AI News Nov 6, 2026
Language: DE / EN / ZH / FR / ES / PT / JA / KO
AI-generated: summaries written by AI from the linked sources. How we use AI
Key Takeaways
- Nonprofit METR spent eleven weeks auditing Anthropic's internal research data and concluded the pace of model development has roughly doubled, with about 80 percent of the speedup …
- AI physics startup PSI says a system running a non-public Anthropic model proved the Aaronson-Ambainis conjecture, a quantum computing question open since 2008. The claim, if verif…
- Ricursive Intelligence is releasing Ripple-65B Preview, a 65-billion-parameter open-weight model under Apache 2.0 that the company says tops open-source U.S. models on the Terminal…
- OpenAI: $11B (Unknown)
- Palo Alto Networks → Clutch ($500M)
- For agent workflows, explicitly decide what context to store, where it should live, and when to reuse it rather than letting agents accumulate context unchecked…
Why Today Matters
- METR Anthropic auditMETR's eleven-week audit of Anthropic's internal data concluded model development pace has roughly doubled, with about 80 percent of the speedup driven by AI-assisted research. Anthropic redacted the most informative evidence, leaving the public reliant on METR's judgment.
- Open-weight model raceTwo open-weight releases land the same day: Ricursive's 65-billion-parameter Ripple-65B under Apache 2.0, built at $32 million R&D cost, and Thinking Machines' Inkling 2 Small, a 594-billion-parameter mixture-of-experts model the company calls the strongest western open-weight model.
- AI fundingThinking Machines is reportedly raising $5 billion led by Accel at a $40 billion pre-money valuation, while OpenAI has reportedly raised an additional $11 billion in an unclosed round valuing it at $1.25 trillion. Both figures remain unconfirmed until the rounds close.
- Data center politicsAfter voters backed data center restrictions on 40 of 41 midterm ballot measures, OpenAI announced host-community commitments including no NDAs with public officials. The pledge addresses scrutiny of how data center projects use NDAs to conceal discussions.
- AI mathematicsPSI says a non-public Anthropic model proved the Aaronson-Ambainis conjecture, open since 2008; if verified, it would mark a notable formal math result by an AI system. The claim is PSI's and awaits independent verification.
What are the top AI breakthroughs?
This Nov 6, 2026 covers 10 curated AI news items spanning technology, research, and product developments. Nonprofit METR spent eleven weeks auditing Anthropic's internal research data and concluded the pace of model developmen...
Nonprofit METR spent eleven weeks auditing Anthropic's internal research data and concluded the pace…
Nonprofit METR spent eleven weeks auditing Anthropic's internal research data and concluded the pace of model development has roughly doubled, with about 80 percent of the speedup driven by AI-assisted research. Anthropic redacted the most informative evidence, so the public must rely on METR's judgment 'much more than would be ideal,' and there are skeptical voices inside the company.
AI physics startup PSI says a system running a non-public Anthropic model proved the Aaronson-Ambain…
AI physics startup PSI says a system running a non-public Anthropic model proved the Aaronson-Ambainis conjecture, a quantum computing question open since 2008. The claim, if verified, would mark a notable case of an AI system contributing to a formal mathematical result.
Ricursive Intelligence is releasing Ripple-65B Preview, a 65-billion-parameter open-weight model und…
Ricursive Intelligence is releasing Ripple-65B Preview, a 65-billion-parameter open-weight model under Apache 2.0 that the company says tops open-source U.S. models on the TerminalBench 2.1 coding benchmark, at a total R&D cost of $32 million. On other benchmarks it falls behind Thinking Machines' newly released Inkling 2 Small.
Thinking Machines, the startup from former OpenAI CTO Mira Murati, is releasing Inkling 2 Small, a m…
Thinking Machines, the startup from former OpenAI CTO Mira Murati, is releasing Inkling 2 Small, a mixture-of-experts model with 594 billion parameters. The company claims it is the strongest western open-weight model, though better agentic performance comes with sharply higher token usage.
Ant Group released Ling-3.2-flash, a native multimodal model accepting text, images, and video, with…
Ant Group released Ling-3.2-flash, a native multimodal model accepting text, images, and video, with 560 billion total parameters, 25 billion active per inference step, and a one-million-token context window. Ant says it targets complex visual understanding, multistep agent tasks, 3D understanding, and medical applications.
Baidu open-sourced ERNIE-5.5-Flash-Base, a raw pretraining checkpoint with 280 billion total and 14 …
Baidu open-sourced ERNIE-5.5-Flash-Base, a raw pretraining checkpoint with 280 billion total and 14 billion active parameters, 128K training context, and 1M-token support, intended for further training. Alibaba's Taobao MNN team separately open-sourced MNN-Pocket-0.3B, a 173MB INT4 on-device model it says decodes about 350 tokens per second offline on an iPhone 18 Pro.
