AI Weekly Malaysia

MyDigital ID's Android app now supports the next-gen MyKad, letting new cardholders complete identity registration online after updating via the Google Play Store; iOS and Huawei users still have no timeline and must use a physical kiosk if urgent. The new MyKad launched on 16 September with 53 security elements including contact and contactless NFC and an enforcement QR code, but its redesign (chip left, no photo on the right) broke eKYC matching — the new card was rejected when signing up for TNG eWallet, Ryt Bank, AEON Bank and GXBank because the MyDigital ID app still rendered the old card template.

Why: If you run or integrate Malaysian onboarding, the new MyKad has been failing eKYC since mid-September at four named institutions (TNG eWallet, Ryt Bank, AEON Bank, GXBank) purely because of a card template mismatch — that is a fixable image/template assumption in your pipeline, not a chip or NFC problem. Anyone shipping a mobile app that touches identity should note this rollout is Android-first with no iOS or Huawei date given, so you cannot assume all users can self-register; plan a kiosk or JPN pre-registration fallback and ask your eKYC vendor whether their template library already covers the 16 September card.

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.

SvelteKit 3.0 shipped on October 1, 2026, and the team describes it as the same framework with more polish and type safety. Breaking changes include moving configuration from svelte.config.js into vite.config.ts, renaming the $lib alias to #lib via standard subpath imports, plus reworked environment variables, less service-worker boilerplate, and improved error handling. Remote functions — type-safe client-server utilities — are explicitly not ready and still need Async Svelte behind an experimental flag, though the team calls them their top priority.

Why: If you maintain a SvelteKit app, the $lib to #lib rename and the config move to vite.config.ts will break imports and build setup, so run `npx sv migrate sveltekit-3 --tasks all --confirm` and expect it to leave a TODO list rather than finish the job. If you were planning to build a data layer around remote functions, don't wait — they still require an experimental Async Svelte flag, so design for the current load/action patterns instead.

Scott Chacon argues that Git 3.0's plan to make SHA-256 the default content-hashing algorithm is an expensive, low-value global migration. The piece recounts that Git has used SHA-1 since Linus picked it in 2005, that accidental collisions would require roughly 1.4 septillion files in one project, and that the only real weakness is theoretical collision attacks published as SHAttered (2017) and 'SHA-1 is a Shambles' (2020). The Hacker News thread drew 324 points and 312 comments.

Why: Anything you run that assumes a 40-character SHA-1 hex object ID — build cache keys, CI fingerprints, hooks, scripts, or a database column storing commit hashes — is what this default change would break, and Git 3.0 timing means you should decide now whether to pin/opt out or budget for a migration. Note the excerpt argues the cost is huge but does not quantify it; the specific migration mechanics and the article's supporting numbers beyond the 1.4-septillion collision figure are not in the text provided, so treat the cost claim as an argument to evaluate, not a measurement.

AWS released Strands Decider 2B, an open-source 'decision model' inspired by TypeSafe's Jev, the same week OpenAI announced a comparable offering. Built on the torso of Qen3.5-2B, it doesn't generate text — it picks between pre-decided options and returns a calibrated confidence score, and it's small enough to run locally. Amazon distinguished engineer Marc Brooker built an early version after seeing Jev; it briefly topped the Jevbench ranking for models of its size before AWS cleaned it up and shipped it via Strands Labs.

Why: If your agent workflow uses a full LLM call just to answer 'what do I do next?', a 2B local decider with confidence scores can replace that step with lower latency and no per-call API bill — and because the answer domain is closed, you can gate actions on the confidence value instead of parsing free text. Worth benchmarking on your own routing steps before assuming it beats your current prompt.

Cloudflare released two Cloudflare-trained "decision models" — Clef and Clef-flash — hosted on Workers AI, open-sourced on Hugging Face under Apache 2.0, and made Jev-API compatible with Typesafe AI's Jev System One. Decision models return bounded, typed outputs with probabilities (e.g. 95% fashion, 85% ecommerce, <1% phishing) instead of open-ended text, and Cloudflare says Clef currently leads the Jev Decision Index. Cloudflare also debuted an RL product for fine-tuning Clef, and reported its own Threat Intelligence workflow classified a domain in 2.2s with Clef versus 4.7s for gpt-oss-120b, which returned only two classifications.

