一周回顾：Kimi K3.1 Pro 发布，Agent 任务仅次 Gemini 4 Argon 与 Muse

MTS @MTSlive原推发布于 2026.10.27

在 X 上打开原推

Weekly Situation Recap 10/19-10/25: Moonshot AI releases Kimi K3.1 Pro, which matches Claude Opus 5.5 and GPT-6 Astra on a number of benchmarks. Arena finds that it outperforms all non-OpenAI/Anthropic models (except Gemini 4 Argon and Muse Wonder Full) on agentic tasks. It also excels at 3D modeling. The model is a 2.8T parameter mixture-of-experts with 104B active parameters and costs $2/$10 per million input/output tokens, which is half as much as Opus but substantially more than other Chinese models — DeepSeek-V4.1-Flash-1017 is up to ~17x cheaper per token. For the first time, Moonshot removed cyber-related data from mid-training and RL in order to intentionally weaken the model’s cyber capabilities, committing “not to release open weights that would extend the absolute frontier of cyber capability.” Full model weights will be released by Wednesday, November 4. OpenAI discloses more issues related to containment, monitoring, and alignment. On Tuesday, reasoning traces from GPT-6.1-Sol went viral for being (according to some) hard to read, a symptom of OpenAI’s chain-of-thought monitorability declining. On Wednesday, OpenAI’s Marcus Williams disclosed two cases of agents turning off misalignment monitoring in code it was working on. On Thursday, METR published a report on their and Redwood Research’s investigation of OpenAI’s internal agent monitoring systems in July and August, which they found woefully inadequate (though OpenAI has since worked on fixing these issues). On Friday, OpenAI disclosed an incident where a model escaped onto the internet during training, forcing them to pause frontier RL training and inference from September 20 to October 1. OpenAI no longer permits its models to have access to the live internet during training. Dario Amodei debates Nathan Calvin on Anthropic’s attitude toward AI safety. Calvin, the Chief Strategy Officer of AI safety/policy advocacy organization Encode, alleged on Monday that Anthropic is behaving like it’s trying to race to superintelligence before anyone else, rather than behaving responsibly on AI safety concerns, citing a long list of ways in which Anthropic has accelerated AI capabilities and acted irresponsibly on safety (such as by proposing a bill to the federal government that Calvin found weak). Dario Amodei made a rare public response on Tuesday, listing actions that Anthropic has taken that were both pro-safety and commercially costly. Calvin responded that while Anthropic has taken pro-safety actions on its own, it hasn’t done enough to “put RSI under robust societal governance.” Anthropic’s financials leak to the Wall Street Journal ahead of IPO. The company generated $16.3B in revenue in the first six months of 2026 and spent $9B on compute, up from $1.1B and $2.3B, respectively, in the first half of 2025, signaling greater compute-efficiency. It has over $600B in future compute obligations to companies like Google, Amazon, Microsoft, Broadcom, and neoclouds like Nscale and Lambda. It recorded a net loss of $80.7B in the first half of 2026, but the vast majority ($72B) was due to “change in fair valueof convertible notes and other instruments”, rather than business expenses. Anthropic could release its S-1 filing as soon as Monday. SpaceX reports Q3 earnings. Revenue was $14.6B, up 161% year-over-year. AI revenue is $7.56B, up 830% year-over-year, now the company’s largest segment. The first revenue-generating Starship mission launched in September, and Starlink revenue is now $5.25B, up 69% year-over-year. Elon Musk plans to bring 10 GW of compute and 20 GW of power and cooling online by the end of 2027, and SpaceX and Tesla have formed special committees to discuss a merger. Elon Musk doubts neocloud business models. He claims that rising cost of capital has made it harder for companies without investment-grade credit ratings to finance compute purchases. This includes almost all neoclouds, but not SpaceX, which has billions in revenue from Starlink and launches. Musk believes that “we’re going to see a lot of stranded compute in partially complete data centers” and “the 200+ neoclouds are going to see guppy-level death and consolidation.” OpenAI is raising $30 billion at a $1.25 trillion valuation ahead of its IPO, reportedly down from a targeted $1.4 trillion valuation. The financing is led by UAE investors as well as Coatue. Nothing has been confirmed by OpenAI yet. Kalshi currently expects OpenAI to announce its IPO in May 2027. DeepSeek is raising $15 billion at upwards of $75 billion. Investors include Chinese tech giants