Daily AI News — 7 Nov 2026

OpenAI has closed roughly $11 billion in new funding, the first portion of a round that values the company at about $1.25 trillion, Bloomberg reported. The cash was wired on Friday, with the round expected to grow toward $20 billion in coming weeks. Backers include Coatue, T. Rowe Price, SoftBank, Sequoia, Founders Fund, Andreessen Horowitz and a UAE investor syndicate. OpenAI had initially aimed for $30 billion at a $1.4 trillion valuation before trimming the target amid a broader public-market selloff in the AI sector and a delayed IPO.

The size and the step-down both matter: OpenAI is still raising historic sums, but a scaled-back target signals investors are pricing AI more cautiously after a rocky stretch. UAE money in the syndicate underlines how Gulf capital keeps anchoring frontier-AI balance sheets.

Founders raising now should read the room: even OpenAI trimmed its ask and valuation. Benchmark your round against a market that rewards real revenue and disciplined burn over a headline number.

The independent nonprofit Model Evaluation and Threat Research (METR) estimated that Anthropic is improving its models' ability to handle well-specified tasks roughly twice as fast as a year ago, according to an analysis shared with the Associated Press. METR researchers, given access to internal data and staff interviews, attributed much of the acceleration to AI itself — Anthropic says Claude increasingly helps build newer, more capable versions. The finding lands even as CEO Dario Amodei has publicly urged the industry to slow the technology's pace.

A credible third party putting numbers on self-accelerating development is a milestone for the whole 'recursive self-improvement' debate. If AI meaningfully compounds AI research, capability timelines — and the safety work that must keep up — both get shorter.

Teams planning multi-year AI roadmaps should shorten their planning horizons: capabilities you're building around today may be outdated in months. Budget for faster model refreshes and re-evaluation rather than locking in a single generation.

China's Moonshot AI published the open weights of Kimi K3.1 Pro on Hugging Face, a mixture-of-experts model with 2.8 trillion total parameters (104 billion active per token) and a one-million-token context window. Partners including Databricks, vLLM and SGLang provided day-zero serving support across Nvidia, AMD and Google hardware. Moonshot reports it rivals top closed models on coding and agentic benchmarks. The license is permissive for internal use, but model-as-a-service vendors above $20 million in revenue must negotiate a separate deal, and large products must display the Kimi brand.

An openly downloadable model at this scale narrows the gap between open and closed frontier AI, giving enterprises a self-hostable option that keeps data in-house. The revenue-capped license, though, shows 'open weights' increasingly comes with commercial strings attached.

A company with strict data-residency rules can self-host Kimi K3.1 Pro for internal coding and agent workflows — keeping sensitive code off third-party APIs — provided it stays within the license's internal-use terms.

Google DeepMind said its AI models, working under human direction, uncovered two weaknesses in legacy 1024-bit RSA encryption, announced November 6. One enables forgery of digital signatures; the other is a new algorithm that lowers the cost of batch-factoring related moduli, meaning attacks on these older keys may take far less compute than cryptographers assumed. DeepMind stressed the limits clearly: modern deployments using 2048-bit or larger keys remain secure against this specific threat, so current standard systems face no immediate risk.

AI is starting to do original cryptanalysis, not just code. Even when today's standard keys stay safe, AI-accelerated attacks on the math at the core of security are a warning shot for how fast the defensive clock may need to move.

Security and infrastructure teams should audit for any lingering 1024-bit RSA keys — in old certificates, embedded devices or legacy systems — and rotate them to 2048-bit or larger, treating this as concrete motivation to finish deprecation.

Meta has indefinitely paused new commercial terms for business accounts that were set to take effect November 20, after an outcry from publishers and injunctions abroad. A court in Norway granted a temporary injunction at the request of 180 Norwegian media companies, and a Sao Paulo judge barred Meta from using journalistic content to train AI without authorization while requiring the company not to retaliate against publishers who refuse. Critics said the terms would have let Meta scrape vast business-account content, without payment, to build competing AI services.

This is the clearest sign yet that external courts and regulators — not labs' own policies — are setting the real limits on AI training data. The non-retaliation ruling matters most: it lets publishers say no without being buried on Meta's platforms.

Publishers and content businesses should document their scraping objections now and watch these cases: the Norway and Brazil precedents give leverage to demand licensing deals rather than accept blanket, unpaid data-use terms.

