Cloud Dialogues

(apologies - audio is not up to our usual standard, but this version is the best I could get it)
Guest: Lena Hall (Senior Director of Developer Experience at Akamai, formerly DevRel lead at Amazon Web Services and AI/Data Advocacy Director at Microsoft) joins Georgia and Matt to unpack the latest AI developments and what they mean for how we build software.

News Highlights

The episode kicks off with several major AI updates. The US government excluded Anthropic from a supplier list over concerns around WMD and surveillance policies - only for OpenAI to sign a government deal shortly after with similar language, raising questions about whether the decision was policy-driven or political.

Meanwhile, OpenAI released GPT-5.4, a reasoning-focused "thinking model" with tunable reasoning depth. Early feedback suggests stronger accuracy and less verbosity, though it consumes more tokens and is slightly more expensive to run.

The hosts also discuss a pledge from Meta, Microsoft, Google, and Amazon to fund new electricity generation to support AI infrastructure - a move framed as sustainability but widely seen as a practical response to AI's growing energy demand.

Finally, a new partnership between CVS Health and Google Cloud highlights a broader shift in hyperscaler strategy: AWS continuing to focus on horizontal infrastructure while Google invests more heavily in vertical AI solutions such as healthcare.

The Core Discussion: Engineering in the Age of AI

The main conversation explores how AI systems fundamentally challenge traditional software engineering practices.

Unlike deterministic systems, AI outputs exist on a spectrum of quality. A system may be operationally healthy yet still produce incorrect or harmful responses, creating a new category of production issues that are harder to detect and diagnose.

Lena argues that while organizations don't necessarily need entirely new AI platform teams, platform engineering must evolve. Teams need infrastructure for AI observability, evaluation frameworks, fallback mechanisms, and intervention controls. Without this foundation, individual product teams end up solving the same problems repeatedly.

A key takeaway is the need for clearer responsibility across three types of AI failure: capability issues owned by product teams, safety risks defined by leadership, and operational reliability managed by platform engineering.

The group also emphasizes the importance of product-level evaluation, focusing not just on model benchmarks but on whether AI actually works for real users. Effective evaluation frameworks measure capability, safety, and operational reliability, with scrutiny increasing for higher-risk applications.

For organizations adopting AI, Lena recommends a gradual approach: start with assistants for narrowly defined tasks, move to supervised agents, and only introduce autonomous systems once observability and governance are mature.

AI and the Human Factor

The discussion ends with the impact of AI on developers themselves. Engineers are spending less time writing code and more time making high-level decisions about architecture, system behavior, and trade-offs. While this can increase productivity, it also raises cognitive load and shifts responsibility toward more experienced engineers reviewing large volumes of AI-generated output.

Cool AI Pick

Lena's pick is Codex Spark, an ultra-fast model designed for executing well-defined tasks. Her preferred workflow combines reasoning models like GPT-5.4 for planning, then handing execution to Spark - highlighting a broader trend toward specialized models working together in AI development pipelines.