Techmeme
Cursor deployed Composer 2.5, a new coding model trained with reinforcement learning on text feedback to handle longer-horizon reasoning tasks. The training budget reached ~6.3T total compute with 200B active parameters trained on ~56T tokens—a frontier-scale investment comparable to major model labs and a significant competitive positioning move in the coding assistant market.
Separately, Cloudflare's Mythos model excelled at chaining multiple vulnerabilities into working proof-of-concept exploits across 50+ repositories, outperforming other frontier models at multi-bug attack chains. The findings reportedly influenced Anthropic's decision to rewrite the Bun JavaScript runtime from Zig to Rust for memory safety, signaling how AI-powered vulnerability discovery is reshaping infrastructure security priorities and language choices at major projects.
Azeem Azhar, Exponential View
A CloudZero survey of 500 engineering managers found that 71% of enterprises exceeded their AI budgets in 2025, with token consumption exploding 17,000x over four years. Uber and ServiceNow depleted their entire 2026 token budgets within four months, while average monthly AI spend at large US enterprises grew 36% to $85,000 year-over-year. The proportion of organizations planning to invest over $100,000 per month in AI tools more than doubled from 20% in 2024 to 45% in 2025.
Blackstone's President Jon Gray reported that LLM spend among the company's 270 portfolio companies was up 15-fold in Q1 2025 versus Q1 2024, though implementation remains slow. Gray noted the spend "is off a small base," but the 15x growth across transaction processing, legal, and rules-based businesses signals that token anxiety is becoming a structural cost management problem across industries. Only 51% of organizations strongly agree they can track AI ROI effectively, despite 91% claiming overall confidence—a gap that underscores the need for better cost attribution and visibility systems.
Lenny's Newsletter
Token abundance is enabling a fundamental shift in how AI-augmented development workflows allocate resources. According to reporting from Lenny's Newsletter, only approximately 1% of generated tokens reach production, with the majority spent on custom interfaces, interactive dashboards, and living design systems encoded in HTML that improve process legibility without shipping to users. This signals a structural reallocation toward human-AI collaboration infrastructure rather than final product code.
![AINews] How to land a job at a frontier lab (on Pretraining](https://www.latent.space/p/ainews-how-to-land-a-job-at-a-frontier?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2e69d902-1d29-4e8c-834c-41e83b07223f_1194x604.png&open=false){kind=link}
