The publication index
Topics
From agents to inference, follow the recurring themes across Daily Pulses and Weekly Reviews.
Agents
A typed evaluator matched five labels across 500 repeated decisions
Evaluation
OpenAI found up to 14% PPA in AMD GPU IP. Its own benchmark shows why specs are release gates
Security
RBA liaison finds AI use near universal, but only 34% report moderate or significant gains
Inference
Two engineers rewrote a 20M-RPS service with AI. The hard part was knowing when to rewrite.
AI Engineering
Gemini 4 Argon's coding score needs a benchmark audit
Models
The system frontier moved more than the model frontier
AI Infrastructure
Australia wants time-bound reporting after an AI agent crossed the boundary
Memory
Agent governance is becoming the real control plane
Agent Evaluation
A 42.9% actor trained an 82.9% policy by judging the steps after the run
AI Governance
OpenAI found up to 14% PPA in AMD GPU IP. Its own benchmark shows why specs are release gates
Agent Engineering
OpenAI found up to 14% PPA in AMD GPU IP. Its own benchmark shows why specs are release gates
Agent Security
Google found TODO comments triggered unrequested coding-agent work
AI Security
Enveda's reviewer surfaced all 11 known batch issues and held no release authority
Human Oversight
Cursor says a backdoored Claude Code extension reached up to 5,000 users
Post-training
Agent governance is becoming the real control plane
Agent Operations
vLLM 0.32 changed four contracts at once
AI Safety
Australia wants time-bound reporting after an AI agent crossed the boundary
Applied AI
A 42.9% actor trained an 82.9% policy by judging the steps after the run
Model Lifecycle
Gemini 4 Argon's coding score needs a benchmark audit
Observability
Compaction left an estimated 10% of eligible Codex traffic unmonitored
Agent Architecture
OpenAI priced Ultrafast at six times standard. Budget the whole task
AI Architecture
NVIDIA put the agent watchdog outside the host
Distributed Systems
ChatGPT Wallet makes approval a state, not a completed purchase
Model Migration
Gemini 2.5 will reroute on December 1. Treat it as a system migration
Reliability
Changing an agent’s tools now has a cache bill and a replay contract
Reliability Engineering
OpenAI found agents suppressing alerts and removing a monitor
Agent Harnesses
Gemini 2.5 will reroute on December 1. Treat it as a system migration
Agent Identity
Anthropic now bills some refusals with zero output tokens
Agentic Commerce
ChatGPT Wallet makes approval a state, not a completed purchase
AI Applications
Gemini Live can keep talking while its tools are still running
AI Economics
OpenAI's four-model shutdown hides three migrations behind one replacement
AI-Assisted Software Delivery
AWS Emulator gives coding agents a fast loop, but IAM stays outside the proof
API Compatibility
AWS Emulator gives coding agents a fast loop, but IAM stays outside the proof
Context Engineering
Kimi-K3 solved 94% of workflows once. Only 13% survived every run
Cybersecurity
Amazon's VLM grader reached 94% accuracy only after OCR and a small classifier
Decision Systems
Compaction left an estimated 10% of eligible Codex traffic unmonitored
Developer Infrastructure
OpenAI says Pando cut median push latency nearly 80% after relaxing repository-wide consistency
Document AI
ChatGPT Wallet makes approval a state, not a completed purchase
Document Intelligence
Enveda's reviewer surfaced all 11 known batch issues and held no release authority
Enterprise AI
AI teams are starting to measure work, not usage
Formal Verification
Australia wants time-bound reporting after an AI agent crossed the boundary
Inference Infrastructure
vLLM 0.32 changed four contracts at once
MCP
GPT-Rosalind starts billing today. Price the whole scientific run
Measurement
RBA liaison finds AI use near universal, but only 34% report moderate or significant gains
Model Governance
OpenAI found agents suppressing alerts and removing a monitor
Model Routing
AWS Emulator gives coding agents a fast loop, but IAM stays outside the proof
Open-Weight Models
OpenAI's cyber pipeline reached root 200 times. Twenty used the wrong zero-day
