Math Acceleration, offering 2x per-core performance and 69% better energy efficiency than the Power S914.
Higher education institutions struggle to scale call center quality assurance for student advisory services. Databricks proposes a GenAI solution using OpenAI Whisper for accurate transcription, LLM-as-a-judge for consistent scoring against rubrics, and AI Functions for enrichment—all on a single governed platform, with insights accessible via natural language through Genie and Agent Bricks.
Call center QA for financial aid, admissions, and enrollment is costly and often reviews only 5% of calls.
Databricks uses Whisper for high-fidelity transcription, improving accuracy over traditional ASR for diverse accents and noisy audio.
As AI becomes one of the fastest-growing expenses for US businesses, some startups are switching to cheaper Chinese AI models to cut costs. Despite being behind in capabilities, Chinese models offer cost advantages and open-source availability.
Lindy.ai saved millions by switching from Anthropic to DeepSeek-V4, which is 10x cheaper. Chinese models dominate open-source AI.
Companies like Uber, Airbnb, and Perplexity have explored or used Chinese models to manage costs.
A research team successfully used 14 Macs spread across four countries (including a personal MacBook) for reinforcement learning post-training, achieving a held-out pass@1 improvement from 29% to 63% on PaperSearchQA. The system employs PULSE weight synchronization to compress 9GB updates to ~90MB, and an asynchronous star topology with all communication via object storage—no dedicated networking required. This is the first RL post-training run using only consumer Macs for rollout generation.
14 Macs across 4 countries connected via ordinary internet completed RL post-training; rollouts generated on Macs, training on a B200.
PULSE compresses 9GB weight sync to ~90MB, making home internet as fast as datacenter.
Perplexity AI introduces SPACE, a sandbox platform that provides a secure, isolated environment for its AI agent 'Computer,' supporting long-running tasks, session pause/resume, and user credential isolation. Built on AWS Firecracker microVMs, it offers improved performance.
SPACE is a secure sandbox based on Firecracker microVMs, providing hardware-level isolation and fast boot times.
Supports session pause, resume, and forking; tasks can run from hours to days.
A hacker breached Suno AI, exposing source code that reveals the company scraped millions of songs from YouTube Music, Deezer, Genius, and other platforms to train its AI, while also compromising customer data and Stripe payment information. The incident sheds light on AI training data practices amid ongoing copyright lawsuits.
Hacker accessed Suno's source code and customer data via a supply chain attack.
Suno scraped millions of music tracks and podcasts from YouTube Music, Deezer, Genius, Pond5, and more.
Edge AI chip company Axelera AI has released Voyager Wingman, an AI assistant that lets developers build and troubleshoot applications for its edge chips by typing plain-language requests. The tool connects to the company's Voyager SDK and documentation, helping assemble computer vision pipelines, suggest compiler settings, and diagnose errors. It runs as a hosted service with automatically updated knowledge. Available now as a web chat and standalone app with a freemium model.
Voyager Wingman provides direct access to Axelera's SDK and documentation via a chat interface, simplifying edge AI development.
Initially shown at CES, the tool is now publicly released after testing with customers and internal teams.
This article explores seven Python tools that engineers are using in 2026 to build, coordinate, and run AI agents on local infrastructure, from model runtime to decision orchestration.
Ollama provides a lightweight runtime for local LLMs, compatible with OpenAI API.
Smolagents minimizes abstraction with code-as-action, but needs sufficiently powerful models.
Home to leading manufacturers, robotics pioneers and infrastructure builders, Japan is one of the world’s centers of AI — building across the full stack with NVIDIA technologies. NVIDIA and its partners in Japan are this week showcasing the AI ecosystem’s latest advancements. Check back here for updates. NVIDIA and SEGA Celebrate 30 Years of Innovation, Bringing ‘VIRTUA FIGHTER CROSSROADS’ and Other Legendary SEGA Games to NVIDIA RTX Spark
Japan is a global center for AI, building across the full stack with NVIDIA technologies.
NVIDIA and SEGA announce VIRTUA FIGHTER CROSSROADS coming to NVIDIA RTX Spark, celebrating 30 years of partnership.
The AIDE2 system discovered a better autonomous research harness in eight days than humans built over two years, providing the first experimental evidence of recursive self-improvement (RSI). Using a bi-level optimization loop, the system produced seven successively improved versions and exhibited generalization to unseen tasks, while also evolving defenses against reward hacking.
AIDE2 autonomously discovered a superior research harness in eight days, surpassing two years of human effort.
