待翻译：The Sequence Radar - Issue 919: Last Week in AI: Stripe Wants to Own the Token Economy

AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。

Next Week in The Sequence: Learn more about distillation techniques in our knowledge series. To keep you current we will dive into DeepSeek’s new release, the amazing EnvHarness paper released by Google and the AVO paper published by NVIDIA. In the opinion section we discuss the 6th layer of the AI cake: financing. Subscribe and don’t miss out: 📝 Editorial: Last Week in AI: Stripe Wants to Own the Token Economy The most consequential AI announcement this week was not a new frontier model. It was a payments company buying the switchboard. Stripe agreed to acquire OpenRouter, the gateway that routes requests across hundreds of models from dozens of providers. The strategic logic is unusually revealing: tokens are becoming an economic resource, and choosing which model should process each token is becoming a financial decision. For years, AI applications mostly hard-coded one provider. The emerging architecture looks more like a payment network or cloud scheduler: a request arrives, and infrastructure selects the best supplier based on capability, cost, latency, and reliability. The deal looks less like fintech diversification than Stripe expanding its definition of a transaction. That interpretation became even clearer in Stripe’s accompanying investor letter. The company said it began operating on the assumption that January 1, 2026 marked “the beginning of the singularity”—not necessarily science-fiction superintelligence, but a phase change in long-term economic trends. Stripe also highlighted the extraordinary concentration of AI companies already building on its infrastructure. The language is deliberately dramatic, but the behavior matters more: Stripe is assembling payments, billing, token metering, and now model routing into something resembling an economic operating system for AI. Ramp’s launch of Router.com validates the thesis while also showing how quickly the gateway layer may become competitive. Router exposes multiple models through one API and automatically selects the lowest-cost option that clears a required performance threshold. The deeper story is not Ramp versus OpenRouter. It is that model routing is becoming a standard enterprise primitive. Every inference call is turning into a tiny capital-allocation decision. Should this request go to the smartest model? The fastest? The cheapest model that is good enough? Suddenly engineering architecture and CFO cost controls begin to converge. Etched represents the physical layer underneath that emerging market. The inference-chip startup raised an impressive $700 million at a $21 billion valuation and shipped its first rack to Jane Street, which also led the round after testing the hardware. A customer becoming both buyer and lead investor is a stronger signal than another benchmark chart. It suggests specialized inference hardware is moving from promise to production—and that faster, cheaper intelligence can already constitute a financial edge. Finally, DeepSeek is pushing beyond text. Its new experimental multimodal model can reason over images and screenshots, extending the company’s aggressive efficiency-focused approach into visual intelligence. That matters because multimodality changes what an AI system can actually do. Once models can reliably understand interfaces, documents, charts, and visual environments, agents stop being conversational tools and start becoming operators. Taken together, these announcements reveal a stack becoming increasingly legible. DeepSeek supplies intelligence. Etched supplies compute. OpenRouter and Ramp allocate requests. Stripe meters and monetizes the flow. The frontier is no longer just a smarter model. It is an economic system deciding which intelligence to buy, on which silicon, for which task, and at what price. 🔎 AI Research EnvHarness: Awakening Static Worlds for Agent Learning AI Lab: Google Cloud AI Research, Washington University in St. Louis, and University of North Carolina at Chapel Hill Summary: This paper introduces EnvHarness, a programmable layer of plug-in components that dynamically customizes static environments—altering initial states, interaction rules, or chaining tasks—without modifying the underlying simulator logic or human-built verifiers. To automate this customization, the authors also propose EnvRigger, a system that diagnoses an agent’s weaknesses from its execution trajectories to generate targeted environment wrappers, leading to significant performance gains in both skill-based and reinforcement learning. Agent Lightning v1.0: Towards Harnessed Agentic RL AI Lab: Microsoft Summary: This paper introduces Agent Lightning v1.0, a lightweight framework