Building an agentic app means wrestling with PDFs, wikis, tickets, and contracts. This guide walks you through modeling an ontology, constructing a graph from that unstructured data, and writing queries that let your AI reason over it instantly. The result is a searchable knowledge base that scales with your data and improves agent performance.
After a single RDS Postgres could no longer handle both transactional and high‑cardinality analytics, Physical Intelligence migrated to ClickHouse Cloud with managed Postgres. The new stack now supports 10‑100 billion rows, slashing query latency and letting robotics engineers explore data instantly.
AI agents can write correct SQL but still make bad decisions because warehouses don’t embed business rules, metric definitions, or data provenance. Traditional schemas expose types but hide nuances like fee inclusion or currency conversion, leaving agents to guess. Without explicit logic, agents risk recommending actions that hurt ROI.
A new technique lets you spread a budget across multiple categories while preserving the shadow prices that justify the allocation. The result is an optimal split that remains interpretable, letting analysts see exactly why each dollar lands where it does.
IBM Research’s ALTK‑Evolve builds an agentic memory of past trajectories without compressing them, using per‑lesson support counts. The system achieves the same performance as the ACE framework but needs far fewer tokens during inference, cutting LLM inference costs and speeding reasoning workflows.
Meta unveiled Muse Glimmer, a 30‑billion‑parameter open‑source vision‑language model that can run on a single GPU and supports agentic reasoning. Benchmarks show it outperforms comparable models on tasks like MCP Atlas and SWE‑Bench, positioning it as a strong, locally‑deployable alternative in the open‑weight LLM arena.
DeepMind’s WeatherNext model delivers state‑of‑the‑art cyclone track, intensity, and wind forecasts, giving forecasters an extra 24 hours of accuracy, its three‑day predictions match the prior two‑day baseline. The open‑sourced system already helped the NHC predict Hurricane Melissa’s rapid intensification in 2025.
Databricks announced the beta of FILE, a new column type that stores files, images, video, and other unstructured assets directly in Delta Lake tables. By treating these assets as native, governed columns, teams can query and process multimodal data alongside structured rows without performance hits, keeping data open and portable for AI workloads.
PostgreSQL core developer Oleg Bartunov explains how a null‑filled “ugly” table led him and Teodor Sigaev to invent hstore, a Perl‑hash style key‑value column. That experimental datatype proved so useful it evolved into nested hstore and ultimately JSONB in Postgres 9.4, illustrating how real‑world pain drives major DB features.
Prompt engineering taught us to ask better questions, but as LLMs become coding agents and data copilots, merely phrasing prompts no longer guarantees correct results. Specification engineering, writing explicit goals, constraints, validation criteria, and edge‑case handling, ensures outputs are usable, testable, and aligned with business needs. Mastering it will prevent hidden failures and AI gaming.
I replayed 27 tasks from my AI assistant on two models. A 30B model on one RTX 3090 scored 22.8 vs Claude’s 89.4 and broke a quarter of calls; a 122B model on RTX 3090s (256K context) hit 80.0, eliminated errors, and cost $0.00097 per task, ≈787× cheaper. It shows a stack can replace Claude, though it needs a larger model and context.
You’re caught up.
That was all 11 stories in the 11 August issue.