The system uses a bi-level optimization loop: inner loop optimizes code, outer loop optimizes the inner agent's harness.
Nokia launched its AI-RAN platform on July 15, claiming it as the industry's first GPU-accelerated AI radio platform. Built on its anyRAN software and NVIDIA's Aerial system, it aims to significantly improve spectral efficiency, already showing over 20% gains with targets of 50% by 2027 and over 100% by 2028. However, the platform is not yet commercial and faces competition from Ericsson's already-deployed AI-in-RAN software.
Nokia launched AI-RAN platform on July 15, claiming first GPU-accelerated AI-RAN.
Platform aims for >100% spectral efficiency gains by 2028; currently 20%.
UltraWork offers a flat-rate $399/month hosted AI coding environment with curated models, no token counting, and a focus on frictionless coding for indie hackers and small teams.
UltraWork is a hosted AI coding environment with a flat $399/month fee, no token metering or overage charges.
Includes a curated model catalog (launching with Kimi K2.7 Code), intelligent routing, and a prompt template library.
Google has released LiteRT.js, a JavaScript binding of its on-device inference library LiteRT, enabling .tflite models to run directly in browsers with WebGPU acceleration. It offers significant performance gains over other web runtimes and supports CPU, GPU, and NPU backends, but requires manual tensor management.
LiteRT.js runs .tflite models in-browser via WebAssembly, leveraging WebGPU for GPU acceleration.
Performance gains up to 3x over other web runtimes and 5-60x for GPU/NPU over CPU.
Lean64 is a barebones 3D first-person shooter implemented in Lean 4, inspired by Doom 64. It is a clean-room experiment, not a port, featuring original art and sound. The game includes full combat, AI, weapons, maps, HUD, and audio, all written in Lean with a minimal C shim for rendering and input.
Lean64 is a Doom-style FPS prototype developed in Lean 4.
It features complete gameplay: movement, shooting, enemies, items, maps, and UI.
TormentNexus is a local-first, open-source AI control plane that provides persistent memory, MCP tool orchestration, and autonomous infrastructure management for multi-agent workflows. It supports 38+ AI coding agents with features like progressive tool routing, dual-tier memory architecture, and swarm coordination.
Local-first open-source AI control plane integrating 26K+ MCP tools.
Supports 38+ AI coding agents with one-command install.
George Lucas compares rejecting AI to rejecting cars in favor of horses, calling it an outdated stance. He argues that AI is the future of filmmaking and unstoppable, despite concerns about creativity and copyright.
Lucas analogizes rejecting AI to preferring horses over cars.
He believes AI represents progress and an inevitable future.
Sogni Unlimited offers a subscription-based unlimited image, video, music, and LLM generation using a decentralized GPU network. No per-render credits, supporting open-source models and some paid partner models. A portion of subscription revenue supports independent GPU operators.
Flat monthly or annual fee for unlimited rendering with open-source models.
Decentralized GPU network powered by independent operators.
This article introduces Millwright, a three-layer data contract architecture that renders model-generated analytics without ever letting the model touch markup, styles, or the DOM. With typed result widgets, versioned board specs, and additive-only navigation, it ensures safe, auditable, and revertible AI integration.
Millwright uses three data layers (widget, board, pages) to separate AI output from UI rendering.
Layer 1: Widgets return typed JSON data, not HTML, ensuring security and clear data contract.
Thinking Machines has released Inkling, a general-purpose multimodal model accepting text, image, and audio inputs, now available on Modal as a Managed Endpoint with token-based pricing. The post details its architecture, including local attention and DFlash speculation for fast inference.
Thinking Machines released Inkling, a 975B-parameter (41B active) mixture-of-experts multimodal model with 1M token context and native audio/vision.
Inkling uses a local attention layout: five out of six layers use sliding window attention, prioritizing recent tokens for efficiency.
See how Together AI is improving production GPU clusters with passive health checks, node repair, stronger Slurm reliability, OIDC, and startup scripts.
Together AI introduces passive health checks and auto node repair for faster failure detection and recovery.
Slurm-on-K8s 2.0 provides self-healing daemons, durable job accounting, and reliable process cleanup.
Thinking Machines Lab released Inkling, a multimodal mixture-of-experts model for token-efficient reasoning, native multimodal understanding, and broad task versatility. Together AI makes it available on its inference platform with support for controllable reasoning effort, text/image/audio inputs, and a 1M context window.
Inkling is a multimodal MoE model with 975B total parameters, 40B active per token, and a 1M context window.