Production AI
GPU count moved the benchmark. Zephon made the experiment identifiable
Testing
AWS Emulator gives coding agents a fast loop, but IAM stays outside the proof
Adoption Measurement
AI improved patent drafts. Junior judgment did not improve on average
Agent Infrastructure
Meta says tests designed to fail helped catch 74 serious privacy bugs
Agent Memory
Google found TODO comments triggered unrequested coding-agent work
Agent Monitoring
Compaction left an estimated 10% of eligible Codex traffic unmonitored
Agent Platforms
Gemini 4 Argon chains 64K hops. Count the repeated context
Agent Safety
OpenAI found agents suppressing alerts and removing a monitor
Agentic Research
GPU count moved the benchmark. Zephon made the experiment identifiable
AI Agents
Changing an agent’s tools now has a cache bill and a replay contract
AI Coding Agents
Anthropic replaced Git after merges hit 30,000 a day
AI Coding Tools
Cursor says a backdoored Claude Code extension reached up to 5,000 users
AI for Science
Kimi-K3 solved 94% of workflows once. Only 13% survived every run
AI Hardware
OpenAI found up to 14% PPA in AMD GPU IP. Its own benchmark shows why specs are release gates
AI Memory
Google says user devices will hold the keys to persistent AI memory
AI Operations
A 42.9% actor trained an 82.9% policy by judging the steps after the run
AI Platforms
The agent harness is becoming a managed runtime, and policy is moving into the control plane
AI Productivity
RBA liaison finds AI use near universal, but only 34% report moderate or significant gains
AI-Assisted Testing
Meta says tests designed to fail helped catch 74 serious privacy bugs
API Economics
Anthropic now bills some refusals with zero output tokens
API Migration
OpenAI's four-model shutdown hides three migrations behind one replacement
API Operations
OpenAI found agents suppressing alerts and removing a monitor
API Reliability
Gemini 4 Argon chains 64K hops. Count the repeated context
Authorization
Google found TODO comments triggered unrequested coding-agent work
Autonomous Agents
OpenAI gave its always-on agent a read-only research mode
Benchmarking
Gemini 4 Argon's coding score needs a benchmark audit
Browser Agents
Gemini 2.5 will reroute on December 1. Treat it as a system migration
Business Workflow Automation
Ironclad turned 11 workflows into agent tests. The rubric is the release artifact
Capacity Planning
OpenAI returned to shared storage after isolation made every team a capacity planner
Causal Attribution
Fin's 8.6% hard-resolution lift needs a feedback denominator
Cloud Sandboxes
Claude moved Cowork's execution boundary. Rebuild the trust map
Coding Agents
Google found TODO comments triggered unrequested coding-agent work
Computer Use
Ironclad turned 11 workflows into agent tests. The rubric is the release artifact
Consistency Models
OpenAI says Pando cut median push latency nearly 80% after relaxing repository-wide consistency
Context Compression
Gemini 4 Argon chains 64K hops. Count the repeated context
Context Retrieval
Anthropic replaced Git after merges hit 30,000 a day
Contract Operations
Ironclad turned 11 workflows into agent tests. The rubric is the release artifact
Cost Optimization
OpenAI's kill switch can take 35 seconds because authorization is cached
Customer Support AI
Fin's 8.6% hard-resolution lift needs a feedback denominator
Data
AI changed the build cost. Verification now sets the architecture
Data Architecture
OpenAI returned to shared storage after isolation made every team a capacity planner
Data Governance
Claude moved Cowork's execution boundary. Rebuild the trust map
Desktop Integration
Claude moved Cowork's execution boundary. Rebuild the trust map
Distillation
Gemini 4 Argon's coding score needs a benchmark audit
Document Processing
OpenAI priced Ultrafast at six times standard. Budget the whole task
Failure Taxonomy
Anthropic now bills some refusals with zero output tokens
Feedback Systems
Fin's 8.6% hard-resolution lift needs a feedback denominator
Human Factors
Two engineers rewrote a 20M-RPS service with AI. The hard part was knowing when to rewrite.
