AI Weekly Malaysia

Jeff is an independent open-source project offering fine-tunes of Qwen3.5 (0.8B and 2B) and Gemma 4 (E2B) as tiny zero-shot classification models that reuse Jev's request format and return a calibrated probability per option from a single forward pass instead of generated text. The README reports about 22 ms per decision on an RTX PRO 6000 and 28 ms on an Apple M4 Max via MLX, with the 0.8B training in roughly 2 hours and the 2B in about 3.5 hours on one RTX PRO 6000, using synthetic data written by an open model on two DGX Sparks. It is explicitly not affiliated with or endorsed by TypeSafe, the makers of Jev, and the repo shows 298 stars, 8 forks and 6 commits; the Hacker News thread drew 222 points and 71 comments.

Why: If you currently route simple label decisions — support queues, moderation labels, intents, game moves — through a hosted LLM API, this is a concrete alternative: ~22-28 ms per decision on a single GPU or an M4 Max MacBook, no per-token billing and no data leaving the machine. The reported fine-tune result (held-out accuracy 31.7% to 95.8% for voice navigation in under 30 minutes on one GPU) is the number to test against your own labels, since zero-shot accuracy at 0.8B is the stated weak point and the README itself says reasoning will not match a much larger model. For teams in Malaysia, running this on local or consumer hardware removes cloud GPU spend and cross-border data transfer for classification tasks, though you still need to verify the models' licensing and Jev's own terms before swapping them in.

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.

Epoch AI's report, covered by Tom's Hardware, claims the price of AI has fallen by thousands of times in recent years — roughly 50% cheaper every quarter, or about 13x cheaper per year. That pace outruns lithium batteries, DNA sequencing, and even compute riding Moore's Law. The article also notes that vendor loyalty and subscription schemes have limited appeal when prices can fall this fast.

Why: If inference really is deflating ~13x a year, any pricing model that assumes today's per-token or per-seat API cost for a 12-month horizon is wrong by an order of magnitude — that flips build-vs-buy math toward 'buy now, revisit in a quarter' and argues against multi-year vendor commitments or self-hosting to chase cost. Treat the 13x figure as a claim from one report, not a law, and check your own invoice trend before re-architecting.

PostHog published Jeeves, an open-source reasoning classifier built on Qwen3.5-9B with LoRA plus a pointer head, trained with SFT and CISPO, and shipped with full training code and train/dev/test data. It reports 0.889 accuracy on held-out out-of-domain test data (vs Kev-9B 0.822 and Jev 0.857) and 0.935 on JevBench's 231 public items (vs Jev 0.866), using a block-4 diffusion drafter and a Jev-compatible API supporting noul/choice/score questions. Latency is about 0.3 s per request without thinking and a 3.3 s median with thinking on a single H100 at --precision fp8; it runs on CUDA bf16, FP8 on compute capability 8.9+, and Apple Silicon MPS. The thread drew 239 points and 93 comments on Hacker News.

Why: The headline numbers hide a regression: Jeeves scores 0.746 on Transfer (MMLU-Pro and buried state) versus Jev's 0.800, and 0.793 on MMLU versus Jev's 0.900, so reasoning-before-deciding helps on the benchmarks it targets and hurts on general transfer tasks. If you currently fall back to a reasoning model when a Jev-like classifier is uncertain, the 3.3 s median thinking latency versus 0.3 s without means that fallback costs roughly an order of magnitude more wall-clock per request on one H100 at fp8 — decide per pipeline whether you truncate the chain, or keep the calibrated classifier and only reason on the hard slice. Because inference also runs on Apple Silicon in bf16 or FP8, you can benchmark it on a local Mac before paying for cloud GPU time.

A Futurum CEO, quoted by Tom's Hardware, claims AI agents consume about 5x more tokens than human users and that the figure will eventually reach 10x, largely because agents keep re-reading context they have already seen. The article frames this as a KV cache demand problem that compounds existing RAM shortages. The excerpt carries no methodology, benchmark, or per-model breakdown — only the multiplier claims and the cache/RAM framing.