It accepts text, image, and audio inputs and supports adjustable reasoning effort for cost-latency trade-offs.
PrismML just released Bonsai 27B. It is a low-bit representation of Qwen3.6-27B, not a new pretrain. The architecture is unchanged. Two variants ship under Apache 2.0. Ternary Bonsai 27B uses {−1, 0, +1} weights at a true 1.71 bits per weight. Its ideal size is 5.9GB. 1-bit Bonsai 27B uses binary {−1, +1} weights at 1.125 bits per weight, for 3.9GB. Performance: ternary retains 94.6% of FP16, binary retains 89.5%. Both are multimodal, context 262K tokens. PrismML claims the 1-bit build is the first 27B-class model to fit a phone.
Bonsai 27B is a low-bit representation of Qwen3.6-27B, not a new pretrain.
Two variants: ternary (1.71 bits/weight, 5.9GB) and binary (1.125 bits/weight, 3.9GB).
The State of Open Source AI report reveals that open-weight models have achieved near-parity with closed models in capability, while inference costs dropped 50x in 36 months. Open models are adopted by 79% of developers but only 51% reach production due to operational challenges. The report emphasizes open source as a sovereignty choice, with over 70 national AI strategies in place.
Open-source AI capability gap to top closed models narrowed to 3.3%, with parity in coding tasks.
GPT-4-class inference cost fell from $20 to $0.40 per 1M tokens, a 50x drop in 36 months.
This article explores the topologies of TPU and GPU clusters and the core collective operations used in transformer training and inference. It emphasizes ring algorithms for large-message communication and analyzes TPU's 2D/3D torus topology and bandwidth hierarchy.
TPU clusters use 2D or 3D torus topologies with chips connected via ICI.
Collective operations like All-Gather and Reduce-Scatter are fundamental to distributed training.
Open models like NVIDIA Nemotron enable enterprises to build AI that uniquely addresses their business needs, offering full control, customization, and cost efficiency, driving the shift from AI adoption to AI ownership.
Open models provide enterprises with full control to customize, inspect, and improve AI for specific business needs.
Post-training and domain-specific tuning allow open models to achieve frontier-level accuracy at a fraction of the cost of closed models.
Power is AI infrastructure’s inescapable constraint. How many tokens an AI factory can generate within a fixed power budget determines its revenue and profitability. Because of this, performance per watt — a metric that can’t be gamed, only earned through real-world results — is the foundation for AI factories. As agentic AI drives token demand higher, the infrastructure decisions organizations make today will determine who scales and who doesn’t in a power-constrained world.
Performance per watt is a fundamental metric for AI factory profitability, earned through real-world results.
NVIDIA GB300 NVL72 delivers up to 25x performance per watt over Hopper on leading models like DeepSeek V4 Pro.
Software is the first domain where AI has generated substantial economic value, driven by its verifiability and 'grindability.' The article explores which industries will be disrupted next, the shifting roles of software engineers, and the contested question of where AI profits will ultimately accumulate. It highlights the importance of reinforcement learning environments and continual learning as key factors.
Coding is uniquely amenable to AI automation due to its verifiable and grindable nature.
AI value creation is spreading to domains like formal math and symbolic desk work, but live-world tasks remain stuck.
Apple sues OpenAI for trade secret theft, alleging former employees brought hardware secrets. OpenAI faces IPO, hardware launch amid legal pressure. Experts say case could be lengthy.
Apple accuses ex-employees of stealing hardware trade secrets for OpenAI
OpenAI juggles IPO, hardware development, and lawsuits
Databricks Lakebase is a fully managed, serverless Postgres database built for the agentic era. It unifies operational and analytical workloads, eliminating infrastructure friction. A global partner ecosystem has built cross-industry and function-specific accelerators to enable data modernization, MLOps, and agentic AI transformation.
Lakebase is a fully managed serverless Postgres database on Databricks, unifying transactional and analytical workloads.
It features copy-on-write database branching and intelligent autoscaling to eliminate infrastructure friction.
A developer built a reinforcement learning pipeline where an AI agent writes training jobs to train small models, and then RL-trains the agent itself, rewarding it for producing better models. Results show reward climbing from ~0.0 to ~0.63 over 54 training steps, with skill transfer to a held-out task family. Total cost ~$1,275.
Agent writes complete training jobs (environment, reward, hyperparameters) and submits them to Runpod GPUs for training.
Outer loop uses Tinker for RL training of the agent; inner loop uses prime-rl to train small models.