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.

World Labs has signed a definitive agreement to join AMD, following a technical partnership that began last year around model training and inference optimization on AMD GPUs. Dr. Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist working directly with CEO Dr. Lisa Su, while Justin Johnson and Ben Mildenhall continue leading the World Labs team inside AMD to form a frontier research organization. The deal is expected to close by the end of 2026, subject to regulatory approvals, and no price or terms were disclosed.

Why: The concrete signal for builders is that AMD is buying a frontier lab rather than only shipping silicon, and stating an intent to build an end-to-end open ecosystem of hardware, software, platforms, and open models. That is worth tracking if you are weighing AMD GPUs as a training or inference option instead of defaulting to NVIDIA, but there is nothing here you can act on yet: no model names, no licensing terms, no pricing, no dates beyond the end-of-2026 close. Do not re-plan GPU budgets or migration timelines off this post alone; wait for the close and for actual released models or tooling. The one adjacent detail worth noting is World Labs' stated focus on spatial intelligence and simulation for robotics, per its SceniX acquisition discussion, which is a different workload from LLM training.

Tom's Hardware reports that AMD is acquiring World Labs — the lab founded by Fei-Fei Li, described in the headline as an 'ImageNet pioneer' — for $8.2 billion, with Li becoming AMD's chief scientist and the lab brought in-house. That is the entire substance available here: the article body as supplied contains only subscription prompts, navigation and newsletter boilerplate, so there are no stated terms, timelines, model details, or product implications. Treat the $8.2B figure, the chief-scientist appointment and the acquisition itself as headline claims with no supporting reporting in this text.

Why: There is nothing here a builder can act on yet — no statement about whether World Labs' models stay accessible, change licence, or get tied to AMD hardware. The one concrete decision-relevant fact is the $8.2B price for a world-models lab, which is a signal about where chipmakers think spatial/world-model work is heading; if you are building on World Labs' outputs, the honest position is that this text gives you no basis to change anything, and you should wait for the actual terms before planning around it.

This is the first essay in OpenAI's new "Intelligence Age" series, a platform for outside contributors writing about an AGI future; the authors' note states the views are their own, not OpenAI's. The argument: human minds already outrun human execution capacity — Galileo needed a few dozen hands, while the $10B James Webb telescope needed 18 mirror segments at 50-nanometer precision, 300 organizations across 14 countries, and a global economy. The authors cite Nick Bloom's research showing that sustaining Moore's law now takes 18x more researchers than in the early 1970s and that economy-wide effective research effort rose 23-fold since the 1930s, concluding that genius machines may be most valuable doing the monotonous support work rather than the genius work.

Why: The concrete claim to test against your own roadmap is the ratio, not the philosophy: if progress needs 18x more researchers and 23x more research effort to sustain the same rate, the bottleneck is the surrounding grind, not the ideas. That argues for pointing AI agents and tooling at the unglamorous middle of your pipeline — data prep, test scaffolding, migration and review chores — rather than at the tasks you enjoy. The essay is truncated mid-sentence in the supplied text and offers no product, price, or measurement you can act on this week, so treat it as a framing argument for where to spend agent budget, not as news.

Simon Willison quotes cryptographer Matthew Green reacting to Anthropic's recent cryptography work. Green argues the field is mid-transition from EC and RSA public-key algorithms to post-quantum schemes built on newer hard problems — hence the number of standards under consideration such as HAWK — and that this makes it an unusually good moment for AI to get good at cryptanalysis. In the best case, he says, AI failing to break these problems gives real confidence in them and makes the cryptanalysis literature more robust.

Why: There is no Malaysia or Southeast Asia angle in this text, and no detail about what Anthropic actually did or published — it is a single quoted opinion. The one concrete decision-relevant point for builders is the migration context Green names: if you have a post-quantum migration on your roadmap (EC/RSA to newer schemes), his argument is that AI-assisted cryptanalysis during this window is more likely to validate the new problems than to break them, so the standards churn around candidates like HAWK is expected rather than alarming. Treat the 'Anthropic cryptography work' claim as unverified from this item alone.

