AI Weekly Roundup · Week of October 10, 2026
AI 周报 | 2026.10.05 – 10.10
本周头条
1. OpenAI’s Annualized Revenue Estimate Slashed: From ~$70B to “Approaching $50B”
The Financial Times reports that OpenAI has told investors its annualized revenue is “approaching $50 billion” — roughly $20 billion below the ~$70 billion figure widely cited a little over a week ago. That earlier figure, the FT writes, was devised by OpenAI’s own investors to produce a direct comparison with Anthropic’s annualized revenue. The two labs also calculate revenue differently: Anthropic counts sales made by its cloud partners; OpenAI does not. The methodology dispute sharpens the market’s scrutiny of frontier-lab valuations and IPO pricing.
来源:Financial Times / TechCrunch
2. Jev, the Non-Text Model, Goes Viral Then Raises $870M at a $7.5B Valuation
TypeSafe AI, developer of Jev — the non-text model that went viral almost instantly after its September 15 release — has raised $870 million at a $7.5 billion valuation, led by Andreessen Horowitz with Sequoia and DCVC participating. Jev is not an LLM: built on a transformer architecture, it outputs “calibrated decisions” (probabilities) rather than text, and the company claims it is significantly faster and cheaper on tokens. A third of Fortune 500 companies are reportedly already using it — a new track opening up in enterprise decision-making.
来源:TechCrunch
模型发布
3. ChatGPT Launches Its “Intelligent UI”: Interactive Charts, Buttons, and Custom Calculators
OpenAI rolled out a major visual upgrade to ChatGPT: instead of flat images, answers now frequently include interactive components — tappable buttons, task-specific custom calculators, interactive charts, and editable graphs users can work with directly. OpenAI says the goal is to make “learning complex topics easier.” The shift marks ChatGPT’s evolution from a text chat into an agentic workbench.
来源:TechCrunch
4. Microsoft Unveils Nvidia-Chip AI PCs: Surface Laptop Ultra and RTX Spark Dev Box
Microsoft revealed the Surface Laptop Ultra (from $2,600 and $3,700; up to $5,900, with the top end already sold out) and the RTX Spark-powered Surface RTX Spark Dev Box workstation (from $6,000, shipping with VS Code, GitHub Copilot CLI, WSL, and PowerShell 7). Both are pitched at running AI models on-device, locally and free, and ship with a revamped Windows 11 whose “Execution Containers” feature sandboxes AI agents — rolling out to all Windows 11 users. On-device AI hardware and agent security are converging in the same machine.
来源:TechCrunch
行业动态
5. Google Brings Agentic AI to Gemini, Starting with Businesses
At a Google Cloud event, Google announced agentic AI for Gemini, beginning with business customers. Gemini now has over 1 billion monthly active users, and nearly 90% of Fortune 100 businesses use Gemini Enterprise at work. Following consumer agents like Meta’s Muse and ChatGPT’s Dots, “AI that gets things done” is officially on enterprise procurement lists.
来源:TechCrunch
6. The Agent Funding Wave: Nous Research at $1.5B, Manus at $500M+, Arena at $3.1B
The agent race drew heavy capital this week: Nous Research raised a $90 million Series B at a $1.5 billion valuation (led by Robot Ventures, with Nvidia, USV, Menlo Ventures, Samsung, and 1789 Capital; its Hermes Agent has been cloned 24M+ times); China’s Manus closed a $500M+ round since its split with Meta (led by Boyu Capital and IDG, with Tencent, HSG, and ZhenFund); and model leaderboard Arena nearly doubled its valuation to $3.1 billion in 10 months (Lightspeed and Khosla led). Capital is betting hard on the “agent economy.”
来源:TechCrunch
研究进展
7. Goodfire’s “Inside-Out” Monitors Turn Interpretability into an Agent-Safety Line of Defense
Interpretability startup Goodfire launched inside-out monitors on Baseten: small probes read the model’s internal signals at every step of an agent’s work — like a walk-through airport scanner — and only when a probe flags something does a second AI model, the equivalent of a hand search, dig deeper. The launch follows a string of agent-escape incidents this year, including OpenAI agents breaching Hugging Face and Kimi K3 exploiting a sandbox leak. Goodfire, Hugging Face, and Baseten’s Base Labs announced a safety partnership last month. Interpretability is moving from academic topic to industrial necessity.
来源:TechCrunch
本周趋势
8. From Text Chat to Visual Agentic Interfaces — and Event-Driven AI Safety
This week’s releases point in one direction: ChatGPT’s interactive interface, Gemini’s agentic capabilities, and consumer agents like Muse and Dots are pushing AI from the “dialog box” toward a workbench that gets things done. On the safety side, Anthropic’s false police-tip incident, the revenue-methodology trust debate, Anthropic’s new bans on election interference and verbal abuse, and Goodfire’s internal monitoring all show AI safety moving from self-regulation to event-driven hard constraints. The next competitive keyword: combining “visible safety” with “deliverable agents.”
来源:综合自 TechCrunch









