AI reliability depends on the systems around models

Major tech companies are pouring unprecedented capital into AI infrastructure, with The Kobeissi Letter projecting combined spending by Alphabet, Amazon, Meta, Microsoft, and Oracle at around $1.1 trillion by 2027, or about 3.2% of the US GDP. Yet McKinsey research shows that while roughly 65% of organizations are integrating generative AI into workflows, only a small number have seen tangible economic impact.

The gap reflects a shift away from treating bigger foundation models as the main route to progress. Successful deployments increasingly depend on operational layers around models, including retrieval systems, verification cycles, memory formats and reusable skills that adapt raw model capabilities to specific business settings.

Mikhail Arbuzov’s work on “patch-local” reliability frames dependable AI as something engineered within bounded task domains rather than expected to emerge uniformly from larger general-purpose systems. Skillfed, co-created by Arbuzov and Sisong Bei, gives AI agents a way to search for task-specific tools and was downloaded more than 1,000 times in its first week. Internal tests found that Claude Opus 4.5 improved by more than 20 percentage points when given relevant skills, while Claude Opus 4.6 improved by roughly 30% relative to its no-skill baseline when searching a library of about 26,000 real-world skills.

Andrew Ng, Tobi Lütke, Andrej Karpathy and Satya Nadella are cited as backing a similar direction: progress comes from agent workflows, context engineering and changes to business processes. The competitive advantage is shifting toward companies that build the adaptation layer first.