Agents, Voice and Compute

Agent design takes centre stage as OpenAI explains faster voice interaction, while software orchestration and the economics of frontier compute sharpen the stakes.

Briefs

Agent Engineering

• Wheelhouse puts agent fleets to work — Steve Yegge’s technical write-up describes persistent agents planning, building, reviewing and operating his game. It offers one intensive project’s model, but depends on bespoke systems and costly model access.

• Harness engineering for self-improvement — Lilian Weng’s Harness Engineering for Self-Improvement argues that agents can improve through better memory, tools, workflows and tests. This is most convincing where tasks have clear checks, such as coding. Weak evaluators can still reward shortcuts and narrow behaviour.

Voice and Frontier Systems

• GPT-Live streams speech both ways — OpenAI’s engineering write-up says GPT-Live listens and speaks at once, while other models handle deeper reasoning and tools in the background. Its streaming design uses stateful inference, smaller context windows and faster media setup. OpenAI says silent production testing exposed failures that load tests missed, though its performance claims remain the company’s own.

• Frontier compute could cost far more — Dwarkesh Patel’s thought experiment argues that frontier compute could become far pricier if model revenue and capability outpace chip supply. That would favour labs with capital and reserved capacity. The case rests on uncertain assumptions about demand, supply and labour markets.

Search and Developer Tools

• Exa expands its web index — Exa’s index announcement says it now covers 80 billion pages and tracks 1.4 trillion URLs. The company sees this as a route towards Google-scale AI search. Its comparisons with other indexes are difficult to verify independently.

• Grok Build improves session visibility — The Grok Build release update improves model-picker feedback, refreshes change tracking after commits and reports fuller background-task logs. It is a focused maintenance update that makes coding-agent sessions easier to follow.

Technology Law

• OpenAI disputes Apple’s lawsuit claims — In its legal response, OpenAI disputes Apple’s account of a trade-secrets case involving former employees. OpenAI says Apple’s lawyers made inaccurate claims and did not raise the specific allegations before filing. The dispute shows how staff moves can expose gaps in access controls and handling of confidential material. These remain contested claims in active litigation.