Agent Reliability: Why 99% Per Step Is Not a Service Guarantee
Agent workflows compound errors. Pass^k, observable state and safe retries explain why verified completion matters more than a successful demonstration.
LPlutonous6 min read
Latest Analysis
Fast reads on model releases, compute strategy, policy pressure, and the companies fighting over the AI stack.
Mixture of Experts
MoE Models Have Two Sizes. Your Memory Bill Uses the Bigger One.
Active parameters explain only part of sparse AI economics. Historical Mixtral, DeepSeek-V3 and Qwen3 examples separate routing, weight storage and production capacity.
7 min read
Prompt Caching
Prompt Caching Turns Repeated Context Into a Margin Decision
A source-backed analysis of prompt-cache economics, write premiums, hit rates and why a cheaper prefix does not guarantee a cheaper completed task.
7 min read
RAG
RAG Has an Evidence Problem. Search Is Only the First Test.
Why retrieval, citation support and answer correctness need separate acceptance tests, with ALCE evidence and an operating model for document assistants.
6 min read
Small Language Models
Small Models Win When the Task Is Stable Enough to Teach
DistilBERT, TinyStories, Phi and Orca explain why compact AI is a training-data and deployment-economics strategy, with an explicit break-even calculation.
6 min read
Coding Agents
SWE-bench Scores: The Gap Between Passing Tests and Shipping Software
What SWE-bench measures, why OpenAI retired Verified, and how task selection, test quality and agent harnesses change a coding score.
6 min read
Synthetic Data
Synthetic Data: The Acceptance Rule Matters More Than the Label
A 200-example audit exposes the gap between plausible instructions and valid training records. What model-collapse research actually says about synthetic data.
6 min read
AI Reasoning
Test-Time Compute: The Verifier Is Part of the Business Model
More reasoning can improve a result without improving its economics. Historical studies show why allocation, verification and accepted-task cost belong together.
6 min read
Apple
Apple's M6 Macs: Buy Memory for Local AI, Not the Benchmark Headline
The new M6 MacBook Pro and iMac share an AI pitch, but their memory ceilings and software paths make the buying decision more specific.
8 min read
OpenAI
ChatGPT Pro and Business API Credits: Who Gets Them and How to Claim
OpenAI's new API credit benefit can provide $100, $200 or $500 for eligible Pro plans, or up to $500 pooled across paid Business seats. Check ownership, activation, expiry and billing before building.
8 min read
Anthropic
Haiku 4.5 Left Claude’s Model Menu: What Users and Developers Should Do
Anthropic removed Haiku 4.5 from Free, Pro and Max chat menus on October 29. Existing conversations continue on Haiku 5.5, while Claude Code and API users still have a choice.
7 min read
GitHub Copilot
GitHub Copilot's Cloud Handoff Changes Where Agent Work Runs, and Who Pays
GitHub's connected Copilot sessions move between devices, but cloud sandboxes now share an Actions compute allowance. Choose the execution environment before starting work.
8 min read
OpenAI
GPT-6.1 Terra Makes Screen Automation Cheaper to Try. The Pilot Still Has to Pass.
OpenAI's new Terra model costs $1 per million API input tokens and $5 per million output tokens. A bounded screen-workflow pilot can reveal the costs and errors that a benchmark cannot.
8 min read