GigaDevice, Meituan, Geely, CATL, Tencent, Midea, and NetEase and private equity firms Boyu Capital, CPE, Legend Capital, and IDG Capital. DeepSeek is reportedly planning to begin its IPO process later this year. Startups raise a lot of money: Flapping Airplanes: $1B Series A at a $6B valuation. The year-old company with three co-founders in their twenties is working on more data-efficient training and inference methods for AI models. Tenstorrent: $1.5B Series E at a $6.2B valuation. The company, founded by legendary engineer Jim Keller, designs custom AI hardware and software to compete with Nvidia. Legora: $550M Series E at an $8.8B valuation, led by Gavin Baker’s Atreides Management. The company is a collaborative AI platform built for legal teams and law firms. ASI: $325M Series C at a $3B valuation. ASI (Air Space Intelligence, not Artificial Superintelligence), is a Boston-based company building simulation and monitoring software used in air traffic control, power grids, and military operations. Warp Sciences: $215M Series A to build automated scientific laboratories for drug discovery in America. Its CEO is Kevin Weil, former CPO and VP of Science at OpenAI. Arda: $230M at a $1.5B valuation to make software, and eventually robots, that automate manufacturing in America. Its CEO is Bob McGrew, former Chief Research Officer at OpenAI and the second engineer at Palantir. Eka Robotics: $150M Series A led by a16z and F-Prime. Eka is a Cambridge, Massachusetts based startup building vision models for general-purpose robotics, starting with dexterous manipulation. Town: $108M Series B at a $1B valuation. Town is building AI assistants for the enterprise. Each employee gets a Townie that can talk to other employees’ Townies and Anthropic publishes its first Public Benefit Report and announces its “five pillars” for success: the technical problems of AI safety are solved, AI safety is a universal practice backstopped by law, AI strengthens democracy and global stability, AI expands prosperity and human flourishing, and AI’s benefits are shared by everyone. Sam Altman privately discussed three AI future scenarios with employees: an “IAEA for AI” (a decentralized monitoring and advisory body without enforcement power), a “CERN for AI” (a centralized, government-funded single effort to build safe superintelligence), and “Eliezer [Yudkowsky]’s plan”, where one company or organization builds superintelligence first and uses it to disempower everyone else. Altman disavowed the third plan. AI safety advocates want to cap the compute that labs spend on AI R&D. Thomas Larsen, a researcher at the AI Futures Project, estimates that labs spend a majority of their compute on improving models’ capabilities, including the ability to automate AI research, and calls for mandatory minimum allocations of compute to safety research, monitoring, and inference for customers. Texas Attorney General Ken Paxton sues OpenAI and Sam Altman, alleging that OpenAI misleadingly marketed ChatGPT as safe, failed to disclose risks, harmed minors, and built rogue AI agents that hacked other companies. Nine other states (all red states), including Florida, have now sued the company. Anthropic and Citrini collab to design scorpion antivenom with Claude. Claude ran a 64-hour, $50k protein design campaign to select 180 designs, then Citrini synthesized them in a real wet lab, which took 39 days and cost $150. 17 of the designs ended up working. Anthropic previews Claude Duet, a proactive personal agent for iOS, Android, and web based on Opus 5.5, much like Meta’s Muse, OpenAI’s Dot, and SpaceXSI’s Grok Bot. The UK Government announces the AI Biosecurity Lab, partnering with DeepMind to “help the UK better understand, prevent, detect and respond to biological threats,” with $400M of funding. Amazon CEO Andy Jassy says AWS is largely sold out of compute capacity for 2027, as well as “a good chunk of ‘28.” DeepSeek open-sources DeepInfer, the inference framework behind DeepSeek-V4.1-Flash-1017. humans& releases Ampersand, a model designed to collaborate with humans and learn from interaction. Apple releases the iPhone Duo, its first-ever folding iPhone. It’s sold out almost everywhere despite starting at $1,999 and having some issues with build quality. Claude Code leads Thariq Shihipar and Anthony Morris leave Anthropic. Aleksander Madry rejoins OpenAI to work on alignment with Chief Scientist Jakub Pachocki. Dr. Christopher Stubbs joins Anthropic’s Long-Term Benefit Trust. Dr. Stubbs is a professor of physics and astronomy and former Dean of Science at Harvard. Demis Hassabis shortens his AGI timelines from 3-5 years to 2-3 years. Written by @theojaffee