Google expanded real-time speech translation in Google Meet to 63 languages across web and mobile, powered by Gemini 3.5 Live Translate. Instead of configuring one language pair for the whole meeting, Meet now automatically detects the spoken language and translates it into each listener's preferred language in near real-time — even when speakers switch languages mid-call. Participants turn it on individually from the toolbar, with synchronized translated captions. It is available to eligible Workspace and Google AI plan holders, with meeting-length limits depending on tier.

Live, automatic, per-listener translation lowers the oldest barrier to global collaboration. Embedding it directly in the meeting — rather than a separate tool — makes multilingual calls feel ordinary, a meaningful edge for distributed teams and cross-border deals.

A startup with teams in Dubai, Sao Paulo and Seoul can hold one meeting where each person speaks their own language and hears everyone else in theirs, with captions — no interpreter and no English-only default.

Baidu released ERNIE 5.5 Flash Base, an open-weight mixture-of-experts foundation model with 280 billion total parameters and 14 billion active per token, published on Hugging Face and ModelScope. It was pre-trained from scratch on 33.5 trillion tokens with a training context up to 128K and support for up to one million tokens at inference. Baidu shipped it deliberately as a base model — without post-training — so developers decide its purpose, with a post-trained version to follow. It posted an 82.3% pass rate on HumanEval+, and vLLM and SGLang announced immediate deployment support on Nvidia hardware.

A no-strings base model from a major Chinese lab adds to a flood of open-weight releases, giving builders raw material to fine-tune for their own domains. Shipping untuned is a bet that the real value now sits in how teams specialize models, not in the lab's own alignment.

A team with domain-specific data — say legal or medical text — can fine-tune ERNIE 5.5 Flash Base into a specialized assistant, starting from a strong, freely available foundation instead of training a large model from zero.

Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter dense open-weight model built for agentic coding and shipped under the permissive Apache 2.0 license. The company reports 80.5 on Terminal-Bench 2.1 — its self-reported top score among US open-weight models — and 79.1 on SWE-bench Pro v2.0 Hard, beating mixture-of-experts rivals that carry 500 billion to nearly a trillion parameters. It is a general-purpose variant of the model behind Ricursive's AI chip-design platform, trained on a budget the firm puts at $32 million.

A small, dense, truly open model matching giant MoE systems challenges the 'bigger is always better' story. Apache 2.0 means teams can run, modify and ship it commercially without permission — the opposite of the revenue-capped licenses spreading elsewhere.

A startup can self-host Ripple-65B on modest hardware as the engine for an in-house coding agent — reviewing pull requests, fixing bugs, running terminal tasks — and freely fine-tune or embed it in a commercial product.

Goldman Sachs has barred most of its traders and bankers from using advanced AI coding tools, including Anthropic's Claude Code and Microsoft's GitHub, limiting them to engineers after a review of productivity and risk concerns, Bloomberg reported. Internal guidance updated in October frustrated some traders who had used the tools to track how clients were trading, simulate market shifts and draft text and visuals. The directive exempts 'strats' — staff blending software engineering, quant research and trading. Goldman, which built its own AI assistant drawing on OpenAI, Anthropic and Google models, said it makes certain AI tools available to determine the best use cases.

The counter-narrative to AI hype: a sophisticated firm pulling powerful tools from most staff over governance and risk concerns. It shows that in regulated, high-stakes settings, control and auditability can outweigh raw productivity gains.

Any regulated business rolling out coding assistants should define who gets access and why before a blanket deployment — scoping tools to roles with proper review, so convenience doesn't outrun code-quality and compliance controls.

OpenAI published a National Community Compact pledging to bear all electricity and infrastructure costs at future data-center campuses it owns and operates, moving to defuse backlash against AI data centers. The pledge came two days after voters sided with curbs on development, or rejected pro-development measures, in all but one of 41 local ballot questions across 10 states. OpenAI said it will publish construction plans, commission independent noise studies, report site-level water and electricity use, avoid NDAs with public officials, and give residents a say in how sites are built. It also released a policy blueprint urging lawmakers to set similar rules.

AI's buildout now runs straight into local politics — power bills, water, land and secrecy. How labs earn community trust will shape where they can build at all, making transparency pledges a real constraint on the compute that frontier AI depends on.

A local official weighing a data-center proposal can now point to OpenAI's published commitments — full energy costs covered, water and noise reported, no gag orders — as a baseline to demand from any developer before approving a site.