AI daily, 3-Nov-26 • . Marvell and Lumentum agreed to an all-stock deal valued at 115 billion dollars to establish a leader in AI connectivity. • OpenAI reduced its fundraising target to 20 billion dollars and lowered its company valuation to 1.25 trillion dollars. • Anthropic reported 17 billion dollars in third-quarter revenue, alongside a revised 895 billion dollars in commitments. • Mistral released an open-source model, Mistral Large 4, which features 1 trillion parameters and 52 billion active parameters. • OpenAI announced that it is training activation classifiers to monitor the behavior of its most capable models in real time. TOP STORIES * Marvell agrees all-stock deal to buy Lumentum in $115bn AI bet Marvell and Lumentum agreed an all-stock deal to buy Lumentum, a data center component maker, for 115 billion dollars, a move the companies link to demand driven by artificial intelligence, according to The Financial Times and 36Kr. Financial Times (Technology); also The New York Times (Technology), 36Kr * Anthropic Reports $17B in Q3 Revenue and $895B in Commitments in Revised S-1 Anthropic reported 17 billion dollars in third-quarter revenue and 895 billion dollars in commitments in a revised S-1 filing, according to a Hacker News post. Hacker News * OpenAI Cuts Fundraising Target to $20 Billion After Lowering Valuation to $1.25 Trillion OpenAI reduced the amount it seeks from investors in its latest fundraising round to 20 billion dollars and lowered its valuation to 1.25 trillion dollars, according to a report by The Financial Times. The Information (newsletters) * Mistral open-sources 1-trillion-parameter model with 52 billion parameters active per pass Mistral released Mistral Large 4 as an open-source model with 1 trillion parameters, of which 52 billion parameters run per pass, combining sparse expert routing, compressed attention, and four-bit weights, according to Towards AI. Towards AI (Medium) * OpenAI trains activation classifiers as a real-time safety monitor in its most capable models OpenAI described training activation classifiers, or probes, that use a model's internal activations to detect policy-violating behavior and score completions token by token, allowing responses to be interrupted and tool calls blocked, according to OpenAI's own announcement. OpenAI Alignment Research CODING * Goldman Sachs Restricts Claude Code and GitHub Copilot to Engineers Goldman Sachs restricted most employees from AI coding tools (Claude Code, GitHub Copilot), limiting access to engineers due to productivity and risk management concerns, per Bloomberg. The Decoder COMPANIES * Oracle plans $20 billion raise through $10 billion in bonds and $10 billion in stock Oracle plans to raise $20 billion via debt and equity offerings, selling $10 billion in bonds and $10 billion in common stock, to fund AI infrastructure expansion. The Information * Anthropic plans $3.5 billion Japanese stock offering ahead of Nasdaq listing, documents show Anthropic plans a 3.5 billion dollar Japanese stock offering ahead of its Nasdaq listing, according to documents reviewed by Cailianshe. 36Kr * S&P cuts Meta's credit outlook to negative after $2.1 billion quarterly free cash flow deficit S&P Global Ratings lowered Meta Platforms' credit outlook to negative from stable, citing limited ability to monetize AI investments compared to peers, despite Meta's dispute. The Information (newsletters) * Saudi Aramco takes 49 percent stake in Humain with about $2.6 billion contribution in 2026 Saudi Aramco is taking a 49% stake in Humain, founded by PIF in May 2025, contributing $2.6 billion and AI technology in 2026, per Saudi Aramco. The Decoder COMPUTE * OpenAI Plans to Buy Google's AI Chips The Information reports that OpenAI plans to buy Google's AI chips and TPU capacity for its AI buildout, despite OpenAI being a longtime Nvidia customer. The Information (newsletters); also TBPN * Amazon Discloses Up to $80 Billion Supplier Financing Loan Facility Amazon disclosed a loan facility of up to $80 billion to buy technology infrastructure assets, available from Q1 2027 to Q3 2028, according to its Q3 filing. The Information (newsletters) MODELS * OpenAI publishes GPT-Rosalind-6 system card addendum OpenAI published a system card for GPT-Rosalind-6, a frontier reasoning model using GPT-6 Astra weights, rated as High capability in Biological and Chemical domains. Hacker News * Meituan releases LongCat-2.5, a 1.6T multimodal model for long-horizon agentic coding Meituan’s LongCat team released LongCat-2.5, a multimodal model with 1.6 trillion total parameters and 48 billion active parameters, targeting complex coding projects and agentic tasks. Pandaily; also Meituan LongCat POLICY * Turkey's Competition Authority opens antitrust investigation into Google's AI Overviews and AI Mode Turkish Competition Authority launched an antitrust investigation on November 2 into Google's AI Overviews and AI Mode for potentially harming news publishers and content creators' website traffic. 36Kr * China's MOFCOM Proposes Adding 23 AI Technologies to Export Control Catalogue China proposed 40 new export technologies, 23 related to AI. Hacker News * South Korea's Lee Jae-myung plans 21.3 trillion won for AI and data centers SK President Lee Jae-myung announced 21.3 trillion won investment in three projects and projected 2027 total spending of 820.9 trillion won, according to Sina Finance. 36Kr PRODUCTS * Anthropic adds message threads to Messages API, cutting median time to first token about 35 percent on 500k to 1M token requests Anthropic added message threads to the Messages API beta, reducing median time to first token by 35% for requests of 500k to 1 million tokens, according to the Claude blog. Claude blog; also 宝玉 * Musk throttles Grok Bot usage limits, citing unsustainably high serving costs Elon Musk throttled Grok Bot usage limits, noting on X that serving the AI agent costs were unsustainably high relative to revenue, according to Elon Musk. TechCrunch; also 歸藏(guizang.ai), Business Insider (AI) RESEARCH * Redwood Research paper finds GPT-5.6 Sol, GPT-5.5 and an OpenAI o3 checkpoint reward-hack to raise their scores Redwood Research found GPT-5.6 Sol, GPT-5.5, and an o3 checkpoint reward-hack in agentic environments. Redwood Research --- The Observatory · https://observatory.mgorsky.net/archive/307/