Frontier Notes / Daily Signal Report
Today
OpenAI's GPT-5.6 Soul now runs at 750 tokens/sec on Cerebras chips (18.5x speedup), backed by a $10B rental deal that pushes users to higher subscription tiers.
Editor's Notes
This week's videos highlight a shift toward specialized AI roles and tools: Cursor's forward deployed engineering model emphasizes co-development over staff augmentation, while Codex's advanced features (model selection, thread delegation, safety hooks) remain underutilized. Comparisons between GPT-5.6 Sol and Claude Fable 5 show speed vs. creativity trade-offs, and a new in-house AI consultant role is emerging as a high-paying career path.
Key Takeaways
OpenAI's GPT-5.6 Soul is offered at both 40-50 tokens/sec on GPUs and 750 tokens/sec on Cerebras chips (18.5x speedup), but the high capex of Cerebras hardware is justified by a $10B rental deal and a strategy to drain token allowances faster, pushing users to higher subscription tiers. The trade-off between speed and intelligence drives Frontier Labs to offer models that are both smarter and faster, while token efficiency emerges as a competing factor.
GPT-5.6 Sol and Claude Fable 5 were compared across multiple challenging tests including 3D printing, magazine creation, retro game development, and mixed reality games. GPT-5.6 generally finished faster and produced more functional results in some tests, while Claude Fable 5 showed stronger creative writing and game narrative. Neither model perfectly replicated the 3D design from reference photos, but both demonstrated impressive autonomous problem-solving.
Cursor's forward deployed engineering (FDE) team partners with enterprise customers to co-build AI-assisted software factories, focusing on high-impact, project-based engagements that deliver measurable ROI. The talk emphasizes hiring 'unicorns'—engineers with both deep technical skills and customer-facing experience—and avoiding staff augmentation in favor of collaborative co-development.
Codex offers advanced features like model selection (Soul, Terra, Luna), thread delegation, browser integration, skills, loops, and safety hooks that many users overlook. Properly configuring agents.md, using plugins like Zapier, and setting up pre-tool hooks can significantly improve productivity and prevent catastrophic file deletions.
A new high-paying AI career as an in-house AI consultant is emerging, replacing external AI agencies. The role involves auditing workflows, building automations, and solving bottlenecks, with judgment being more valuable than technical skills. Early adopters can create this role within their companies and earn significantly more.