AI & Agents
AI tooling is past adoption and into divergence — outcomes now depend on structural context, not tool access.
Signals
Data from Laura Tacho's research with DX (referenced by Martin Fowler on 2026-02-25) shows 92.6% of developers are now using AI assistants. Developers self-report saving 4 hours per week. 27% of code is written by AI without significant human intervention.
At 92.6% penetration, the frontier has moved from 'will teams use AI' to 'which organizational conditions make AI an amplifier vs. a liability.' The validation/provenance deficit is actively accumulating in production codebases.
Rahul Garg on Martin Fowler's site identifies a 'frustration loop' in AI coding. Knowledge priming: supplying seven context elements. Without priming ~45min of fixes; with priming ~5min.
The default failure mode of AI coding tools is context collapse. Teams treating priming as versioned infrastructure vs throwaway prompting get measurably different outputs.
InfoQ reported OpenAI launched GPT-5.3-Codex-Spark on Cerebras wafer-scale chips — first production model not on Nvidia GPUs.
OpenAI on non-Nvidia silicon breaks the assumption frontier inference is Nvidia-locked. Constraint arbitrage: Cerebras accepts manufacturing complexity for latency advantages.
IEEE Spectrum: Fields Medal winner Viazovska's proofs formally verified through human-AI collaboration in two weeks.
AI-assisted formal verification shifts the verification capacity bottleneck. Early infrastructure closing the generation-verification gap.
The New Stack and The Verge reference OpenAI replacing Anthropic at the Pentagon. Framing implies policy modifications to secure contract.
Constraint arbitrage: procurement rewards firms accepting looser safeguards. Long-run risk is a race to minimum viable safety standard in high-stakes verticals.
Control Surfaces
| Lever | Status | Change | Evidence |
|---|---|---|---|
| AI coding penetration | 92.6% of devs | Near saturation | DX/Tacho via Fowler |
| AI code share | 27% without review | Rising | DX/Tacho via Fowler |
| Compute lock-in | OpenAI on Cerebras | First Nvidia bypass | InfoQ |
| Defense AI posture | OpenAI in, Anthropic out | Divergent safety rewarded | The New Stack |
Watchlist
- ConfirmationCerebras/Codex-Spark independent benchmarks
- InvalidationAnthropic wins defense contracts under existing policies
- Observable.cursor/priming.md adoption on public GitHub repos
Falsifiers
- AI incident rates don't differ by org practice quality
- Cerebras claims fail benchmarking
- OpenAI Pentagon deal didn't require safety policy changes
Key Unknowns
- Actual defect rate differential structured vs ad hoc AI adoption
- Knowledge priming adoption rate across teams
- Pentagon contract policy modifications
Noise Filter
- Running Qwen3.5 27b on 2x3090— Hobbyist benchmarking
- Jan-Code-4B— Incremental model release
- RLC Pro enterprise Linux— Product announcement, off-pillar
Get The Signal daily
Cross-domain structural analysis, delivered every morning.