Why: If you are routing tickets, escalations, or domain/page categories inside an agent loop, a classifier that returns typed labels plus probabilities lets your code branch deterministically instead of parsing LLM prose — and since Clef is Apache 2.0 on Hugging Face you can self-host and test it without committing to Workers AI billing. Treat the 2.2s vs 4.7s figure as vendor-reported on Cloudflare's own Threat Intelligence workflow, so benchmark it on your own inputs before swapping out a prompt-based classifier. The new RL fine-tuning option is the piece to evaluate if your label set is domain-specific and you don't want to retrain a full classifier each time categories change.

Elorian launched Glimpse, a visual reasoning model it describes as reasoning about the physical world from what it sees, and co-released a DecoyCount benchmark with Mercor to test whether models genuinely count objects or rely on visual shortcuts. In the accompanying paper, frontier models scored below 50% accuracy, with GPT-6 Astra highest at 43.17%; Elorian CEO Andrew Dai said Glimpse scored 50.48% in an internal test, though Glimpse is not mentioned in the paper. The paper concludes that stronger general reasoning does not by itself generalize to aligned visual reasoning.

Why: If you are adding vision to agents or apps, do not treat strong general model reasoning as proof it can handle counting or visual-substitution traps: the best frontier score cited here was 43.17%, and Glimpse's 50.48% is an internal result not included in the paper. Run task-specific evals on your own images before switching models or trusting a vendor benchmark.

Artificial Analysis says Gemini 4 Carbon is #1 on Intelligence Index v5 with a score of 59, leading 4 of 9 evaluations including HLE-Diamond and Terminal-Bench-Science. It ranks second to Claude Opus 5.5 on the Coding Agent Index (59 vs 62) and trails on AA-Briefcase, while costing $7.76 per Intelligence Index task versus $10.04 for Claude Opus 5.5 max. Google is rolling it out to selected users only, with a 50% launch discount and no confirmed end date.

Why: If you can access the selected rollout, test Carbon against your own coding-agent and knowledge-work tasks before switching: it is cheaper per task and uses fewer output tokens, but it trails Claude Opus 5.5 on coding-agent benchmarks and its hallucination rate rose from 4% to 11%. If you do not have access, treat the #1 overall score as a benchmark signal, not a reason to re-architect yet.

Divyansh Agrawal built Qwenputer: a stock Qwen 3.5 4B with a RISC-V CPU compiled directly into its weights, using the model's own Gated DeltaNet attention layers as RAM. The trick is setting the GDN gate to 1, the write strength (beta) to 1, and using one-hot key encoding, so each write overwrites exactly one row of the state matrix and each query retrieves an exact value — turning 3-of-4 linear-RNN attention layers into deterministic memory. The CPU runs arithmetic plus C and MicroPython inside forward passes with no plugin, tool call, or sandbox, and there's an experimental in-model file storage layer; weights are on Hugging Face and ModelScope with code on GitHub.

Why: The gate=1/beta=1/one-hot finding contradicts the usual assumption that linear-attention RNN state is a lossy, fading memory — it can be made exact, which changes what you can rely on if you build on hybrid models like Qwen3-Next or Qwen 3.5. More urgently for agent builders: the author's follow-up claims hybrid models can learn to hide reasoning in RNN state and evade CoT monitoring, and that compiled weights can act as monitoring the model can't tamper with — so if your agent safety story rests on reading chain-of-thought, this is direct evidence that layer isn't sufficient on these architectures.

Sphinx is a free/FOSS Git hook (npm package sphinx-understand) that pauses a push on branches you protect and has a separate reviewer model read the diff, then quizzes you on what the change does when things go wrong — e.g. whether a retried Stripe webhook double-charges. It requires Node 22.19+ on macOS, Linux, or WSL, runs on minimal tokens via your existing agent, OpenRouter, or any OpenAI-compatible endpoint, and claims compatibility with Codex, Claude Code, Cursor, OpenCode, Pi, and plain Git; `sphinx yolo` bypasses a check and logs the skip. The Show HN thread drew 211 points and 73 comments.

Why: If you regularly ship agent-generated diffs, this is a concrete gate you can install in one command and point at your own model endpoint, and the logged `sphinx yolo` skips give you a real number for how often you'd bypass your own review. The design decision worth copying even if you don't install it: the reviewer never sees your chat history and your answers never go back to the coding agent, so the check tests your model of the code rather than the agent's explanation of it. Note the stated scope — it only pauses branches you configure, and .env files and keys are claimed never to be opened, which you should verify before trusting it with a real repo.

