Luis Serrano views transformer attention not as a search table, but as words exerting gravity to pull each other across embedding dimensions. Physicist Ricardo Di Sipio extended this analogy, showing that attention matrices mathematically calculate geodesics in curved space-time where contextual words bend the geometry around them.
As generative AI collapses the cost of producing code, software development risks drowning in unmaintainable spaghetti code. Aishwarya Srinivasan notes that while anyone can build a prototype in ten minutes, engineering moats remain anchored in traditional fundamentals, intentional architecture, access security, and knowing what problem is actually being solved.
Delegating work to autonomous sub-agents increases throughput but inflates the total workload because sub-agents execute tasks with less token discipline. Ish Shaw explains that Jevons paradox applies here: falling per-token prices cannot offset the massive token inflation caused by branching agent swarms.
Management skills like task delegation, defining acceptance criteria, and handling misinterpretation transfer directly to directing AI agents. Katie Malone notes that while traditional managers find these supervision dynamics natural, former individual contributors often struggle with the constant reviewing and loss of flow.
Unlike turn-based models that idle until prompted, continuous AI assistants run in the background evaluating calendars, messages, and emails. Delani Kahawala explains that assistants must handle multi-agent fan-outs while concurrently queuing conversational interruptions when an urgent real-world event supersedes an active voice dialogue.
Evaluating agentic systems requires moving past LLM text benchmarking to assess entire execution pipelines under uncertainty. Aishwarya Srinivasan emphasizes that non-determinism compounds across tool calls, API logins, and loops, meaning every system integration point introduces unique failure modes that require custom, end-to-end evaluation.