AI Weekly Malaysia

Latent Space interviews Alex Zhang, an MIT PhD and first author on Recursive Language Models (RLMs), covering GPU kernels and KernelBench, RLMs, 'mismanaged geniuses,' multi-agent swarms, and the idea of harnesses as compositional generalizers. The episode points to concrete signals: Prime Intellect's Prime Agent, described as a self-improving RLM harness using programmatic tool calling, context as a variable, multi-agent messaging, and self-modifiable harness state, was claimed to be first to ~solve ARC-AGI-3 ahead of OpenAI's Astra; and Rulin Shao's Context Language Models (Sep 30, 2026) push the same idea further by learning context policies in model weights with no harness at all. Zhang's framing is that wrapping stronger models in primitive systems leaves capability on the table.

Why: The concrete decision this surfaces for agent builders: if your harness hardcodes how context is assembled, trimmed, and passed between steps, that is the exact layer these researchers argue is underperforming. The pattern to evaluate is context as a variable or file the model edits itself, plus programmatic tool calling and subagent calls instead of fixed orchestration — the episode attributes token efficiency and expressiveness gains to that shift. There is no Malaysia or Southeast Asia angle in this text; treat it purely as an architecture question for what you are building.

turbopuffer published the first post in a series on its upcoming v3 storage architecture, saying it is moving off a vector-primary design in which the ANN index is the primary index all other indexes and query plans revolve around, and making ANN 'just another' secondary index. The recap covers v1 (documents were only an ID and a vector, object storage as source of truth plus tiered NVMe SSD/memory caches, with Cursor and Notion named as early customers) and v2 (strong text and regex search, used by Linear for a syncing engine), and states the vector-primary layout has constrained query plans like GROUP BY and aggregations. No migration timeline, benchmarks, or pricing appear in this first update; the post is framed as setting the stage for following along.

Why: If you are choosing or already running a dedicated vector store, this is a concrete argument that a vector-first index can block SQL-style query plans (GROUP BY, aggregations) and hybrid text/regex work — so if your roadmap includes analytics or filtered aggregations over the same data as your embeddings, weigh that against a general query engine or Postgres+pgvector. Do not schedule anything from this post: it names no release date, no performance numbers, and no migration path, so the 'RIP, vector database' framing is positioning until v3 ships with measured results. Nothing in the text ties this to Malaysian or SEA infrastructure, pricing, or policy, so there is no local angle to act on yet.

This is the 2021 arXiv paper "Thinking Fast and Slow in AI: the Role of Metacognition" by Marianna Bergamaschi Ganapini, Murray Campbell, Francesco Fabiano, Lior Horesh, Jon Lenchner, Andrea Loreggia, Nicholas Mattei, Francesca Rossi, Biplav Srivastava and Kristen Brent Venable, resurfaced on Hacker News (163 points, 67 comments). It proposes a multi-agent architecture where incoming problems are handled either by "system 1" fast agents that react from past experience, or by "system 2" slow agents deliberately activated when optimal solutions are needed beyond what system 1 can deliver, with both backed by a world model and a model of "self" holding past actions and solver skills. The text contains no benchmarks, code, datasets, or results — it is a position/architecture argument, not an implementation.

Why: The concrete thing here is the escalation trigger: the paper's design puts the decision to spend slow reasoning on a separate "self" model that tracks past actions and solver skills, rather than routing everything through one model. If you are building agents, that is the same lever as choosing between a cheap fast model and an expensive reasoning model per request — but the paper gives no measurements, so it won't tell you when escalation pays off. Treat it as a vocabulary source for your own routing design, not as evidence for a specific threshold.