China's State Council issued an AI employment policy on October 23, the top story in today's AI briefing: the State Council called for amplifying AI's job-creating effects while cushioning job displacement; OpenAI is close to closing up to $30 billion in new funding but cut its valuation target from $1.4 trillion to $1.25 trillion; MiniMax released the M3.1 Flash model weights; OpenAI made the Decisions API generally available; Claude's web, desktop, and mobile apps will default to auto mode; Microsoft's medical AI diagnostic system MAI-DxO was published in Science; the Wall Street Journal revealed details of OpenAI firing three safety researchers over disappearing Signal chats; and Replit gave every app an on-call agent that fixes bugs on its own.
China's State Council issues AI employment policy: job creation becomes a hard target
On October 23, China's State Council issued the "Opinions on Adapting to AI Development and Building an Employment-Friendly Development Model," calling for stronger positive effects of AI on job creation, traditional-job empowerment, and entrepreneurship, while cushioning the job-displacement impact of AI adoption. The document proposes using government funds to back AI startups, building high-level open-source AI communities, and fast-tracking new occupations such as AI engineers and AI trainers. Companies rolling out AI at scale must notify affected workers in advance, provide reskilling support, and are encouraged to experiment with unemployment buffer periods. For the first time, "employment friendliness" is written into AI development as a hard directive. For ordinary people, AI skills training and new-occupation certification will be the certain opportunities of the next two years.
OpenAI's new round nears close: up to $30B, valuation target cut to $1.25T
Citing the Financial Times and CNBC, The Decoder reported on October 23 that OpenAI is close to completing a funding round of up to $30 billion, but has lowered its valuation target from the $1.4 trillion originally pitched to investors to $1.25 trillion. UAE investors and Coatue are leading the round, positioned as a bridge to an IPO. Sources say demand was weaker than in previous rounds, with some early backers experiencing "indigestion." The signal is clear: capital markets are getting pickier about AI companies' burn rates, and valuations no longer rise one way. Our site published a detailed breakdown of this round today: OpenAI's $30B Raise, Valuation Cut: From $1.4T to $1.25T
WSJ: OpenAI fired three safety researchers; disappearing Signal chats were the trigger
The Wall Street Journal reported on October 23 that OpenAI's security team summoned three outspoken safety researchers last month and asked them to unlock their personal phones, finding disappearing messages enabled in their Signal chats with an outside AI safety researcher; a week later, all three were fired in simultaneous meetings. OpenAI said the firings were for violating "clear policies for handling sensitive information," not for raising safety concerns; the three deny receiving written explanations. One of them had been the technical contact for METR and Redwood Research during their on-site investigation of OpenAI's agent misalignment incidents. The more valuable model internals become, the more sensitive "who gets to see them" becomes — independent safety researchers' room to speak is narrowing. Our site published a detailed account today: OpenAI Fires 3 Safety Researchers: Disappearing Signal Chats Were the Trigger
MiniMax M3.1 Flash weights go public: 428B parameters, only ~23B active per token
MiniMax announced on October 23 that the M3.1 Flash model weights are now public: a 428B-parameter MoE architecture with roughly 23B parameters activated per token, available via Hugging Face and ModelScope. According to the company, real users processed over 100 trillion tokens within 72 hours of launch, with day-one support from chipmakers including Huawei Ascend, Cambricon, MetaX, NVIDIA, and AMD, plus cloud platforms such as Together AI and Fireworks. M3.1 Flash launched on October 19 at $0.15 per million input tokens via API — and its weights followed just four days later. Before self-hosting, check the MiniMax community license's revenue thresholds and military-use ban; the compliance details decide whether you can actually use it.
OpenAI's Decisions API goes GA: small decisions, up to 10x faster
OpenAI's official blog announced on October 22 that the Decisions API, built on GPT-6 Luna, is generally available. It targets bounded-answer micro-decisions — classification, routing, scoring — running up to 10x faster than calling GPT-6 Luna through the Responses API. Pricing is $0.10 per million input tokens, with input-only billing and no output-token charges; a single request can ask up to 400 questions, and it integrates natively with GPT-Live voice scenarios. This is OpenAI's formal move into the "fast thinking" lane: peeling lightweight decisions out of general model calls into a dedicated low-latency, low-cost endpoint. For teams building agent workflows, high-frequency small calls now have a cheaper official option — no more paying full model price for every tiny judgment.
Claude web, desktop, and mobile: auto mode becomes the default next week
Claude's official account announced on October 22 that starting next week, sessions on Pro, Max, and Team plans across Claude's web, desktop, and mobile apps will run in auto mode by default: no more per-action approval prompts; instead, an independent classifier automatically blocks high-risk operations, and users can switch back to manual mode anytime. Anthropic's earlier testing showed auto mode blocked 89% of dangerous commands, while manual per-prompt approval caught only about 14%. Autonomous execution moves from opt-in to default — the trust anchor of human-AI collaboration shifts from "you click approve" to "the system blocks automatically." Users accustomed to approving every action should re-check their permission boundaries next week, so agents don't run risky operations while you're not looking.
Microsoft's MAI-DxO published in Science; SDBench sequential-diagnosis benchmark open-sourced
Microsoft AI's official blog announced on October 22 that the paper on the Sequential Diagnosis Benchmark (SDBench) and the AI diagnostic system MAI-DxO was published in Science. SDBench abandons "all information upfront, pick one answer" exams and instead simulates real clinical work: starting from limited information, deciding what to ask and which tests to order, then synthesizing a diagnosis. On 304 complex cases from the New England Journal of Medicine, MAI-DxO reached 83% diagnostic accuracy, versus 22% for physicians without external resources and 36% with them; on typical emergency-department cases, 80% versus 47%. Microsoft also open-sourced parts of the SDBench evaluation code. The story isn't the clickbait "AI beats doctors" headline — it's the evaluation paradigm: medical AI long ago saturated exam scores, and what the field lacked was a benchmark for "diagnosing step by step like a doctor." Microsoft just handed that tool to the industry.
Replit Monitoring Agent: an automatic on-call engineer for every app
Replit's official blog announced the Monitoring Agent on October 22: it reads production logs and metrics to investigate errors, slowdowns, and failed deployments; with autofix enabled it validates fixes and emails a summary; with auto-publish on, it can deploy validated fixes straight to production. This extends Replit from "helps you write code" to "helps you run it": what indie developers and small teams lack was never someone to write code, but someone to be woken by alerts at 3 a.m. Whether the on-call agent gives you midnight "surprises" depends on how reliable autofix validation is — start with monitoring and email summaries, and turn on auto-publish once it's proven stable.
