Applied AI & Agent Systems

Cheaper models move the bottleneck upward: value sits in the layer that turns capability into supervised work across tools and systems.

For document-heavy and monitoring-intensive work, one experienced professional can supervise a small agent team—provided the verification layer is real.

MIT, McKinsey, and BCG each find that only a small minority of firms generate material AI returns. Their advantage is organisational.

Frontier models now perform within a narrow band of each other on reasoning benchmarks. The remaining quality gap in agentic AI is not in the models; it is in the human capacity to brief them properly.

MIT research finds 95% of enterprise AI pilots show no measurable P&L impact, the Bank of England has named AI a systemic financial-stability risk, and reliability math shows a 20-step agent workflow succeeds only 36% of the time. The evidence does not suggest caution; it demands it.

An agent can maintain the archive. The reader still has to reconstruct, challenge and apply what it contains.

AI leaders turn each deployment into reusable institutional capability. Laggards keep funding isolated pilots that must rebuild the path to production.

Perplexity’s Portable Computer moves routine agent work onto owned hardware and sends only the difficult tail to frontier models. That makes cloud inference more selective, not less valuable.

Sparse models promise broad capacity at bounded compute, but corporate value depends on memory, explicit controls, and measured customisation.

MIT research finds 95% of enterprise AI pilots show no measurable P&L impact, the Bank of England has named AI a systemic financial-stability risk, and reliability math shows a 20-step agent workflow succeeds only 36% of the time. The evidence does not suggest caution; it demands it.

Frontier models now perform within a narrow band of each other on reasoning benchmarks. The remaining quality gap in agentic AI is not in the models; it is in the human capacity to brief them properly.

MIT, McKinsey, and BCG each find that only a small minority of firms generate material AI returns. Their advantage is organisational.

Finding an old note saves time. Bringing several notes into a useful new relationship creates something that was not in the archive before.

The useful change in Karpathy’s LLM wiki is not retrieval. The agent maintains the synthesis so the archive can compound between sessions.

OpenClaw’s architecture supports the operating-system metaphor. Its isolation model, economics, and evidence base are not there yet.

The early labour-market effect is visible in fewer hires rather than mass layoffs. That saves cost now and weakens the future talent pipeline.

For document-heavy and monitoring-intensive work, one experienced professional can supervise a small agent team—provided the verification layer is real.

Cheaper models move the bottleneck upward: value sits in the layer that turns capability into supervised work across tools and systems.

A permanent spending cap mistakes the agent for the unit of risk. Authority should attach to a validated workflow and reset when that workflow changes.

A model catalogue can shape a hardware decision before procurement begins. NVIDIA’s proposed acquisition puts the engineering behind those defaults in focus.

Hundreds of agents coordinated a real intrusion. They succeeded through shared infrastructure, exposed credentials and safeguards that were absent or bypassed.

Fable 5 leads on capability but has yet to become a default business model. The constraint is the shrinking value of intelligence above the required threshold.

OpenAI did not merely report a cyber-capability result for Astra. It disclosed the controls it believes must surround the model.

Local models are improving while memory suppliers favour server products. That combination may divide AI access between hardware owners and renters.

Uber did not reprice by raising fares alone. It repriced by moving the take rate, quietly, over four years. The AI labs filing to go public have the same lever and a better cost curve.

OpenAI cut GPT-5.6 Luna by 80% three weeks after launch. The move changes buyer economics, but it does not disclose the cost of serving an agent.

A model scoring 52 on the Artificial Analysis index now runs on hardware you can buy outright, under a licence nobody can revoke. The question has moved from capability to price.

Western markets keep discovering the same thing every four weeks. Chinese labs ship frontier open-weight models on a cadence, and the surprise itself is now the story.

Commercial frontier models refused the attack payloads Hugging Face needed to analyse. GLM 5.2 completed the work on the defender's own hardware.

Productive work with agents comes from a well-managed loop: preserve context, iterate in small steps, and hold the scope.

Useful agent work depends on context, scope, checkpoints, and clear standards—the ordinary disciplines of management.

An invitation-only UK pilot for a managed personal AI assistant, designed around ordinary Telegram conversations.

A private Markdown research desk with dossiers, Pathfinder topic trails, citation exports, and source-grounded answers.

A reproducible control plane for my private two-node NVIDIA DGX Spark cluster, serving one local DeepSeek model for bounded agent work.

A private household assistant for family logistics, with local memory, Telegram access, and approval before it acts outside the home.

My self-hosted OpenClaw agent: local memory, explicit permissions, stable project workspaces, and bounded delegation.

A governance blueprint for an enterprise AI chief of staff, with secure context, bounded memory, and approval-gated action.