Why: If the 5x-to-10x token multiplier holds for agentic workloads, your per-seat agent pricing, free-tier limits, and API cost forecasts built on human-chat token volumes are understated by roughly an order of magnitude, and the re-reading pattern means prefix/prompt caching — not just cheaper models — is where the savings sit. The linked KV cache and RAM shortage angle is a second-order decision: self-hosted or reserved GPU memory for agent workloads is likely to get more expensive before it gets cheaper. No Malaysian or Southeast Asian detail appears in the text, so treat this as a general cost and infrastructure planning signal, not a local policy or funding item.

A blog post on insufferable.dev argues the competitive dynamic between Western and Chinese AI labs has flipped: instead of Western labs accusing Chinese labs of distilling their models, Western labs are now quietly adopting Chinese inference optimizations. It cites DeepSeek's KV cache work — MLA at roughly 15x compression, then Compressed Sparse Attention and Heavily Compressed Attention, and DeepSeek-V4.1-Flash with CSA2, cross-layer cache reuse and FP4 caching bringing the global KV cache to 890 bytes per token, roughly 437x below DeepSeek-V1 — and claims Claude Opus 5.5 and GPT-6.1 Sol shipped with these techniques, with Opus 5.5 cutting cache-read pricing 60% versus Opus 5. The excerpt is truncated mid-sentence, and the pricing claims and model-release details are asserted by the author without cited primary sources.

Why: If the cache-read price cuts described here are real, the cost of running long-context coding and agent sessions shifts from output tokens toward a much cheaper cache-read line item, which changes how you'd budget and architect retrieval-heavy agents. But the article gives no links to DeepSeek's papers or to Anthropic/OpenAI pricing pages, so before repricing anything, verify the 890 bytes-per-token figure and the claimed 60% Opus cache-read reduction against the vendors' own docs — the HN thread (349 points, 368 comments) is a better starting point than the post itself.

AMD announced it will acquire World Labs, the world-model startup founded by Fei-Fei Li in 2024, for $8.2 billion, with the deal disclosed on September 28, 2026. Li will join AMD as executive vice president and chief scientist, and World Labs framed the move as necessary because AI development requires 'close collaboration across model research, systems and compute.' The two companies already had an inference optimization-and-training partnership formed last year, and World Labs' first product is Marble, pitched for creating entertainment experiences.

Why: If you build on Marble or are prototyping world-model/3D-generation features, the vendor behind that tool now sits inside a chip company, so expect roadmap, API and pricing decisions to be driven by AMD's hardware interests rather than a standalone lab's. For teams weighing non-Nvidia inference stacks, AMD explicitly says frontier workloads like World Labs' will shape its chip roadmap, which is a signal to watch ROCm/inference tooling rather than to act on today. The text contains no Malaysia or Southeast Asia detail, so any local cost or availability impact is not something this article supports.

Cloudflare opened a closed beta for a self-serve OHTTP Gateway, a paid add-on customers enable on their zone to receive Oblivious HTTP traffic without seeing client IP addresses or TLS fingerprints. OHTTP splits requests across two independently operated hops — a relay that blindly forwards encrypted requests and a gateway that decapsulates them — so no single party sees both client identifiers and request contents; Cloudflare also renamed its 2022 'Privacy Gateway' product to 'Cloudflare OHTTP Relay'. Named users of the pattern include Flo Health's app Anonymous Mode and Apple's Private Cloud Compute, which uses OHTTP to disassociate AI inference requests from user identities.

Why: The specific decision: if your servers already sit behind Cloudflare, you previously could not pair them with a Cloudflare-operated relay without collapsing the relay/gateway separation of trust — this gateway is the missing half, so the choice becomes pairing a non-Cloudflare relay with the Cloudflare gateway versus staying with your current setup. Because it is a waitlist closed beta on a paid add-on with no published price, treat it as a watch-list item and do not architect around it this quarter; the actionable step now is deciding whether an IP-free request path is worth a two-hop latency and vendor-pairing cost for any feature where you currently log client IPs. No Malaysia- or SEA-specific detail appears in this text, so the relevance is only as general infrastructure local apps could adopt.