for harnessed agentic reinforcement learning where the deploy-time harness directly manages the environment interaction loop during post-training. Using approximately 3,500 lines of code, the system addresses unique challenges like retokenization and dynamic sample counts, successfully improving a coding agent’s performance on SWE-bench by 14.6%. THE PROBLEM IS THE PROBLEM: TOWARDS SCALABLE MATHEMATICAL DISCOVERY AI Lab: Carnegie Mellon University and Anysphere Co. Summary: The authors propose the Find, Attempt, and Recommend (FAR) pipeline, which shifts AI assistance from solving pre-selected math problems to automatically extracting and filtering open conjectures from large literature corpora. Tested on combinatorics literature, the system recovered thousands of open problems and produced 77 publishable artifacts, including new proofs and counterexamples. AVO: Agentic Variation Operators for Autonomous Evolutionary Search AI Lab: NVIDIA Summary: This paper introduces Agentic Variation Operators (AVO), which replace traditional evolutionary search mechanisms with an autonomous coding agent that plans, implements, evaluates, and debugs code edits. Over a 7-day autonomous evolution period, AVO generated multi-head attention kernels for NVIDIA Blackwell GPUs that outperformed state-of-the-art expert-engineered implementations like cuDNN and FlashAttention-4. PTXBench: Benchmark and Adapt LLMs for GPU Kernel Optimization with Architecture-specific PTX AI Lab: Stanford University, RadixArk, and Carnegie Mellon University Summary: This paper presents PTXBench, a benchmark designed to evaluate and improve large language models’ ability to generate optimized GPU kernels using architecture-specific PTX instructions. Through targeted supervised fine-tuning conditioned on execution feedback, the authors show that LLMs can improve low-level optimization capabilities, though performance remains uneven across hardware and complex workloads. Demystifying Agent Skills: Why They Work-Until They Don’t AI Lab: Princeton University, UC San Diego, University of Southern California, Johns Hopkins University, and Stanford University Summary: This study analyzes the mechanisms of LLM agent skills, finding that they primarily succeed by acting as procedural anchors that stabilize execution rather than by injecting missing factual knowledge. Through contrastive trajectory analysis, the authors reveal that skills can also fail due to poor retrieval precision from large candidate pools or when guidance is misapplied to incompatible contexts. 🤖 AI Tech Releases DeepSeek-V4-Flash-Vision-Exp DeepSeek unveiled an experimental multimodal model with impressive performance. Sonic 3.6 Cartesia released Sonic 3.6, easily leading the voice leaderboards. 📡10 AI News You Need to Know About Stripe confirmed it has agreed to acquire OpenRouter, the gateway routing across 400+ models from 80+ providers, in a deal reported at roughly $7.5B. Micro1's gross annual run rate went from $100M to $500M in eight months, with roughly 60 to 70 percent of that retained as net, as demand for expert-generated training data keeps outrunning supply. Etched raised $700M at a $21B valuation led by Jane Street, double its July mark, and named Jane Street its first customer after shipping an inference rack last month. Anthropic’s annualized revenue run rate hit $65B at the end of July, up sevenfold from year-end, on preliminary Q2 revenue above $11.5B, ahead of an expected IPO. Ramp launched Router.com, a single endpoint that sends each request to the cheapest model clearing a set performance bar, built on the router Ramp ran internally for three years and free through the end of 2026. Groq closed a $350M Series A led by Disruptive with planned Nvidia participation at a $3.5B valuation, completing its shift from LPU chipmaker to Nvidia-powered inference neocloud running 13 data centers. Temporal is in talks to raise about $500M at a pre-money valuation of at least $12B, more than double its February mark, for its durable-execution platform that lets agent workflows resume after failures rather than restart. Nvidia will invest $1.5B in SB Energy and guarantee up to $105B in lease payments to become the exclusive compute provider at the PORTS-Pike campus in Ohio, which SB Energy will build and operate under a 20-year lease to OpenAI. Fractile is in advanced talks to raise about $600M at a $6.5B pre-money valuation, more than six times its May mark, on the strength of an initial deal to sell roughly $250M of chips to Anthropic that will not ship until 2027. Starcloud raised a $250M Series A extension at a $2.3B post-money valuation led by Manhattan West with Nvidia and Cisco joining, funding a new Woodinville factory and the Starcloud-3 orbital data center spacecraft slated to fly on Starship.