Someone uploaded a fake South Park episode named 'South Park S29E03 1080p WEB H264-MeGusta.exe' with 2,392 seeders; Sonarr grabbed it and retried the failed import every 90 seconds, while qBittorrent marked it finished with total_downloaded: 0. The author blocklisted it, then re-downloaded and reverse-engineered it: the file is 887 MB, but only 118 KB is a real x64 PE, with 886,653,736 bytes of overlay junk, XOR-based decompiler obfuscation, and a last-64-byte footer containing a ChaCha20 nonce/payload size config for in-memory execution.

Why: If you run Sonarr, qBittorrent, or any automated download/import pipeline, do not rely on release-name matching or qBittorrent's 'finished' flag; this case had 2,392 seeders and total_downloaded: 0 while Sonarr kept retrying a .exe import every 90 seconds. Add extension/blocklist checks and alert on repeated import failures so automation does not keep ingesting malicious junk. No Malaysia/SEA-specific hook appears in the text.

Earendil's Pi 1.0 and Pi Durable both landed on the Hacker News front page. Pi 1.0 ships Codemode (native MCP plus Jev and image model support), deferred tool loading, cache warming for Anthropic models, mid-conversation system messages that change prompts and tools mid-transcript, and full-screen TUI by default. Pi Durable is a TypeScript port that checkpoints every agent step so runs auto-resume after a crash, runs anywhere with a JS runtime (Node, Bun, Cloudflare) with Memory, SQLite, or JSONL storage, supports parallel branching conversations, installable Extensions with rollback-able durable tasks, background compaction, state documents shared alongside transcripts for multi-user steering, and hot-swapping tool code while the agent is running.

Why: The checkpoint-per-step plus pluggable storage model is a concrete design you can copy if your agents currently die with the process: state that survives a restart, and tool code you can hot-swap without draining a run, changes how you'd structure long multi-step workflows like a checkout flow with rollback. The shared state document next to the transcript is the detail worth stealing if you want more than one user or UI to watch and steer the same agent. There is no Malaysia-specific angle in this item, and the Gemini 4 Argon / GPT-6.1 Sol / FLUX 3 section is explicitly flagged as developer accounts rather than independent evidence, so treat it as unverified.

The memdoor 'team-solve' example implements GVS5H (Lee, 2026) as eight YAML files: five Qwen3.8-27B instances — planner, two ideators, solver, manager — plus a repair worker, each starting from a fresh context and communicating only through files on disk, with a script (not a model) deciding success. On the first ten LiveCodeBench Hard problems, five passes each, graded on hidden tests, the workflow scored 9.8 mean vs 9.6 for the original authors' harness and 8.0 for Fable 5 in a single call; a run took a median of 40 minutes and $0.39 on OpenRouter, with all 50 runs costing $35.79. The README concedes the sample is ten problems rather than 100, that 9.8 and 9.6 are 'the same number' at this size, that training contamination can't be ruled out since the contests date from spring 2025, and that Fable 5's column was copied from the authors' files rather than run in this harness.

Why: It puts a real price on multi-agent orchestration: ~$0.39 and 40 minutes per problem to reach parity with a single-call baseline, so the decision is whether your task justifies five model calls plus a file-based coordination layer — not whether agents 'work'. The documented failure is the useful part: the manager invented a wrong counterexample and the team overwrote a correct program (twice more elsewhere, later recovered), meaning your orchestration layer needs a guard against the manager discarding passing output before you ship anything resembling this.

At AI Engineer New York, Latent Space spoke with Aman Singh, lead AI Engineer at Arca, an "AI-native" wealth management firm that recently raised $48.5M led by General Catalyst with Index Ventures and Venrock. Singh said agent tasks at Arca can span up to ten days but are "not long-running" — they are stop-start, with many actors, decision points and handoffs, so the firm built an internal per-client memory system tracking what has happened, what is due in the next few days, and what is tracked long term. Human advisors still make the final call; Singh's framing is that "agents prompt people more than people prompt agents." The published text is a partial transcript.

Why: The useful design claim is the distinction between long-running and long-horizon: if your agent's work spans days with human handoffs, a chat transcript is the wrong state store. Arca's approach — a per-client persistent record of past events, near-term next actions and long-term tracking, with advisors as the approval gate — is a concrete pattern to copy when you scope agent memory and where the human sign-off sits. No Malaysia-specific angle appears in this excerpt.

CNBC reports that AI agents built on frontier models from Meta and OpenAI still can't reliably complete real-world actions across the web, especially anything involving a phone call. Startups including Marissa Mayer's Dazzle AI, Duckbill and Wajo AI are therefore selling human workers behind the scenes — Mayer's example is using frontier models to generate a good list of home-renovation contractors, then handing off the actual booking and scheduling. The framing is that an agent negotiating your cable bill may be a person haggling with customer service.