A firm rented four Nvidia H200 GPUs at $13,200 per month to independently test DeepSeek's claim of being '80x cheaper', and the rental alone reportedly doubled what the firm was already paying for Claude. The same write-up notes that security flaws forced the team to keep their code offline during the test. The article body itself did not load in the supplied text, so no benchmark results, token throughput, or final verdict are available here.

Why: The only concrete numbers we have are the cost side: $13,200/month for four H200s versus an existing Claude bill that this doubled, plus a security constraint that kept code off the network entirely. If you are weighing self-hosted or rented-GPU inference against API spend, this is a reminder that the comparison is rental + ops + isolation overhead, not just per-token price — and that the '80x cheaper' figure is still unverified here. Because no results are in the text, don't cite this as evidence either way yet; wait for the actual measurements.

TechCrunch reports, citing a source with knowledge of the deal, that AI inference provider Modal Labs is close to a $750 million round led by Accel at a $15.75 billion valuation — more than triple the $4.65 billion it hit in its $355 million raise just four months earlier. The article notes the wider inference market is repricing fast: Baseten is reportedly nearing a round at a $26 billion valuation (double its June number), Fireworks said in July its annualized revenue hit $1 billion (5x year over year), and Fal has also talked to investors. It also flags the catch: revenue is growing quickly but margins are thin because acquiring or leasing compute stays expensive.

Why: This is a capital-markets signal, not a product change — no pricing, API, or capability change is stated, so there is nothing to migrate or re-architect today. The one decision-relevant detail is the thin-margin caveat alongside Fireworks' $1B annualized revenue: if you are choosing an inference provider for a production or agent workload, expect competition to keep going up while compute costs keep the providers' margins thin, which makes multi-provider abstraction and exit cost worth designing for before you commit. Nothing here is Malaysia-specific; treat it as context on the vendors a Malaysian team might depend on, not as local market news.

Tom's Hardware published a full interview transcript with OpenAI's VP of Hardware, Richard Ho, about 'Jalapeño,' the inference ASIC OpenAI revealed at Hot Chips in August 2026. The one concrete claim in the available excerpt is that the chip leaned heavily on AI to achieve an 'incredibly short design window.' The rest of the retrieved text is site navigation, membership prompts, and interview framing — no process node, die size, throughput, power, or cost figures are present.

Why: There is not enough in this excerpt for a builder to change anything: no performance, power, or price numbers, so no basis to revise inference-cost assumptions for API-dependent apps. If the full transcript eventually includes design-cycle or efficiency figures, that is what would matter — the speed of custom inference silicon is a leading indicator for what you pay per token. Until then, treat this as an announcement of a transcript, not of a measurable change.

Tom's Hardware reports that a $5,245 prebuilt PC with an RTX 5090 had its power connectors melt after the machine sat boxed for a year, with both Digital Storm (the system builder) and PNY (the card vendor) denying warranty claims. The stated reasons for denial were expired coverage and the presence of third-party cables. The article body available here is almost entirely site navigation and subscription boilerplate, so the headline is the only substantive detail — no dates, cable models, photos, or vendor statements are included in the supplied text.

Why: If you buy a prebuilt GPU box for local model inference or rendering, this is a concrete reminder that the warranty clock starts at purchase, not at first power-on — a machine left boxed for a year can burn its coverage before it ever runs. It also means the cable you plug in matters: swapping in a third-party 12VHPWR/12V-2x6 cable gives the vendor a stated reason to deny a melted-connector claim, so unbox, inspect, and test the system with the supplied cable during the coverage window rather than shelving it.

A critique by Deric Tilson and Adam Stein argues that Harvard-affiliated studies linking nuclear power plants to cancer use a proximity-score methodology that also yields absurd results: Costco warehouses 'cause' 120,687 annual cancer deaths, private colleges 83,782, Harvard 5,842, and MLB fields 19 times as many cancer deaths as MLS pitches. They replicated and expanded the analysis, claiming the method produces increased cancer risk/mortality no matter which landmark is used, and recap the underlying papers led by Yazan Alwadi at Harvard T.H. Chan, from a Dec 2025 Massachusetts study using a 120 km radius to national studies using 200 km radii.

