World models get practical, storage gets tight, robotaxis hit the freeway

🧩 The Gist

OpenAI is rolling out GPT-5.1 with warmer, more capable conversation and new tone controls, paired with updated safety evaluations for sensitive use cases. At the same time, AI’s infrastructure footprint is biting the supply chain, with enterprise hard drives reportedly booked out for two years and hyperscalers pivoting to QLC SSDs to cope. On the applied side, Waymo began using freeways in Los Angeles, San Francisco, and Phoenix, which TechCrunch says could cut ride times by half. Researchers and builders shipped new world-model tooling and local-run apps, from a public launch of Marble to a JAX codebase for scalable training, while startups and enterprises showcased how AI is reshaping work and hiring.

🚀 Key Highlights

• OpenAI says GPT-5.1 is rolling out to paid users, with models that are warmer and more capable plus new ways to customize ChatGPT’s tone and style.
• An OpenAI system card addendum updates safety metrics for GPT-5.1 Instant and Thinking, including new evaluations for mental health and emotional reliance.
• Tom’s Hardware reports enterprise HDD capacity is effectively sold out for two years, with cloud providers switching to QLC NAND to avoid long lead times.
• Waymo robotaxis will take freeways in Los Angeles, San Francisco, and Phoenix, a change TechCrunch says could reduce ride times by up to 50 percent.
• World Labs introduced Marble, a frontier multimodal world model that is available to everyone starting today.
• Jasmine, a JAX-based world modeling codebase, claims order-of-magnitude faster reproduction on a CoinRun case study, reproducible training, and scaling from single hosts to hundreds of accelerators.
• Show HN: Gerbil is an open source desktop app for running local LLMs via llama.cpp, with integrations for Open WebUI, SillyTavern, ComfyUI, StableUI, and KoboldAI Lite.

🎯 Strategic Takeaways

• Infrastructure and supply
  • Storage is the new bottleneck, not just GPUs. Two-year nearline HDD backlogs and a pivot to QLC SSDs signal tighter planning for data-heavy training and retrieval workloads.
• Models and safety
  • Feature velocity is paired with governance. GPT-5.1’s rollout comes with fresh safety metrics for high-risk domains, a sign that evaluation coverage is expanding alongside capability.
• Autonomy in the wild
  • Freeway driving in three major metros suggests robotaxi operations are moving from constrained city grids to faster corridors, improving utility and rider perception.
• Open tooling and R&D
  • World models are getting easier to train and use. Public releases like Marble and performant codebases like Jasmine lower the barrier for robotics, simulation, and agent research.
• Enterprise adoption and trust
  • A retail case study from Neuro highlights time and cost savings with ChatGPT Business, while OpenAI’s stance against demands for private chat logs underscores rising expectations for data privacy.
• Talent and ecosystem
  • Hiring for forward deployed engineers at OmniAI, focused on AI-first lending infrastructure, shows demand for engineers who can ship integrations, OCR, and data aggregation in regulated workflows.

🧠 Worth Reading

• Jasmine: A Simple, Performant and Scalable JAX-based World Modeling Codebase
  The paper introduces a JAX codebase that emphasizes fully reproducible training, flexible sharding, and an order-of-magnitude speedup on a public case study compared to prior open implementations. For teams experimenting with world models, the practical takeaway is clear: standardized, scalable training infrastructure can accelerate iteration while keeping experiments reproducible.