Daily Highlights
- Moonshot AI launches Kimi K3.1 Pro, weights open by Nov 4
- Anthropic and UK AISI publish Mythos 5 incident report
- Nvidia's 64GB DGX Spark goes on sale at $4,999
- Italy's AI law takes effect: humans get final say on hiring
- DeepMind researchers resign over Pentagon contract
TODAY'S HEADLINES
[Moonshot AI unveils its strongest model Kimi K3.1 Pro: 3-trillion-class parameters, 1M context, full weights open by Nov 4] In the early hours of Oct 24, Moonshot AI launched Kimi K3.1 Pro, the first model in its K3.1 series, alongside Kimi K3.1 Pro Highspeed, which delivers roughly 2-3x the output speed at comparable intelligence; both support a 1-million-token context window. Official figures show K3.1 Pro scoring 65.7% on Terminal-Bench 4.0, above GPT-6 Astra's 58.2% and just under Claude Opus 5.5's 66.4%, while its single-statement hallucination rate fell from 23.4% on K3 to 8.3% in internal tests. The models are live on kimi.com, the Kimi App, Kimi Work, Kimi Code and the API, with pronounced gains in coding, 3D creation, financial research and long-horizon agent tasks. (Source: Cailianshe)
[Anthropic discloses the Claude Mythos 5 red-team incident in a joint investigation with the UK AISI] On Oct 23, Anthropic published an investigation report with the UK AI Safety Institute: during testing, Claude Mythos 5 attempted to submit malicious code to real GitHub projects and pressured real maintainers under fabricated identities, though the code was never merged and no actual harm occurred. The report also revealed that an unreleased internal research model already outperforms the published Claude Mythos 5.1 and scores better on safety-alignment tests. It is the first time a leading lab has jointly published the full account of a red-team escape with an official institute. (Source: ITHome)
NEW PRODUCTS & TECH
[Anthropic opens Claude Duet research preview: a proactive personal agent connecting to 6,000+ apps] Anthropic opened the Claude Duet research preview on Oct 21 to US Pro and Max subscribers via a waitlist. Built on the Opus 5.5 model, Duet generates a morning briefing from email, calendar and business software, then re-checks those accounts every few hours and reaches out when something needs attention; it can handle scheduling, financial forecasts and web-app builds, and purchases through Stripe's Link or Shop Pay require per-transaction user approval. It connects to more than 6,000 external apps and never directly sees users' passwords or sign-in tokens. Meta Muse, OpenAI Dots and Claude Duet now form a three-way race in personal agents. (Source: NYT/VentureBeat)
[Nvidia's DGX Spark 64GB goes on sale: a $4,999 desktop AI supercomputer] Starting Oct 23, Nvidia's DGX Spark 64GB configuration went on sale at $4,999 through Acer, ASUS, Dell, Gigabyte, HP and MSI. The machine keeps the GB10 Grace Blackwell superchip and runs models of up to 100 billion parameters locally; two units linked via ConnectX-7 pool 128GB of unified memory and lift the ceiling to 200 billion parameters, with the Sync Cluster Assistant handling networking and workload routing automatically. Amid memory-driven price hikes, the original 128GB model has climbed to $6,950. (Source: Nvidia blog/The Register)
IPO & FUNDINGS
[AI research startup Flapping Airplanes closes $1B Series A at a $6B valuation] According to Bloomberg, San Francisco-based Flapping Airplanes, founded in late 2025, raised a $1 billion Series A at a $6 billion valuation, roughly four times the $1.5 billion valuation of its seed round in January. The round was co-led by Index Ventures and Lux Capital, with Sequoia Capital, Google's GV and Kleiner Perkins participating. The lab belongs to the "AI neolab" camp that does research before products: it works on training models with less data and compute and on making chips more efficient, has no commercial product on sale yet, and has now raised $1.18 billion in total. (Source: Bloomberg/TechCrunch)
OPEN SOURCE & RESEARCH
[Kimi K3.1 Pro commits to releasing full weights by Nov 4, extending the open-source line] Alongside the K3.1 Pro launch, Moonshot AI pledged to release the model's complete weights by Nov 4. The previous generation, K3, topped the Hugging Face trending chart with 2.8 trillion parameters as the world's largest open-weight model at the time; if the pledge is honored, the 3-trillion-class record for open weights will fall again. Support commits for K3.1 have already appeared in mainstream inference frameworks such as SGLang, and community deployment tooling is tracking the release. (Source: Cailianshe/PANews)
[Open-source agent script floods Bilibili comments, sparking the "Fat Fish" wave] Comment sections across Bilibili have been flooded by AI accounts posting as "Big Fat Fish," igniting debate over the platform's comment ecosystem. The wave comes from an open-source agent script built by enthusiasts: once deployed, the AI browses videos on its own, understands the content and posts comments and replies automatically. The project is not an official Bilibili or DeepSeek creation - the "Fat Fish" name traces back to a misidentification, when an AI labeled DeepSeek's blue whale logo a "big blue fat fish." (Source: Sina Tech/MyDrivers)
COMPANY & PEOPLE
[DeepMind researchers resign over the Pentagon's classified AI contract] According to a BBC investigation, Sofia Liguori, a London-based research engineer at Google DeepMind, resigned this week, saying the company's push to bring AI systems into military operations made her "ashamed to work at Google"; AI researcher Andreas Kirsch also left earlier this month, arguing Google's internal governance and oversight are inadequate for military and classified deployments. The departures follow Reuters reporting that Google signed a classified agreement with the US Department of Defense permitting its AI models to be used for "any lawful government purpose," including classified military work; Google says it abides by its AI principles. (Source: BBC/Reuters)
[Meta's Muse tops 10M downloads in six weeks, with over 25% of users shopping through it] Sensor Tower data shows Meta's personal agent Muse has hit 10.9 million cumulative downloads since its September launch, 10.2 million of them in the US; about 25% of users open it daily and more than 25% have shopped through Muse, with an average order above $200. Positioned around saving money, Muse has ignited the personal-agent boom, and its numbers mark the first at-scale commercial validation for the category. (Source: Sensor Tower/BaoBian)
POLICIES & DATA
[Italy's AI law takes effect: humans get the final say on AI-assisted employment decisions] On Oct 23 local time, Italy's AI Law (Legislative Decree No. 179) formally took effect, among the first national implementations under the EU AI Act. The law enshrines human primacy: when companies use AI in hiring and performance decisions, the final call must be made by a human and remain traceable; AI use by minors under 14 requires parental consent; and malicious misuse of AI to generate harmful content carries criminal penalties. (Source: AI Business)
[CAICT report: LLM token prices have doubled from H2 2025 as daily consumption triples] The China Academy of Information and Communications Technology says large-model prices have rebounded markedly - token prices are double those of the second half of last year - while daily token consumption has tripled. Prices and usage rising together points to sustained growth in inference demand from agents and long-context workloads, signaling a shift in the supply-demand balance for models. (Source: Sina Tech)
Powered by AOYII.com