Why: If you build or review geospatial risk models, this is a concrete sanity-check: run the same proximity-score + regression + attributable-fraction pipeline against a neutral landmark (the article's Costco example) before trusting a claim. Otherwise, for Malaysian developers, SaaS founders, and AI/ML learners, there is no direct product, policy, or platform action here.

Tom's Hardware reports that OpenAI's custom Jalapeño inference ASIC is intended for OpenAI's internal use, with the company leaving the door open to a broader rollout. OpenAI is quoted as saying it will have its "hands full" with Jalapeño for "a good long time." The supplied page text contains only site navigation, membership prompts, and newsletter boilerplate — no chip specs, performance numbers, manufacturing partner, pricing, availability date, or benchmark data.

Why: Nothing here changes a build decision yet: there is no published throughput, price, or third-party availability for Jalapeño, so it should not factor into inference-cost planning or vendor selection for anyone outside OpenAI. If you are modelling API price drops or self-hosting economics, treat this as an unquantified internal roadmap statement and keep using current published pricing until OpenAI ships numbers.

This is a Tom's Hardware Premium roundup post for its themed "AI Chip Design Week," headlined by a full, unredacted interview with OpenAI hardware VP Richard Ho about Jalapeño, the Broadcom co-developed inference ASIC OpenAI debuted at Hot Chips 2026. The teaser claims the interview covers OpenAI's use of AI in its own chip design process, beating Nvidia on efficiency, and correcting the misconception that Jalapeño only runs OpenAI models. Tom's Hardware says the series is free for a limited time, extended through Monday, October 5, 2026, and the post also flags AI agent safety as a topic covered that week.

Why: The post itself contains no benchmarks, pricing, specs, or dates beyond the free-access window ending Monday, October 5, 2026 — so the only decision it supports is whether to read the unredacted Ho transcript before it goes back behind the paywall. If you're evaluating custom inference ASICs versus GPUs for cost-per-token, the interview is the item of interest; the roundup page is just a pointer. There is no Malaysia or Southeast Asia angle anywhere in this text, so don't read one into it.

CNBC's Jim Cramer said AMD's run is not over after the chipmaker briefly crossed $1 trillion in market value, having rallied nearly 13% in a week and about 30% in September 2026 alone. He credits the next leg to agentic AI driving CPU demand, pointing to Meta's launch of Muse, its personal AI agent platform, as a partial catalyst. The piece contains no product launches, pricing, benchmarks, or roadmap dates from AMD — it is market commentary plus a prediction.

Why: This is a stock-and-narrative item, not an engineering one. The only concrete claim a builder can act on is Cramer's assertion that demand shifted this year from AI model training toward inference and agentics, making CPUs the bottleneck — and that claim comes with zero measurements, no token/CPU ratios, no cost per agent run, and no AMD product detail. So don't re-plan your inference hardware or cloud spend on it; treat it as a signal that CPU capacity for agent workloads is being talked up by investors, and wait for actual pricing or performance numbers before changing anything.

Tom's Hardware is flagging a $520 discount on the MSI Stealth A18 AI+ gaming laptop, configured with an RTX 5090 GPU, 64GB DDR5, a 2TB SSD, a 12-core AMD Ryzen AI 9 CPU, and an 18-inch UHD+ 120Hz display. The article body is almost entirely paywall, newsletter, and membership boilerplate — the actual sale price, retailer, and expiry date are not present in the supplied text.

Why: This is a consumer deal post, not a product change, so nobody's build pipeline or tooling decision changes because of it. The only decision it supports is a hardware purchase, and the text withholds the one number you'd need to make it: the final price after the $520 cut. If you're weighing a 64GB/RTX 5090 laptop as a local inference or fine-tuning box, you'd have to check the retailer page yourself — nothing here lets you compare it against cloud GPU spend.