TLDRocket
4 August 2026
The AI industry is grappling with a fundamental tension between capability and control. On one front, organizations are drowning in R&D waste—over a third of companies fritter away 25 to 40 percent of their development budgets on projects that never ship, with failed initiatives costing over a million dollars median—yet they're deploying AI mainly for execution tasks like data analysis rather than the early ideation phase where intelligence could prevent disaster. Meanwhile, the model-building stack is splintering. Alibaba's Qwen 3.8-Max launched with 2.4 trillion parameters and a 1 million token context window, but developers are skeptical about whether promised open weights materialized and whether self-benchmarked performance claims hold up to third-party scrutiny. Anthropic, meanwhile, is vertically integrating application layers alongside its models, pushing inference margins from 38–40 percent to over 70 percent and forcing specialized agent labs to consider building their own models just to stay competitive. The practical layer is moving faster: Nvidia's NOOA collapses agent development into single Python classes, Liquid AI's 2.6-billion-parameter LFM2.5 runs capable agents locally on consumer devices, and Cloudflare's open-sourced computer package gives agents their own virtual filesystem—solving an industry-wide shortage of compute capacity for concurrent agents. Yet none of this solves the deeper problem: as Daphne Koller points out, AI excels at molecular design but can't fix drug discovery's real bottleneck, which is knowing which disease mechanisms are worth targeting. Domain expertise still amplifies what models can do; they don't replace it.
Read the full briefing →