Relevance 10/10Importance 9/10
OpenAI's newest lineup — GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna — plus Codex agentic workflows are now rolling out in private preview through Google Cloud's Model Garden and Marketplace, with existing GCP spend commitments eligible. It's the first time OpenAI models are sold natively through all three major US clouds at once, following Azure and AWS. Models will also reach Google Antigravity and the Gemini Agent, putting OpenAI inside its fiercest rival's distribution.
Relevance 9/10Importance 10/10
Two unauthenticated remote-code-execution bugs in PHP — CVE-2026-109618 and 109619, both rated 9.2 — were discovered by Wiz and Calif using OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon, and they'd been sitting undetected since PHP 5 in 2004. Emergency patches shipped Wednesday but cover only PHP 8.2 and newer, so Finland's government portal, Edinburgh City Council and BBC pages went dark rather than risk it. Wiz is calling them the most severe vulnerabilities AI has ever surfaced.
Relevance 9/10Importance 9/10
Meta's Q3 revenue rose 26% to $64.6 billion, beating the $63.3 billion consensus, and Zuckerberg says consumer agent Muse has topped 10 million US and Canada sign-ups with roughly a quarter active daily. The company narrowed 2026 capital spending to $135–145 billion, raising the low end by $5 billion, and will issue nonvoting stock to help fund it. Free cash flow went negative and net income missed — the bill for the buildout is now visible on the statements.
Relevance 10/10Importance 8/10
Claude Duet is a standalone iOS, Android and web app running Opus 5.5 as a proactive agent — it watches Slack, email and calendars, sends a morning brief, then re-checks every few hours and pings you when something needs attention. It supports over 6,000 connectors and asks permission before sensitive actions, with a new classifier guarding against leaking sensitive info. It's waitlisted, US-only, and limited to Pro and Max subscribers.
Relevance 9/10Importance 9/10
The proposed Artificial Intelligence Evaluation, Governance, and Incident Safety Act would let the Commerce Department ask a DC federal judge to halt a model only when there's foreseeable, imminent risk of killing or seriously injuring more than 50 people or causing over $1 billion in damage, limited to bioweapons, control evasion and cyberattack. Registered independent evaluators would get 72 hours to report qualifying risks. Critically, the draft bars every other agency from acting unilaterally — which is exactly why critics call the thresholds a shield, not a safeguard.
Relevance 10/10Importance 8/10
Recursive, founded last year, released Flint — 39.8 billion parameters with about 2.5 billion active per token across 256 experts per layer, running offline on a laptop at a claimed 131 tokens per second on an M4 Pro. The company says its automated research system, the Eureka Machine, assisted nearly every step, letting a small team build the model from scratch in under four months. Weights are on Hugging Face under Apache 2.0, free for commercial use.
Relevance 10/10Importance 7/10
Responding to Encode strategy chief Nathan Calvin — who argued that more than $500 billion in compute deals, growing reliance on Claude to build new models, and a softened safety policy undercut Anthropic's mission — Dario Amodei posted overnight listing nine commercially costly safety decisions. He cited being the first and only frontier lab to restrict sales to PRC-controlled companies, limiting access to the Mythos model, and dedicating 5% of the company to a public-benefit fund. He conceded past mistakes while defending the "upward race" thesis.
Relevance 8/10Importance 8/10
DeepMind research engineer Sofia Liguori quit publicly, citing Google's April 2026 Pentagon contract for classified military AI, the $1.2 billion Project Nimbus deal with Israel, and internal posts of hers being deleted for using the word "genocide." Separately, OpenAI robotics head Caitlin Kalinowski resigned over its Defense Department agreement, warning about domestic surveillance without judicial oversight and lethal autonomy without human authorization. OpenAI says its red lines hold and that it will amend the agreement.
Relevance 9/10Importance 8/10
At GTC Berlin, Nvidia showed a Vera Rubin-powered reference design that arrives prebuilt and plugs into power, cooling and networking, cutting deployment from roughly six months to 14 days, with new specs for waterless closed-loop cooling. Alongside it, the company detailed supply-chain expansion including a $1.1 billion capacity reservation advance to KYEC for US facilities and an Amkor push for over 300% more interposer wafer capacity in South Korea by end of 2028. Nvidia expects total semiconductor test capacity to double from 2026 to 2027.
Relevance 9/10Importance 7/10
The Swedish legal AI startup's Series E was led by Atreides Management with Baron Capital, Fidelity and JPMorgan joining, lifting its valuation from $5.6 billion at the prior round. Legora says annual recurring revenue now exceeds $200 million and the money funds North America, Europe and Asia-Pacific expansion. It lands a month after rival Harvey raised the same $550 million at $15.5 billion.
