Weekly 3×3: How hedge funds compete for AI talent. Inference architecture disaggregates. Agent evals focus on reliability.
Hedge funds compete for AI talent as inference architecture becomes increasingly disaggregated and agentic AI evaluation shifts from capability to reliability.
MARKETS
Hedge funds turn AI expertise into a competitive hiring advantage
Top quantitative and multi-strategy hedge funds are increasingly competing for AI talent. Millennium has introduced a dedicated AI track within its internship programme, embedding AI engineers into investment teams rather than treating AI as a separate technology function. Read more: Business Insider
Point72 applies AI to weather forecasting for commodity trading
Point72 has hired AI specialist Alex Alifimoff to lead weather prediction efforts for its global macro business, applying machine learning to improve commodity and macro trading decisions. Read more: Financial News London
AI financing enters a new phase of market scrutiny
The Bank of England highlighted growing financial system exposure to AI through three channels: rising valuations of AI-related companies, increasing leverage linked to AI investment, and the operational risks created as financial institutions deploy AI systems. Read more: Bank of England
MACHINES
AMD and Cerebras target ultra-low-latency AI inference
AMD and Cerebras have announced a partnership combining AMD's rack-scale systems with Cerebras' wafer-scale architecture to optimise AI inference workloads around latency, throughput, and efficiency. Read more: Cerebras.ai
Cerebras expands AI compute footprint in Europe
Cerebras announced plans to build 200MW of European AI compute capacity by the end of 2027, reflecting the growing competition for regional AI infrastructure and energy capacity. Read more: Cerebras.ai
AI runtime infrastructure emerges as a new software layer
New research is exploring "AI runtime infrastructure" — systems that monitor, optimise, and intervene during agent execution to improve reliability, latency, token efficiency, and safety. Read more: arxiv.org/abs/2603.00495
MODELS
AI agents learn to become more efficient by doing less
A recent paper proposes that coding agents should estimate the minimum amount of information required before executing a task, rather than repeatedly exploring unnecessary context. The approach introduces the idea of "minimum-sufficient execution", allowing agents to expand their reasoning only when required. Read more: arxiv.org/abs/2607.13034
Agent evaluation begins to focus on reliability
Research into autonomous systems is increasingly focusing on monitoring, alignment, and behavioural control. The challenge is moving from "can the model complete a task?" to "can the system reliably operate over long horizons without unexpected behaviour?" Read more: arxiv.org/abs/2605.24216
Agentic AI adoption accelerates
Research using OpenAI Codex usage data suggests that agentic AI is moving into more and more workflows, with users deploying multiple agents and assigning more complex tasks. Read more: arxiv.org/abs/2606.26959