MI5 issued a "Security Service Espionage Alert" on September 30, 2026 stating that the China General Technology Research Institute (CGTRI, also called the China Academy of General Technology) exists primarily to fund research that improves Chinese Ministry of State Security technical espionage capability, including work on AI, cybersecurity, covert communications systems, and steganography. The alert says more than 100 U.K.-linked academics have contributed to CGTRI-funded projects, in some cases without knowing the funding source, and urges U.K. institutions to immediately review ongoing or planned CGTRI collaboration and trace funding sources on Chinese research partnerships. It warns continued collaboration could be prosecuted under the U.K. National Security Act 2023; the Chinese embassy in the U.K. called the accusations "imaginary and purely fabricated" and said it lodged formal representations.

Why: This is a U.K.-scoped alert, so it creates no direct obligation for Malaysian institutions or companies — but it is a concrete example of funding provenance becoming a due-diligence item. If you take research grants, host visiting scholars, or co-author with overseas institutions, the operational takeaway from the alert is narrow and specific: ask who the ultimate funder is and whether CGTRI/CAGT appears anywhere in the chain, because the alert says contributors have been funded without knowing it. If your work is U.K.-linked or you have U.K. partners or staff, the alert's National Security Act 2023 prosecution warning is the part that changes behaviour; if you are not U.K.-linked, treat it as a signal about how funder disclosure is trending, not as a rule you must now follow.

Import AI 474 leads with a paper by Michael Levin arguing for a non-physicalist model of mind, framed as "mind:body is as math:physics" — minds are patterns in a hypothetical "platonic" space of arbitrary complexity that "ingress" into physical bodies, which act as interfaces. Levin cites synthetic morphology and xenobots (biorobots made of frog cells) as suggestive examples, and explicitly notes there are no home-run experiments backing the claim. The newsletter title also names "TPUs in space" and a Zhipu "outer RSI loop," but the supplied text covers only the Levin item.

Why: There is no build decision here: no model, benchmark, API, price, or tooling change appears in the text, so nothing for a working developer to adopt or avoid this week. The one usable takeaway is terminological — Levin treats agency as a graded property of patterns (from facts about integers and geometric shapes up through "kinds of minds") rather than a binary, which is a prompt to write down what you actually mean by "agent" in your own system specs instead of assuming the word is self-defining.

Quanta Magazine reports that traveling waves of electrical activity across the cortex are more structured than the simple planar oscillations neuroscientists traditionally assumed — including source waves emanating from a point, sink waves converging on one, and vortex-like spiral waves. A study published in April 2026 in Nature Communications, from a team including Joshua Jacobs and Anup Das at the University of Chicago, measured these patterns with intracranial electrodes and found different behavioral tasks produced distinct wave patterns; Earl K. Miller, a cognitive neuroscientist at MIT, is quoted saying the work has moved the field from 'Are they relevant?' to calling traveling waves 'a major motif of how the cortex processes information.' The piece also notes the brain spends roughly half its energetic resources maintaining the electrochemical gradients that keep neurons ready to fire.

Why: There is no direct action for builders here: no code, dataset, model, API, or benchmark, and no Malaysia or Southeast Asia angle. The only practical link is for people working on neuromorphic or brain-inspired computing, and this text offers zero implementation detail they could use — treat it as background reading, not a signal to change a roadmap.

Ars Technica published an item headlined 'AI finally beat the best Stratego player in history and did it on a budget,' covering an AI system that reportedly defeated the strongest known Stratego player. However, the text supplied here contains no article body at all — only Conde Nast cookie-consent and privacy-preference boilerplate. No model name, method, compute figures, match score, or author is present in the source text, so the headline's 'on a budget' claim cannot be verified or quantified from what we have.

Why: Nothing here is actionable yet: there is no model name, no cost figure behind the 'budget' claim, no code, and no benchmark detail, so you cannot evaluate or reuse anything from this item as delivered. Treat the headline as an unverified pointer and do not cite the 'budget' framing in a deck or blog post until someone reads the actual paper or article. If you are building agents that must act under hidden information, this is worth a follow-up read, not a decision this week.

The supplied page content contains only Ars Technica's consent-management and cookie-notice text; the actual article body about Google watermarking AI-designed proteins is not present. The only substantive information available is the headline itself and the publication timestamp of 2026-09-30. No method, paper, tool, model name, or result is described in the text.

Why: There is nothing here a builder can act on: no technique, no released tool, no API, no measurement. Anyone wanting to know whether AI-generated protein sequences can be marked and later detected will have to read the original article, because this excerpt does not say how the watermarking works or whether it is robust. Do not cite this as evidence of a usable watermarking capability yet.