OpenAI's Altman Signs Energy and Investment Deals in the UAE — Amazon to Sell Trainium Outside AWS, Oaktree's Marks Calls Skepticism "Healthy"

Your evening AI news roundup for November 2. Today's stories centered on money and chips. OpenAI kept stacking up deals in the Middle East, while Amazon is moving to sell its own chips beyond its cloud. Meanwhile, a prominent investor is urging everyone to slow down. There's plenty of developer-focused news as well.

OpenAI's Altman Signs Energy and Investment Deals in the UAE

According to Bloomberg, OpenAI CEO Sam Altman visited Abu Dhabi and met with Crown Prince Sheikh Khaled bin Mohamed Al Nahyan. OpenAI has signed multiple agreements to deploy its technology with UAE energy companies, investment firms, and government agencies — a move that significantly deepens ties as the company seeks billions of dollars in funding from the country's state-backed investors.

AI development requires both power for data centers and the massive capital to fund it. It's becoming increasingly clear that major US-based AI players are looking to source both directly from the Middle East. As for the details of the meeting and the size of the deals, we'll wait for official announcements.

Amazon to Sell Its In-House AI Chip "Trainium" Outside AWS

Amazon reportedly plans to start selling Trainium, its in-house AI chip, outside of AWS — a story that gained traction on Hacker News. For now it's report-based, and details such as the sales model, timing, and pricing remain unknown.

If it materializes, a cloud provider would be entering the NVIDIA-dependent AI semiconductor market as a chipmaker in its own right. For companies struggling with soaring inference costs, it would mean more options — a development worth watching.

Oaktree's Howard Marks: "AI Stumbles Broke the Sense of Invincibility. Skepticism Is Healthy"

Howard Marks, co-founder of Oaktree Capital and known for his analysis of market cycles, told Bloomberg TV in an interview that stumbles in AI-related markets have deflated, to some extent, the "bubble of invincibility" surrounding the sector. Marks, who declared he was on "bubble watch" last year, said, "More people are saying let's think a little, let's be a little skeptical. That's a very healthy thing."

With AI stocks continuing to correct, it's telling that a prominent investor has framed the wariness as "healthy." How the market settles after the first wave of euphoria is something we'll be watching closely for a while.

An MTP-Head Variant of JetBrains' "Mellum2.1" — Decoding 1.4–1.8x Faster with Verification

A version of JetBrains' code-specialized model "Mellum2.1-12B-A2.5B-Thinking" equipped with a Multi-token Prediction (MTP) head has been released, and it's making waves in the local LLM community.

The MTP head is a single additional Transformer block trained alongside the main model that predicts the "next-next" token after the one just sampled. Inference engines use these predictions as drafts for speculative decoding, and because the main model verifies every token, the output stays identical while decoding gets 1.4–1.8x faster. The draft acceptance rate is reportedly around 90%. The head is stored in `model-mtp.safetensors` and can be enabled via `--speculative-config` in vLLM 0.33 and later, or with a flag in transformers 5.20 and later.

The base model is identical to the existing Mellum2.1 Thinking, so if you're running code completion locally, it looks worth trying.

Rocky Linux Creator Releases a "Not Linux" Kernel Built with AI in 11 Days

Gregory Kurtzer, the developer behind Rocky Linux, has published a new kernel called "ginux" on GitHub, built in just 11 days using AI. Its unique positioning — billing itself as "Not Linux" — has drawn attention on Hacker News. It's also an interesting experiment in how far AI code generation can reach toward system software development at a practical scale. For details on the design philosophy and internals, refer to the repository's public information for now.

China Officially Issues Its First Full-Fledged AI Export Regulation — Ministry of Commerce Announcement No. 48

The story reported on the 22nd of last month — that China was moving to place its top AI models and chip designs under export controls — has now landed as a formal announcement. The Cloud Security Alliance (CSA), a nonprofit, has published a research note analyzing China's Ministry of Commerce (Mofcom) Announcement No. 48 as "China's first cutting-edge AI export regulation."

Unlike the US-led export controls, this appears aimed at "technology protection" — safeguarding the country's AI competitiveness from flowing abroad. As for the scope of coverage and practical implications, we'll wait for CSA's analysis and how enforcement plays out.

arXiv Skim: Pruning Attention Heads Suppresses Verbatim Memorization

The problem of LLMs memorizing and regurgitating their training data has long raised concerns from privacy and intellectual property standpoints. "Fewer Heads, Fewer Memories," published on arXiv today, builds on the theoretical insight that attention heads function as associative memories and that memory capacity grows in proportion to the number of heads, and shows that pruning redundant attention heads suppresses verbatim memorization.

The results suggest it may be possible to reduce the risk of memorization leakage without retraining the entire model — a compelling demonstration of how "what a model remembers and where" can be handled at the structural level.