Google rolled out Gemini Omni 1.2 Flash, adding native 15-second video creation, improved reference …
Google rolled out Gemini Omni 1.2 Flash, adding native 15-second video creation, improved reference adherence, and multi-turn conversational editing. Google says the update lowers costs by 30% and improves consistency for professional projects.
Following midterm elections in which voters backed data center restrictions on 40 of 41 ballot measu…
Following midterm elections in which voters backed data center restrictions on 40 of 41 ballot measures, OpenAI announced commitments to host communities, including that it won't use NDAs with public officials. The move comes as data center projects face scrutiny over their use of NDAs to hide discussions.
Perplexity is launching 'Teammates,' always-on agents that work through ongoing projects on their ow…
Perplexity is launching 'Teammates,' always-on agents that work through ongoing projects on their own virtual machines. The feature requires the $200-per-month Max plan, well above similar offerings from OpenAI or Anthropic, in an increasingly crowded always-on agent market.
OpenAI introduced BioBasicsBench, a 100-task benchmark measuring how well AI agents interpret raw ex…
OpenAI introduced BioBasicsBench, a 100-task benchmark measuring how well AI agents interpret raw experimental data and handle everyday wet-lab science. The benchmark targets a gap in evaluating agents on practical laboratory work rather than textbook-style questions.
Top AI Videos This Week
2 curated YouTube videos about AI developments.
Based on a transcript excerpt: Fireship explains the Exponential Time Hypothesis (ETH) and reports t…

Based on a transcript excerpt: Fireship explains the Exponential Time Hypothesis (ETH) and reports that an unnamed OpenAI model reportedly wrote an algorithm solving 3-SAT faster than exponential time, potentially breaking a 27-year-old hypothesis that over 800 published papers depend on. The video covers what 3-SAT and ETH mean, how the model reportedly broke it using a technique no human thought to try, and why those papers now have a problem.
Based on a transcript excerpt: Wes Roth discusses METR's independent audit of Anthropic on automatin…

Based on a transcript excerpt: Wes Roth discusses METR's independent audit of Anthropic on automating AI research, noting that Anthropic redacted much of the evidence METR found most informative and that an internal model reportedly broke all of Anthropic's internal benchmarks. The video examines what the redactions may signal about acceleration in AI R&D and Anthropic's updated S1 filings on recursive self-improvement.
What are the latest AI investment signals?
Latest AI investment signals: 5 funding rounds, 4 market updates, and 1 M&A transactions.
Primary Market – Funding Rounds
| Company | Amount | Round | Investors |
|---|---|---|---|
| OpenAI | $11B | Unknown | |
| Thinking Machines Lab | $5B | Unknown | Accel |
| Sand.ai | $200M | Series A+ | Cornerstone Capital, Wangsu Science & Technology |
| Physicl | $10M | Unknown | |
| Kalepa | $75M | Series B | Updata |
Secondary Market – Market Updates
M&A – Mergers & Acquisitions
| Acquirer | Target | Deal Value | Deal Type |
|---|---|---|---|
| Palo Alto Networks | Clutch | $500M | Acquisition |
What are practical AI tips this week?
4 practical AI tips curated from Reddit communities and expert blogs. NVIDIA's memory strategy for AI agents focuses on deciding what context to store, where it belongs, ...
Productivity
NVIDIA's memory strategy for AI agents focuses on deciding what context to store, where it belongs, and when to reuse it, since agents accumulate context as they read, reason, and use tools.
For agent workflows, explicitly decide what context to store, where it should live, and when to reuse it rather than letting agents accumulate context unchecked.
Productivity
OpenAI's CTO of Applications Vijaye Raji said more code gets thrown away now that coding is cheap; the piece covers how to decide what a team keeps versus scraps, and identifying constraints.
When coding is cheap, plan for discarding more code: decide deliberately what your team keeps versus scraps, and identify the constraints in your own work.
Productivity
Claude can analyze bank transactions through Plaid; Anthropic promises read-only access and limits on AI training, but disconnecting your bank doesn't erase financial information from previous chats.
If you connect financial accounts to an AI assistant, note that disconnecting the account does not erase financial information already present in previous chats.
Productivity
OpenAI's GPT-6 Cyber can develop working exploits and reportedly shows major improvements in following authorized testing boundaries, but its safety results vary across benchmarks, raising questions about how training, access restrictions, and external controls work together in real deployments.
When deploying models for security testing, combine access restrictions and external controls with model training, since safety results reportedly vary across benchmarks.