Newsletter Archive

Evaluating frontier AI models has become prohibitively expensive: a single GAIA benchmark run costs $2,829, agent leaderboards spend $40K per evaluation round, and scientific ML architecture sweeps require thousands of H100-hours. Cost spreads of 33× on identical tasks reveal that evaluation methodology—not just model capability—now drives budget decisions.

Hugging Face · 17 min read Research

OpenAI, DeepMind, and Intel execs are signaling that inference compute is becoming the critical bottleneck in AI infrastructure, not training. CPU demand is surging as enterprises refresh aging hardware and shift budgets from GPUs to inference optimization — a five-to-six year cycle inflection point that could reshape how companies architect AI systems.

Latent Space · 10 min read Market

OpenAI has secured its target AI computing capacity in the US years earlier than planned, clearing the path for accelerated data center expansion and larger model training runs.

Bloomberg Tech · 1 min read Industry

A 21-day live deployment of 3,505 language-model trading agents on real Ethereum revealed that reliability emerges not from the base model but from the operating layer around it—prompt compilation, typed controls, policy validation, and execution guards. The system processed 7.5M agent invocations and $20M in trading volume with 99.9% settlement success, but only after targeted fixes (e.g., reducing fabricated sell rules from 57% to 3%) that text-only benchmarks never caught.

arXiv AI · 3 min (abstract) Research

Cursor released a public beta TypeScript SDK that gives developers programmatic access to its coding agents, enabling deployment via CI/CD pipelines, backend services, and embedded workflows rather than just interactive use in the IDE.

MarkTechPost · 6 min read Tools

SoftBank is spinning out a new AI and robotics company called Roze with a reported $100 billion valuation targeting a U.S. IPO, consolidating its bets in autonomous systems and artificial intelligence into a standalone public entity.

CNBC Tech · 1 min read Industry

US tech giants plan to spend $725 billion on capital expenditures this year, with the majority going to AI data center infrastructure. This massive spending surge reflects the industry's bet-the-company shift toward AI capabilities and has downstream implications for GPU scarcity, cloud pricing, and compute availability.

Bloomberg Tech · 1 min read Market

Australia's financial regulator (APRA) warned banks and superannuation trustees that AI governance is dangerously immature, finding boards lack understanding of model risks and over-rely on vendor claims. The warning follows a review of major institutions deploying AI in loan processing, fraud detection, and customer service—forcing financial firms to overhaul oversight practices.

AI News · 3 min read Policy