AI Compute & Chips Microsoft's Maia 300 Publishes Rack Numbers for Its Own Inference Silicon Microsoft announced Maia 300 on 12 October 2026 and framed it as a complete rack system. The chip carries eight HBM4 stacks inside a 950 W package, the rack holds 96 accelerators in a 167-ki
LLMs & Models Meta's Muse Wonder Adds a Memory Layer to Its Agent Models Meta launched Muse Wonder and Muse Wonder Full on 12 October 2026 with a 1M-token context and a thirteen-page model card published the same day. Active Memory carries details between conversat
AI Agents Automating Invoice Intake With Checks a Person Can Verify Invoice intake automation reads a supplier invoice, extracts the fields and routes each one with rules you can inspect. The design question sits in the checks between extraction and posting:
AI Compute & Chips SparseEngine Puts Fifteen Sparse Attention Methods Behind One Lifecycle Contract A research team published SparseEngine, an inference engine designed around sparse attention from the start. It runs 15 sparse methods across four families on a shared
Coding Agents & Dev Tools ArtCraft's AI-Built Apps Bring Open-Source Pressure to Adobe ArtCraft’s creative-software lineup has grown from seven apps in launch coverage to twelve on its current directory. PhotoCraft advertises broad PSD support; the project still labels it earl
LLMs & Models Ai2's Olmo-core 3 Keeps Every Expert on the GPU and Throughput Nearly Flat Ai2 replaced fully sharded data parallelism with DDP in Olmo-core 3, held a 47-billion-parameter mixture of experts at 52,000 tokens per second per GPU and grew the expert pool from 8 to 128
AI Compute & Chips Scality's AI Inference Factory Keeps the Context Cache in Object Storage Over RDMA Scality shipped a validated stack for running open-weight models on customer-owned servers, and its storage layer holds the context cache that normally lives inside GPU memory.
LLMs & Models Microsoft-Decision-1 Scores the Choices an Agent Has to Make Microsoft put a 9-billion-parameter decision-scoring model in Foundry that returns a calibrated probability for each option in a fixed list, and priced it at $0.042 per million input tokens.
AI Agents Designing an AI Concierge That Qualifies Visitors With Facts You Publish A website concierge answers, asks and recommends from your own catalogue and rules, quotes prices instead of inventing them, and hands the conversation to a person with a summary that can be
AI Compute & Chips AMD's Trillion-Dollar Mark Rests on the CPU Inside Personal Agents AMD closed at $649.42 on 6 October 2026 and passed a trillion dollars in market value while both flagship personal agents disclose the CPU underneath them: two cores per Muse sandbox, nine p
AI Compute & Chips The Rack Is Now a Power Contract: Density Numbers Decide Where Compute Lands Rack density moved from about 16 kW to 27 kW in a single year, and the newest AI racks are specified up to 246 kW. That makes the electricity room, the fluid loop and the grid queue the deci
AI Compute & Chips DGX Spark at 64GB: What $4,999 of Local Compute Buys a Small Team NVIDIA's 64GB DGX Spark configuration lands on 23 October from $4,999, runs models up to 100 billion parameters on the desk, and clusters two units into 128GB of pooled memory. The interesti
LLMs & Models Gemini 4 Argon Arrives as a Gated Release With Frontier Numbers Google published frontier numbers for Gemini 4 Argon and opened access to a vetted group of cyber defenders first. The benchmarks are real and self-reported, and the gate is the part that sh
AI Agents How a Private RAG Assistant Answers From Your Own Documents A private RAG assistant has three separate jobs: retrieve the right passage, attach the passage to the claim, and refuse when retrieval comes back empty. The refusal is an engineering output
AI Agents What an AI Agent Harness Contains, and Why the Model Layer Stays Swappable An agent harness is the loop around the model: context, tool calls, session state and permissions. The model inside it is the part that changes on somebody else's calendar, so the harness is
LLMs & Models Dialect Fine-Tuning Starts With Data Curation: What NVIDIA's Nemotron Recipe Shows NVIDIA published a full recipe for adapting its Nemotron 3.5 ASR model to Saudi Arabic dialects, and the numbers are honest enough to be useful: a default UTMOS quality threshold of 3.0 woul
AI Compute & Chips Anthropic's $11.6B Akamai Commitment: The Commercial Economics of the CPU Serving Layer Akamai's $11.6 billion, seven-year commitment with Anthropic on September 24, 2026 is for CPU workloads — serving, orchestration, agents, delivery — not GPU training. The warrant ties Anthro
AI Agents DeepSeek Just Published the Blueprint for Training Agents at Scale DeepSeek's DSec report shows what agent training actually needs from infrastructure: an elastic sandbox factory with tiered isolation, on-demand image delivery, and RL co-design.
LLMs & Models The Browser Becomes an AI Distribution Point: What Mozilla and Mistral Actually Changed Mistral Small 4 is now inside Firefox Smart Window, with France first for official French-language support. The real story is distribution — and whether user choice survives the integration.
AI Agents n8n Agents Change the Automation Contract: From Steps to Goals n8n Agents shift the automation contract from laying out steps to describing an outcome. The session log becomes the audit trail; one turn becomes the budget.
AI Agents The OpenAI Agents API Makes the Harness a Managed Product OpenAI's Agents API turns the Codex harness into a managed service with hosted sandboxes. Netics on why the interesting boundary is the sandbox, not the model.
AI Agents Salesforce on Google Cloud: The Agent Stack Gets a Shared Foundation Salesforce and Google Cloud are unifying infrastructure and agents: Hyperforce on Google Cloud, Gemini Enterprise with Salesforce data via MCP, and Zero Copy data sharing. Netics on what is
AI Compute & Chips NVIDIA Topograph Treats GPU Placement as a Live Topology Problem NVIDIA's open-source Topograph discovers cluster topology from cloud APIs or on-prem fabrics and publishes it as Kubernetes labels, Slurm config, or Slinky ConfigMaps. Netics on why topology
LLMs & Models GPT-6 Prompt Caching Makes Latency and Cost an Engineering Choice OpenAI's GPT-6 prompt caching brings dashboards, diagnostics, breakpoints, and prewarming. Netics on why caching is an economic layer agents must be architected around, not a vendor convenie
AI Compute & Chips Benchmarking LLM Inference Needs a Real Load Client NVIDIA's AIPerf replaces GenAI-Perf with a multiprocess load client that stops being the bottleneck. Netics on why LLM inference numbers are only worth as much as the client generating them.
