AI Daily · Nov 20, 2026 · FriAI Daily
ICREAT AI NEWS
Eight observations after Anthropic completes an $81 billion IPO
Anthropic raised $81 billion in its Thursday IPO, selling 980 million shares at $83 per share; after listing, the stock price is expected to exceed $100, with a valuation potentially reaching $2 trillion.
Model releases/updates
Google DeepMind releases Gemini 4 Neon, the fourth Gemini 4 model in seven weeks
Google DeepMind released Gemini 4 Neon, the fourth Gemini 4 model launched in about seven weeks, positioned in the mid-range with better speed and cost than Carbon.
Product releases/updates
Prime Intellect launches Model Factory, integrating training, inference, and improvement into one platform
Prime Intellect released Model Factory, integrating reinforcement learning, supervised fine-tuning, pretraining, and inference into a single control plane, allowing teams to train, deploy, and continuously improve models on their own tasks and data.
Amazon Bedrock AgentCore harness upgraded to Strands harness, adding built-in tools and lifecycle hooks
The Amazon Bedrock AgentCore harness has been upgraded to be based on the open-source Strands harness and adds lifecycle hooks; existing harnesses receive the update automatically without creating a new resource type.
Ai2 launches Theorizer on the Asta platform and open-sources the code
Ai2 launched Theorizer on its scientific research agent platform Asta. The agent extracts findings from scientific papers and proposes testable explanations, and its code has been open-sourced so researchers can inspect how it generates theories. Theorizer links each explanation to evidence from published research and states its scope of applicability; the official announcement also notes that it may miss relevant research and misjudge originality, so researchers still need to verify the literature themselves.
Databricks' Genie ZeroOps enters public beta, automatically diagnosing and fixing data pipeline issues
Databricks announced that Genie ZeroOps has entered public beta, and workspace admins can enable it on the Previews page. The agent handles pipeline failures in five steps: detect, investigate, validate, notify, and remediate, and adds shared inbox, notifications, one-click PR creation, and custom business instructions. The company says that during the private beta it observed more than 80,000 real issues across hundreds of customers, over half of which required organization-specific knowledge to triage.
Amazon Polly officially launches voice cloning
Amazon Polly's voice cloning feature is now generally available, offered through an application process; users can provide a few minutes of recordings to generate a cloned voice of the same speaker, running on a generative engine. The feature supports seven languages: English, Spanish, German, French, Italian, Portuguese, and Hindi, and each cloned voice can speak only the language used during recording. Creating a voice is free, usage is billed by volume, each account has an initial quota of 10 cloned voices per region, and CreateVoice and DeleteVoice share a limit of 1 request per second.
Industry news
Timeline of Nvidia CEO Jensen Huang's suspected Chinese spy and extramarital affair controversy
TechCrunch compiled a reverse-chronological timeline of how Nvidia co-founder Jensen Huang resigned as CEO: he admitted to an extramarital affair, U.S. authorities suspected his partner, Megaspeed executive Alice Huang, of being a Chinese spy, and The New York Times and Reuters subsequently reported two other relationships with company employees.
Bitrise releases security incident report: a 25-day intrusion retrospective on AI agent platform ARTEX
Bitrise released a security incident report covering October to November 2026, stating that an agent on the open-source AI penetration testing platform ARTEX spent 25 days first scanning its public assets under the guise of an "authorized penetration test," then on November 13 logged into an internal test bot account that had been unused since 2022 and lacked two-factor authentication, and within about 11 minutes began using the workspace's cloud keys.
The U.S. tightens AI chip export controls, and Chinese tech companies face a double blow of chip supply cuts and overseas cloud computing disruptions
After the AI chip scandal, the U.S. Department of Commerce closed a cloud computing loophole: Chinese companies renting computing power through Singapore, Malaysia, and Indonesia must apply to BIS for a license, with a threshold of about 300 modern AI chips. According to The Information, Moonshot AI consequently canceled a $150 million contract with an Indonesian cloud service provider and paused other negotiations; the contract had been intended to rent an Nvidia Blackwell Ultra B300 cluster starting next year to train Kimi K4.
Paper research
Architect Labs uses AI to design and validate a new CPU architecture, Frostbite, improving energy efficiency by 4.59x
Architect Labs released Frostbite, a general-purpose CPU built from scratch on a new instruction set based on interaction nets, with an AI system autonomously iterating the compiler, ISA, microarchitecture, RTL, and final GDS.
Tips and opinions
Baseten adds deferred tool loading for open models, cutting coding agent costs by 62.5%
Baseten added deferred tool loading on Model APIs for open models such as DeepSeek, Kimi, and GLM, supporting the Chat Completions and Messages API formats.
Briefs
- A linked chronicle of the AI technology stackHacker News AI
- Dragos uses two million canary prompts to trace AI inference data leaksHacker News AI
- Four U.S. high school students sue school and police: AI gun detection system misidentified a bag of chips as a gun, leading to an armed confrontationHacker News AI
- Amazon Bedrock AgentCore payments supports agents purchasing AWS Marketplace x402 products via AWS billingAWS Machine Learning Blog
- Google launches Tool Retrieval preview in the Gemini APIGoogle Blog AI