The Sequence Radar - Issue #958: Last Week in AI: Cheap Intelligence, Expensive Ambitions
Next Week in The Sequence:
TL;DR
- DeepSeek released DeepInfer, an open-source inference framework designed to maximize output from expensive AI hardware.
- NVIDIA announced DSX POD, a modular design for AI factories that drastically reduces deployment time.
- Meta reported strong revenue growth, attributing it to AI improvements in advertising, while also detailing massive capital spending on AI infrastructure.
- SpaceX's AI division, SpaceXSI, showed significant revenue growth, surpassing its Starlink business, despite heavy AI capital expenditure.
- Legora raised $550 million for its legal AI software, indicating a focus on workflow automation and cost savings for customers.
- New AI research includes self-evolving coding agents (Self-HPO), co-adaptive language models (Cola2), unified perception models for autonomous vehicles (Foundation4D), and autonomous research systems for 3D Gaussian Splatting (SplatLab).
- Other research areas explored include synthesizing computer-use agent environments (WildAWM), evaluating Retrieval Augmented Generation compositionality (LASAGNA), advancing coding in low-resource languages, and using LLMs for large-scale formal software verification.
- Recent AI tech releases include DeepSeek DeepInfer, NVIDIA's Run:ai open source platform, Google Cloud's Agent Compute Engine, and NVIDIA's Nemotron 3.5 Super VL and Bolt edge models.
- Key AI news includes major funding rounds for Legora, Warp Sciences, OpenAI, and DeepSeek, the release of Recursive's Flint coding model, OpenAI's acquisition of Orchestra Bio, and significant compute deals between Nscale and Anthropic, and SpaceX's AI growth.