Research & Deep Dives
Research papers, reports, scenarios, and technical findings translated into practical implications for builders and operators.
Open Research & Deep DivesA study of ChatDev traces finds that agentic coding systems spend most tokens on review and repeated context passing, not initial code generation. The finding shifts optimization toward handoffs, critique loops, and context efficiency.
A Perplexity field study compares conversational search with autonomous agent execution and finds large time savings, lower dissatisfaction, and broader task scope. The comparison shows why autonomy changes both the economics and ambition of knowledge work.
A consumer research paper finds that LLMs can match human purchase-intent surveys when they answer in free text before being mapped back to Likert ratings. The result favors better survey design over treating synthetic respondents as wholesale human replacements.
A DeepMind report argues that human-level AGI may be followed by several pathways toward superintelligence, each constrained by data, compute, embodiment, and coordination bottlenecks. It reframes AGI as an intermediate capability threshold rather than a stable endpoint.
A post-training paper shows how interpretability tools can audit preference data, expose unwanted learning signals, and reshape rewards before models absorb bad habits. The method turns interpretability into a practical intervention in data curation and reward design.
Self-Harness shows that agents can use execution traces, targeted edits, and regression tests to improve the scaffolding around their own model. The important constraint is preserving task performance while the runtime changes itself.
Paul David's dynamo analogy explains why general-purpose technologies can spread through business before appearing clearly in productivity statistics. The lesson is to distinguish implementation lag from evidence that the technology lacks value.
A paper separates the model that improves an agent's harness from the model that must use it. Cheap models may propose useful scaffold changes, but task performance still depends on an executor capable enough to load and follow them.