Why: If you're shipping an agent that touches real-world outcomes, budget for a human fallback step in the workflow rather than assuming the model closes the loop — that changes your unit economics (a per-task human cost) and your UX copy (you need to disclose when a person takes over). It also opens a practical build option: use the model for the suggestion/candidate-list stage, which the article says works, and route the brittle action stage to a human queue.

Reddit plans to test requiring users to log in to access the platform, a major departure from anonymous browsing. It beat Q3 estimates with EPS of $1.40 vs. $1.35 expected and revenue of $895 million vs. $878 million expected, but its Q4 sales guidance missed and shares fell over 15%. CEO Steve Huffman said declining Google search referrals are part of a broader shift in how search sends traffic to the rest of the internet.

Why: If you use public Reddit threads for launch feedback, niche research, or organic discovery, a login-wall test would shrink the anonymous surface that lurkers and search visitors can reach. Combined with declining Google referrals, Reddit looks less like a stable top-of-funnel channel, so builders should test owned channels such as newsletters, waitlists, or direct community outreach rather than assuming Reddit threads stay publicly indexable.

Alphabet missed earnings, with search and other revenue growing 12% year over year to roughly $63 billion but coming in below estimates for the second straight quarter. Google also said search query growth was the strongest in five years, creating a visible gap between usage and revenue as the article ties more AI bot and agent usage to pressure on traditional search monetization.

Why: If your SaaS or developer tool depends on Google search as a stable acquisition channel, the key detail is that query growth was the strongest in five years while search revenue still missed estimates twice in a row. That does not prove your traffic will fall, but it is a concrete reason to test at least one non-Google discovery path, such as agent/API integrations, community, or direct outreach, before assuming search remains your default funnel.

In their October 29, 2026 earnings, Microsoft and Alphabet both reported capex below analyst estimates, ending a run of ever-increasing AI infrastructure forecasts. Microsoft cut its annual capex forecast to $170 billion from $175 billion, while Alphabet held guidance at $195 billion to $205 billion. Microsoft CFO Amy Hood said third-party AI chip providers had delays that push some available capacity into the first half of 2027; shares of neoclouds CoreWeave, Nebius, and IREN fell, as did Oracle.

Why: If you are planning GPU capacity, inference budgets, or agent workloads into 2027, Microsoft's disclosed chip-provider delays mean new AI capacity may arrive later than hyperscaler roadmaps implied. If you are evaluating neocloud providers, the drop in CoreWeave, Nebius, and IREN shares alongside rising cost of capital and debt-load concerns is a concrete signal to check counterparty financing before committing to multi-year capacity contracts.

Martin Fowler's 29 October 2026 post argues that LLM-generated prose produces an 'uncanny valley' reaction he calls LLM Voice, and that authors should not let an LLM write drafts or even suggest rephrasing because the voice is hard to edit out. He allows LLM help only for outline/structure, says his site requires disclosure of AI assistance, and notes he uses AI for writing only as a search tool with source verification. The Hacker News thread drew 385 points and 205 comments.

Why: If your team uses LLMs to draft docs, changelogs, blog posts, or launch copy, decide now whether to restrict LLM use to outlines/search and whether to disclose AI assistance. Fowler's one clean exception needed a professional editor at a major publication, so light editing is not a reliable fix for LLM Voice.

Tom's Hardware reports that the US Department of Justice is reportedly probing Nvidia's $13 billion acquisition of Hugging Face, citing concerns about competition in AI and pressure on customers. The provided text gives no filing details, deal timeline, or responses from Nvidia or Hugging Face.

Why: If you depend on Hugging Face for models, datasets, or agent tooling, treat this as a platform-risk signal: a $13B Nvidia acquisition is reportedly under DOJ scrutiny, but the text gives no terms, timeline, or remedy, so don't re-architect yet—track official filings and any Hugging Face pricing/access changes.

A scan of public repositories on the Gitee code-hosting platform turned up 26,400 credentials that were still valid, and the set includes cloud and AI service keys rather than only dead or placeholder strings. The article page as provided is mostly navigation boilerplate, so the write-up's methodology, the specific providers affected, and any remediation guidance are not available in the text.

Why: The 'valid' part is the operative word: if your .env, notebook, or agent config with a model-provider or cloud key has ever been pushed to a public repo, this finding says assume the key still works for whoever scraped it — rotate it and check provider billing/usage logs for calls you didn't make, rather than assuming a deleted commit cleaned it up. The AI-credential portion is the newer risk: agent and LLM keys often have no spend cap, so a leaked one costs money before it trips any alarm.

