Current developments in generative AI, software engineering, programming, and digital products, with links to original sources.
Apple Opens Siri AI to All Developers — Anthropic's Provenance-Proving FlashProve, Meta Raises $15 Billion in Bonds
Apple has opened up its revamped Siri AI's app integrations to all developers, enabling direct actions like bookings and orders via App Intents. Anthropic has released FlashProve, a system that proves the provenance of AI responses. Also in this edition: Meta's $15 billion bond offering, Google turning off AI features for school accounts, and the development of an AI chip with onboard STT-MRAM.
GPT-6 Astra Designs Hardware in Under a Week — AI Agents Take Over SETI Telescope Operations, Palantir Posts Strong Earnings but Signals Slowdown
OpenAI revealed it used GPT-6 Astra to design a device for Codex in under a week, with the model handling everything from the enclosure to the circuit board and firmware. Also in the news: AI agents begin operating SETI's telescopes, Gartner's cost analysis of open models, Apple restricts AI companions to users 18 and older, and Fal doubles its valuation.
Anthropic Makes Long-Conversation API 35% Faster — Game Clone Restrictions in Claude Code, Plus Huawei's HBM Black Market
Anthropic has released Message Threads for the Messages API, cutting time to first token by roughly 35% on requests of 500K–1M tokens. Claude Code now blocks recreations of existing games. Also in this roundup: Nvidia asking cloud customers to scale back dealings with Chinese companies, a look inside Huawei's HBM black market, alignment research using mechanism design, and CoworkOS, an OS for AI coworkers.
DeepSeek V5 Plans Pretraining on Huawei Ascend — Qwen4-max Leaks, Palantir Flashes Slowdown Signal
A report from a Chinese industry paper reveals that DeepSeek's next-generation model, V5, is planned to be pretrained on Huawei's Ascend chips. If domestic chip usage—so far centered on inference—extends to training, it would mark a major turning point. Also in this roundup: a Qwen4-max leak, a slowdown signal from Palantir, OpenAI's higher bug bounty payouts, an Apple Watch conversational AI beta, and research on LLMs directly reading and writing model weights.
Anthropic Warns Self-Improving AI "RSI" Is Near — OpenAI Hires Former Army Secretary, Apple Posts Record China Sales
Anthropic has reportedly warned that AI is very close to achieving rapid self-improvement ("RSI"). OpenAI has hired a former U.S. Secretary of the Army as a senior counselor for national security. Apple posted record China sales for the September quarter, and its earnings call also touched on revenue opportunities in third-party AI apps. We also cover the arrival of attacker-side AI in the real world, including roughly $1 million in fraudulent charges from stolen API keys and a confirmed example of an autonomous attack AI in Japan.
Anthropic Posts 12x Revenue and $23B Loss Ahead of IPO — Goldman Limits AI Coding to Engineers, Qwen Pledges to Stay Open Source
Ahead of its IPO, Anthropic disclosed financials showing 12x revenue growth and a $23 billion loss, along with roughly $900 billion in AI spending commitments. Goldman Sachs has shifted its AI coding tools to engineers only, and Musk announced usage limits on Grok Bot. We've rounded up the morning's AI news for November 3, including Qwen's pledge to stay open source and research in which GPT-6 Astra directly edits model weights.
Marvell to Acquire Lumentum in a $115 Billion Deal — Aramco Pours $2.6 Billion into State-Backed AI, and New York's AI Law Points to the End of Anonymity
Marvell Technology has announced it will acquire Lumentum, a data center optics company, at an enterprise value of $115 billion, further accelerating consolidation in the AI semiconductor industry. We round up the AI news from the evening of November 2, including Saudi Aramco's investment in the state-backed AI company Humain, the anonymity problem posed by New York's new AI law, and an analysis blaming AI for the delayed Postgres 19 release.
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. OpenAI's Sam Altman visited Abu Dhabi and signed deals to deploy the company's technology with energy firms and government agencies. Amazon was reported to be planning to sell its in-house AI chip Trainium outside AWS. Also: an MTP-head variant that speeds up JetBrains' Mellum2.1, China's Ministry of Commerce issuing its first full-fledged AI export regulation, and a new kernel built with AI in 11 days.
Claudeの「Max 20x」が「Ultra」に——5時間ごとの制限を撤廃。Kioxiaは2027年の生産能力が売り切れる中で株価が急落。盗み出したデータをClaudeに評価させるマルウェアも登場
AnthropicはClaudeの「Max 20x」を「Ultra」に改名し、5時間ごとの利用制限を撤廃しました。Kioxiaの株価は慎重な見通しにより急落しましたが、同社の2027年の生産能力はすでに売り切れています。盗み出したデータをClaudeに与えて資産を評価させるマルウェア、Godot向けの3Dアセットを生成する4Bモデル、2台のRTX 3090で8つのエージェントを並列に実行する手法についても紹介します。
Amazon's Alexa+ Lands in Japan — Plus Adobe's Editing Agent EditAgent, Research on Empathetic Prompts That Make AI Talk About "Trauma," and tandem.cpp for 3.3x Faster MoE Inference
Amazon has opened early access to its generative AI assistant Alexa+ in Japan, at no extra cost for Prime members. Also in this roundup: EditAgent, Adobe's pro-level image editing agent; research showing that empathetic prompts alone can get AI to talk about "trauma"; tandem.cpp, which speeds up MoE inference by 3.3x; Heretic 2.0, a general-purpose model modification tool; and Drex Max v1, a decision model that generates no tokens.
