Microsoft's Overhauled Copilot Exposes How Far the Super App Race Has Left Every Competitor Behind

The Day's Thesis

▶

Signal of the Day: The UK's AI Security Institute found that standard benchmarks underestimate frontier AI agent capabilities by ~60%, with success rates on software engineering tasks jumping 25 percentage points when token budgets increase tenfold.

The 30-Second Read:

• UK AISI: frontier AI progress is ~60% steeper than benchmarks show, reshaping capability timelines for every institutional risk model
• Bridgewater and Mira Murati's Thinking Machines Lab report 84.7% accuracy on private financial benchmarks with a Qwen3-235B fine-tune — at one-fourteenth the cost of GPT/Claude
• Canada commits C$500M to Red Chris copper-gold mine expansion; copper closes the week at US$6.22/lb, up 2.59%
• Epoch AI: June 2026 saw ~1,500 high-severity CVEs reported, 3.5× the prior monthly record, driven by AI bug-hunting programs

The day's through-line is systematic mismeasurement: AI capabilities are being underestimated by benchmarks, AI costs are being undercut by fine-tuned open-weight models, and AI-driven security tooling is surfacing vulnerabilities at a pace that compliance infrastructure was not built to absorb. Copper's rally and Canada's mine capex — critical infrastructure for data centers and EVs alike — ground this measurement problem in physical supply.

---

AI & Research Frontier

Standard AI benchmarks are systematically wrong, and the error margin is large enough to invalidate most published capability timelines.

The UK's AI Security Institute studied seven benchmarks and found that compute-capped evaluations — token budget (the ceiling on how much computation a model can use per task) constrained — understate actual agent performance by roughly 60% at the frontier.

A tenfold token budget increase pushed software engineering task success rates up ~25 percentage points. The practical consequence: every enterprise risk model, regulatory framework, and competitive roadmap built on published benchmark scores has been calibrated against a floor, not a ceiling.