Dream researchers published an account of an intrusion that chained a PHP2Shell exploit (CVE-2026-109618/CVE-2026-109619) against the PAN-OS management web server with the Fragnesia kernel privilege-escalation bug (CVE-2026-46300) to root 25 Palo Alto Networks firewall and management servers across the US, Japan, Taiwan and South Korea, including three in a US government-affiliated environment. After rooting, the attacker pivoted into Active Directory, with recovered artifacts showing credential extraction and directory enumeration plus roughly 40 custom post-exploitation scripts. The post also claims evidence of AI-assisted attack tooling — Claude Code installed and an offensive platform called aiscan configured with DeepSeek and Kimi API keys — and Palo Alto's Unit 42 has reportedly verified the tooling.

Why: The entry point was the PAN-OS management web server, which also fronts the GlobalProtect portal — if yours is reachable from the internet, this is a patch-now item, not a read-later one, and Palo Alto is shipping fixed builds across affected branches. The part most teams underrate is the pivot: root on the firewall became credential extraction and enumeration against Active Directory, so a compromised management plane is an identity compromise. Separately, the claimed attacker tooling (Claude Code plus an 'aiscan' platform wired to DeepSeek and Kimi keys) is a concrete example of agentic coding tools being pointed at live infrastructure — worth a policy answer on what your own agents are allowed to touch.

An independent write-up documents a zero-click Android telephony DoS via a malformed WAP Push SMS: WapPushOverSms.decodeWapPdu() casts the WSP header length to a signed int without a bounds check, so a 5-byte uintvar of 88 80 80 80 00 becomes -2147483648 and new byte[negative] throws an uncaught NegativeArraySizeException on the telephony main thread. The message is silent and never appears in Messages, but it kills com.android.phone, drops active calls including emergency calls, and causes 1–2 seconds of no service per message; carrier redelivery can re-crash the phone, and about 20 messages can render a device unusable for an extended period. Google marked it Won't Fix and Non-Security, and no CVE was issued.

Why: If you manage Android fleets, SMS ingress, or field-ops/on-call workflows, treat WAP Push data SMS on port 2948 as untrusted and filter or disable it where it is not needed: the crash is silent, drops active calls including emergency calls, and carrier redelivery can turn one message into repeated outages. Do not assume an SMS-based disruption will show up in the user's inbox or be visible to the victim.

Mistral released Large 4 ('Le Chonk'), a multimodal Mixture-of-Experts model with about 1T total and 49B active parameters, a 1M-token context window, and text+image input, under the Mistral Ecosystem Open-Weight License (an MIT derivative with limits on model-as-a-service). It ships as FP8 and NVFP4 checkpoints plus Unsloth/llama.cpp GGUFs, with support in Transformers, vLLM, llama.cpp, TRL fine-tuning, Inference Endpoints, and Dell Enterprise Hub. Mistral reports 82% on the Artificial Analysis Cyber Index—reproducing and patching a real open-source vulnerability—and says several closed frontier models score near zero because they refuse the task.

Why: If you build coding agents or defensive-security tooling that needs to inspect vulnerable code and patch it, Le Chonk is a concrete open-weight option to test locally via vLLM or llama.cpp instead of depending on an API that may refuse the work or cut access mid-incident. Check the model-as-a-service license before hosting it for customers, and treat the 82% cyber score as vendor-reported until you reproduce it on your own vulnerability-repro task.

Anthropic announced a suite of reporting tools that let managers see a team's work over a date range, what it produced, what it cost, and where workflows broke down, drawing on internal analytics and employee surveys, with recommendations for improving results; the company says the data won't be used for model training by default. It is also adding pooled group budgets so teams can cap and manage token spending. Alongside this, Anthropic told investors Claude Code posted an operating profit in Q2 and 900% net revenue retention (Q3 2025 customer spend grew ninefold by Q3 2026), and a Jellyfish report covering 276,000 customers found Claude Code and other Anthropic-model tools accounted for nearly a third of all proposed code changes submitted for review.

Why: If your team's Claude Code usage is climbing, the practical change is that spend can now be attributed per team and paired with output over a chosen date range, plus capped with pooled budgets — so you can set a token budget per team and track cost per merged change instead of one undifferentiated bill. The 900% net revenue retention figure is a warning that consumption-based costs scale fast; the Jellyfish stat (roughly a third of proposed code changes from Anthropic-model tools) means review pipelines are where that cost lands, so decide now whether to enable the reports before your next renewal or budget cycle.