Prompt-to-production ML, codemaps, and GPU Datalog point to pragmatic AI

🧩 The Gist

This week’s updates skew practical, with new tools that turn natural language into working ML systems and help developers actually understand code. Plexe pitches an AI data scientist that automates model building, evaluation, and deployment from a plain‑English brief. Cognition’s Codemaps aims to give humans and AI a shared, accurate view of code paths. On the platform side, OpenAI spotlights Brazil’s broad AI adoption, Apple leans on 3D Gaussian splatting for lifelike Personas, and a Datalog-on-GPU writeup breaks down data structures that unlock speed.

🚀 Key Highlights

• Plexe launches a platform that creates production ML models from natural language, then handles feature engineering, experimentation, evaluation, and deployment.
• The system runs experiments across architectures from logistic regression to neural nets, delivers error analysis and robustness testing, and supports monitoring with automatic retraining.
• Data connectors include Postgres, Snowflake, S3, and direct uploads, with deployments for online and batch inference using a multi‑agent pipeline.
• Cognition’s Codemaps generates a map of any codebase or snippet to illuminate code paths, offering consistent explanations so AI can teach users about the system.
• OpenAI highlights Brazil as one of the most engaged AI markets, citing usage across classrooms, farms, and small businesses for learning, creation, and innovation.
• A CNET piece notes Apple uses 3D Gaussian splatting for Vision Pro Personas and converting photos to 3D.
• A Substack breakdown of an ASPLOS ’25 paper explains GPULOG, which uses hash‑indexed sorted arrays to optimize Datalog on GPUs and aims to outperform engines like Soufflé.

🎯 Strategic Takeaways

• Builder productivity: Natural‑language model building plus automated evaluation and retraining can compress ML project cycles, shifting teams toward problem framing and data quality.
• Developer experience: Codemaps’ shared human‑AI understanding tackles code comprehension, making AI assistants more reliable by grounding them in explicit code paths.
• Compute and data structures: Purpose‑built GPU data layouts for logic engines, like hash‑indexed sorted arrays in GPULOG, signal a broader performance play for analytics and program analysis.
• Human interfaces: Apple’s use of 3D Gaussian splatting underlines a push to more lifelike presence and 3D content, relevant for collaboration and immersive computing.
• Market adoption: Brazil’s cross‑sector uptake shows generative AI traction is not limited to traditional tech hubs, expanding where applied value is clear.

🧠 Worth Reading

• Optimizing Datalog for the GPU (ASPLOS ’25, GPULOG): The article outlines how GPULOG applies hash‑indexed sorted arrays to accelerate Datalog evaluation on GPUs, with reported gains over engines like Soufflé. Practical takeaway: careful data structure design on modern hardware can materially speed up rule engines and static analysis workflows.