Human-AI Collaboration
AI improved patent drafts. Junior judgment did not improve on average
Identity and Authorization
Meta’s agent handed calls to people. Delegation needs its own control plane
Inference Economics
OpenAI priced Ultrafast at six times standard. Budget the whole task
Inference Serving
OpenAI says Pando cut median push latency nearly 80% after relaxing repository-wide consistency
Inference Systems
ReplaySSM cuts state writes, not state reads
Information Retrieval
Claude moved Cowork's execution boundary. Rebuild the trust map
Key Management
Google says user devices will hold the keys to persistent AI memory
Latency Engineering
OpenAI priced Ultrafast at six times standard. Budget the whole task
LLM-as-a-Judge
Anthropic’s test agent filed a false homicide tip. Evaluation needs a blast radius
Local Inference
Anthropic now bills some refusals with zero output tokens
Long-Context Inference
Gemini 4 Argon chains 64K hops. Count the repeated context
ML Experimentation
GPU count moved the benchmark. Zephon made the experiment identifiable
MLOps
A typed evaluator matched five labels across 500 repeated decisions
Model Architecture
Chain-of-thought can mislead the monitor, trust is moving to external evidence
Model Deprecations
OpenAI gave its always-on agent a read-only research mode
Model Economics
GPT-Rosalind starts billing today. Price the whole scientific run
Model Evaluation
Gemini 4 Argon's coding score needs a benchmark audit
Model Integrity
Cursor says a backdoored Claude Code extension reached up to 5,000 users
Model Migrations
Anthropic replaced Git after merges hit 30,000 a day
Model Post-Training
Fin's 8.6% hard-resolution lift needs a feedback denominator
Model Testing
Ironclad turned 11 workflows into agent tests. The rubric is the release artifact
Multimodal Evaluation
Amazon's VLM grader reached 94% accuracy only after OCR and a small classifier
Performance Engineering
Gemini 2.5 will reroute on December 1. Treat it as a system migration
Policy
The personal-agent OS has arrived, and the security boundary sits outside the model
Policy Enforcement
NVIDIA put the agent watchdog outside the host
Privacy Engineering
Google says user devices will hold the keys to persistent AI memory
Process Compliance
Compaction left an estimated 10% of eligible Codex traffic unmonitored
Product
AI changed the build cost. Verification now sets the architecture
Production Evaluation
Fin's 8.6% hard-resolution lift needs a feedback denominator
Professional Expertise
AI improved patent drafts. Junior judgment did not improve on average
Prompt Caching
Changing an agent’s tools now has a cache bill and a replay contract
Quantization
ReplaySSM cuts state writes, not state reads
Reasoning Budgets
Gemini 4 Argon chains 64K hops. Count the repeated context
Regulated AI
Enveda's reviewer surfaced all 11 known batch issues and held no release authority
Reinforcement Learning
A 42.9% actor trained an 82.9% policy by judging the steps after the run
Release Engineering
vLLM 0.32 changed four contracts at once
Research Reproducibility
GPT-Rosalind starts billing today. Price the whole scientific run
Retrieval
Siemens put about 50 rules behind its lead-qualification agent
Runtime Isolation
NVIDIA put the agent watchdog outside the host
Scientific AI
GPT-Rosalind starts billing today. Price the whole scientific run
Scientific Evaluation
RBA liaison finds AI use near universal, but only 34% report moderate or significant gains
Scientific Workflows
Amazon's VLM grader reached 94% accuracy only after OCR and a small classifier
Software Engineering
Gemini 4 Argon's coding score needs a benchmark audit
Software Reliability
Meta says tests designed to fail helped catch 74 serious privacy bugs
Software Supply Chain
Cursor says a backdoored Claude Code extension reached up to 5,000 users
Software Testing
OpenAI found up to 14% PPA in AMD GPU IP. Its own benchmark shows why specs are release gates
Source Control
Anthropic replaced Git after merges hit 30,000 a day
Speculative Decoding
ReplaySSM cuts state writes, not state reads
State-Space Models
ReplaySSM cuts state writes, not state reads
Stateful Testing
Ironclad turned 11 workflows into agent tests. The rubric is the release artifact
Synthetic Data
ExtractBench Pro found the documents your average score hides
System Design
Two engineers rewrote a 20M-RPS service with AI. The hard part was knowing when to rewrite.
Time-Series Forecasting
ExtractBench Pro found the documents your average score hides
Tool Calling
Meta says tests designed to fail helped catch 74 serious privacy bugs
Trusted Execution
Google says user devices will hold the keys to persistent AI memory
Voice Agents
ExtractBench Pro found the documents your average score hides
Voice AI
Gemini Live can keep talking while its tools are still running
Workforce Design
AI improved patent drafts. Junior judgment did not improve on average