Analysis
210 articles
Analysis and opinion on where AI is going and what it means for a business that has to make decisions now.
Latest
Showing 1 to 20 of 210
Why Does Your AI Agent Keep Forgetting, And What Is That Costing You?
An agent has a context window, not a memory. Unbounded transcript replay causes drift and memory poisoning. Here is what to store, what to expire and how to test it.
What Does Real AI Agent ROI Look Like, Beyond The Vendor Case Study?
Real agent ROI is measured in one unit of work: volume, touch time, rework and cycle time. Here is a defensible calculation and the costs most cases leave out.
Should Your AI Agent Harness Be Open Source, Or Locked To One Vendor?
Vendor lock-in rarely lives in the model. It hides in memory, tool schemas, sandboxes and eval data. Five checks tell you whether you can actually leave.
Your AI Agent Has Employee-Level Access, So Who Is Watching What It Does?
An agent with employee-level access is an insider-risk problem, not a model problem. Scope every agent like a new starter: identity, least privilege, gates and expiry.
Can You Run AI Agents Fully On-Premises Without Handing Your Data To A Cloud Vendor?
Oxide's VM-per-session pattern makes self-hosted agents practical, Ollama shipped local speech, and our two DGX Sparks sustain 35 tokens a second.
LangChain's Open-Source Design Agent: What It Takes to Make AI Look On-Brand
LangChain open-sourced Content Studio, the Slack design agent behind 3,000+ assets. What the harness gets right, what a request costs, and whether to self-host.
How Granola Grew AI Search Users 4,600%: The AEO Playbook Explained
The full breakdown of Granola's 4,600% AI search growth: how the AEO system works and how to run it.
What Is Google's SLEEP Framework? LLM Memory Consolidation Explained
Google's SLEEP framework lets a deployed LLM fold new knowledge into its weights without catastrophic forgetting. How consolidation and dreaming work, what the benchmarks show, and what is still unproven.
Claude for Legal Artifacts: What the Gallery Does and What It Cannot Do
Anthropic's Claude for Legal gallery ships 19 ready-made artifacts and 60+ connectors. What they do, where privilege, verification and confidentiality bite, what the Legal Agent Benchmark really says, and how the same pattern runs on open-weight models inside your own environment.
Liquid AI LFM3: Why A 1.7B Model Is Now A Business Option
Liquid AI's LFM3 family puts six specialised models on one 2B backbone, and the text model runs in about 1 GB at 4-bit. Here is what the benchmarks actually show, where a small model beats a frontier model on cost and custody, and the four conditions where on-device AI pays for itself.
MiniCPM-OCR 1.0: A Free 3.5B Pair That Reads Messy Scans And Flags Its Own Mistakes
OpenBMB's MiniCPM-OCR 1.0 is a 1.3B open-weight parser that leads real-world document benchmarks, and Jev is a 2B checker that scores every region so you know exactly what a human must review. What the numbers actually say, where the launch claims overreach, and how to run the pair on your own hardware.
Local AI vs Subscriptions: What 25 Seats Really Cost in 2026
At 25 seats, cloud AI subscriptions cost $6,000 to $9,000 a year before API usage. A 128GB DGX Spark lists at $6,950 and pays back in about 15 months. Where local wins, where cloud still wins, and the hybrid routing pattern that works in practice.
Lovable Moved 70 Million AI-Built Apps Off Next.js: What the 98.5% Lift Means
Lovable migrated 910,000 lines from Next.js to TanStack Start on Cloudflare Workers and now runs 70 million AI-built apps. Inside the numbers, how TanStack Start actually works, the 11-minute outage, and the 98.5% jump in AI-tool traffic that came from pre-rendering alone.
The alpha in served intelligence: what 155 trillion tokens a day really tells you
DeepSeek open sourced DeepInfer and disclosed one cluster serving 155 trillion tokens a day at US$38 of revenue per GPU-hour. We checked the maths against its public list prices and it closes. The margin is in the KV cache, not the model, and that changes where the opportunity sits for Australian teams.
Is Complex Agentic Memory Always Needed? Why Naive RAG Beats Agentic Memory Systems
A raw-history RAG baseline beat 12 agentic memory systems in six of seven benchmark settings, at a fraction of the ingestion cost. Here is what the evidence shows and when a memory layer still earns its place.
Kimi K3.1 Pro vs Claude Fable 5.1 and GPT-6 Astra: Open Weights Have Caught the Frontier on Paper
Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, two points above Claude Fable 5.1 and GPT-6 Astra, at $1.98 per task against $7.63 and $3.26. Here is what the open-weights milestone proves, and what the benchmark audit says it does not.
Do No-Code AI Agent Platforms Beat a Custom Built Agent?
No-code agent platforms are fast and cheap to start, then credits scale with volume. We compare both paths on cost, integration depth and governance.
What Happens When an AI Agent Job Runs for Hours in the Background?
MCP Tasks replaced the long-held connection with a durable task handle. Here is the lifecycle, the support gap, and how to design a job that survives a crash.
Should You Let AI Agents Run Your HR and Finance Workflows?
Workday and Google now put agents inside HR and payroll. The limiting factor is not the model, it is the process map and the approval chain.
Can Local AI Agents Really Replace Your Cloud Chatbot Subscriptions?
We ran the numbers on $20 seats versus a $6,950 local agent box. Payback lands at 15 months for 25 seats, and the trade-offs are not only financial.
What's actually useful in AI, every Friday
We follow everything happening in AI and send what's worth your time: new models, AI agents, and how to put them to work. One email a week, unsubscribe anytime.