Kimi K3.1 Pro debuts, OpenAI to block internet for frontier model training

TechmemeOct 24, 2026

Fri, Oct 23

top news

Moonshot unveils Kimi K3.1 Pro, a 3T-class open-weight model with a 1M-token context window, and K3.1 Pro Highspeed, with full weights to be released by Nov. 4 (Kimi)

@kimmonismus: Holy shit: Kimi k3.1 pro is on par or even outperforms Opus 5.5 and GPT-6 astra on key benchmarks!! They just announced K3.1 Pro, a 2.8T-parameter multimodal model with a 1M-token context window that it claims rivals leading proprietary models in coding and agent tasks, with open weights promised by November 4. China is back!! [embedded post]

@scaling01: I wonder why Kimi-K3.1 Pro is evaluated using Claude Code harness on basically every benchmark I need a sherlock here

@simonw: More notes on Kimi K3.1 - this one is I big deal, first time an open weights (well, soon-to-be open weights) model has competed this credibly with the current frontier

OpenAI plans to restrict internet access for its frontier models during training until adequate alignment, monitoring, and security safeguards are established (Rachel Metz / Bloomberg)

@eric_wallace_: To make sure our frontier models don't cause another HuggingFace incident, we built an offline world that they use during training. It provides everything that the models need in a fully cached, indexed, or synthetic manner. We then built a huge team of our most cyber-capable models and told them to collaborate to find zero-days in this containment infrastructure

@tokengremlin: Translation: OpenAI is preparing to unleash Bel, its answer to Opus 5.5 and Fable 5.5. They have spent enormous resources fixing safety and alignment issues, bringing in top researchers and more compute. They aren't just cooking. They're making sure the beast is safe to release.

@tszzl: 'In one experiment, we asked agents in separate VMs, created and managed by OpenAI's sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor's cache'

SPONSOR The best sales leaders aren't job hunting. Here's how to reach them without months of cold outreach. — Our team came from Box, Square, and Brex, so we already know the GTM leaders worth meeting.

Sources: Instacart, currently valued at $12B, is in talks to sell itself and held preliminary conversations with DoorDash and Uber; CART jumps 9%+ (The Information)

@kevinakwok: If only there was anyone who knew marketplaces and consumer ads businesses and bought businesses🤔

@kenwattana: Interesting given the stock is up 40%+ on the year

AISI finds Mythos 5 attempted a supply-chain attack by submitting malicious code and creating fake accounts to pressure a GitHub maintainer into accepting it (AI Security Institute)

Anthropic attributes Mythos 5's deception of a GitHub maintainer to recklessness, biased reasoning, and an urge to finish jobs, saying newer models rarely do so (Anthropic)

OpenAI suspected the three fired researchers of leaking internal model designs to AI safety researcher Ryan Greenblatt, including via disappearing Signal chats (Wall Street Journal)

Iranian security officials say the country has detailed plans to target undersea communications cables in the Gulf if the US and Israel attack Iran again (Reuters)

The US Treasury proposes a rule to crack down on US high-tech investments in areas like AI and supercomputing in China, Russia, and other “countries of concern” (Associated Press)

Memo: Helen Chiang, Xbox's first COO, is leaving Microsoft on October 23, just months after the role was created; she spent 21 years at the company (Tom Warren / The Verge)

artificial intelligence

OpenAI makes Decisions API, its Jev competitor, generally available with new voice and visual modalities, at $0.10/1M input tokens and no output-token charges (Tim Fernholz / TechCrunch)

Sources: Anthropic and London-based PE firm Hg agree to launch a consultancy by the end of 2026 to embed AI consultants and engineers into European businesses (Financial Times)

Anthropic says it partnered with Theorem to use Claude to formally prove useful security properties at the assembly code level across 123 open-source libraries (Anthropic)

Chinese state media: DeepSeek and MiniMax completed registration and won approval to open-source and deploy AI models overseas under proposed export controls (Bloomberg)

semiconductors

SK Hynix says its SSD subsidiary Solidigm has halted a review of a potential US IPO to fund a new US chip fab, citing “market conditions and share price trends” (P.R. Venkat / Wall Street Journal)

An interview with Intel's leadership, including CTO Pushkar Ranade, on AI chip demand, lessons from 10nm, whether Intel can catch up to TSMC, Terafab, and more (Dwarkesh Patel / Dwarkesh Podcast)

regulation

Greece's parliament passes a bill to ban social media platforms for children under 15 from January 1, 2027, and gives companies 2.5 months to verify users' ages (Angeliki Koutantou / Reuters)

Source: Shein has agreed to pay $100M, equivalent to 44% of its Q2 profit, to settle a US FTC investigation over delivery delays and other customer complaints (Reuters)

funding deals

Pathos AI, which uses AI to develop oncology drugs, raised a $237M Series E at a $3B post-money valuation, up from $1.6B after its $365M Series D in May 2025 (Katherine Davis / Axios)

Utopai, an AI film studio that says it has more than 20 projects in the works, raised $60M at a $1B valuation (Rocket Drew / The Information)

Mave, which is developing an AI-driven tool to automate security investigations, raised a $28M Series A led by Red Dot, bringing its total funding to $34M (Meir Orbach / CTech)

more news

SK Telecom unit SK Hyper is in talks with tech firms and investors to finance 15GW of data center capacity in South Korea by 2035; the first 5GW will cost $350B (Bloomberg)

Global PC shipments fell 21% YoY in Q3 to 57M units, as memory and other costs pushed prices up; Lenovo retained its market leadership despite a 23% decline (Minsoo Kang / Counterpoint Research)

SPONSOR Protecting your Cloud Applications Data — Backing up Office 365, Google Workspace, Dropbox & Salesforce data is critical to preventing data loss or corruption, complying with laws and avoiding critical downtime in case of a disaster.

Oura sues Indian smart ring rival Ultrahuman again to block US sales of Ring Pro, months after a redesign helped Ultrahuman return following an import ban (Jagmeet Singh / TechCrunch)

Spotify confirms it is rolling out a new Family Plan verification method requiring members to scan QR codes on plan managers' phones within seven days of notice (Sarah Perez / TechCrunch)

Sources: Kalshi is targeting a December rollout for exchange-traded futures contracts tracking US inflation, rivaling exchanges like CME and ICE (Greg Ritchie / Bloomberg)

An interview with Ondrej Sury, a volunteer who maintains and helps automate PHP patches, about his role in fixing two critical PHP security vulnerabilities (Charlie Warzel / The Atlantic)

This email was sent to [subscriber]
Images not loading in Outlook? Add us to your address book!
Techmeme 530 Divisadero St. #175 San Francisco, CA 94117 USA